Add files using upload-large-folder tool

.gitattributes

@@ -939,3 +939,145 @@ preprint/preprint__4480d0108b1ac86fbd7de14d6ed11178e827bbb675fa636a1e0079b7242e2
 preprint/preprint__446a6d1639de149e2cdadfff5059c93994fb6bf575933900ed819a058513438a/preprint__446a6d1639de149e2cdadfff5059c93994fb6bf575933900ed819a058513438a_layouts.pdf filter=lfs diff=lfs merge=lfs -text
 preprint/preprint__44bd8119853e2fcc7fb044a9f959e97bc3b0440d36ba5953911b7e794c249100/preprint__44bd8119853e2fcc7fb044a9f959e97bc3b0440d36ba5953911b7e794c249100_layouts.pdf filter=lfs diff=lfs merge=lfs -text
 preprint/preprint__447ea8ea6e006d78bc64133f35a1e45dcbd8cb9ac5fc966fd3285782d3d6b1a8/preprint__447ea8ea6e006d78bc64133f35a1e45dcbd8cb9ac5fc966fd3285782d3d6b1a8_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4434c62cfd8163014f10888e9efa5b5818b5458a31008d13e8e813f04a96a9ab/preprint__4434c62cfd8163014f10888e9efa5b5818b5458a31008d13e8e813f04a96a9ab_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__441781baa1b2dedeba52491b135464cc88bacc38a173fe0d5b9f563ad3051d30/preprint__441781baa1b2dedeba52491b135464cc88bacc38a173fe0d5b9f563ad3051d30_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__44ac714467bf52e85c6116012157571d06f97bede43922ad63033d05e6843059/preprint__44ac714467bf52e85c6116012157571d06f97bede43922ad63033d05e6843059_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__45142058222c4ee18640002c7658fa05a38dbf6116a638ff9ce04deab9bdf06e/preprint__45142058222c4ee18640002c7658fa05a38dbf6116a638ff9ce04deab9bdf06e_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__44e0278cb1f5606a8d5bc36783fcb7feb6bb37b64915576d51c9046779a60453/preprint__44e0278cb1f5606a8d5bc36783fcb7feb6bb37b64915576d51c9046779a60453_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4578638c25f2519598bb564eee3e670e13f13833272c7a212079b454235d8d83/preprint__4578638c25f2519598bb564eee3e670e13f13833272c7a212079b454235d8d83_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__45a2b61c455ce309b9de56e26c6c7a9de051fae70d6c381e16584a6ec95d1ecf/preprint__45a2b61c455ce309b9de56e26c6c7a9de051fae70d6c381e16584a6ec95d1ecf_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__45396c5958b245233b6b2c31915a68bf2e0759229a7b1a5f4508fa98479cf166/preprint__45396c5958b245233b6b2c31915a68bf2e0759229a7b1a5f4508fa98479cf166_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__452e4d81e3a8904ec4a17925194ea479d8c2d33f82037bb4ad06d1a819814bd2/preprint__452e4d81e3a8904ec4a17925194ea479d8c2d33f82037bb4ad06d1a819814bd2_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__45b9e21e2ac6cd83cfa1298ab2404ec4b26128e261f8ca48526b3b795a7dc6aa/preprint__45b9e21e2ac6cd83cfa1298ab2404ec4b26128e261f8ca48526b3b795a7dc6aa_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__450135318d771e1ad02d941c44ec73a4211958ceb573b2e79287eceef29d0695/preprint__450135318d771e1ad02d941c44ec73a4211958ceb573b2e79287eceef29d0695_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4489bae4076d4fb9e786e28c2752440d96d0c0303c22928d1e44a5bef3eb7650/preprint__4489bae4076d4fb9e786e28c2752440d96d0c0303c22928d1e44a5bef3eb7650_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4559b53214897807ac2b802d84edaf78240f9915a277408759e9d64a62441f20/preprint__4559b53214897807ac2b802d84edaf78240f9915a277408759e9d64a62441f20_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4543fa1b84b3ec7a9d517333d863bea2a0bc4a51cc0d819dc9d684364175d727/preprint__4543fa1b84b3ec7a9d517333d863bea2a0bc4a51cc0d819dc9d684364175d727_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4581dc25f72e60dee12903d05362de1d6b78894117cff28f5b1da8e4e217deba/preprint__4581dc25f72e60dee12903d05362de1d6b78894117cff28f5b1da8e4e217deba_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4598f2e68537767015cb37ef6f87d4a18fffe6639428a32ee5f3060ef36b01ee/preprint__4598f2e68537767015cb37ef6f87d4a18fffe6639428a32ee5f3060ef36b01ee_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__44ec7d85d28d6d1fa10c508c3a2bec2976d9b428ded224567e985d7c4af76073/preprint__44ec7d85d28d6d1fa10c508c3a2bec2976d9b428ded224567e985d7c4af76073_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__45d244e36e4d8f66fc4cadf66b3b17bfed6588d966d3c855a2c323fa65c2f636/preprint__45d244e36e4d8f66fc4cadf66b3b17bfed6588d966d3c855a2c323fa65c2f636_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4573c773b81886349773bc086529f947395ee238028721e46c6c4ca5942f5dea/preprint__4573c773b81886349773bc086529f947395ee238028721e46c6c4ca5942f5dea_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__45d3762c6a4c021ea1fc547c5c795247698c99a0f1c7316d961871b5372ac612/preprint__45d3762c6a4c021ea1fc547c5c795247698c99a0f1c7316d961871b5372ac612_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__45cffc023f744cdd2bfc2b7ef9dba0edc7ad24903d3623e919acfa17c09b22c2/preprint__45cffc023f744cdd2bfc2b7ef9dba0edc7ad24903d3623e919acfa17c09b22c2_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__45d2500ca5b1760923390d301e8d0da12bac50057a0004d927257df9a246ad05/preprint__45d2500ca5b1760923390d301e8d0da12bac50057a0004d927257df9a246ad05_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__45ddd94af6aa870471d01122d881c72c9faad905980985dec3a9883641e317a1/preprint__45ddd94af6aa870471d01122d881c72c9faad905980985dec3a9883641e317a1_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__460f01fd436a445849077f508bb6ffc06e7cf9e45bd85132567550f22de9b4dc/preprint__460f01fd436a445849077f508bb6ffc06e7cf9e45bd85132567550f22de9b4dc_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__45ebfc65a7e0fbed4c91731906e87d84ebf7641a808f46e753b49f895747be08/preprint__45ebfc65a7e0fbed4c91731906e87d84ebf7641a808f46e753b49f895747be08_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__45f74ca9709712e156fe6f899d732c207d96d2bd47f1119c60253d0e71a68fbb/preprint__45f74ca9709712e156fe6f899d732c207d96d2bd47f1119c60253d0e71a68fbb_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__46c123409754c654bca4fa07a7aa0a40165fe55a38c2eec2afc939d4848ef074/preprint__46c123409754c654bca4fa07a7aa0a40165fe55a38c2eec2afc939d4848ef074_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__46102d80359574387f0af062d82638202a31b11b1dde9fbc73206dcb6c29fec7/preprint__46102d80359574387f0af062d82638202a31b11b1dde9fbc73206dcb6c29fec7_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__46c5e377957065df21fad84d74c43b2ad9b74854f4d2e72e13f5245b50600182/preprint__46c5e377957065df21fad84d74c43b2ad9b74854f4d2e72e13f5245b50600182_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__46dbe139afbe9502b6366856fea0cb41bd4112355e2113d8a4b79534be3d8dfa/preprint__46dbe139afbe9502b6366856fea0cb41bd4112355e2113d8a4b79534be3d8dfa_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__45ea2661197d5d5736515b46b0318adabc36086c80cd793293570089373f0128/preprint__45ea2661197d5d5736515b46b0318adabc36086c80cd793293570089373f0128_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4645ad16049dd4fc2cd1b4196e7916ebfa6330fe839887d0943c214aa553fd17/preprint__4645ad16049dd4fc2cd1b4196e7916ebfa6330fe839887d0943c214aa553fd17_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__46864738baad318750f7c9eaed66d5f03d849bfd1b6f86065cb70c233d2c7966/preprint__46864738baad318750f7c9eaed66d5f03d849bfd1b6f86065cb70c233d2c7966_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__466d647ed649392203b30a56d7a0b09858cbcb0ca4057c33667200d42a1734b1/preprint__466d647ed649392203b30a56d7a0b09858cbcb0ca4057c33667200d42a1734b1_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__468538d3c2aea2868454af7c70e616ef531f1eb57948eec4bff4ae93abdbaba3/preprint__468538d3c2aea2868454af7c70e616ef531f1eb57948eec4bff4ae93abdbaba3_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__46f8225776634ac8d37f18dc2239d5e80c2d86fb9f163fa79002408b3beec78a/preprint__46f8225776634ac8d37f18dc2239d5e80c2d86fb9f163fa79002408b3beec78a_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__46e74ba786ecb86ff45b271c15118553dbc62d7953da4b92ed1ae631d19f3d92/preprint__46e74ba786ecb86ff45b271c15118553dbc62d7953da4b92ed1ae631d19f3d92_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__46ce64b3b34a664743f5eef77ee4c2fea25fa579f31ae4df63f8aa5e71375807/preprint__46ce64b3b34a664743f5eef77ee4c2fea25fa579f31ae4df63f8aa5e71375807_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__464166c7614ce71840865ad4fdf5ff6fab4d616871595069cfe6aee6eb009000/preprint__464166c7614ce71840865ad4fdf5ff6fab4d616871595069cfe6aee6eb009000_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4729d651d9bc3f79db56669b6ae29aeaf6369592caa95533432a672b8cd99b2e/preprint__4729d651d9bc3f79db56669b6ae29aeaf6369592caa95533432a672b8cd99b2e_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__475a995b6c3d17751d969c5c3205dfef8a97553e39a974721d15f3f6002b781a/preprint__475a995b6c3d17751d969c5c3205dfef8a97553e39a974721d15f3f6002b781a_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__47ab6ca0792629f9715dc12d6fe1d83a450d00e3b61202ec7f238b1279ceda78/preprint__47ab6ca0792629f9715dc12d6fe1d83a450d00e3b61202ec7f238b1279ceda78_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4766e2f4895c6e61a3829ea1ab6c7459322f57c53071d0c646ba70feb8da723d/preprint__4766e2f4895c6e61a3829ea1ab6c7459322f57c53071d0c646ba70feb8da723d_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__470832070d078d21bc9cd0d6bedf7ee4625f667c0ee705bea255c22adc8392a2/preprint__470832070d078d21bc9cd0d6bedf7ee4625f667c0ee705bea255c22adc8392a2_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__480f3b232d2c623f4f1c4e7f68c215376692cfa37c148324e0a6ccb5cfdf0b05/preprint__480f3b232d2c623f4f1c4e7f68c215376692cfa37c148324e0a6ccb5cfdf0b05_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__476d34b5686337db26de5c8c1e64d86058ec7527a2ac502b29c765017df714fa/preprint__476d34b5686337db26de5c8c1e64d86058ec7527a2ac502b29c765017df714fa_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__470c984db0b52a3daf119787d0aa0ca233371b8e7ffee24badd95bed1895219a/preprint__470c984db0b52a3daf119787d0aa0ca233371b8e7ffee24badd95bed1895219a_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4776ba26c85422c733956275893db4f5cc2c44e398c8f6918e879643059eafc7/preprint__4776ba26c85422c733956275893db4f5cc2c44e398c8f6918e879643059eafc7_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__480fe05899d1520cb4b6e72a4a2f2de5769529009fcec55abe3ac7fda0ce6eaa/preprint__480fe05899d1520cb4b6e72a4a2f2de5769529009fcec55abe3ac7fda0ce6eaa_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__486c1efee492ebae01b70be8d249ce174803a4e1d880fe55bf307bc77f5d24e2/preprint__486c1efee492ebae01b70be8d249ce174803a4e1d880fe55bf307bc77f5d24e2_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4819fa60e2ce209b34a157450ed28888546a35fcfa4cdfb7ed96c2ae1ed60b19/preprint__4819fa60e2ce209b34a157450ed28888546a35fcfa4cdfb7ed96c2ae1ed60b19_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4844999f9c48fab95f677e12ff57f18204d1027644dc64a7ffea5ccbbcee4859/preprint__4844999f9c48fab95f677e12ff57f18204d1027644dc64a7ffea5ccbbcee4859_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__483a7c25f93b46e958c7647bac54d6284eb0e0588af3ad7439c754c5ccd549c1/preprint__483a7c25f93b46e958c7647bac54d6284eb0e0588af3ad7439c754c5ccd549c1_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__47648711f4c7bd454a6688fa978ba86510c475dee95543be94e19975642463f2/preprint__47648711f4c7bd454a6688fa978ba86510c475dee95543be94e19975642463f2_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__485373ce90d30c4730644ea71cd7df747a03d71497007b34dcac392a1a08f4a3/preprint__485373ce90d30c4730644ea71cd7df747a03d71497007b34dcac392a1a08f4a3_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__47ed76f73bd4fdabb270d5144335729ac54d53f99282c632c26d276655ed3abc/preprint__47ed76f73bd4fdabb270d5144335729ac54d53f99282c632c26d276655ed3abc_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__481e74eddbc811fda69ed68e86d7c0484896cd9e4768cbe36b0ddb638b669048/preprint__481e74eddbc811fda69ed68e86d7c0484896cd9e4768cbe36b0ddb638b669048_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__48961379f14baa7d8ee3a9307154d8bf615c946407f087937d7589936c43fad7/preprint__48961379f14baa7d8ee3a9307154d8bf615c946407f087937d7589936c43fad7_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__491c12817894379960e50a02a9c4da0c41dd179e729ff4fd51850bf7dd1a6580/preprint__491c12817894379960e50a02a9c4da0c41dd179e729ff4fd51850bf7dd1a6580_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4d070a94e98d044c38c8dc00ba8ee1398286cbe3174cef7d5bbef45c1ae097b6/preprint__4d070a94e98d044c38c8dc00ba8ee1398286cbe3174cef7d5bbef45c1ae097b6_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4ce2f9d621aca8295f6d6b72cd400088a0542527d15c7016d8ced35f16204951/preprint__4ce2f9d621aca8295f6d6b72cd400088a0542527d15c7016d8ced35f16204951_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4cb3ef22f4e83005d8b129fcfcf6dcedecca443fde82f0edb21bfe8de2d48f38/preprint__4cb3ef22f4e83005d8b129fcfcf6dcedecca443fde82f0edb21bfe8de2d48f38_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4c256191cecaab5089f417b929c81d599c8fa875ed06c3066a9cafa714df38b0/preprint__4c256191cecaab5089f417b929c81d599c8fa875ed06c3066a9cafa714df38b0_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4d0b72b215f47540c00d43254f05ea5257370e936a593adc3a345a5e880b8cfb/preprint__4d0b72b215f47540c00d43254f05ea5257370e936a593adc3a345a5e880b8cfb_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4d79fb021050680bdb93bab9de12ec6bc5f441c6cbcb77e7118f95f79311981b/preprint__4d79fb021050680bdb93bab9de12ec6bc5f441c6cbcb77e7118f95f79311981b_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4d9d847dadfe7fbfac248b53be7c97532a519b9276157d3652ab7e1981f432bb/preprint__4d9d847dadfe7fbfac248b53be7c97532a519b9276157d3652ab7e1981f432bb_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4d70ac2a0d5c788f11e79c4fd5ab497ed911acc17ee68059aa0806750d3e06d4/preprint__4d70ac2a0d5c788f11e79c4fd5ab497ed911acc17ee68059aa0806750d3e06d4_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4da78f37130ddd04233fe6c9c0c1004583ddf85a00c3d657add7c0ee35632e0f/preprint__4da78f37130ddd04233fe6c9c0c1004583ddf85a00c3d657add7c0ee35632e0f_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4e5be71ace5bab4eeb4fe6cdcf235ce6640c18e37ddd4ebbc511e0cf238434d6/preprint__4e5be71ace5bab4eeb4fe6cdcf235ce6640c18e37ddd4ebbc511e0cf238434d6_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4e03f3f5e88bf396d3599ff96036d6e740e80f50453e4796f91ef91cbaad6f65/preprint__4e03f3f5e88bf396d3599ff96036d6e740e80f50453e4796f91ef91cbaad6f65_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4e4a848992cc6f4f19b227182accd709f88b94ee21297d25cd65326d17e6ca20/preprint__4e4a848992cc6f4f19b227182accd709f88b94ee21297d25cd65326d17e6ca20_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4e5035952f4492d0000a619da37d2cfff02c336c5602037482b79617e7f068ca/preprint__4e5035952f4492d0000a619da37d2cfff02c336c5602037482b79617e7f068ca_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4e7e7cd0b26c25c706aaa28d49aeb326467fe52d66961bfac897ef1e9f88c8b8/preprint__4e7e7cd0b26c25c706aaa28d49aeb326467fe52d66961bfac897ef1e9f88c8b8_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4e54bc2068e78b98e4b080a53e760fcfa37d4645d269ec1bb31faa57532ff75b/preprint__4e54bc2068e78b98e4b080a53e760fcfa37d4645d269ec1bb31faa57532ff75b_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4ec18687be9919fc1de5ae25a5e6560952e41f3a864533abe3fe3796f91dc08d/preprint__4ec18687be9919fc1de5ae25a5e6560952e41f3a864533abe3fe3796f91dc08d_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4f263872a2fd160a6114aed126f24d2794e5ffbb7055d60f1754ad52247fe9eb/preprint__4f263872a2fd160a6114aed126f24d2794e5ffbb7055d60f1754ad52247fe9eb_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4ff19a920dd4e831c245a2d5426e5b535538b015063016d9ae744260a1735310/preprint__4ff19a920dd4e831c245a2d5426e5b535538b015063016d9ae744260a1735310_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4f40b76b54027d52a076b375a067fc5c8342f20d1eaaac10ba9dc1bc81c7e3d5/preprint__4f40b76b54027d52a076b375a067fc5c8342f20d1eaaac10ba9dc1bc81c7e3d5_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4e94a0d01e50c4cdcf80023b216df16a108434142b51e933f5569cf2fe3d0d52/preprint__4e94a0d01e50c4cdcf80023b216df16a108434142b51e933f5569cf2fe3d0d52_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__50368ac3866e301b878863107e83ced6677b15bf10b96f214d076af3e340dbd0/preprint__50368ac3866e301b878863107e83ced6677b15bf10b96f214d076af3e340dbd0_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4fddf41d1eef1d1a06ff64e628d1c7c64af25d3ffb7dc614bc0b4de5fba9d419/preprint__4fddf41d1eef1d1a06ff64e628d1c7c64af25d3ffb7dc614bc0b4de5fba9d419_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4e87dfe716a1e58488d95ba07d3b2c47541da4647c40e4d9d5f57ab4f2cab83e/preprint__4e87dfe716a1e58488d95ba07d3b2c47541da4647c40e4d9d5f57ab4f2cab83e_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4f598871c3e4850af0af011ccf79c62164f1f9fef4885ca48e396f67f178329c/preprint__4f598871c3e4850af0af011ccf79c62164f1f9fef4885ca48e396f67f178329c_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4ffec46a419eec0c0da68d2d7a45fe5a699b40590987fd5440d5a849c33b8887/preprint__4ffec46a419eec0c0da68d2d7a45fe5a699b40590987fd5440d5a849c33b8887_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4e1bf66d5e6a1b351dc2cd0d447d9af3bb23e981795b493729c154d5245d911a/preprint__4e1bf66d5e6a1b351dc2cd0d447d9af3bb23e981795b493729c154d5245d911a_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4841b1fcf83ab880bb3afb296fa2b3ae0c5a97728da9b5ceacdd382aa94e3ebd/preprint__4841b1fcf83ab880bb3afb296fa2b3ae0c5a97728da9b5ceacdd382aa94e3ebd_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__488dd5e7f45f267645d4ecafeb0063d839323925821422c8cf4a42428a5d1df4/preprint__488dd5e7f45f267645d4ecafeb0063d839323925821422c8cf4a42428a5d1df4_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__487268b60d60a57ddf72c596fba8ead5ff013eeab984aee8c25ad6580afe57e1/preprint__487268b60d60a57ddf72c596fba8ead5ff013eeab984aee8c25ad6580afe57e1_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__48fdc4dc3cce1523cd0346c874c274f187408b5762031fb3d036d6b4eed72790/preprint__48fdc4dc3cce1523cd0346c874c274f187408b5762031fb3d036d6b4eed72790_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__492c2e2da2205da347da47f3bd48120eebbc4d54f4a2561ed8178325a9f67295/preprint__492c2e2da2205da347da47f3bd48120eebbc4d54f4a2561ed8178325a9f67295_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4917d2ec82456765310e4fd0e81bd9c6564f10de32d29557ebc886beb9b7789e/preprint__4917d2ec82456765310e4fd0e81bd9c6564f10de32d29557ebc886beb9b7789e_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__487dd727461ac5ce1675c5210f0d2bb09de73dbbde2a7e84d0c10a7c04a6fbfd/preprint__487dd727461ac5ce1675c5210f0d2bb09de73dbbde2a7e84d0c10a7c04a6fbfd_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__48de5430c259c6879fe0e3cc8610c93e4199c05191d114610b87fb8ee3f4f045/preprint__48de5430c259c6879fe0e3cc8610c93e4199c05191d114610b87fb8ee3f4f045_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__494df694103eb21d7d8b6b6f23f85841f9345d4d21a755d4cc450568c8886572/preprint__494df694103eb21d7d8b6b6f23f85841f9345d4d21a755d4cc450568c8886572_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__48b464b07d1a9fb2491f3443e735e9f38e6b8abe90e64efd27de8474ea209aaf/preprint__48b464b07d1a9fb2491f3443e735e9f38e6b8abe90e64efd27de8474ea209aaf_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4972e0ad04103763b904c933c377b3165204b3cb580d558632a39a982760b49d/preprint__4972e0ad04103763b904c933c377b3165204b3cb580d558632a39a982760b49d_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__498fd5c2917d09467144a0e7e2f07487cf520745f96cfa0378c448526a87f516/preprint__498fd5c2917d09467144a0e7e2f07487cf520745f96cfa0378c448526a87f516_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__492f4e7d71f2240b0f75b1158e8039beec54b3b758cf0e997ee76ce7c0176923/preprint__492f4e7d71f2240b0f75b1158e8039beec54b3b758cf0e997ee76ce7c0176923_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__49cce96390898a8bd75d5e488fcc64159b662630a8e87557765d2c3200c3c4a2/preprint__49cce96390898a8bd75d5e488fcc64159b662630a8e87557765d2c3200c3c4a2_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4960479d174da1f29727dcd759d659b303a9ccc37c7af432394097be8c4b00a2/preprint__4960479d174da1f29727dcd759d659b303a9ccc37c7af432394097be8c4b00a2_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4a0d9ab801df4bc282768981f75bca37f9f1fed6c762c8d12980f643e4142502/preprint__4a0d9ab801df4bc282768981f75bca37f9f1fed6c762c8d12980f643e4142502_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4997c403ccbdf63e34792b3427909377712af8591f37fed23a643d8e1d3c6575/preprint__4997c403ccbdf63e34792b3427909377712af8591f37fed23a643d8e1d3c6575_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__491dd82e431bd9db89911af3357d8436b6d13fcb4ae9cd9d6e41df61934cee6d/preprint__491dd82e431bd9db89911af3357d8436b6d13fcb4ae9cd9d6e41df61934cee6d_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4a167c7e692e546129e9a4ce1df46c5d312fd6f76aecbb43cb3d8f227702a53b/preprint__4a167c7e692e546129e9a4ce1df46c5d312fd6f76aecbb43cb3d8f227702a53b_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__49734382434ae8d95e5356a6547f15d6b7e4d786ecf1d052537abb1c3502f80a/preprint__49734382434ae8d95e5356a6547f15d6b7e4d786ecf1d052537abb1c3502f80a_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__49ae9c421de8e97b50ba8ab5843915e96f6994ff78d1456b8bfa263174de9070/preprint__49ae9c421de8e97b50ba8ab5843915e96f6994ff78d1456b8bfa263174de9070_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__49e7ce1d267093271a3a5cbe1a2ba5ccdc9494afc41ef0d65af6e91215a2df71/preprint__49e7ce1d267093271a3a5cbe1a2ba5ccdc9494afc41ef0d65af6e91215a2df71_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4aa17d8e584d3ecad03dd62afb4f8dd955d6793e50db788cc4ee4d0445cbbdd5/preprint__4aa17d8e584d3ecad03dd62afb4f8dd955d6793e50db788cc4ee4d0445cbbdd5_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4a9ca118232a7edbf3d10bcce79c90c94afcf2db8e4c5b282c603fb2ae2f2a80/preprint__4a9ca118232a7edbf3d10bcce79c90c94afcf2db8e4c5b282c603fb2ae2f2a80_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__497d2d0bb5b25d54da2c1d0e465c103ea8853351a764d5075b67bfadab276049/preprint__497d2d0bb5b25d54da2c1d0e465c103ea8853351a764d5075b67bfadab276049_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4aa9bc2ceb510b05dd9418739af000cf42ed874dbdaab74077f98f8cbe0df170/preprint__4aa9bc2ceb510b05dd9418739af000cf42ed874dbdaab74077f98f8cbe0df170_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4b4b25c4c607211c9e5893b47a3971be09042f56c3bafa6baff0706b9da18911/preprint__4b4b25c4c607211c9e5893b47a3971be09042f56c3bafa6baff0706b9da18911_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4b7e1ccdb2df650fb756f467b343474aeaebd8d793961920b32ea4a33be9bacd/preprint__4b7e1ccdb2df650fb756f467b343474aeaebd8d793961920b32ea4a33be9bacd_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4b6207b1b6e3e00fd78241e94465c4f0533e8a521023fc51c52b9e42d71c7ec6/preprint__4b6207b1b6e3e00fd78241e94465c4f0533e8a521023fc51c52b9e42d71c7ec6_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4b85599767ba20cada880da264fd2e36bc19aa824f35c6dcf80f838e96c5caa2/preprint__4b85599767ba20cada880da264fd2e36bc19aa824f35c6dcf80f838e96c5caa2_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4b7c88bbedec50765fb40bbba0db265be8a99905388fdba94abf500d3978b2e0/preprint__4b7c88bbedec50765fb40bbba0db265be8a99905388fdba94abf500d3978b2e0_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4b4c750c607343569402a16b901303aa2e3f8c68cac4f4b48001af5f7147a94b/preprint__4b4c750c607343569402a16b901303aa2e3f8c68cac4f4b48001af5f7147a94b_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__49914922f8e5b5facb404599c10560c52c19daa2dd35d42dc738abadf3d8688b/preprint__49914922f8e5b5facb404599c10560c52c19daa2dd35d42dc738abadf3d8688b_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4ba9a92303b664366f73f0ae958a8568c2b488177b6ecf2f1e414337c58d6632/preprint__4ba9a92303b664366f73f0ae958a8568c2b488177b6ecf2f1e414337c58d6632_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4a284858b92ba57449505dc9844a88f53191d9601ff8a411bd8b8b6ea1ec380d/preprint__4a284858b92ba57449505dc9844a88f53191d9601ff8a411bd8b8b6ea1ec380d_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4bad90e156cd7674a1ed298f896e33bdde0feca2ff771177791c36cf30c0d798/preprint__4bad90e156cd7674a1ed298f896e33bdde0feca2ff771177791c36cf30c0d798_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4b3a66a4579d1623dd4fe8f22d75df21486176a2f353b14fba285703f8e16145/preprint__4b3a66a4579d1623dd4fe8f22d75df21486176a2f353b14fba285703f8e16145_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4b885056629a835494bd68e96970ceb6998489b7dde8c55c90089c133450ced3/preprint__4b885056629a835494bd68e96970ceb6998489b7dde8c55c90089c133450ced3_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4bad99da6edb5275abb0dfb383ec12ab0d9aa5d140b700116c07d5b2e982d11a/preprint__4bad99da6edb5275abb0dfb383ec12ab0d9aa5d140b700116c07d5b2e982d11a_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4beeafdab5ca185d3eec0e3fc33bf21ca59c35b59844e4107d8f1ac53e96b4ca/preprint__4beeafdab5ca185d3eec0e3fc33bf21ca59c35b59844e4107d8f1ac53e96b4ca_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4affc2f1e2a2106729bed6846e7c1dcde0d1d8176759b048413f89056fc43f23/preprint__4affc2f1e2a2106729bed6846e7c1dcde0d1d8176759b048413f89056fc43f23_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4bf431b6506b56cf089e4c8c8d34e971d8739a75fcbef213d5ad6515162061aa/preprint__4bf431b6506b56cf089e4c8c8d34e971d8739a75fcbef213d5ad6515162061aa_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4b8bc4f8afb1bb1bfc32abc11059aabb40f336b781c58949a04852118844514a/preprint__4b8bc4f8afb1bb1bfc32abc11059aabb40f336b781c58949a04852118844514a_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4bb83cebfeef9653c529247dfda89d9f5d1955eeb87106dcbe2918c878a00766/preprint__4bb83cebfeef9653c529247dfda89d9f5d1955eeb87106dcbe2918c878a00766_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4cf534931c7f6d43bf9bc1d952950a6a5a0883109f8849da2c222540e2700ec3/preprint__4cf534931c7f6d43bf9bc1d952950a6a5a0883109f8849da2c222540e2700ec3_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4cadb082c6a83b30d03e558acb1848a2dc2dd61684114e47a32c4356640e3c07/preprint__4cadb082c6a83b30d03e558acb1848a2dc2dd61684114e47a32c4356640e3c07_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4cea15a0e0be552d6ef41196ffedbffaf9922bdafac31dc04485b46b987f9f51/preprint__4cea15a0e0be552d6ef41196ffedbffaf9922bdafac31dc04485b46b987f9f51_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4c902ed890e3f1190db3f1844582aa1ad2b3d1c98441ed1414ab9727292183a7/preprint__4c902ed890e3f1190db3f1844582aa1ad2b3d1c98441ed1414ab9727292183a7_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4c1a32eb7bfc2ac8add83f6e8018407e752e6b3d0d32318da291388c989adbf6/preprint__4c1a32eb7bfc2ac8add83f6e8018407e752e6b3d0d32318da291388c989adbf6_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4bdad8e4dcc966c1c61df987a6be6c34ab176fd113ddb750db47957cd1c7573b/preprint__4bdad8e4dcc966c1c61df987a6be6c34ab176fd113ddb750db47957cd1c7573b_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__501c283c7a4f4fb91a9d990f610e4c87a4c45628c1711acffd1c9100060c5828/preprint__501c283c7a4f4fb91a9d990f610e4c87a4c45628c1711acffd1c9100060c5828_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__504089f7e0d794a97954804d468c2685ebb1d46b6c4ba3b6b40579c69707ac82/preprint__504089f7e0d794a97954804d468c2685ebb1d46b6c4ba3b6b40579c69707ac82_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4ff71ab05187a480b15d98adc388fd242e9d69f37974c5be3606cab4fc511458/preprint__4ff71ab05187a480b15d98adc388fd242e9d69f37974c5be3606cab4fc511458_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4fdd27fd21b36138c9c66108914bd00f0ddf57e4133570a629417c14ff7fed9d/preprint__4fdd27fd21b36138c9c66108914bd00f0ddf57e4133570a629417c14ff7fed9d_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__508712a3e27d7adf20378db70d6d03c536940fdd1834d79298bd85f8c095e3e9/preprint__508712a3e27d7adf20378db70d6d03c536940fdd1834d79298bd85f8c095e3e9_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__508bbbd43d46eb8268292a204fa36e2d2fe0e7b78f8e67c5aace7c2a1aae55fc/preprint__508bbbd43d46eb8268292a204fa36e2d2fe0e7b78f8e67c5aace7c2a1aae55fc_layouts.pdf filter=lfs diff=lfs merge=lfs -text
+preprint/preprint__4ffb7662bd5443653eee4e5bd2bd42044e10cd6a630d0b46c0e4fa72d9f94e34/preprint__4ffb7662bd5443653eee4e5bd2bd42044e10cd6a630d0b46c0e4fa72d9f94e34_layouts.pdf filter=lfs diff=lfs merge=lfs -text

peer_reviews/supplementary_0_Transparent Peer Review file__6d1f1284a3af25f662a14484e7f1e7495d11a0328d008eb6ec4dba657744c3fc/images_list.json

@@ -0,0 +1,153 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_0.jpg",
+    "caption": "Figure S5. NMIIi induces neurite outgrowth in dissociated RGCs",
+    "footnote": [],
+    "bbox": [
+      [
+        113,
+        141,
+        600,
+        293
+      ]
+    ],
+    "page_idx": 8
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_1.jpg",
+    "caption": "Figure S10. Assessment of non-neuronal cells at sciatic nerve injury site after NMII treatment",
+    "footnote": [],
+    "bbox": [
+      [
+        117,
+        92,
+        777,
+        352
+      ]
+    ],
+    "page_idx": 9
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_6.jpg",
+    "caption": "Figure 6. NMIIi applied at the sciatic nerve injury site promotes sensory axon regeneration and functional recovery",
+    "footnote": [],
+    "bbox": [
+      [
+        117,
+        192,
+        688,
+        776
+      ]
+    ],
+    "page_idx": 10
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_2.jpg",
+    "caption": "Figure (not included in MS). Evaluation of motor function recovery with 10 mM and 25 mM doses of NMIIi at the sciatic nerve injury site.",
+    "footnote": [],
+    "bbox": [],
+    "page_idx": 12
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_5.jpg",
+    "caption": "Figure 5. NMll at sciatic nerve injury site accelerates motor function recovery in mice",
+    "footnote": [],
+    "bbox": [
+      [
+        130,
+        686,
+        860,
+        864
+      ]
+    ],
+    "page_idx": 12
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_3.jpg",
+    "caption": "Figure S8. Neuromuscular junctions days 7 and 28 post-sciatic nerve crush injury Representative images of lumbrical muscles at day 7 (A) and day 28 (B) post-crush injury from mice treated with local application of either vehicle or NMII, labeled with anti-neurofilament (NF), anti-synaptophysin (SYP), or \\(\\alpha\\) -bunqarotoxin (BTX). (C) Quantification of presynaptic markers (NF; SYP) overlapping with postsynaptic marker (BTX), normalized to the number of postsynaptic sites. Statistics: one-way ANOVA with Tukey's post hoc test, SEM error bars. Data collected from 5 mice per group on day 7, and 4 mice per group on day 28.",
+    "footnote": [],
+    "bbox": [
+      [
+        115,
+        93,
+        812,
+        465
+      ]
+    ],
+    "page_idx": 14
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_4.jpg",
+    "caption": "Figure (not included in MS). Plasma pharmacokinetic parameters of NMIIi2 following IP administration at doses of 3, 10, and \\(30mg / kg\\) in male mice",
+    "footnote": [],
+    "bbox": [
+      [
+        123,
+        656,
+        437,
+        831
+      ]
+    ],
+    "page_idx": 16
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_5.jpg",
+    "caption": "Figure (not included in MS). NMIIi treatment does not decrease MYH9 level",
+    "footnote": [],
+    "bbox": [
+      [
+        115,
+        240,
+        377,
+        415
+      ]
+    ],
+    "page_idx": 17
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_6.jpg",
+    "caption": "Figure S8. Evaluation of the hydrogel's intrinsic regenerative effects after nerve injury (D) Quantification of luminance ratio of left/right paws post sciatic nerve crush with and without hydrogel. Data collected from 8 mice. Stats: two-way ANOVA with Tukey's post hoc test, SEM error bars",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        551,
+        610,
+        730
+      ]
+    ],
+    "page_idx": 21
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_6E.jpg",
+    "caption": "Figure 6E. Pearson correlation analysis of the paw luminance ratio and pinprick behavior scores on day 14 post-injury. Data collected from 10 male mice per group.",
+    "footnote": [],
+    "bbox": [
+      [
+        130,
+        99,
+        395,
+        280
+      ]
+    ],
+    "page_idx": 22
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_7.jpg",
+    "caption": "Fig. S7A. Regenerating axons in distal nerve 7 days post-sciatic nerve crush injury (A) Representative images of mouse sciatic nerves labeled with an antibody against Neurofilament (NF) and Choline Acetyltransferase (ChAT) on day 7 post-sciatic nerve crush injury. White dashed lines indicate the boundary of the crush injury site, and arrowhead indicates tips of regenerated axons.",
+    "footnote": [],
+    "bbox": [],
+    "page_idx": 24
+  }
+]

peer_reviews/supplementary_0_Transparent Peer Review file__6d1f1284a3af25f662a14484e7f1e7495d11a0328d008eb6ec4dba657744c3fc/supplementary_0_Transparent Peer Review file__6d1f1284a3af25f662a14484e7f1e7495d11a0328d008eb6ec4dba657744c3fc.mmd

@@ -0,0 +1,593 @@
+
+# nature portfolio  
+
+Peer Review File  
+
+# NON-MUSCLE MYOSIN II INHIBITION AT THE SITE OF AXON INJURY INCREASES AXON REGENERATION  
+
+Corresponding Author: Professor Clifford Woolf  
+
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+Version 0:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author)  
+
+Heo and colleagues explored the effect of a blebbistatin analogue on axon growth and nerve regeneration both in vitro and in vivo. After screening over 4000 compounds, they identified 10 main candidates for improving axon outgrowth, focusing on blebbistatin, a known NMII inhibitor. However, due to its poor biocompatibility and its photosensitivity, they proceeded with an analogue called NMII2. They tested the drug in vitro on various cellular models and in vivo in a mouse model of sciatic nerve regeneration following a crush injury.  
+
+They evaluated neurite elongation in vitro and in vivo and conducted studies on functional recovery.  
+
+The work is well- structured and thoughtfully designed and however the novelty is limited given that the effect of blebbistatin in neuronal outgrowth has been described and the proposed mechanism is not surprising and not very well developed. The regenerative phenotype observed is statistically significant but not especially exciting compared to other approaches.  
+
+Major points  
+
+- The choice to include the MYH9 knockdown model in this manuscript is unclear. It has long been known that inhibition of NMII leads to increased axonal growth through a well-documented mechanism. Validation of this model in the present article adds little to the existing literature or to the quality of this manuscript  
+
+The mechanism has not been well studied neither at the cellular nor signalling level. Cell-specific unbiased approaches could reveal novel interesting mechanisms.  
+
+- It is unclear why the authors tested non-permissive substrates in culture to mimic CNS regeneration and then they focused on peripheral regeneration in vivo. The authors should clarify this aspect. Did they test regeneration in the CNS? This would add to the manuscript's impact.  
+
+- With the regards to the in vivo drug delivery exp, which represent the main novelty of the work, there are several aspects to consider:  
+
+i) Local delivery of the NMII inhibitor might affect not only the injured axons but also the entire non-neuronal component present in the sciatic nerve. Since this is a chronic treatment (over several days), there could be side effects on glial or immune cells that need to be considered, especially since these could potentially lead to systemic issues or beneficial effects. These cell populations should be evaluated, at least through histochemistry, to assess their condition. ii) Similarly, while the machine learning method is powerful, it is not sufficient to demonstrate complete and physiological functional recovery. In the peripheral system, there are various neuronal populations involved at different levels in sensation and perception, which should be analyzed independently.  
+
+Additional points  
+
+- The phenotypic screening should be implemented with more procedural details. Specifically:  
+
+i) Why were two different cell lines used between the first and second screen? The authors should clarify this in the text. ii) Some of these compounds may cause toxicity that is not necessarily reflected in neurite length but in cell survival. Was neuronal survival assessed at the concentrations used?
+
+<--- Page Split --->
+
+- The authors should clarify how the concentration of \(100 \mu M\) was established for in vivo use, given that the drug is being tested for the first time to test nerve regeneration.  
+
+- There are many mechanisms that have been studied and could underlie the results observed by the authors in this work, beyond the Rho/ROCK pathway (such as increased microtubule stability, alterations in interactions with the environment via adhesion molecules, growth cone dynamics, changes in intrinsic forces, etc.). The authors should, even briefly, expand the discussion to include these aspects  
+
+- The authors should briefly explain why silicone tubes were chosen for local drug application as opposed to other methods.  
+
+(Remarks on code availability)  
+
+## Reviewer #2  
+
+(Remarks to the Author)  
+
+- What are the noteworthy results?  
+
+Heo et al. in the work entitled "NON- MUSCLE MYOSIN II INHIBITION AT THE SITE OF AXON INJURY INCREASES AXON REGENERATION" employed an unbiased screen of axonal growth promotors on a growth inhibitory environment (CSPG) and found several compounds that increased the outgrowth of human iPSC derived motor neurons. Of these, the highest hit was blebbistatin, which inhibits NMII. Although used previously in similar conditions as shown here, this work examined a new NMII2 that is more bioavailable than previous versions. This was shown to work to extend dendrites proximal and axons distal to the lesion site following laser dissection of spot cultures. More importantly, this work showed for the first time in vivo local NMII2 treatment of the injured sciatic nerve leads to increased regeneration within the nerve and the muscle of the foot, increased synaptic connections and functional recovery.  
+
+- Will the work be of significance to the field and related fields? How does it compare to the established literature? If the work is not original, please provide relevant references.  
+
+Blebbistatin and NMII inhibition have been found previously to be involved in axonal growth of human NPCs (10.3389/fcell.2021.719636), RGCs and the optic nerve (10.1016/j.celrep.2020.107537), and embryonic DRG (10.1073/pnas.1011258108). Here for the first time a new NMII2 that is more bioavailable than previous versions was examined and found to promote axonal regeneration when placed at the site of injury, the sciatic nerve, leading to faster functional recovery. This work takes the field into clinical relevance. It can also be noted that many of the in vitro models are typically considered immature neurons (which tend to be more growth-permissive) except for the adult DRG work, so the in vivo work shows that in a mature neuronal state, the NMII2 works effectively.  
+
+- Does the work support the conclusions and claims, or is additional evidence needed? This work states that NMMII works well when given at the lesion site inducing accelerated regeneration and functional recovery. However, one could strengthen these conclusions with some additions. Firstly, in Figure 6 the co-localization is difficult to see with current magnifications. Why are the NF and Syn in the same color? Please provide higher magnifications of the co-localization analysis and have the antibodies split into separate channels/colors along with the merged overlap. This analysis was from D14 when the greatest behavioral differences were seen. It should also be provided at the end timepoint (D28) and possibly from D7 where no behavioral differences are observed. The assumption would be that there would be less colocalization overall at D7 because there wouldn't be enough growth into the muscle by this timepoint. As for the D28 at the end of the study where the behavior is the same between the treated and vehicle groups it would be important to know if the growth and co-localization are the same or if this is due to compensatory mechanisms that arise in the vehicle group. This would further support the correlation graph in Figure 6, which is a little weak given that two of the five samples sit at zero for colocalization yet have very different behavioral outcomes. Is the correlation work from D14? This should be stated somewhere. Further, it should be spoken about why this work is important given there is functional recovery at the end of the study similar to the treated group. What is the benefit of the accelerated recovery? In the introduction it is stated, "After a traumatic nerve injury, mature motor neurons only regenerate at a rate of 1-4 mm per day, the slowness of which results in only \(\sim 45\%\) of injured axons achieving anatomical reinnervation of their muscle targets with, in consequence, limited functional recovery (1-7). Discovering therapeutic targets for the effective treatment of axon injury is therefore critical to ensure a more complete recovery of motor and other disturbed functions." However, it appears that in Figure 6 the controls reach the same level of functional recovery as the treated group at D28. This may just be a difference between rodents and humans, either way this should be discussed. Some of the additions suggested here may help address this issue. Given the work done to show that NMII works in various cell types it would make sense to also look at the sciatic nerves shown in Fig. S6 for sensory neuronal regeneration with SCG10. This would take the in vitro work into the in vivo relevance. As there is not any increased mechanistic understanding greater than what was published before, I do believe these additions are important to show the novelty and relevance of this work.  
+
+Lastly, it should be stated why \(10mg / kg\) of NMII2 was used systemically and whether there were any obvious things that were associated with weight loss, such as a decrease in organ size or discoloration. Also how long is it thought for the hydrogel release of NMII2 at the lesion site, immediate or over days, weeks? This will be of interest to the readers.  
+
+- Are there any flaws in the data analysis, interpretation and conclusions? - Do these prohibit publication or require revision? I do not see any flaws just some additions that would strengthen the conclusions as I stated above.  
+
+- Is the methodology sound? Does the work meet the expected standards in your field?
+
+<--- Page Split --->
+
+The methodology is the standard of the field.  
+
+- Is there enough detail provided in the methods for the work to be reproduced? There are some minor additions needed in the methods: 
+- The correlation analysis for Figure 6 is not listed nor is it stated if this is a Pearson or Spermann correlation. 
+- The post hoc tests are not listed by name, it would be helpful if the authors listed if they chose a more conservative or liberal post hoc analysis method. 
+- \(10\mathrm{mg / kg}\) twice a day for 3 days via i.p. of the NMMlll was used for systemic application, but why was this dose/timing/method chosen? This should be listed given the side effects that were observed.  
+
+(Remarks on code availability) I could not access the code when I tried to click on the GitHub link.  
+
+Reviewer #3  
+
+(Remarks to the Author)  
+
+The manuscript entitled "Non- muscle myosin II inhibition at the site of axon injury increases axon regeneration" is a study in which the authors produce a novel myosin II inhibitor (NMIli2; analog of blebbistatin) and show it's efficacy in induce axon growth in tissue culture and after peripheral nerve injury in adult mice.  
+
+Following a phenotypic screen in human iPSC derived motor neurons, the authors identified 10 compounds that induced neurite outgrowth on a CSPG extracellular matrix environment. Blebbistatin was found to be the strongest inducer of growth, with and without laser cut injury, however given it's issues with photosensitivity and bioavailability, the authors produced a more selective non- muscle myosin II inhibitor, termed NMIli2. As non- muscle myosin II is a motor protein that normally stabilises microtubules, use of NMIli2 would be hypothesised to destabilise MTs thus inducing axon elongation/extension.  
+
+To confirm the findings from their screen as well as to confirm the efficacy of NMIli2, the authors grew human iPSC- derived motor neurons, cortical- like neurons, and human cortical organoids on either laminin or CSPGs; as well as adult mouse DRGs and RGCs on laminin only. The authors also compared their findings with NMIli2 with a lentiviral knockdown of MYH9 (encoding the heavy chain of NMIli)in human iPSC- derived motor neurons, demonstrating axon growth with the KD with confirmed KD \((- 55\%)\) with a western blot. The authors then tested NMIli2 in vivo in a mouse model of peripheral nerve injury, showing that local/direct administration complexed to a collagen hydrogel, compared with intraperitoneal administration or control, was effective at inducing ChAT- labelled axon growth, synaptic reinnervation and better hindpaw plantar placement.  
+
+While the findings of this study are very interesting for the field of axon regeneration specifically in the PNS, there are some areas of the study that require clarification and/or further information.  
+
+Major questions:  
+
+- Why were the mouse cellular experiments using adult RGCs and adult DRGs not performed on a CSPG matrix? CSPG substrates were used in the motor and cortical-like neuron experiments with success. All of these cell types have been shown to respond to CSPGs with retraction and/or overall stunted growth, especially in adult ages. Some studies have shown that the lack of growth on these substrates is due to integrin are downregulated and or subsequent inactivation, leaving adult cells unable to respond to the CSPG matrix. Given the mechanism of action of destabilising MTs, do the authors think that NMIli2 would induce the growth in the mouse cells on CSPG regardless of integrin regulation? If not, can the authors discuss the mechanism of action in the CNS neurons that differs in the PNS neurons?  
+
+- The axon growth in vitro suggests that blebbistatin and NMIli2 both induce neurite/axon elongation as well as branching. The branching effect is indeed an interesting finding, which would in some cases compete with elongation. Can the authors discuss the mechanisms of action by which the compound can drive both an elongation and a branching phenotype?  
+
+- In the MYH9 experiments, a western blot was performed to indicate knockdown of MYH9. Have the authors looked at changes in MYH9 in their NMIli2 treated cells? Would they expect a greater effect, or a less specific effect? Can the authors perform immunofluorescence with the MYH9 antibody to confirm that the knockdown cells are the cells with increased axon growth?  
+
+- Hydrogels were used in the peripheral nerve injury experiments for drug delivery. What are the biokinetics of these hydrogels? Are they absorbable, and do they have a regenerative effect alone, when used long term for nerve repair? There was recovery of function with vehicle at 28 days equal to NMIli2 treated animals, did the authors expect this?  
+
+- ChAT immunohistochemistry was used to examine motor axon regeneration in the sciatic nerve. Is there any possibility of tracing axons, rather than relying on antibody staining for regenerative growth? How was the injury confirmed, as the methods only state that the sural nerve was crushed for 20 sec with a hemostat? The sural nerve is primarily a sensory nerve, so I query why this was used to assess motor axon regeneration.  
+
+- Likewise, to assess motor recovery, hindpaw placement was assessed. Plantar placement is mainly used to assess sensory recovery, or possibly sensorimotor feedback following injury and repair. While the results found are promising, how
+
+<--- Page Split --->
+
+do the authors discount the potential for sensory axon changes inducing this effect? Have the authors looked at sensory axon repair in this context? Very little detail is included in the methods section for this test. Can further detail be included (length of walking tract, timing of test, etc)  
+
+- did the authors test for changes in Rho/ROCK activity in their in vitro experiments?  
+
+Minor issues:  
+
+- Page 9, paragraph 2: why is the term 'frustrated' used for total internal reflection? 'Uninjured' misspelled as 'uninured'.  
+- Page 17, Methods, Western Blot: 'Proteins' should be 'Protein'.  
+- Page 17, Methods, Mouse surgical procedures and behavioural test: There is an unfinished sentence 'Sciatic nerve crush surgery and subsequent local application were performed aseptically under 2% isoflurane anesthesia followed by'?  
+- Page 18, Methods, Mouse surgical procedures and behavioural test: 'To assess functional recovery, lumbrical muscles were dissected form the paws' This need more detail for a functional recovery assessment.  
+- Page 19, Methods, Statistics: Please add what post-hoc tests used.  
+
+(Remarks on code availability)  
+
+Version 1:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author)  
+
+The authors have addressed some of my comments and concerns. However, the limited impact of the intervention on the peripheral regenerative phenotype (in comparison many other interventions are superiors to this and no in vivo work in CNS injury models has been performed), the very limited novelty and the lack of more comprehensive mechanistic insight do hamper my enthusiasm for this study.  
+
+(Remarks on code availability)  
+
+Reviewer #2  
+
+(Remarks to the Author)  
+
+The authors have thoroughly addressed my questions and those of the other reviewers. My only suggestion is that the authors be careful in their mention of sensory neuroregeneration, which does not lead to functional recovery. The majority of sensory neurons are indeed made up of nociceptors, which were examined here, but when it comes to motor recovery, it is more the mechanoreceptors (TrkB) and proprioceptors (TrKC) that would be of interest that was indicated by reviewer #3. Therefore, please add this caveat. NF, while bringing up myelinated neurons, would not be specific to myelinated sensory neurons alone, which is what the comment on 281 of page 11 makes confusing. These slight clarification changes are all that are required for acceptance.  
+
+(Remarks on code availability)  
+
+Reviewer #3  
+
+(Remarks to the Author)  
+
+Thanks to the authors for responding to my comments and queries and amending their manuscript. This manuscript is now suitable for publication.  
+
+(Remarks on code availability)
+
+<--- Page Split --->
+
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source.  
+
+The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.  
+
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+
+## RESPONSE TO REVIEWER COMMENTS  
+
+We thank the reviewers for their constructive comments and valuable suggestions. We address each comment raised below, which include the new experiments and analyses recommended.  
+
+We have in consequence modified many of the figures, as summarized here. All the modified text in the revised manuscript is highlighted in yellow.  
+
+## Revised figure panels including new data requested by the reviewers:  
+
+Former Fig. 5 has been moved to supplementary figure (now Fig. S6).  
+
+Fig. 5c and 5g (was Fig. 6 before) now include additional sample numbers for both vehicle and treated groups.  
+
+Added a new Fig. 6, illustrating the effect of NMIIi on sensory neuron regeneration.  
+
+Added new panels to Fig. S1 (panels i- k) to present results of cell viability at the concentrations used in the compound screen.  
+
+Added a new figure panel, Fig. S5j, to depict RGC regeneration under CSPG- coated conditions.  
+
+Added a new panel to supplementary figure, Fig. S8 (panel a- c), to show nerve innervation in muscles at days 7 and 28 post- injury.  
+
+Added a new panel to supplementary figure, Fig. S8 (panel d), to illustrate effect of hydrogel itself on regeneration.  
+
+Added a new supplementary figure, Fig. S10, to demonstrates that NMIIi2 specifically targets neurons.  
+
+Added a revised p- value with a more stringent approach and updated them in each figure panel.
+
+<--- Page Split --->
+
+## 1 Reviewer #1:  
+
+2 Heo and colleagues explored the effect of a blebbistatin analogue on axon growth and nerve 3 regeneration both in vitro and in vivo. After screening over 4000 compounds, they identified 10 4 main candidates for improving axon outgrowth, focusing on blebbistatin, a known NMII inhibitor. 5 However, due to its poor biocompatibility and its photosensitivity, they proceeded with an 6 analogue called NMIIi2. They tested the drug in vitro on various cellular models and in vivo in a 7 mouse model of sciatic nerve regeneration following a crush injury.  
+
+8 They evaluated neurite elongation in vitro and in vivo and conducted studies on functional recovery.  
+
+10 The work is well- structured and thoughtfully designed and however the novelty is limited given 11 that the effect of blebbistatin in neuronal outgrowth has been described and the proposed 12 mechanism is not surprising and not very well developed. The regenerative phenotype observed 13 is statistically significant but not especially exciting compared to other approaches.  
+
+14 We appreciate the reviewer's recommendations. To enhance the novelty and discussion of 15 potential mechanisms underlying our finding, we have performed the suggested experiments 16 and analyses.  
+
+17  
+
+18 Major points  
+
+19 - The choice to include the MYH9 knockdown model in this manuscript is unclear. It has long 20 been known that inhibition of NMII leads to increased axonal growth through a well- documented 21 mechanism. Validation of this model in the present article adds little to the existing literature or 22 to the quality of this manuscript  
+
+23 We agree that it is well- established that NMII inhibition promotes axonal growth. Including 24 MYH9 knockdown experiments was, however, crucial to confirm the specificity of our findings 25 and to ensure that the observed effects were directly mediated by NMII inhibition, but have now 26 moved this figure to Supplementary Figure S6 since this is not a novel finding.  
+
+27  
+
+28 The mechanism has not been well studied neither at the cellular nor signalling level. Cell- 29 specific unbiased approaches could reveal novel interesting mechanisms.  
+
+30 We appreciate the reviewer's suggestion regarding cell- specific unbiased approaches and 31 acknowledge that the mechanism of NMII inhibition has not been fully elucidated at a cellular or 32 signaling level. Our current study is thought primarily focused on the screening platform and 33 evaluating the functional outcomes of the new NMII analogue and its therapeutic potential for 34 promoting axonal regeneration by local application to an injured nerve rather than on the cellular 35 mechanism of NMII.  
+
+36 In attempt to provide an understanding of signaling changes induced by NMII inhibition, we 37 performed a bulk RNA- seq analysis of motor neurons treated with the NMII inhibitor blebbistatin 38 on CSPG substrates, which provided some insight into molecular changes induced by NMII 39 inhibition. Pathway analysis of the data revealed elevated activity in Rho- related pathways, 40 integrin- mediated cell surface interactions, and extracellular matrix (ECM) degradation. These
+
+<--- Page Split --->
+
+41 findings imply that NMII inhibition of cell bodies increase integrin levels and enhance cellular interactions with the ECM in CSPG- rich environments, potentially facilitating regeneration in the central nervous system. However, we decided not to include this data in the revised MS for the following reasons:  
+
+45 - Activity- driven vs. Transcriptional changes: The phenotype observed in our study 46 manifests very rapidly after treatment, as shown in Figures 3D and 3E, suggesting that activity- 47 dependent or post- translational mechanisms rather than transcriptional changes likely drive the 48 early/initial regenerative effects. Including the bulk RNA sequencing data could misrepresent the 49 temporal dynamics of NMII's effects by overemphasizing transcriptional changes that may not 50 directly contribute to the regeneration phenotype.  
+
+51 - Localized mechanisms at injured axonal sites: Our findings in Figures 4 and 5 strongly 52 suggest that NMII inhibition exerts its effects locally at the injured axonal sites and not in the cell 53 body. Bulk RNA sequencing lacks the resolution to pinpoint localized axonal mechanisms. Our 54 focus is on the site- specific changes critical to regeneration.  
+
+55 We agree that cell- specific approaches, such as single- cell RNA sequencing or proteomics, 56 could provide a more granular and cellular understanding of NMII's effects. These approaches 57 represent logical steps for future investigations but are beyond the scope of the current study. 58 We have added a discussion of this point in the revised manuscript to emphasize the need for 59 future exploration of mechanisms with these approaches.  
+
+60  
+
+61 - It is unclear why the authors tested non- permissive substrates in culture to mimic CNS 62 regeneration and then they focused on peripheral regeneration in vivo. The authors should 63 clarify this aspect. Did they test regeneration in the CNS? This would add to the manuscript's 64 impact.  
+
+65 The primary outcome of this study was that localized NMII2 treatment to an injured nerve 66 promotes the regeneration of motor axons, even though the primary human motor neuron 67 screening platform was based on growth on CSPG. The CSPG substrate was selected because 68 this revealed pro- regenerative hits much more reliably than screening the neurons on laminin, 69 where axon growth occurs by itself.  
+
+70 We did, however, compare the effects of NMII in vitro on neurite outgrowth phenotypes on 71 CSPG and laminin substrates for human iPSC derived motor, sensory and cortical neurons as 72 well as primary mouse retinal ganglion cells. These findings reveal that NMII treatment 73 promotes neurite outgrowth in CNS and PNS neurons in both CSPG and laminin conditions. We 74 have refined the text to ensure greater clarity.  
+
+75  
+
+76 Did they test regeneration in the CNS? This would add to the manuscript's impact.  
+
+77 As our in vivo data indicates that NMII is not suitable for systemic delivery (Fig. S9), and local 78 delivery in an in vivo CNS- injury model presents significant technical challenges, we opted to 79 test this in vitro for this study. Specifically, we have now conducted new experiments to examine 80 the response of adult mouse dissociated RGCs in CSPG- coated conditions to mimic CNS 81 regeneration. RGCs exhibit very limited growth in non- permissive environments, but NMII2
+
+<--- Page Split --->
+
+treatment enable growth under these conditions (Fig. S5J) (and below) and now specify that future in vivo CNS studies are needed.  
+
+![](images/Figure_unknown_0.jpg)
+
+<center>Figure S5. NMIIi induces neurite outgrowth in dissociated RGCs </center>  
+
+Representative RGC images of RBPMs- labeled neurons treated with DMSO, and \(10 \mu \mathrm{M}\) NMIIi (not included in fig) and quantification of neurite length on CSPG substrate for 24 hours (J). Data collected from 5 female mice. Statistics: unpaired two- tailed Student's \(t\) - test, SEM error bars.  
+
+- With the regards to the in vivo drug delivery exp, which represent the main novelty of the work, there are several aspects to consider:  
+
+i) Local delivery of the NMII inhibitor might affect not only the injured axons but also the entire non-neuronal component present in the sciatic nerve. Since this is a chronic treatment (over several days), there could be side effects on glial or immune cells that need to be considered, especially since these could potentially lead to systemic issues or beneficial effects. These cell populations should be evaluated, at least through histochemistry, to assess their condition.  
+
+To address this point, we conducted additional immunohistochemistry experiments to evaluate if there are changes in Schwann cells (S100+), immune cells (CCR2+), and activated fibroblasts (FAP+) following local NMIIi administration at the sciatic nerve injury site in mice. We assessed changes on day 3 post injury, a time point where these non-neuronal cells actively react to a nerve crush injury (1). However, we found that none of these cell types showed any major changes in the NMIIi-treated group compared to the vehicle-treated group (Fig. S10A- S10C). This suggests that NMIIi specifically targets axons. We have incorporated this data into the revised manuscript (Fig. S10). However, further investigations using flow cytometry or single-cell sequencing will be needed for a more comprehensive assessment of this issue. Given a recent study examining the role of NMII in T cells (2), it would be interesting in the future to explore its effects on other immune cell populations following local treatment.
+
+<--- Page Split --->
+![](images/Figure_unknown_1.jpg)
+
+<center>Figure S10. Assessment of non-neuronal cells at sciatic nerve injury site after NMII treatment </center>  
+
+(A- C) Images of sciatic nerves at day 3 post-injury in mice treated with a local application of either vehicle or NMII, labeled with anti- S100 (red) (A), anti- FAP (green) (B), and CCR2 (red) and/or DAPI (blue) (C).  
+
+ii) Similarly, while the machine learning method is powerful, it is not sufficient to demonstrate complete and physiological functional recovery. In the peripheral system, there are various neuronal populations involved at different levels in sensation and perception, which should be analyzed independently.  
+
+We have now investigated changes in sensory neuron regeneration following a sciatic nerve crush injury at various time points with or without local NMII application. Specifically, as Reviewer 2 suggested, we examined SCG10/STMN2, a microtubule- destabilized protein that is highly expressed after injury and widely used as an injury marker. On day 3 post- injury, SCG10/STMN2 intensity was noticeably higher in NMII2- treated nerves compared to the vehicle- treated mice (Fig. 6A and 6C). We also observed that CGRP- positive nociceptors are more prominently labeled in the NMII2- treated group, while the vehicle- treated group show more limited regeneration (Fig. 6A and 6C). We also further assessed regeneration of neurofilament (NF)- labeled axons and found that these also regenerate faster after NMII2 treatment (Fig. 6B), revealing enhanced regeneration across multiple sensory neuronal subtypes.  
+
+We also assessed sensory recovery, using a skin pinprick assay, as described previously (3). Briefly, we measured mouse nociceptive reactions to pinprick stimulation of the lateral area of the paw, which the sciatic nerve innervates prior to a crush injury. In line with the histology data, the NMII2- treated group exhibited a significantly improved sensory response to pinprick by day 7 post- injury, compared to the vehicle- treated mice (Fig. 6D). These data now reveal that NMII promotes functional regeneration of sensory neurons. The data also show that sensory and motor components recover at different times: sensory recovery is evident by day 7, whereas motor recovery only becomes prominent by day 14. This temporal difference underscores the complementary nature of our methods, with the machine- learning approach being particularly
+
+<--- Page Split --->
+
+effective for evaluating motor function recovery, as revealed by the positive correlation between paw luminance ratio and lumbrical muscle reinnervation (Fig. 5G) on day 14 (Fig. 6D) in Fig. 6E. These additional findings demonstrate that NMIIi2 treatment promotes functional recovery in vivo across multiple neuronal populations in the peripheral system. We include these results in the revised manuscript.  
+
+![](images/Figure_6.jpg)
+
+<center>Figure 6. NMIIi applied at the sciatic nerve injury site promotes sensory axon regeneration and functional recovery </center>  
+
+(A and B) Representative images of mouse sciatic nerves day 3 post- injury treated with local application of either vehicle or NMIIi and labeled with anti- CGRP (green), anti- statthmin (magenta) (A) and neurofilament (green) (B), white dashed lines indicate the boundary of the crush injury site. (C) Quantification of CGRP and stathmin intensity from regenerated axons. Data collected from 3 mice per group and normalized to vehicle condition. Statistics: unpaired
+
+<--- Page Split --->
+
+two- tailed Student's t- test, SEM error bars. (D) Score of pinprick tests following local application of vehicle or NMIIi at baseline and day 4, 7, and 14 post- injury. Statistics: two- way ANOVA with Tukey's post hoc test, SEM error bars, 15 mice per condition. (E) Pearson correlation analysis of the paw luminance ratio and pinprick behavior scores on day 14 post- injury. Data collected from 10 male mice per group.  
+
+Additional points  
+
+- The phenotypic screening should be implemented with more procedural details. Specifically:  
+
+i) Why were two different cell lines used between the first and second screen? The authors should clarify this in the text.  
+
+For the primary screen we tested compounds at a single concentration. After identifying hit compounds, we conducted a secondary screen using a range of concentrations to evaluate their dose- dependent effects. To reduce biological variability, we performed the primary and secondary screens using three different human cell lines: SAH- 0047 from a 46- year- old female donor, LiPSC- GR1.1 from a male newborn donor, and 11a from a 36- year- old male donor. This decision was made to minimize potential biological variability and confirm that the observed effects were not cell line- specific and are representative of human motor neurons across diverse genetic and demographic backgrounds. We have clarified this in the revised manuscript.  
+
+ii) Some of these compounds may cause toxicity that is not necessarily reflected in neuter length but in cell survival. Was neuronal survival assessed at the concentrations used?  
+
+We agree that compound- induced toxicity is an important issue. To address this we evaluated cell viability following treatment with the top compounds identified in the screen, using a viability/cytotoxicity assay with Calcein- AM and ethidium homodimer- 1 (EthD- 1) dyes (4). Treatment with blebbistatin our lead compound, neither changed cell viability nor induced cytotoxicity, indicating that it is safe at the tested concentrations (Fig. S1I- K).  
+
+Cells treated with two ROCK1/ROCK2 inhibitors (Fasudil and Y27632), a JAK2 inhibitor (CEP- 33779), or a voltage- gated calcium channel blocker (Benidipine- HCl) exhibited reduced EthD- 1 intensity (Fig. S1K), suggesting that while these compounds promote neurite outgrowth at the tested concentrations, they may also impact cell survival. These findings provide additional insight into the safety profile of the screened compounds and are included in the revised manuscript (Fig. S1I- K).  
+
+## Figure S1. Assessment of cell viability/cytotoxicity of top 9 compounds from primary screen  
+
+(I) Representative images of neurons cultured on a laminin substrate labeled with Calcein and EthD-1 dye at 24h NMIIi2 post-treatment, and controls. (J and K) Quantification of Calcein (J) and EthD-1 (K) intensity from cells treated with the top hit compounds. Data normalized to DMSO, and experiments done in three replicates. Statistics: Welch's test with Dunnett's multiple comparison test, SEM error bars.
+
+<--- Page Split --->
+![](images/Figure_unknown_2.jpg)
+  
+
+- The authors should clarify how the concentration of \(100 \mu \mathrm{M}\) was established for in vivo use, given that the drug is being tested for the first time to test nerve regeneration.  
+
+For the in vivo studies, we initially tested at a high concentration (10 mM) locally at the injury site. However, this either had no effect or slightly exacerbated motor deficits compared to vehicle-treated conditions, which are likely due to off-target effects. Subsequently, we optimized the dosage by testing a range of lower concentrations and observed that \(100 \mu \mathrm{M}\) provided the best balance between efficacy and safety. At this concentration, NMIIi2 significantly enhanced axonal regeneration and functional recovery without causing observable side effects.  
+
+![](images/Figure_5.jpg)
+
+<center>Figure (not included in MS). Evaluation of motor function recovery with 10 mM and 25 mM doses of NMIIi at the sciatic nerve injury site. </center>  
+
+(A) Effects of local NMIIi2 treatment at high concentrations on paw luminance ratio after sciatic nerve crush injury. Paw luminance ratio (injured/non-injured) was measured for 10 minutes at each time point indicated on x-axis. NMIIi2 at 10 mM did not improve functional recovery compared to the vehicle group, while 25 mM NMIIi2 slightly but not significantly worsen luminance ratio. Data are presented as mean ± SEM. Statistical analysis was conducted using two-way ANOVA with Tukey's post hoc comparisons. n.s., not significant.
+
+<--- Page Split --->
+
+- There are many mechanisms that have been studied and could underlie the results observed by the authors in this work, beyond the Rho/ROCK pathway (such as increased microtubule stability, alterations in interactions with the environment via adhesion molecules, growth cone dynamics, changes in intrinsic forces, etc.). The authors should, even briefly, expand the discussion to include these aspects  
+
+To include potential mechanisms underlying NMII's action in promoting regeneration we have now extended the discussion.  
+
+We suspect that in injured axons a Rho/ROCK- mediated activation of myosin II inhibits axon growth by modifying actomyosin contractile bundles at the growth cone. Disruption of these actomyosin arcs at the growth cone leads to microtubule destabilization, altering growth cone polarity and mechanical stress on the extracellular matrix, leading to elongation.  
+
+Rho/ROCK- independent changes in myosin II may also play a role in regeneration through local myosin II assembly. Myosin II activity is regulated through mechanisms distinct from its assembly (5). While myosin II reduces cell- surface curvature via a change in actomyosin tension, once recruited, the cell surface regulates myosin II stabilization and F- actin binding in a Rho/ROCK independent pathway. This process explains how NMII inhibitors can simultaneously enhance branch protrusion and axon elongation. Our study reveals that NMII inhibitors promote branch protrusion and altered growth cone dynamics, as evident by the thin growth cones lacking lamellipodia. Our findings suggest that NMII inhibitors modulate myosin II activity both via a ROCK- dependent pathway and by simultaneously promoting myosin II stabilization and F- actin binding through a ROCK- independent pathway, forming a positive feedback loop. However, further studies are required to confirm and expand upon these mechanisms.  
+
+- The authors should briefly explain why silicone tubes were chosen for local drug application as opposed to other methods.  
+
+Silicone- based implantations are biocompatible and widely used in medical research. Silicone tubes were chosen for their ability to provide localized, controlled drug delivery while minimizing systemic exposure, making them ideal for isolating NMIIi2's effects on nerve regeneration. We have added this information to the Method section.  
+
+## Reviewer #2:  
+
+- What are the noteworthy results?  
+
+Heo et al. in the work entitled "NON- MUSCLE MYOSIN II INHIBITION AT THE SITE OF AXON INJURY INCREASES AXON REGENERATION" employed an unbiased screen of axonal growth promotors on a growth inhibitory environment (CSPG) and found several compounds that increased the outgrowth of human iPSC derived motor neurons. Of these, the highest hit was blebbistatin, which inhibits NMII. Although used previously in similar conditions as shown here, this work examined a new NMIIi2 that is more bioavailable than previous versions. This was shown to work to extend dendrites proximal and axons distal to the lesion site following laser dissection of spot cultures. More importantly, this work showed for the first time in vivo local
+
+<--- Page Split --->
+
+NMll2 treatment of the injured sciatic nerve leads to increased regeneration within the nerve and the muscle of the foot, increased synaptic connections and functional recovery.  
+
+- Will the work be of significance to the field and related fields? How does it compare to the established literature? If the work is not original, please provide relevant references.  
+
+Blebbistatin and NMll inhibition have been found previously to be involved in axonal growth of human NPCs (10.3389/fcell.2021.719636), RGCs and the optic nerve (10.1016/j.celrep.2020.107537), and embryonic DRG (10.1073/pnas.1011258108). Here for the first time a new NMll2 that is more bioavailable than previous versions was examined and found to promote axonal regeneration when placed at the site of injury, the sciatic nerve, leading to faster functional recovery. This work takes the field into clinical relevance. It can also be noted that many of the in vitro models are typically considered immature neurons (which tend to be more growth-permissive) except for the adult DRG work, so the in vivo work shows that in a mature neuronal state, the NMll2 works effectively.  
+
+We thank the reviewer for the positive feedback, particularly the translational relevance and novelty of our findings with NMll2.  
+
+- Does the work support the conclusions and claims, or is additional evidence needed?  
+
+This work states that NMlll works well when given at the lesion site inducing accelerated regeneration and functional recovery. However, one could strengthen these conclusions with some additions. Firstly, in Figure 6 the co-localization is difficult to see with current magnifications. Why are the NF and Syn in the same color? Please provide higher magnifications of the co-localization analysis and have the antibodies split into separate channels/colors along with the merged overlap.  
+
+To address this, we have now performed additional experiments and include higher magnification images of the co- localization analysis in Fig 5B in the revised manuscript. The images show separate channels/colors for neurofilament (NF), synaptophysin (SYP), and α- bungarotoxin (BTX) alongside a merged overlap, ensuring a clearer representation of pre- and post- synaptic co- localization at the NMJ.  
+
+![](images/Figure_unknown_3.jpg)
+
+<center>Figure 5. NMll at sciatic nerve injury site accelerates motor function recovery in mice </center>
+
+<--- Page Split --->
+
+(B) Representative images of lumbrical muscles day 14 post-injury from mice treated by local application of vehicle or NMIIi labeled with anti-neurofilament (NF), anti-synaptophysin (SYP) or a-bungarotoxin (BTX). (C) Co-localization of pre/post-synaptic markers per muscle in mice with intact sciatic nerve (non-injury) and 14 days post-nerve injury. Number of presynaptic markers that overlap with the postsynaptic marker, \(\alpha\) -bungarotoxin normalized to post-synapse ( \(\alpha\) -bungarotoxin) number. Stats: Kruskal–Wallis test with Dunn's multiple comparison test, SEM error bars. Data collected from 13 male mice/group. (G) Pearson correlation between the paw luminance ratio and percentage of pre/post-synaptic colocalization from the same mouse shown in panel (C). Data collected from 16 male mice per group.  
+
+This analysis was from D14 when the greatest behavioral differences were seen. It should also be provided at the end timepoint (D28) and possibly from D7 where no behavioral differences are observed. The assumption would be that there would be less colocalization overall at D7 because there wouldn't be enough growth into the muscle by this timepoint. As for the D28 at the end of the study where the behavior is the same between the treated and vehicle groups it would be important to know if the growth and co-localization are the same or if this is due to compensatory mechanisms that arise in the vehicle group. This would further support the correlation graph in Figure 6, which is a little weak given that two of the five samples sit at zero for colocalization yet have very different behavioral outcomes. Is the correlation work from D14? This should be stated somewhere.  
+
+To provide a more comprehensive analysis, we have now performed additional experiments examining pre- and post- synaptic colocalization at the NMJ at days 7 and 28 post- injury. At day 7, the nerves had not yet reached the muscle in either the vehicle or treatment group (Fig. S8A and S8C), resulting in a lack of nerve innervation (Fig. S7). However, by day 28, nerve regeneration was complete, and pre- and post- synaptic colocalization in the muscle reached a similar level in NMIIi2- treated and non- injured conditions (Fig. S8B, and S8C). This indicates that successful reinnervation occurred in the vehicle- treated mice by day 28 (Fig. 5F) rather than compensatory mechanisms.  
+
+We also increased the sample size (16 mice in total) to analyze the correlation between colocalization and behavioral outcomes, focusing on D14, when the greatest differences were observed. This information is clarified in the main text and the Figure 5 legend. These additional data strengthen the conclusion that NMIIi2 treatment accelerates functional recovery by promoting neuromuscular junction reinnervation.  
+
+Nerve transection injury models typically show minimal functional recovery (3). In contrast, nerve crush injury enables nerve regeneration and a much, quicker recovery of motor and sensory function and is therefore, frequently used in regeneration studies (3). Our finding emphasizes the importance of faster axonal regeneration for enhancing functional recovery.
+
+<--- Page Split --->
+![](images/Figure_unknown_4.jpg)
+
+<center>Figure S8. Neuromuscular junctions days 7 and 28 post-sciatic nerve crush injury Representative images of lumbrical muscles at day 7 (A) and day 28 (B) post-crush injury from mice treated with local application of either vehicle or NMII, labeled with anti-neurofilament (NF), anti-synaptophysin (SYP), or \(\alpha\) -bunqarotoxin (BTX). (C) Quantification of presynaptic markers (NF; SYP) overlapping with postsynaptic marker (BTX), normalized to the number of postsynaptic sites. Statistics: one-way ANOVA with Tukey's post hoc test, SEM error bars. Data collected from 5 mice per group on day 7, and 4 mice per group on day 28. </center>  
+
+Further, it should be spoken about why this work is important given there is functional recovery at the end of the study similar to the treated group. What is the benefit of the accelerated recovery? In the introduction it is stated, "After a traumatic nerve injury, mature motor neurons only regenerate at a rate of 1- 4 mm per day, the slowness of which results in only \(\sim 45\%\) of injured axons achieving anatomical reinnervation of their muscle targets with, in consequence, limited functional recovery (1- 7). Discovering therapeutic targets for the effective treatment of axon injury is therefore critical to ensure a more complete recovery of motor and other disturbed functions." However, it appears that in Figure 6 the controls reach the same level of functional recovery as the treated group at D28. This may just be a difference between rodents and humans, either way this should be discussed. Some of the additions suggested here may help address this issue.  
+
+While it is correct that the vehicle- treated group eventually achieves similar levels of functional recovery at day 28, our findings underscore the significant benefit of an accelerated recovery in the NMII2- treated group. The reason we used the crush injury model is that it enables a more precise assessment of axonal regeneration and functional recovery.
+
+<--- Page Split --->
+
+The accelerated recovery observed with NMIIi2 treatment reflects both an enhanced regenerative capacity of the injured axons and a timely reinnervation of neuromuscular junctions, which aligns with the translational goal of developing therapies for nerve injuries. While rodents and humans differ considerably in axonal length and regenerative timelines, the accelerated recovery in our model highlights the therapeutic potential of NMII inhibition in reducing the time to functional recovery, a crucial outcome in clinical scenarios to minimize the secondary complications that often arise due to prolonged functional impairment. We have expanded these points in the revised manuscript to emphasize the translational significance of the findings and the broader implications of accelerated recovery in a clinical setting.  
+
+Given the work done to show that NMIIl works in various cell types it would make sense to also look at the sciatic nerves shown in Fig. S6 for sensory neuronal regeneration with SCG10. This would take the in vitro work into the in vivo relevance. As there is not any increased mechanistic understanding greater than what was published before, I do believe these additions are important to show the novelty and relevance of this work.  
+
+As detailed in our response to Reviewer #1, we have now performed additional experiments to examine sensory neuronal regeneration in vivo using SCG10 as a marker and functionally observed an acceleration in sensory response recovery. These results, presented in Figure 6, in the revised MS, demonstrate that NMIIi2 treatment significantly enhances sensory axonal regeneration compared to vehicle- treated groups, further bridging the in vitro findings with in vivo relevance.  
+
+Lastly, it should be stated why 10mg/kg of NMIIi2 was used systemically and whether there were any obvious things that were associated with weight loss, such as a decrease in organ size or discoloration. Also how long is it thought for the hydrogel release of NMIIi2 at the lesion site, immediate or over days, weeks? This will be of interest to the readers.  
+
+The systemic dose of \(10mg / kg\) NMIIi2 was selected based on our pharmacokinetic data, which showed that \(10mg / kg\) achieved a higher plasma concentration compared to \(3mg / kg\) , with minimal additional benefit when increased to \(30mg / kg\) .  
+
+![](images/Figure_unknown_5.jpg)
+
+<center>Figure (not included in MS). Plasma pharmacokinetic parameters of NMIIi2 following IP administration at doses of 3, 10, and \(30mg / kg\) in male mice </center>
+
+<--- Page Split --->
+
+We initially administered \(10 \text{mg / kg}\) of the drug twice daily for three consecutive days following a crush injury. This resulted, however, in significant weight loss in the treated mice (Fig. S9C). When the drug was withheld for three subsequent days, the weight loss improved by day 7 post- injury (Fig. S9C). Conversely, resuming the drug for another three days in the same animals led to further aggravation of weight loss (Fig. S9C), suggesting that systemic administration causes notable side effects. Besides the weight loss, no other obvious side effects were observed. While organ size was not measured, locomotion and general behaviors (grooming, rearing, scratching) remained normal.  
+
+Regarding the hydrogel- based local delivery, we now confirm that the hydrogel/NMll12 mixture remained in place through visual inspection during surgery. Our immunohistological analyses suggest that the hydrogel effectively released NMll2 for approximately 3- 7 days post- application. Once the regenerating axons grew beyond the hydrogel- embedded lesion site, the localized effects of NMll2 diminished, highlighting the importance of early, targeted local intervention. These findings reinforce the therapeutic potential of localized NMll inhibition while addressing the limitations of systemic delivery. We incorporate this into the discussion in the revised manuscript.  
+
+There are some minor additions needed in the methods:  
+
+- The correlation analysis for Figure 6 is not listed nor is it stated if this is a Pearson or Spermann correlation.  
+
+We have edited the Methods section to deal with this.  
+
+- The post hoc tests are not listed by name, it would be helpful if the authors listed if they chose a more conservative or liberal post hoc analysis method.  
+
+We used two- way ANOVA followed by Tukey's post hoc test. We have included this information in the revised MS.  
+
+- 10mg/kg twice a day for 3 days via i.p. of the NMllii was used for systemic application, but why was this dose/timing/method chosen? This should be listed given the side effects that were observed.  
+
+We have included this information in the Methods section.  
+
+Reviewer #2 (Remarks on code availability): I could not access the code when I tried to click on the GitHub link.  
+
+The code is now available on GitHub.  
+
+https://github.com/selwynjayakar/Multi- image- neurite- analysis
+
+<--- Page Split --->
+
+## 412 Reviewer #3:  
+
+413 The manuscript entitled "Non- muscle myosin II inhibition at the site of axon injury increases 414 axon regeneration" is a study in which the authors produce a novel myosin II inhibitor (NMIli2; 415 analog of blebbistatin) and show it's efficacy in induce axon growth in tissue culture and after 416 peripheral nerve injury in adult mice.  
+
+417  
+
+418 Following a phenotypic screen in human iPSC derived motor neurons, the authors identified 10 419 compounds that induced neurite outgrowth on a CSPG extracellular matrix environment. 420 Blebbistatin was found to be the strongest inducer of growth, with and without laser cut injury, 421 however given it's issues with photosensitivity and bioavailability, the authors produced a more 422 selective non- muscle myosin II inhibitor, termed NMIli2. As non- muscle myosin II is a motor 423 protein that normally stabilises microtubules, use of NMIli2 would be hypothesised to destabilise 424 MTs thus inducing axon elongation/extension.  
+
+427 To confirm the findings from their screen as well as to confirm the efficacy of NMIli2, the authors 427 grew human iPSC- derived motor neurons, cortical- like neurons, and human cortical organoids 428 on either laminin or CSPGs; as well as adult mouse DRGs and RGCs on laminin only. The 429 authors also compared their findings with NMIli2 with a lentiviral knockdown of MYH9 (encoding 430 the heavy chain of NMIli) in human iPSC- derived motor neurons, demonstrating axon growth 431 with the KD with confirmed KD (~55%) with a western blot. The authors then tested NMIli2 in 432 vivo in a mouse model of peripheral nerve injury, showing that local/direct administration 433 complexed to a collagen hydrogel, compared with intraperitoneal administration or control, was 434 effective at inducing ChAT- labelled axon growth, synaptic reinnervation and better hindpaw 435 plantar placement.  
+
+436 While the findings of this study are very interesting for the field of axon regeneration specifically 437 in the PNS, there are some areas of the study that require clarification and/or further 438 information.  
+
+439 Major questions:  
+
+- Why were the mouse cellular experiments using adult RGCs and adult DRGs not performed on a CSPG matrix?  
+
+442 We have conducted additional experiments using CSPG-coated surfaces with cells isolated 443 from 6 additional mice to evaluate RGC growth. While neurite outgrowth was more restricted in 444 this inhibitory environment compared to laminin-coated conditions, we consistently observed 445 that NMIli2 treatment significantly increased total branch length compared to control conditions 446 (see response to Reviewer #1: line number: 77).  
+
+447 Regarding DRG neurons it is more physiologically relevant to evaluate their regeneration in a 448 peripheral- like environment (laminin- coated surfaces) rather than a CNS- like environment 449 (CSPG- coated surfaces).
+
+<--- Page Split --->
+
+CSPG substrates were used in the motor and cortical- like neuron experiments with success. All of these cell types have been shown to respond to CSPGs with retraction and/or overall stunted growth, especially in adult ages. Some studies have shown that the lack of growth on these substrates is due to integrin are downregulated and or subsequent inactivation, leaving adult cells unable to respond to the CSPG matrix. Given the mechanism of action of destabilising MTs, do the authors think that NMIIi2 would induce the growth in the mouse cells on CSPG regardless of integrin regulation? If not, can the authors discuss the mechanism of action in the CNS neurons that differs in the PNS neurons?  
+
+We speculate that NMIIi2 may promote growth in central neurons on CSPG substrates independently of integrin regulation. However, in the PNS, integrins likely play a more prominent role in regeneration, as described by the Roca- Cusachs molecular clutch model. This model highlights two key contributions of integrins to growth cone dynamics: the regulation of actin retrograde flow and the generation of mechanical forces through integrin- ECM interactions at the growth cone (6, 7). In contrast, CNS neurons are surrounded by a perineuronal net (PNN) composed of CSPGs (8). These differences may explain why integrins have a more substantial impact on promoting regeneration in the PNS compared to the CNS.  
+
+We have expanded the discussion in the revised manuscript to address these potential differences of NMII inhibition across different neuronal subtypes.  
+
+- The axon growth in vitro suggests that blebbistatin and NMIIi2 both induce neurite/axon elongation as well as branching. The branching effect is indeed an interesting finding, which would in some cases compete with elongation. Can the authors discuss the mechanisms of action by which the compound can drive both an elongation and a branching phenotype?  
+
+We propose two possible mechanisms to explain the role of NMIIi in supporting regeneration.  
+
+First, NMIIi treatment appears to facilitate two different growth modes following nerve injury. Under non- injured conditions, neurons on CSPG- coated surfaces generally cannot extend neurites (Fig. 2E). However, with NMIIi treatment, axonal projection occurs, suggesting that changes in actomyosin force are a key factor driving microtubule destabilization and enabling neuronal elongation. In contrast, on laminin- coated surfaces, where axons are already extending, NMIIi treatment enhances both branch protrusion and elongation of newly formed branches (Fig. 2F), emphasizing the critical role of actin polymerization and integrin- ECM interactions at the growth cone. Under injured conditions, we hypothesize that both mechanisms are simultaneously activated.  
+
+Second, NMII inhibition may accelerate neuronal growth after injury while potentially modulating neural activity. During nerve regeneration, two distinct growth modes are observed: (1) an elongation mode characterized by typical axonal morphology, and (2) a branching mode distinguished by thin, numerous projections (9). During branch arborization, neuronal activity diminishes as it is distributed across multiple branches (10). Once branches form, their fate is determined during the maturation phase, where they either stabilize and grow or undergo retraction (10). Surviving neurons eventually restore neuronal activity, which is essential for functional recovery (10). Based on these previous findings, we suspect that neuronal activity changes in response to NMIIi treatment may contribute to the recovery, but this will need to be shown directly in future studies.
+
+<--- Page Split --->
+
+- In the MYH9 experiments, a western blot was performed to indicate knockdown of MYH9. Have the authors looked at changes in MYH9 in their NMIIi2 treated cells? Would they expect a greater effect, or a less specific effect?  
+
+To address this, we compared MYH9 levels in cells treated with NMIIi2 for 24 hours with control cells using a Western blot analysis. Unlike MYH9 knockdown, the relative intensity of MYH9 remained unchanged in the NMIIi2-treated cells.  
+
+![](images/Figure_unknown_6.jpg)
+
+<center>Figure (not included in MS). NMIIi treatment does not decrease MYH9 level </center>  
+
+Western blot of motor neurons treated with either DMSO or NMIIi2 labeled with anti- MYH9 and anti- GAPDH, and quantification of MYH9 intensity normalized to GAPDH. Experiments conducted in two replicates, statistical analysis unpaired two- tailed Student's \(t\) - test and SEM error bars.  
+
+This suggests that the effects of NMIIi2 are not mediated through changes in MYH9 expression but rather through an inhibition of NMII activity, which likely reduces actomyosin contractility and facilitates cytoskeletal remodeling, enhancing axonal growth.  
+
+This outcome aligns with the known mechanism of action of blebbistatin as a myosin II ATPase inhibitor, which targets the enzymatic activity of non- muscle myosin II without altering its expression (11, 12). We decided not to include this data in the revised manuscript, as a more comprehensive analysis is necessary to fully elucidate how NMII activity, rather than expression, contributes to the observed effects. We have incorporated a discussion of this in the revised manuscript.  
+
+Can the authors perform immunofluorescence with the MYH9 antibody to confirm that the knockdown cells are the cells with increased axon growth?  
+
+While we did not perform immunofluorescence for MYH9 in this study, we confirmed the efficiency of MYH9 knockdown through a Western blot analysis, which showed a significant reduction in MYH9 protein levels after knockdown compared to controls (Fig. S6A and S6B). The increased axonal growth observed in these cells aligns with the known functional consequences of MYH9 suppression, as previously described.
+
+<--- Page Split --->
+
+We recognize that immunofluorescence could validate the link between MYH9 knockdown and enhanced axonal growth. However, our attempts to use commercially available MYH9 antibodies were unfortunately unsuccessful due to their nonspecific binding, which prevented reliable detection. Additionally, the technical challenges of distinguishing axonal growth at single- cell resolution in dense cultures further complicated this approach. We acknowledge the importance of directly visualizing the spatial correlation in MYH9 knockdown axons and have included this in the Discussion of the revised manuscript.  
+
+- Hydrogels were used in the peripheral nerve injury experiments for drug delivery. What are the biokinetics of these hydrogels? Are they absorbable, and do they have a regenerative effect alone, when used long term for nerve repair?  
+
+For the peripheral nerve injury experiments we used bovine type I collagen as a drug delivery system for NMIIi2 given its biocompatibility. Based on our observations and the existing literature, we estimate that the hydrogel remains effective for approximately 3- 7 days post- application, coinciding with the critical early phase of nerve regeneration (see response to Reviewer #2: line number: 383).  
+
+Regarding regenerative effects of the hydrogel itself, previous studies have suggested that collagen- based hydrogels provide structural support and a favorable extracellular matrix environment, potentially aiding in nerve repair when used long- term. To verify this, we performed additional experiments comparing conditions with and without the hydrogel in the absence of NMIIi2. However, we observed minimal differences between these conditions (Fig. S8D), indicating that the hydrogel alone does not have a significant regenerative effect. These findings highlight the utility of the hydrogel as a delivery vehicle for NMIIi2 rather than as an independent regenerative agent and have included this in the revised manuscript.  
+
+![](images/Figure_6E.jpg)
+
+<center>Figure S8. Evaluation of the hydrogel's intrinsic regenerative effects after nerve injury (D) Quantification of luminance ratio of left/right paws post sciatic nerve crush with and without hydrogel. Data collected from 8 mice. Stats: two-way ANOVA with Tukey's post hoc test, SEM error bars </center>  
+
+There was recovery of function with vehicle at 28 days equal to NMIIi2 treated animals, did the authors expect this?
+
+<--- Page Split --->
+
+557 Please see our responses to Reviewer #2 (line number: 338).  
+
+558  
+
+- ChAT immunohistochemistry was used to examine motor axon regeneration in the sciatic nerve. Is there any possibility of tracing axons, rather than relying on antibody staining for regenerative growth? How was the injury confirmed, as the methods only state that the sural nerve was crushed for 20 sec with a hemostat? The sural nerve is primarily a sensory nerve, so I query why this was used to assess motor axon regeneration.  
+
+We selected ChAT staining to specifically identify regenerating motor axons, as this provides robust and reliable labeling of cholinergic neurons. Dye- based tracing approaches, such as Dil, would face challenges in differentiating motor neurons from sensory cell types. This is particularly relevant given our new data that NMII2 also enhances sensory neuron regeneration (Fig. 6). Therefore, we do not anticipate that dye tracing would provide information beyond what can be achieved using immunostaining with specific motor and sensory markers.  
+
+Regarding injury confirmation, the injury was visually confirmed during the surgery and further validated through immunohistochemical analysis (example image of Fig. 6A) and the functional recovery assays (Fig. 5F and Fig. 6D). The precise injury site was clearly identifiable as an enlarged swollen area caused by the hemostat crush (example image in Fig. 6A)  
+
+We apologize for the confusion in the Methods section regarding the location of the injury. To clarify, the injury was targeted at the sciatic nerve trunk, which includes the sural, common peroneal, and tibial nerves (refer to schematic in Fig. 5A). We have clarified this point in the revised text.  
+
+- Likewise, to assess motor recovery, hindpaw placement was assessed. Plantar placement is mainly used to assess sensory recovery, or possibly sensorimotor feedback following injury and repair. While the results found are promising, how do the authors discount the potential for sensory axon changes inducing this effect? Have the authors looked at sensory axon repair in this context?  
+
+We agree that sensory neuron involvement could contribute to the observed effects. To address this, we have now systematically studied the effect of NMII2 on sensory neurons (Fig. 6); further details are provided in our response to Reviewer #1 (line number: 119).  
+
+Additionally, our machine- learning approach assesses hindpaw placement by quantifying luminance intensity, which reflects weight- bearing and gait changes, which primarily captures loss and recovery of motor function, as supported by the positive correlation between paw luminance ratio and lumbrical muscle reinnervation (Fig. 5G) but not the sensory reinnervation (Fig. 6E).
+
+<--- Page Split --->
+![](images/Figure_unknown_7.jpg)
+
+<center>Figure 6E. Pearson correlation analysis of the paw luminance ratio and pinprick behavior scores on day 14 post-injury. Data collected from 10 male mice per group. </center>  
+
+Very little detail is included in the methods section for this test. Can further details be included (length of walking tract, timing of test, etc)  
+
+We now include additional information, such as the length of the walking track (10 minutes) and the specific equipment and methods used to calculate the luminance ratio. The timing of the tests is detailed on the x- axis of each time- course plot. All recordings were conducted between 1:00 PM and 3:00 PM for all animals.  
+
+- did the authors test for changes in Rho/ROCK activity in their in vitro experiments?  
+
+To address this, we analyzed growth cone morphology under ROCK and NMII inhibited conditions, as this is indicative of alterations in Rho/ROCK and/or myosin II activity (13- 15). Cytoskeletal components were labeled with \(\alpha\) - tubulin and F- actin to visualize growth cone structure. In the control condition, growth cones displayed the presence of both lamellipodia and filopodia. However, treatment with NMIIi2 resulted in a thin, elongated growth cone morphology without a lamellopodium. We also observed similar growth cone phenotypes in response to treatment with ROCK inhibitors, Y- 27632 or fasudil, as those observed with NMII inhibitors. These results suggest that myosin II activity in growth cones may be regulated through a ROCK- dependent pathway. This observation aligns with previous studies indicating that the regulation of myosin II activity is ROCK dependent. However, we decided not to include this data in the revised MS, since this paper primarily focused on evaluating the functional outcomes of the new NMIIi analogue as we have mentioned above (line number 32- 35).  
+
+## Figure (not included in MS). Assessment of growth cone morphology under ROCK and NMII inhibited conditions  
+
+(left) Representative images of neurons labeled with anti- TUBB3, F- actin, and DAPI at 24h NMIIi2, Y- 27632, Fasudil post- treatment, and controls. (right) Quantification of F- actin intensity of growth cone. Data normalized to DMSO, and experiments done in three biological replicates. Statistics: one- way ANOVA with Sidak's multiple comparison test, SEM error bars.
+
+<--- Page Split --->
+![PLACEHOLDER_25_0]
+  
+
+623  
+
+624 Minor issues:  
+
+- Page 9, paragraph 2: why is the term 'frustrated' used for total internal reflection? 'Uninjured' misspelled as 'uninured'.  
+
+We use the term 'frustrated total internal reflection' rather than 'total internal reflection' because it describes a modification or disruption of total internal reflection caused by surface contact. The technique is detailed in our original paper (16).  
+
+We apologize for the typographical error where 'uninjured' was misspelled as 'uninured.' This has been corrected in the revised manuscript.  
+
+- Page 17, Methods, Western Blot: 'Proteins' should be 'Protein'  
+
+We have corrected the text. It is now on page 19 of the revised manuscript.  
+
+- Page 17, Methods, Mouse surgical procedures and behavioural test: There is an unfinished sentence 'Sciatic nerve crush surgery and subsequent local application were performed aseptically under 2% isoflurane anesthesia followed by'?  
+
+We have corrected this. It's now on page 20 of the revised manuscript.  
+
+- Page 18, Methods, Mouse surgical procedures and behavioural test: 'To assess functional recovery, lumbrical muscles were dissected form the paws' This need more detail for a functional recovery assessment.  
+
+- Page 19, Methods, Statistics: Please add what post-hoc tests used.  
+
+We have added more information to the Method section (page 20- 22 of the revised MS).
+
+<--- Page Split --->
+
+REFERENCESS1. Qian T, Wang P, Chen Q, Yi S, Liu Q, Wang H, et al. The dynamic changes of main cell types in the microenvironment of sciatic nerves following sciatic nerve injury and the influence of let- 7 on their distribution. RSC Adv. 2018;8(72):41181- 91.2. Yang Y, Wen D, Lin F, Song X, Pang R, Sun W, et al. Suppression of non- muscle myosin II boosts T cell cytotoxicity against tumors. Sci Adv. 2024;10(44):eadp0631.3. Ma CH, Omura T, Cobos EJ, Latremoliere A, Ghasemlou N, Brenner GJ, et al. Accelerating axonal growth promotes motor recovery after peripheral nerve injury in mice. J Clin Invest. 2011;121(11):4332- 47.4. Dravid A, Raos B, Svirskis D, O'Carroll SJ. Optimised techniques for high- throughput screening of differentiated SH- SY5Y cells and application for neurite outgrowth assays. Sci Rep. 2021;11(1):23935.5. Elliott H, Fischer RS, Myers KA, Desai RA, Gao L, Chen CS, et al. Myosin II controls cellular branching morphogenesis and migration in three dimensions by minimizing cell- surface curvature. Nat Cell Biol. 2015;17(2):137- 47.6. Swaminathan V, Waterman CM. The molecular clutch model for mechanotransduction evolves. Nat Cell Biol. 2016;18(5):459- 61.7. Vicente- Manzanares M, Ma X, Adelstein RS, Horwitz AR. Non- muscle myosin II takes centre stage in cell adhesion and migration. Nat Rev Mol Cell Biol. 2009;10(11):778- 90.8. Soleman S, Filippov MA, Dityatev A, Fawcett JW. Targeting the neural extracellular matrix in neurological disorders. Neuroscience. 2013;253:194- 213.9. Kerschensteiner M, Schwab ME, Lichtman JW, Misgeld T. In vivo imaging of axonal degeneration and regeneration in the injured spinal cord. Nat Med. 2005;11(5):572- 7.10. Kalil K, Dent EW. Branch management: mechanisms of axon branching in the developing vertebrate CNS. Nat Rev Neurosci. 2014;15(1):7- 18.11. Kovacs M, Toth J, Hetenyi C, Malnasi- Csizmadia A, Sellers JR. Mechanism of blebbistatin inhibition of myosin II. J Biol Chem. 2004;279(34):35557- 63.12. Straight AF, Cheung A, Limouze J, Chen I, Westwood NJ, Sellers JR, et al. Dissecting temporal and spatial control of cytokinesis with a myosin II Inhibitor. Science. 2003;299(5613):1743- 7.13. Dent EW, Gupton SL, Gertler FB. The growth cone cytoskeleton in axon outgrowth and guidance. Cold Spring Harb Perspect Biol. 2011;3(3).14. Turney SG, Bridgman PC. Laminin stimulates and guides axonal outgrowth via growth cone myosin II activity. Nat Neurosci. 2005;8(6):717- 9.15. Woo S, Gomez TM. Rac1 and RhoA promote neurite outgrowth through formation and stabilization of growth cone point contacts. J Neurosci. 2006;26(5):1418- 28.16. Zhang Z, Roberson DP, Kotoda M, Boivin B, Bohnslav JP, Gonzalez- Cano R, et al. Automated preclinical detection of mechanical pain hypersensitivity and analgesia. Pain. 2022;163(12):2326- 36.
+
+<--- Page Split --->
+
+## RESPONSE TO REVIEWER COMMENTS  
+
+We thank the reviewers and have addressed the key points raised by Reviewer 2 by incorporating them into the Results and Discussion sections, and updating a figure, as summarized here. The revised text in the MS is highlighted in yellow.  
+
+## Revised figure panel including new data:  
+
+Fig. S7A now includes an additional panel displaying NF- positive axons on day 7 post- injury  
+
+## Reviewer #2:  
+
+The authors have thoroughly addressed my questions and those of the other reviewers. My only suggestion is that the authors be careful in their mention of sensory neuroregeneration, which does not lead to functional recovery. The majority of sensory neurons are indeed made up of nociceptors, which were examined here, but when it comes to motor recovery, it is more the mechanoreceptors (TrkB) and proprioceptors (TrKC) that would be of interest that was indicated by reviewer #3. Therefore, please add this caveat. NF, while bringing up myelinated neurons, would not be specific to myelinated sensory neurons alone, which is what the comment on 281 of page 11 makes confusing. These slight clarification changes are all that are required for acceptance.  
+
+We thank the reviewer for raising this important point. Mechanoreceptor and proprioceptor sensory neurons play a crucial role in sensorimotor feedback and motor function recovery after nerve injury (1- 3). Our paw luminance data likely reflects motor behavior primarily driven by motor axon regeneration but may also capture contributions from the regeneration of those proprioceptor and mechanoreceptor sensory neurons. An additional NF panel in Fig. S7A now shows that NF- positive but ChAT- negative axons are present in the regenerating nerve after local NMII treatment, suggesting that proprioceptor or mechanoreceptor sensory neurons also regrow faster. These findings imply that the motor function recovery observed after NMII could be the result of an accelerated regeneration of both motor and sensory neurons. Identification of the specific actions of NMII on those different sensory subtypes which contribute to motor function recovery in future studies, would provide valuable insights into the utility of this approach for promoting a full recovery of motor function.
+
+<--- Page Split --->
+![PLACEHOLDER_28_0]
+
+<center>Fig. S7A. Regenerating axons in distal nerve 7 days post-sciatic nerve crush injury (A) Representative images of mouse sciatic nerves labeled with an antibody against Neurofilament (NF) and Choline Acetyltransferase (ChAT) on day 7 post-sciatic nerve crush injury. White dashed lines indicate the boundary of the crush injury site, and arrowhead indicates tips of regenerated axons. </center>  
+
+## REFERENCES  
+
+1. de Nooij JC, Zampieri N. The making of a proprioceptor: a tale of two identities. Trends Neurosci. 2023;46(12):1083-94.  
+2. Oliver KM, Florez-Paz DM, Badea TC, Mentis GZ, Menon V, de Nooij JC. Molecular correlates of muscle spindle and Golgi tendon organ afferents. Nat Commun. 2021;12(1):1451.  
+3. Wu H, Petitpre C, Fontanet P, Sharma A, Bellardita C, Quadros RM, et al. Distinct subtypes of proprioceptive dorsal root ganglion neurons regulate adaptive proprioception in mice. Nat Commun. 2021;12(1):1026.
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__6d1f1284a3af25f662a14484e7f1e7495d11a0328d008eb6ec4dba657744c3fc/supplementary_0_Transparent Peer Review file__6d1f1284a3af25f662a14484e7f1e7495d11a0328d008eb6ec4dba657744c3fc_det.mmd

@@ -0,0 +1,809 @@
+<|ref|>title<|/ref|><|det|>[[72, 53, 295, 80]]<|/det|>
+# nature portfolio  
+
+<|ref|>text<|/ref|><|det|>[[75, 97, 296, 119]]<|/det|>
+Peer Review File  
+
+<|ref|>title<|/ref|><|det|>[[72, 160, 905, 210]]<|/det|>
+# NON-MUSCLE MYOSIN II INHIBITION AT THE SITE OF AXON INJURY INCREASES AXON REGENERATION  
+
+<|ref|>text<|/ref|><|det|>[[72, 224, 480, 240]]<|/det|>
+Corresponding Author: Professor Clifford Woolf  
+
+<|ref|>text<|/ref|><|det|>[[72, 274, 864, 289]]<|/det|>
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+<|ref|>text<|/ref|><|det|>[[72, 326, 144, 340]]<|/det|>
+Version 0:  
+
+<|ref|>text<|/ref|><|det|>[[72, 353, 220, 367]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[72, 379, 160, 393]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[72, 404, 238, 417]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 418, 920, 485]]<|/det|>
+Heo and colleagues explored the effect of a blebbistatin analogue on axon growth and nerve regeneration both in vitro and in vivo. After screening over 4000 compounds, they identified 10 main candidates for improving axon outgrowth, focusing on blebbistatin, a known NMII inhibitor. However, due to its poor biocompatibility and its photosensitivity, they proceeded with an analogue called NMII2. They tested the drug in vitro on various cellular models and in vivo in a mouse model of sciatic nerve regeneration following a crush injury.  
+
+<|ref|>text<|/ref|><|det|>[[72, 484, 754, 498]]<|/det|>
+They evaluated neurite elongation in vitro and in vivo and conducted studies on functional recovery.  
+
+<|ref|>text<|/ref|><|det|>[[72, 509, 916, 550]]<|/det|>
+The work is well- structured and thoughtfully designed and however the novelty is limited given that the effect of blebbistatin in neuronal outgrowth has been described and the proposed mechanism is not surprising and not very well developed. The regenerative phenotype observed is statistically significant but not especially exciting compared to other approaches.  
+
+<|ref|>text<|/ref|><|det|>[[72, 562, 159, 575]]<|/det|>
+Major points  
+
+<|ref|>text<|/ref|><|det|>[[72, 587, 921, 628]]<|/det|>
+- The choice to include the MYH9 knockdown model in this manuscript is unclear. It has long been known that inhibition of NMII leads to increased axonal growth through a well-documented mechanism. Validation of this model in the present article adds little to the existing literature or to the quality of this manuscript  
+
+<|ref|>text<|/ref|><|det|>[[72, 627, 900, 653]]<|/det|>
+The mechanism has not been well studied neither at the cellular nor signalling level. Cell-specific unbiased approaches could reveal novel interesting mechanisms.  
+
+<|ref|>text<|/ref|><|det|>[[72, 665, 920, 705]]<|/det|>
+- It is unclear why the authors tested non-permissive substrates in culture to mimic CNS regeneration and then they focused on peripheral regeneration in vivo. The authors should clarify this aspect. Did they test regeneration in the CNS? This would add to the manuscript's impact.  
+
+<|ref|>text<|/ref|><|det|>[[72, 717, 920, 744]]<|/det|>
+- With the regards to the in vivo drug delivery exp, which represent the main novelty of the work, there are several aspects to consider:  
+
+<|ref|>text<|/ref|><|det|>[[72, 744, 920, 836]]<|/det|>
+i) Local delivery of the NMII inhibitor might affect not only the injured axons but also the entire non-neuronal component present in the sciatic nerve. Since this is a chronic treatment (over several days), there could be side effects on glial or immune cells that need to be considered, especially since these could potentially lead to systemic issues or beneficial effects. These cell populations should be evaluated, at least through histochemistry, to assess their condition. ii) Similarly, while the machine learning method is powerful, it is not sufficient to demonstrate complete and physiological functional recovery. In the peripheral system, there are various neuronal populations involved at different levels in sensation and perception, which should be analyzed independently.  
+
+<|ref|>text<|/ref|><|det|>[[72, 848, 190, 861]]<|/det|>
+Additional points  
+
+<|ref|>text<|/ref|><|det|>[[72, 874, 716, 888]]<|/det|>
+- The phenotypic screening should be implemented with more procedural details. Specifically:  
+
+<|ref|>text<|/ref|><|det|>[[72, 888, 900, 926]]<|/det|>
+i) Why were two different cell lines used between the first and second screen? The authors should clarify this in the text. ii) Some of these compounds may cause toxicity that is not necessarily reflected in neurite length but in cell survival. Was neuronal survival assessed at the concentrations used?
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 46, 894, 75]]<|/det|>
+- The authors should clarify how the concentration of \(100 \mu M\) was established for in vivo use, given that the drug is being tested for the first time to test nerve regeneration.  
+
+<|ref|>text<|/ref|><|det|>[[72, 85, 920, 140]]<|/det|>
+- There are many mechanisms that have been studied and could underlie the results observed by the authors in this work, beyond the Rho/ROCK pathway (such as increased microtubule stability, alterations in interactions with the environment via adhesion molecules, growth cone dynamics, changes in intrinsic forces, etc.). The authors should, even briefly, expand the discussion to include these aspects  
+
+<|ref|>text<|/ref|><|det|>[[72, 150, 911, 167]]<|/det|>
+- The authors should briefly explain why silicone tubes were chosen for local drug application as opposed to other methods.  
+
+<|ref|>text<|/ref|><|det|>[[73, 190, 283, 204]]<|/det|>
+(Remarks on code availability)  
+
+<|ref|>sub_title<|/ref|><|det|>[[73, 227, 162, 241]]<|/det|>
+## Reviewer #2  
+
+<|ref|>text<|/ref|><|det|>[[73, 254, 238, 268]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[73, 268, 308, 281]]<|/det|>
+- What are the noteworthy results?  
+
+<|ref|>text<|/ref|><|det|>[[72, 281, 918, 386]]<|/det|>
+Heo et al. in the work entitled "NON- MUSCLE MYOSIN II INHIBITION AT THE SITE OF AXON INJURY INCREASES AXON REGENERATION" employed an unbiased screen of axonal growth promotors on a growth inhibitory environment (CSPG) and found several compounds that increased the outgrowth of human iPSC derived motor neurons. Of these, the highest hit was blebbistatin, which inhibits NMII. Although used previously in similar conditions as shown here, this work examined a new NMII2 that is more bioavailable than previous versions. This was shown to work to extend dendrites proximal and axons distal to the lesion site following laser dissection of spot cultures. More importantly, this work showed for the first time in vivo local NMII2 treatment of the injured sciatic nerve leads to increased regeneration within the nerve and the muscle of the foot, increased synaptic connections and functional recovery.  
+
+<|ref|>text<|/ref|><|det|>[[72, 397, 920, 425]]<|/det|>
+- Will the work be of significance to the field and related fields? How does it compare to the established literature? If the work is not original, please provide relevant references.  
+
+<|ref|>text<|/ref|><|det|>[[72, 425, 916, 516]]<|/det|>
+Blebbistatin and NMII inhibition have been found previously to be involved in axonal growth of human NPCs (10.3389/fcell.2021.719636), RGCs and the optic nerve (10.1016/j.celrep.2020.107537), and embryonic DRG (10.1073/pnas.1011258108). Here for the first time a new NMII2 that is more bioavailable than previous versions was examined and found to promote axonal regeneration when placed at the site of injury, the sciatic nerve, leading to faster functional recovery. This work takes the field into clinical relevance. It can also be noted that many of the in vitro models are typically considered immature neurons (which tend to be more growth-permissive) except for the adult DRG work, so the in vivo work shows that in a mature neuronal state, the NMII2 works effectively.  
+
+<|ref|>text<|/ref|><|det|>[[72, 528, 920, 842]]<|/det|>
+- Does the work support the conclusions and claims, or is additional evidence needed? This work states that NMMII works well when given at the lesion site inducing accelerated regeneration and functional recovery. However, one could strengthen these conclusions with some additions. Firstly, in Figure 6 the co-localization is difficult to see with current magnifications. Why are the NF and Syn in the same color? Please provide higher magnifications of the co-localization analysis and have the antibodies split into separate channels/colors along with the merged overlap. This analysis was from D14 when the greatest behavioral differences were seen. It should also be provided at the end timepoint (D28) and possibly from D7 where no behavioral differences are observed. The assumption would be that there would be less colocalization overall at D7 because there wouldn't be enough growth into the muscle by this timepoint. As for the D28 at the end of the study where the behavior is the same between the treated and vehicle groups it would be important to know if the growth and co-localization are the same or if this is due to compensatory mechanisms that arise in the vehicle group. This would further support the correlation graph in Figure 6, which is a little weak given that two of the five samples sit at zero for colocalization yet have very different behavioral outcomes. Is the correlation work from D14? This should be stated somewhere. Further, it should be spoken about why this work is important given there is functional recovery at the end of the study similar to the treated group. What is the benefit of the accelerated recovery? In the introduction it is stated, "After a traumatic nerve injury, mature motor neurons only regenerate at a rate of 1-4 mm per day, the slowness of which results in only \(\sim 45\%\) of injured axons achieving anatomical reinnervation of their muscle targets with, in consequence, limited functional recovery (1-7). Discovering therapeutic targets for the effective treatment of axon injury is therefore critical to ensure a more complete recovery of motor and other disturbed functions." However, it appears that in Figure 6 the controls reach the same level of functional recovery as the treated group at D28. This may just be a difference between rodents and humans, either way this should be discussed. Some of the additions suggested here may help address this issue. Given the work done to show that NMII works in various cell types it would make sense to also look at the sciatic nerves shown in Fig. S6 for sensory neuronal regeneration with SCG10. This would take the in vitro work into the in vivo relevance. As there is not any increased mechanistic understanding greater than what was published before, I do believe these additions are important to show the novelty and relevance of this work.  
+
+<|ref|>text<|/ref|><|det|>[[72, 840, 896, 881]]<|/det|>
+Lastly, it should be stated why \(10mg / kg\) of NMII2 was used systemically and whether there were any obvious things that were associated with weight loss, such as a decrease in organ size or discoloration. Also how long is it thought for the hydrogel release of NMII2 at the lesion site, immediate or over days, weeks? This will be of interest to the readers.  
+
+<|ref|>text<|/ref|><|det|>[[72, 892, 920, 920]]<|/det|>
+- Are there any flaws in the data analysis, interpretation and conclusions? - Do these prohibit publication or require revision? I do not see any flaws just some additions that would strengthen the conclusions as I stated above.  
+
+<|ref|>text<|/ref|><|det|>[[72, 930, 664, 945]]<|/det|>
+- Is the methodology sound? Does the work meet the expected standards in your field?
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[74, 47, 377, 60]]<|/det|>
+The methodology is the standard of the field.  
+
+<|ref|>text<|/ref|><|det|>[[73, 73, 880, 168]]<|/det|>
+- Is there enough detail provided in the methods for the work to be reproduced? There are some minor additions needed in the methods: 
+- The correlation analysis for Figure 6 is not listed nor is it stated if this is a Pearson or Spermann correlation. 
+- The post hoc tests are not listed by name, it would be helpful if the authors listed if they chose a more conservative or liberal post hoc analysis method. 
+- \(10\mathrm{mg / kg}\) twice a day for 3 days via i.p. of the NMMlll was used for systemic application, but why was this dose/timing/method chosen? This should be listed given the side effects that were observed.  
+
+<|ref|>text<|/ref|><|det|>[[73, 191, 525, 218]]<|/det|>
+(Remarks on code availability) I could not access the code when I tried to click on the GitHub link.  
+
+<|ref|>text<|/ref|><|det|>[[73, 230, 161, 243]]<|/det|>
+Reviewer #3  
+
+<|ref|>text<|/ref|><|det|>[[73, 257, 238, 270]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[73, 270, 917, 309]]<|/det|>
+The manuscript entitled "Non- muscle myosin II inhibition at the site of axon injury increases axon regeneration" is a study in which the authors produce a novel myosin II inhibitor (NMIli2; analog of blebbistatin) and show it's efficacy in induce axon growth in tissue culture and after peripheral nerve injury in adult mice.  
+
+<|ref|>text<|/ref|><|det|>[[73, 320, 920, 387]]<|/det|>
+Following a phenotypic screen in human iPSC derived motor neurons, the authors identified 10 compounds that induced neurite outgrowth on a CSPG extracellular matrix environment. Blebbistatin was found to be the strongest inducer of growth, with and without laser cut injury, however given it's issues with photosensitivity and bioavailability, the authors produced a more selective non- muscle myosin II inhibitor, termed NMIli2. As non- muscle myosin II is a motor protein that normally stabilises microtubules, use of NMIli2 would be hypothesised to destabilise MTs thus inducing axon elongation/extension.  
+
+<|ref|>text<|/ref|><|det|>[[72, 398, 923, 503]]<|/det|>
+To confirm the findings from their screen as well as to confirm the efficacy of NMIli2, the authors grew human iPSC- derived motor neurons, cortical- like neurons, and human cortical organoids on either laminin or CSPGs; as well as adult mouse DRGs and RGCs on laminin only. The authors also compared their findings with NMIli2 with a lentiviral knockdown of MYH9 (encoding the heavy chain of NMIli)in human iPSC- derived motor neurons, demonstrating axon growth with the KD with confirmed KD \((- 55\%)\) with a western blot. The authors then tested NMIli2 in vivo in a mouse model of peripheral nerve injury, showing that local/direct administration complexed to a collagen hydrogel, compared with intraperitoneal administration or control, was effective at inducing ChAT- labelled axon growth, synaptic reinnervation and better hindpaw plantar placement.  
+
+<|ref|>text<|/ref|><|det|>[[72, 515, 911, 543]]<|/det|>
+While the findings of this study are very interesting for the field of axon regeneration specifically in the PNS, there are some areas of the study that require clarification and/or further information.  
+
+<|ref|>text<|/ref|><|det|>[[73, 555, 188, 568]]<|/det|>
+Major questions:  
+
+<|ref|>text<|/ref|><|det|>[[72, 580, 910, 672]]<|/det|>
+- Why were the mouse cellular experiments using adult RGCs and adult DRGs not performed on a CSPG matrix? CSPG substrates were used in the motor and cortical-like neuron experiments with success. All of these cell types have been shown to respond to CSPGs with retraction and/or overall stunted growth, especially in adult ages. Some studies have shown that the lack of growth on these substrates is due to integrin are downregulated and or subsequent inactivation, leaving adult cells unable to respond to the CSPG matrix. Given the mechanism of action of destabilising MTs, do the authors think that NMIli2 would induce the growth in the mouse cells on CSPG regardless of integrin regulation? If not, can the authors discuss the mechanism of action in the CNS neurons that differs in the PNS neurons?  
+
+<|ref|>text<|/ref|><|det|>[[72, 685, 915, 724]]<|/det|>
+- The axon growth in vitro suggests that blebbistatin and NMIli2 both induce neurite/axon elongation as well as branching. The branching effect is indeed an interesting finding, which would in some cases compete with elongation. Can the authors discuss the mechanisms of action by which the compound can drive both an elongation and a branching phenotype?  
+
+<|ref|>text<|/ref|><|det|>[[72, 736, 916, 789]]<|/det|>
+- In the MYH9 experiments, a western blot was performed to indicate knockdown of MYH9. Have the authors looked at changes in MYH9 in their NMIli2 treated cells? Would they expect a greater effect, or a less specific effect? Can the authors perform immunofluorescence with the MYH9 antibody to confirm that the knockdown cells are the cells with increased axon growth?  
+
+<|ref|>text<|/ref|><|det|>[[72, 802, 916, 841]]<|/det|>
+- Hydrogels were used in the peripheral nerve injury experiments for drug delivery. What are the biokinetics of these hydrogels? Are they absorbable, and do they have a regenerative effect alone, when used long term for nerve repair? There was recovery of function with vehicle at 28 days equal to NMIli2 treated animals, did the authors expect this?  
+
+<|ref|>text<|/ref|><|det|>[[72, 854, 915, 907]]<|/det|>
+- ChAT immunohistochemistry was used to examine motor axon regeneration in the sciatic nerve. Is there any possibility of tracing axons, rather than relying on antibody staining for regenerative growth? How was the injury confirmed, as the methods only state that the sural nerve was crushed for 20 sec with a hemostat? The sural nerve is primarily a sensory nerve, so I query why this was used to assess motor axon regeneration.  
+
+<|ref|>text<|/ref|><|det|>[[72, 919, 915, 946]]<|/det|>
+- Likewise, to assess motor recovery, hindpaw placement was assessed. Plantar placement is mainly used to assess sensory recovery, or possibly sensorimotor feedback following injury and repair. While the results found are promising, how
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 46, 895, 88]]<|/det|>
+do the authors discount the potential for sensory axon changes inducing this effect? Have the authors looked at sensory axon repair in this context? Very little detail is included in the methods section for this test. Can further detail be included (length of walking tract, timing of test, etc)  
+
+<|ref|>text<|/ref|><|det|>[[73, 98, 644, 114]]<|/det|>
+- did the authors test for changes in Rho/ROCK activity in their in vitro experiments?  
+
+<|ref|>text<|/ref|><|det|>[[73, 138, 165, 152]]<|/det|>
+Minor issues:  
+
+<|ref|>text<|/ref|><|det|>[[72, 175, 910, 270]]<|/det|>
+- Page 9, paragraph 2: why is the term 'frustrated' used for total internal reflection? 'Uninjured' misspelled as 'uninured'.  
+- Page 17, Methods, Western Blot: 'Proteins' should be 'Protein'.  
+- Page 17, Methods, Mouse surgical procedures and behavioural test: There is an unfinished sentence 'Sciatic nerve crush surgery and subsequent local application were performed aseptically under 2% isoflurane anesthesia followed by'?  
+- Page 18, Methods, Mouse surgical procedures and behavioural test: 'To assess functional recovery, lumbrical muscles were dissected form the paws' This need more detail for a functional recovery assessment.  
+- Page 19, Methods, Statistics: Please add what post-hoc tests used.  
+
+<|ref|>text<|/ref|><|det|>[[73, 360, 282, 374]]<|/det|>
+(Remarks on code availability)  
+
+<|ref|>text<|/ref|><|det|>[[73, 399, 145, 412]]<|/det|>
+Version 1:  
+
+<|ref|>text<|/ref|><|det|>[[73, 424, 220, 438]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[73, 450, 160, 463]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[73, 476, 238, 490]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 490, 923, 544]]<|/det|>
+The authors have addressed some of my comments and concerns. However, the limited impact of the intervention on the peripheral regenerative phenotype (in comparison many other interventions are superiors to this and no in vivo work in CNS injury models has been performed), the very limited novelty and the lack of more comprehensive mechanistic insight do hamper my enthusiasm for this study.  
+
+<|ref|>text<|/ref|><|det|>[[73, 555, 282, 569]]<|/det|>
+(Remarks on code availability)  
+
+<|ref|>text<|/ref|><|det|>[[73, 593, 161, 606]]<|/det|>
+Reviewer #2  
+
+<|ref|>text<|/ref|><|det|>[[73, 619, 238, 632]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 632, 920, 724]]<|/det|>
+The authors have thoroughly addressed my questions and those of the other reviewers. My only suggestion is that the authors be careful in their mention of sensory neuroregeneration, which does not lead to functional recovery. The majority of sensory neurons are indeed made up of nociceptors, which were examined here, but when it comes to motor recovery, it is more the mechanoreceptors (TrkB) and proprioceptors (TrKC) that would be of interest that was indicated by reviewer #3. Therefore, please add this caveat. NF, while bringing up myelinated neurons, would not be specific to myelinated sensory neurons alone, which is what the comment on 281 of page 11 makes confusing. These slight clarification changes are all that are required for acceptance.  
+
+<|ref|>text<|/ref|><|det|>[[73, 736, 282, 750]]<|/det|>
+(Remarks on code availability)  
+
+<|ref|>text<|/ref|><|det|>[[73, 775, 161, 789]]<|/det|>
+Reviewer #3  
+
+<|ref|>text<|/ref|><|det|>[[73, 802, 238, 815]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 815, 909, 842]]<|/det|>
+Thanks to the authors for responding to my comments and queries and amending their manuscript. This manuscript is now suitable for publication.  
+
+<|ref|>text<|/ref|><|det|>[[73, 854, 282, 868]]<|/det|>
+(Remarks on code availability)
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 45, 916, 99]]<|/det|>
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+<|ref|>text<|/ref|><|det|>[[72, 99, 796, 113]]<|/det|>
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source.  
+
+<|ref|>text<|/ref|><|det|>[[72, 112, 910, 165]]<|/det|>
+The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.  
+
+<|ref|>text<|/ref|><|det|>[[72, 165, 618, 179]]<|/det|>
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[115, 88, 459, 107]]<|/det|>
+## RESPONSE TO REVIEWER COMMENTS  
+
+<|ref|>text<|/ref|><|det|>[[115, 116, 875, 152]]<|/det|>
+We thank the reviewers for their constructive comments and valuable suggestions. We address each comment raised below, which include the new experiments and analyses recommended.  
+
+<|ref|>text<|/ref|><|det|>[[115, 160, 860, 196]]<|/det|>
+We have in consequence modified many of the figures, as summarized here. All the modified text in the revised manuscript is highlighted in yellow.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 205, 713, 224]]<|/det|>
+## Revised figure panels including new data requested by the reviewers:  
+
+<|ref|>text<|/ref|><|det|>[[115, 232, 680, 251]]<|/det|>
+Former Fig. 5 has been moved to supplementary figure (now Fig. S6).  
+
+<|ref|>text<|/ref|><|det|>[[115, 259, 880, 296]]<|/det|>
+Fig. 5c and 5g (was Fig. 6 before) now include additional sample numbers for both vehicle and treated groups.  
+
+<|ref|>text<|/ref|><|det|>[[115, 303, 775, 323]]<|/det|>
+Added a new Fig. 6, illustrating the effect of NMIIi on sensory neuron regeneration.  
+
+<|ref|>text<|/ref|><|det|>[[115, 331, 881, 367]]<|/det|>
+Added new panels to Fig. S1 (panels i- k) to present results of cell viability at the concentrations used in the compound screen.  
+
+<|ref|>text<|/ref|><|det|>[[115, 375, 880, 394]]<|/det|>
+Added a new figure panel, Fig. S5j, to depict RGC regeneration under CSPG- coated conditions.  
+
+<|ref|>text<|/ref|><|det|>[[115, 402, 866, 439]]<|/det|>
+Added a new panel to supplementary figure, Fig. S8 (panel a- c), to show nerve innervation in muscles at days 7 and 28 post- injury.  
+
+<|ref|>text<|/ref|><|det|>[[115, 447, 857, 483]]<|/det|>
+Added a new panel to supplementary figure, Fig. S8 (panel d), to illustrate effect of hydrogel itself on regeneration.  
+
+<|ref|>text<|/ref|><|det|>[[115, 491, 860, 527]]<|/det|>
+Added a new supplementary figure, Fig. S10, to demonstrates that NMIIi2 specifically targets neurons.  
+
+<|ref|>text<|/ref|><|det|>[[115, 536, 835, 572]]<|/det|>
+Added a revised p- value with a more stringent approach and updated them in each figure panel.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[74, 90, 228, 108]]<|/det|>
+## 1 Reviewer #1:  
+
+<|ref|>text<|/ref|><|det|>[[70, 116, 877, 222]]<|/det|>
+2 Heo and colleagues explored the effect of a blebbistatin analogue on axon growth and nerve 3 regeneration both in vitro and in vivo. After screening over 4000 compounds, they identified 10 4 main candidates for improving axon outgrowth, focusing on blebbistatin, a known NMII inhibitor. 5 However, due to its poor biocompatibility and its photosensitivity, they proceeded with an 6 analogue called NMIIi2. They tested the drug in vitro on various cellular models and in vivo in a 7 mouse model of sciatic nerve regeneration following a crush injury.  
+
+<|ref|>text<|/ref|><|det|>[[70, 230, 833, 265]]<|/det|>
+8 They evaluated neurite elongation in vitro and in vivo and conducted studies on functional recovery.  
+
+<|ref|>text<|/ref|><|det|>[[70, 274, 884, 344]]<|/det|>
+10 The work is well- structured and thoughtfully designed and however the novelty is limited given 11 that the effect of blebbistatin in neuronal outgrowth has been described and the proposed 12 mechanism is not surprising and not very well developed. The regenerative phenotype observed 13 is statistically significant but not especially exciting compared to other approaches.  
+
+<|ref|>text<|/ref|><|det|>[[70, 353, 857, 405]]<|/det|>
+14 We appreciate the reviewer's recommendations. To enhance the novelty and discussion of 15 potential mechanisms underlying our finding, we have performed the suggested experiments 16 and analyses.  
+
+<|ref|>text<|/ref|><|det|>[[70, 415, 88, 429]]<|/det|>
+17  
+
+<|ref|>text<|/ref|><|det|>[[70, 443, 216, 460]]<|/det|>
+18 Major points  
+
+<|ref|>text<|/ref|><|det|>[[70, 469, 880, 539]]<|/det|>
+19 - The choice to include the MYH9 knockdown model in this manuscript is unclear. It has long 20 been known that inhibition of NMII leads to increased axonal growth through a well- documented 21 mechanism. Validation of this model in the present article adds little to the existing literature or 22 to the quality of this manuscript  
+
+<|ref|>text<|/ref|><|det|>[[70, 549, 875, 618]]<|/det|>
+23 We agree that it is well- established that NMII inhibition promotes axonal growth. Including 24 MYH9 knockdown experiments was, however, crucial to confirm the specificity of our findings 25 and to ensure that the observed effects were directly mediated by NMII inhibition, but have now 26 moved this figure to Supplementary Figure S6 since this is not a novel finding.  
+
+<|ref|>text<|/ref|><|det|>[[70, 628, 88, 642]]<|/det|>
+27  
+
+<|ref|>text<|/ref|><|det|>[[70, 655, 828, 690]]<|/det|>
+28 The mechanism has not been well studied neither at the cellular nor signalling level. Cell- 29 specific unbiased approaches could reveal novel interesting mechanisms.  
+
+<|ref|>text<|/ref|><|det|>[[70, 699, 880, 802]]<|/det|>
+30 We appreciate the reviewer's suggestion regarding cell- specific unbiased approaches and 31 acknowledge that the mechanism of NMII inhibition has not been fully elucidated at a cellular or 32 signaling level. Our current study is thought primarily focused on the screening platform and 33 evaluating the functional outcomes of the new NMII analogue and its therapeutic potential for 34 promoting axonal regeneration by local application to an injured nerve rather than on the cellular 35 mechanism of NMII.  
+
+<|ref|>text<|/ref|><|det|>[[70, 812, 877, 900]]<|/det|>
+36 In attempt to provide an understanding of signaling changes induced by NMII inhibition, we 37 performed a bulk RNA- seq analysis of motor neurons treated with the NMII inhibitor blebbistatin 38 on CSPG substrates, which provided some insight into molecular changes induced by NMII 39 inhibition. Pathway analysis of the data revealed elevated activity in Rho- related pathways, 40 integrin- mediated cell surface interactions, and extracellular matrix (ECM) degradation. These
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[65, 88, 877, 160]]<|/det|>
+41 findings imply that NMII inhibition of cell bodies increase integrin levels and enhance cellular interactions with the ECM in CSPG- rich environments, potentially facilitating regeneration in the central nervous system. However, we decided not to include this data in the revised MS for the following reasons:  
+
+<|ref|>text<|/ref|><|det|>[[65, 168, 881, 273]]<|/det|>
+45 - Activity- driven vs. Transcriptional changes: The phenotype observed in our study 46 manifests very rapidly after treatment, as shown in Figures 3D and 3E, suggesting that activity- 47 dependent or post- translational mechanisms rather than transcriptional changes likely drive the 48 early/initial regenerative effects. Including the bulk RNA sequencing data could misrepresent the 49 temporal dynamics of NMII's effects by overemphasizing transcriptional changes that may not 50 directly contribute to the regeneration phenotype.  
+
+<|ref|>text<|/ref|><|det|>[[65, 281, 877, 351]]<|/det|>
+51 - Localized mechanisms at injured axonal sites: Our findings in Figures 4 and 5 strongly 52 suggest that NMII inhibition exerts its effects locally at the injured axonal sites and not in the cell 53 body. Bulk RNA sequencing lacks the resolution to pinpoint localized axonal mechanisms. Our 54 focus is on the site- specific changes critical to regeneration.  
+
+<|ref|>text<|/ref|><|det|>[[65, 360, 867, 448]]<|/det|>
+55 We agree that cell- specific approaches, such as single- cell RNA sequencing or proteomics, 56 could provide a more granular and cellular understanding of NMII's effects. These approaches 57 represent logical steps for future investigations but are beyond the scope of the current study. 58 We have added a discussion of this point in the revised manuscript to emphasize the need for 59 future exploration of mechanisms with these approaches.  
+
+<|ref|>text<|/ref|><|det|>[[65, 460, 88, 473]]<|/det|>
+60  
+
+<|ref|>text<|/ref|><|det|>[[65, 485, 860, 554]]<|/det|>
+61 - It is unclear why the authors tested non- permissive substrates in culture to mimic CNS 62 regeneration and then they focused on peripheral regeneration in vivo. The authors should 63 clarify this aspect. Did they test regeneration in the CNS? This would add to the manuscript's 64 impact.  
+
+<|ref|>text<|/ref|><|det|>[[65, 563, 880, 650]]<|/det|>
+65 The primary outcome of this study was that localized NMII2 treatment to an injured nerve 66 promotes the regeneration of motor axons, even though the primary human motor neuron 67 screening platform was based on growth on CSPG. The CSPG substrate was selected because 68 this revealed pro- regenerative hits much more reliably than screening the neurons on laminin, 69 where axon growth occurs by itself.  
+
+<|ref|>text<|/ref|><|det|>[[65, 659, 880, 745]]<|/det|>
+70 We did, however, compare the effects of NMII in vitro on neurite outgrowth phenotypes on 71 CSPG and laminin substrates for human iPSC derived motor, sensory and cortical neurons as 72 well as primary mouse retinal ganglion cells. These findings reveal that NMII treatment 73 promotes neurite outgrowth in CNS and PNS neurons in both CSPG and laminin conditions. We 74 have refined the text to ensure greater clarity.  
+
+<|ref|>text<|/ref|><|det|>[[65, 757, 88, 770]]<|/det|>
+75  
+
+<|ref|>text<|/ref|><|det|>[[65, 783, 768, 801]]<|/det|>
+76 Did they test regeneration in the CNS? This would add to the manuscript's impact.  
+
+<|ref|>text<|/ref|><|det|>[[65, 810, 880, 897]]<|/det|>
+77 As our in vivo data indicates that NMII is not suitable for systemic delivery (Fig. S9), and local 78 delivery in an in vivo CNS- injury model presents significant technical challenges, we opted to 79 test this in vitro for this study. Specifically, we have now conducted new experiments to examine 80 the response of adult mouse dissociated RGCs in CSPG- coated conditions to mimic CNS 81 regeneration. RGCs exhibit very limited growth in non- permissive environments, but NMII2
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[65, 90, 850, 125]]<|/det|>
+treatment enable growth under these conditions (Fig. S5J) (and below) and now specify that future in vivo CNS studies are needed.  
+
+<|ref|>image<|/ref|><|det|>[[113, 141, 600, 293]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[115, 324, 662, 341]]<|/det|>
+<center>Figure S5. NMIIi induces neurite outgrowth in dissociated RGCs </center>  
+
+<|ref|>text<|/ref|><|det|>[[115, 341, 884, 391]]<|/det|>
+Representative RGC images of RBPMs- labeled neurons treated with DMSO, and \(10 \mu \mathrm{M}\) NMIIi (not included in fig) and quantification of neurite length on CSPG substrate for 24 hours (J). Data collected from 5 female mice. Statistics: unpaired two- tailed Student's \(t\) - test, SEM error bars.  
+
+<|ref|>text<|/ref|><|det|>[[115, 415, 839, 450]]<|/det|>
+- With the regards to the in vivo drug delivery exp, which represent the main novelty of the work, there are several aspects to consider:  
+
+<|ref|>text<|/ref|><|det|>[[115, 459, 872, 545]]<|/det|>
+i) Local delivery of the NMII inhibitor might affect not only the injured axons but also the entire non-neuronal component present in the sciatic nerve. Since this is a chronic treatment (over several days), there could be side effects on glial or immune cells that need to be considered, especially since these could potentially lead to systemic issues or beneficial effects. These cell populations should be evaluated, at least through histochemistry, to assess their condition.  
+
+<|ref|>text<|/ref|><|det|>[[112, 555, 881, 744]]<|/det|>
+To address this point, we conducted additional immunohistochemistry experiments to evaluate if there are changes in Schwann cells (S100+), immune cells (CCR2+), and activated fibroblasts (FAP+) following local NMIIi administration at the sciatic nerve injury site in mice. We assessed changes on day 3 post injury, a time point where these non-neuronal cells actively react to a nerve crush injury (1). However, we found that none of these cell types showed any major changes in the NMIIi-treated group compared to the vehicle-treated group (Fig. S10A- S10C). This suggests that NMIIi specifically targets axons. We have incorporated this data into the revised manuscript (Fig. S10). However, further investigations using flow cytometry or single-cell sequencing will be needed for a more comprehensive assessment of this issue. Given a recent study examining the role of NMII in T cells (2), it would be interesting in the future to explore its effects on other immune cell populations following local treatment.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[117, 92, 777, 352]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[113, 365, 830, 397]]<|/det|>
+<center>Figure S10. Assessment of non-neuronal cells at sciatic nerve injury site after NMII treatment </center>  
+
+<|ref|>text<|/ref|><|det|>[[115, 397, 860, 448]]<|/det|>
+(A- C) Images of sciatic nerves at day 3 post-injury in mice treated with a local application of either vehicle or NMII, labeled with anti- S100 (red) (A), anti- FAP (green) (B), and CCR2 (red) and/or DAPI (blue) (C).  
+
+<|ref|>text<|/ref|><|det|>[[115, 472, 863, 541]]<|/det|>
+ii) Similarly, while the machine learning method is powerful, it is not sufficient to demonstrate complete and physiological functional recovery. In the peripheral system, there are various neuronal populations involved at different levels in sensation and perception, which should be analyzed independently.  
+
+<|ref|>text<|/ref|><|det|>[[113, 551, 872, 741]]<|/det|>
+We have now investigated changes in sensory neuron regeneration following a sciatic nerve crush injury at various time points with or without local NMII application. Specifically, as Reviewer 2 suggested, we examined SCG10/STMN2, a microtubule- destabilized protein that is highly expressed after injury and widely used as an injury marker. On day 3 post- injury, SCG10/STMN2 intensity was noticeably higher in NMII2- treated nerves compared to the vehicle- treated mice (Fig. 6A and 6C). We also observed that CGRP- positive nociceptors are more prominently labeled in the NMII2- treated group, while the vehicle- treated group show more limited regeneration (Fig. 6A and 6C). We also further assessed regeneration of neurofilament (NF)- labeled axons and found that these also regenerate faster after NMII2 treatment (Fig. 6B), revealing enhanced regeneration across multiple sensory neuronal subtypes.  
+
+<|ref|>text<|/ref|><|det|>[[113, 750, 875, 907]]<|/det|>
+We also assessed sensory recovery, using a skin pinprick assay, as described previously (3). Briefly, we measured mouse nociceptive reactions to pinprick stimulation of the lateral area of the paw, which the sciatic nerve innervates prior to a crush injury. In line with the histology data, the NMII2- treated group exhibited a significantly improved sensory response to pinprick by day 7 post- injury, compared to the vehicle- treated mice (Fig. 6D). These data now reveal that NMII promotes functional regeneration of sensory neurons. The data also show that sensory and motor components recover at different times: sensory recovery is evident by day 7, whereas motor recovery only becomes prominent by day 14. This temporal difference underscores the complementary nature of our methods, with the machine- learning approach being particularly
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[60, 90, 880, 177]]<|/det|>
+effective for evaluating motor function recovery, as revealed by the positive correlation between paw luminance ratio and lumbrical muscle reinnervation (Fig. 5G) on day 14 (Fig. 6D) in Fig. 6E. These additional findings demonstrate that NMIIi2 treatment promotes functional recovery in vivo across multiple neuronal populations in the peripheral system. We include these results in the revised manuscript.  
+
+<|ref|>image<|/ref|><|det|>[[117, 192, 688, 776]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[113, 781, 775, 814]]<|/det|>
+<center>Figure 6. NMIIi applied at the sciatic nerve injury site promotes sensory axon regeneration and functional recovery </center>  
+
+<|ref|>text<|/ref|><|det|>[[113, 814, 861, 894]]<|/det|>
+(A and B) Representative images of mouse sciatic nerves day 3 post- injury treated with local application of either vehicle or NMIIi and labeled with anti- CGRP (green), anti- statthmin (magenta) (A) and neurofilament (green) (B), white dashed lines indicate the boundary of the crush injury site. (C) Quantification of CGRP and stathmin intensity from regenerated axons. Data collected from 3 mice per group and normalized to vehicle condition. Statistics: unpaired
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[113, 88, 875, 172]]<|/det|>
+two- tailed Student's t- test, SEM error bars. (D) Score of pinprick tests following local application of vehicle or NMIIi at baseline and day 4, 7, and 14 post- injury. Statistics: two- way ANOVA with Tukey's post hoc test, SEM error bars, 15 mice per condition. (E) Pearson correlation analysis of the paw luminance ratio and pinprick behavior scores on day 14 post- injury. Data collected from 10 male mice per group.  
+
+<|ref|>text<|/ref|><|det|>[[113, 196, 250, 212]]<|/det|>
+Additional points  
+
+<|ref|>text<|/ref|><|det|>[[113, 223, 864, 241]]<|/det|>
+- The phenotypic screening should be implemented with more procedural details. Specifically:  
+
+<|ref|>text<|/ref|><|det|>[[113, 250, 840, 285]]<|/det|>
+i) Why were two different cell lines used between the first and second screen? The authors should clarify this in the text.  
+
+<|ref|>text<|/ref|><|det|>[[113, 294, 884, 435]]<|/det|>
+For the primary screen we tested compounds at a single concentration. After identifying hit compounds, we conducted a secondary screen using a range of concentrations to evaluate their dose- dependent effects. To reduce biological variability, we performed the primary and secondary screens using three different human cell lines: SAH- 0047 from a 46- year- old female donor, LiPSC- GR1.1 from a male newborn donor, and 11a from a 36- year- old male donor. This decision was made to minimize potential biological variability and confirm that the observed effects were not cell line- specific and are representative of human motor neurons across diverse genetic and demographic backgrounds. We have clarified this in the revised manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[113, 471, 835, 506]]<|/det|>
+ii) Some of these compounds may cause toxicity that is not necessarily reflected in neuter length but in cell survival. Was neuronal survival assessed at the concentrations used?  
+
+<|ref|>text<|/ref|><|det|>[[113, 515, 861, 602]]<|/det|>
+We agree that compound- induced toxicity is an important issue. To address this we evaluated cell viability following treatment with the top compounds identified in the screen, using a viability/cytotoxicity assay with Calcein- AM and ethidium homodimer- 1 (EthD- 1) dyes (4). Treatment with blebbistatin our lead compound, neither changed cell viability nor induced cytotoxicity, indicating that it is safe at the tested concentrations (Fig. S1I- K).  
+
+<|ref|>text<|/ref|><|det|>[[113, 611, 870, 715]]<|/det|>
+Cells treated with two ROCK1/ROCK2 inhibitors (Fasudil and Y27632), a JAK2 inhibitor (CEP- 33779), or a voltage- gated calcium channel blocker (Benidipine- HCl) exhibited reduced EthD- 1 intensity (Fig. S1K), suggesting that while these compounds promote neurite outgrowth at the tested concentrations, they may also impact cell survival. These findings provide additional insight into the safety profile of the screened compounds and are included in the revised manuscript (Fig. S1I- K).  
+
+<|ref|>sub_title<|/ref|><|det|>[[113, 752, 840, 784]]<|/det|>
+## Figure S1. Assessment of cell viability/cytotoxicity of top 9 compounds from primary screen  
+
+<|ref|>text<|/ref|><|det|>[[113, 785, 880, 864]]<|/det|>
+(I) Representative images of neurons cultured on a laminin substrate labeled with Calcein and EthD-1 dye at 24h NMIIi2 post-treatment, and controls. (J and K) Quantification of Calcein (J) and EthD-1 (K) intensity from cells treated with the top hit compounds. Data normalized to DMSO, and experiments done in three replicates. Statistics: Welch's test with Dunnett's multiple comparison test, SEM error bars.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[115, 93, 757, 380]]<|/det|>  
+
+<|ref|>text<|/ref|><|det|>[[115, 400, 850, 437]]<|/det|>
+- The authors should clarify how the concentration of \(100 \mu \mathrm{M}\) was established for in vivo use, given that the drug is being tested for the first time to test nerve regeneration.  
+
+<|ref|>text<|/ref|><|det|>[[115, 444, 876, 549]]<|/det|>
+For the in vivo studies, we initially tested at a high concentration (10 mM) locally at the injury site. However, this either had no effect or slightly exacerbated motor deficits compared to vehicle-treated conditions, which are likely due to off-target effects. Subsequently, we optimized the dosage by testing a range of lower concentrations and observed that \(100 \mu \mathrm{M}\) provided the best balance between efficacy and safety. At this concentration, NMIIi2 significantly enhanced axonal regeneration and functional recovery without causing observable side effects.  
+
+<|ref|>image<|/ref|><|det|>[[118, 567, 636, 747]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[115, 765, 850, 799]]<|/det|>
+<center>Figure (not included in MS). Evaluation of motor function recovery with 10 mM and 25 mM doses of NMIIi at the sciatic nerve injury site. </center>  
+
+<|ref|>text<|/ref|><|det|>[[115, 799, 868, 896]]<|/det|>
+(A) Effects of local NMIIi2 treatment at high concentrations on paw luminance ratio after sciatic nerve crush injury. Paw luminance ratio (injured/non-injured) was measured for 10 minutes at each time point indicated on x-axis. NMIIi2 at 10 mM did not improve functional recovery compared to the vehicle group, while 25 mM NMIIi2 slightly but not significantly worsen luminance ratio. Data are presented as mean ± SEM. Statistical analysis was conducted using two-way ANOVA with Tukey's post hoc comparisons. n.s., not significant.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[113, 89, 865, 175]]<|/det|>
+- There are many mechanisms that have been studied and could underlie the results observed by the authors in this work, beyond the Rho/ROCK pathway (such as increased microtubule stability, alterations in interactions with the environment via adhesion molecules, growth cone dynamics, changes in intrinsic forces, etc.). The authors should, even briefly, expand the discussion to include these aspects  
+
+<|ref|>text<|/ref|><|det|>[[113, 185, 859, 220]]<|/det|>
+To include potential mechanisms underlying NMII's action in promoting regeneration we have now extended the discussion.  
+
+<|ref|>text<|/ref|><|det|>[[113, 230, 860, 300]]<|/det|>
+We suspect that in injured axons a Rho/ROCK- mediated activation of myosin II inhibits axon growth by modifying actomyosin contractile bundles at the growth cone. Disruption of these actomyosin arcs at the growth cone leads to microtubule destabilization, altering growth cone polarity and mechanical stress on the extracellular matrix, leading to elongation.  
+
+<|ref|>text<|/ref|><|det|>[[113, 309, 880, 499]]<|/det|>
+Rho/ROCK- independent changes in myosin II may also play a role in regeneration through local myosin II assembly. Myosin II activity is regulated through mechanisms distinct from its assembly (5). While myosin II reduces cell- surface curvature via a change in actomyosin tension, once recruited, the cell surface regulates myosin II stabilization and F- actin binding in a Rho/ROCK independent pathway. This process explains how NMII inhibitors can simultaneously enhance branch protrusion and axon elongation. Our study reveals that NMII inhibitors promote branch protrusion and altered growth cone dynamics, as evident by the thin growth cones lacking lamellipodia. Our findings suggest that NMII inhibitors modulate myosin II activity both via a ROCK- dependent pathway and by simultaneously promoting myosin II stabilization and F- actin binding through a ROCK- independent pathway, forming a positive feedback loop. However, further studies are required to confirm and expand upon these mechanisms.  
+
+<|ref|>text<|/ref|><|det|>[[115, 536, 875, 571]]<|/det|>
+- The authors should briefly explain why silicone tubes were chosen for local drug application as opposed to other methods.  
+
+<|ref|>text<|/ref|><|det|>[[115, 580, 876, 650]]<|/det|>
+Silicone- based implantations are biocompatible and widely used in medical research. Silicone tubes were chosen for their ability to provide localized, controlled drug delivery while minimizing systemic exposure, making them ideal for isolating NMIIi2's effects on nerve regeneration. We have added this information to the Method section.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 688, 228, 704]]<|/det|>
+## Reviewer #2:  
+
+<|ref|>text<|/ref|><|det|>[[115, 714, 395, 731]]<|/det|>
+- What are the noteworthy results?  
+
+<|ref|>text<|/ref|><|det|>[[115, 741, 881, 880]]<|/det|>
+Heo et al. in the work entitled "NON- MUSCLE MYOSIN II INHIBITION AT THE SITE OF AXON INJURY INCREASES AXON REGENERATION" employed an unbiased screen of axonal growth promotors on a growth inhibitory environment (CSPG) and found several compounds that increased the outgrowth of human iPSC derived motor neurons. Of these, the highest hit was blebbistatin, which inhibits NMII. Although used previously in similar conditions as shown here, this work examined a new NMIIi2 that is more bioavailable than previous versions. This was shown to work to extend dendrites proximal and axons distal to the lesion site following laser dissection of spot cultures. More importantly, this work showed for the first time in vivo local
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[113, 89, 856, 125]]<|/det|>
+NMll2 treatment of the injured sciatic nerve leads to increased regeneration within the nerve and the muscle of the foot, increased synaptic connections and functional recovery.  
+
+<|ref|>text<|/ref|><|det|>[[113, 160, 841, 196]]<|/det|>
+- Will the work be of significance to the field and related fields? How does it compare to the established literature? If the work is not original, please provide relevant references.  
+
+<|ref|>text<|/ref|><|det|>[[113, 205, 880, 360]]<|/det|>
+Blebbistatin and NMll inhibition have been found previously to be involved in axonal growth of human NPCs (10.3389/fcell.2021.719636), RGCs and the optic nerve (10.1016/j.celrep.2020.107537), and embryonic DRG (10.1073/pnas.1011258108). Here for the first time a new NMll2 that is more bioavailable than previous versions was examined and found to promote axonal regeneration when placed at the site of injury, the sciatic nerve, leading to faster functional recovery. This work takes the field into clinical relevance. It can also be noted that many of the in vitro models are typically considered immature neurons (which tend to be more growth-permissive) except for the adult DRG work, so the in vivo work shows that in a mature neuronal state, the NMll2 works effectively.  
+
+<|ref|>text<|/ref|><|det|>[[113, 370, 844, 405]]<|/det|>
+We thank the reviewer for the positive feedback, particularly the translational relevance and novelty of our findings with NMll2.  
+
+<|ref|>text<|/ref|><|det|>[[113, 443, 808, 460]]<|/det|>
+- Does the work support the conclusions and claims, or is additional evidence needed?  
+
+<|ref|>text<|/ref|><|det|>[[113, 469, 855, 574]]<|/det|>
+This work states that NMlll works well when given at the lesion site inducing accelerated regeneration and functional recovery. However, one could strengthen these conclusions with some additions. Firstly, in Figure 6 the co-localization is difficult to see with current magnifications. Why are the NF and Syn in the same color? Please provide higher magnifications of the co-localization analysis and have the antibodies split into separate channels/colors along with the merged overlap.  
+
+<|ref|>text<|/ref|><|det|>[[113, 584, 868, 670]]<|/det|>
+To address this, we have now performed additional experiments and include higher magnification images of the co- localization analysis in Fig 5B in the revised manuscript. The images show separate channels/colors for neurofilament (NF), synaptophysin (SYP), and α- bungarotoxin (BTX) alongside a merged overlap, ensuring a clearer representation of pre- and post- synaptic co- localization at the NMJ.  
+
+<|ref|>image<|/ref|><|det|>[[130, 686, 860, 864]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[112, 875, 848, 893]]<|/det|>
+<center>Figure 5. NMll at sciatic nerve injury site accelerates motor function recovery in mice </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[112, 90, 880, 237]]<|/det|>
+(B) Representative images of lumbrical muscles day 14 post-injury from mice treated by local application of vehicle or NMIIi labeled with anti-neurofilament (NF), anti-synaptophysin (SYP) or a-bungarotoxin (BTX). (C) Co-localization of pre/post-synaptic markers per muscle in mice with intact sciatic nerve (non-injury) and 14 days post-nerve injury. Number of presynaptic markers that overlap with the postsynaptic marker, \(\alpha\) -bungarotoxin normalized to post-synapse ( \(\alpha\) -bungarotoxin) number. Stats: Kruskal–Wallis test with Dunn's multiple comparison test, SEM error bars. Data collected from 13 male mice/group. (G) Pearson correlation between the paw luminance ratio and percentage of pre/post-synaptic colocalization from the same mouse shown in panel (C). Data collected from 16 male mice per group.  
+
+<|ref|>text<|/ref|><|det|>[[113, 250, 880, 422]]<|/det|>
+This analysis was from D14 when the greatest behavioral differences were seen. It should also be provided at the end timepoint (D28) and possibly from D7 where no behavioral differences are observed. The assumption would be that there would be less colocalization overall at D7 because there wouldn't be enough growth into the muscle by this timepoint. As for the D28 at the end of the study where the behavior is the same between the treated and vehicle groups it would be important to know if the growth and co-localization are the same or if this is due to compensatory mechanisms that arise in the vehicle group. This would further support the correlation graph in Figure 6, which is a little weak given that two of the five samples sit at zero for colocalization yet have very different behavioral outcomes. Is the correlation work from D14? This should be stated somewhere.  
+
+<|ref|>text<|/ref|><|det|>[[113, 432, 874, 570]]<|/det|>
+To provide a more comprehensive analysis, we have now performed additional experiments examining pre- and post- synaptic colocalization at the NMJ at days 7 and 28 post- injury. At day 7, the nerves had not yet reached the muscle in either the vehicle or treatment group (Fig. S8A and S8C), resulting in a lack of nerve innervation (Fig. S7). However, by day 28, nerve regeneration was complete, and pre- and post- synaptic colocalization in the muscle reached a similar level in NMIIi2- treated and non- injured conditions (Fig. S8B, and S8C). This indicates that successful reinnervation occurred in the vehicle- treated mice by day 28 (Fig. 5F) rather than compensatory mechanisms.  
+
+<|ref|>text<|/ref|><|det|>[[113, 580, 878, 666]]<|/det|>
+We also increased the sample size (16 mice in total) to analyze the correlation between colocalization and behavioral outcomes, focusing on D14, when the greatest differences were observed. This information is clarified in the main text and the Figure 5 legend. These additional data strengthen the conclusion that NMIIi2 treatment accelerates functional recovery by promoting neuromuscular junction reinnervation.  
+
+<|ref|>text<|/ref|><|det|>[[112, 676, 844, 744]]<|/det|>
+Nerve transection injury models typically show minimal functional recovery (3). In contrast, nerve crush injury enables nerve regeneration and a much, quicker recovery of motor and sensory function and is therefore, frequently used in regeneration studies (3). Our finding emphasizes the importance of faster axonal regeneration for enhancing functional recovery.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[115, 93, 812, 465]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[115, 476, 872, 590]]<|/det|>
+<center>Figure S8. Neuromuscular junctions days 7 and 28 post-sciatic nerve crush injury Representative images of lumbrical muscles at day 7 (A) and day 28 (B) post-crush injury from mice treated with local application of either vehicle or NMII, labeled with anti-neurofilament (NF), anti-synaptophysin (SYP), or \(\alpha\) -bunqarotoxin (BTX). (C) Quantification of presynaptic markers (NF; SYP) overlapping with postsynaptic marker (BTX), normalized to the number of postsynaptic sites. Statistics: one-way ANOVA with Tukey's post hoc test, SEM error bars. Data collected from 5 mice per group on day 7, and 4 mice per group on day 28. </center>  
+
+<|ref|>text<|/ref|><|det|>[[115, 614, 879, 805]]<|/det|>
+Further, it should be spoken about why this work is important given there is functional recovery at the end of the study similar to the treated group. What is the benefit of the accelerated recovery? In the introduction it is stated, "After a traumatic nerve injury, mature motor neurons only regenerate at a rate of 1- 4 mm per day, the slowness of which results in only \(\sim 45\%\) of injured axons achieving anatomical reinnervation of their muscle targets with, in consequence, limited functional recovery (1- 7). Discovering therapeutic targets for the effective treatment of axon injury is therefore critical to ensure a more complete recovery of motor and other disturbed functions." However, it appears that in Figure 6 the controls reach the same level of functional recovery as the treated group at D28. This may just be a difference between rodents and humans, either way this should be discussed. Some of the additions suggested here may help address this issue.  
+
+<|ref|>text<|/ref|><|det|>[[115, 815, 875, 884]]<|/det|>
+While it is correct that the vehicle- treated group eventually achieves similar levels of functional recovery at day 28, our findings underscore the significant benefit of an accelerated recovery in the NMII2- treated group. The reason we used the crush injury model is that it enables a more precise assessment of axonal regeneration and functional recovery.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[113, 88, 870, 245]]<|/det|>
+The accelerated recovery observed with NMIIi2 treatment reflects both an enhanced regenerative capacity of the injured axons and a timely reinnervation of neuromuscular junctions, which aligns with the translational goal of developing therapies for nerve injuries. While rodents and humans differ considerably in axonal length and regenerative timelines, the accelerated recovery in our model highlights the therapeutic potential of NMII inhibition in reducing the time to functional recovery, a crucial outcome in clinical scenarios to minimize the secondary complications that often arise due to prolonged functional impairment. We have expanded these points in the revised manuscript to emphasize the translational significance of the findings and the broader implications of accelerated recovery in a clinical setting.  
+
+<|ref|>text<|/ref|><|det|>[[113, 282, 880, 369]]<|/det|>
+Given the work done to show that NMIIl works in various cell types it would make sense to also look at the sciatic nerves shown in Fig. S6 for sensory neuronal regeneration with SCG10. This would take the in vitro work into the in vivo relevance. As there is not any increased mechanistic understanding greater than what was published before, I do believe these additions are important to show the novelty and relevance of this work.  
+
+<|ref|>text<|/ref|><|det|>[[113, 377, 880, 483]]<|/det|>
+As detailed in our response to Reviewer #1, we have now performed additional experiments to examine sensory neuronal regeneration in vivo using SCG10 as a marker and functionally observed an acceleration in sensory response recovery. These results, presented in Figure 6, in the revised MS, demonstrate that NMIIi2 treatment significantly enhances sensory axonal regeneration compared to vehicle- treated groups, further bridging the in vitro findings with in vivo relevance.  
+
+<|ref|>text<|/ref|><|det|>[[113, 520, 870, 589]]<|/det|>
+Lastly, it should be stated why 10mg/kg of NMIIi2 was used systemically and whether there were any obvious things that were associated with weight loss, such as a decrease in organ size or discoloration. Also how long is it thought for the hydrogel release of NMIIi2 at the lesion site, immediate or over days, weeks? This will be of interest to the readers.  
+
+<|ref|>text<|/ref|><|det|>[[113, 598, 875, 650]]<|/det|>
+The systemic dose of \(10mg / kg\) NMIIi2 was selected based on our pharmacokinetic data, which showed that \(10mg / kg\) achieved a higher plasma concentration compared to \(3mg / kg\) , with minimal additional benefit when increased to \(30mg / kg\) .  
+
+<|ref|>image<|/ref|><|det|>[[123, 656, 437, 831]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[113, 845, 860, 881]]<|/det|>
+<center>Figure (not included in MS). Plasma pharmacokinetic parameters of NMIIi2 following IP administration at doses of 3, 10, and \(30mg / kg\) in male mice </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[111, 88, 880, 230]]<|/det|>
+We initially administered \(10 \text{mg / kg}\) of the drug twice daily for three consecutive days following a crush injury. This resulted, however, in significant weight loss in the treated mice (Fig. S9C). When the drug was withheld for three subsequent days, the weight loss improved by day 7 post- injury (Fig. S9C). Conversely, resuming the drug for another three days in the same animals led to further aggravation of weight loss (Fig. S9C), suggesting that systemic administration causes notable side effects. Besides the weight loss, no other obvious side effects were observed. While organ size was not measured, locomotion and general behaviors (grooming, rearing, scratching) remained normal.  
+
+<|ref|>text<|/ref|><|det|>[[112, 237, 872, 377]]<|/det|>
+Regarding the hydrogel- based local delivery, we now confirm that the hydrogel/NMll12 mixture remained in place through visual inspection during surgery. Our immunohistological analyses suggest that the hydrogel effectively released NMll2 for approximately 3- 7 days post- application. Once the regenerating axons grew beyond the hydrogel- embedded lesion site, the localized effects of NMll2 diminished, highlighting the importance of early, targeted local intervention. These findings reinforce the therapeutic potential of localized NMll inhibition while addressing the limitations of systemic delivery. We incorporate this into the discussion in the revised manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[113, 412, 564, 430]]<|/det|>
+There are some minor additions needed in the methods:  
+
+<|ref|>text<|/ref|><|det|>[[112, 439, 802, 475]]<|/det|>
+- The correlation analysis for Figure 6 is not listed nor is it stated if this is a Pearson or Spermann correlation.  
+
+<|ref|>text<|/ref|><|det|>[[113, 484, 541, 503]]<|/det|>
+We have edited the Methods section to deal with this.  
+
+<|ref|>text<|/ref|><|det|>[[112, 540, 874, 575]]<|/det|>
+- The post hoc tests are not listed by name, it would be helpful if the authors listed if they chose a more conservative or liberal post hoc analysis method.  
+
+<|ref|>text<|/ref|><|det|>[[112, 583, 872, 619]]<|/det|>
+We used two- way ANOVA followed by Tukey's post hoc test. We have included this information in the revised MS.  
+
+<|ref|>text<|/ref|><|det|>[[112, 655, 877, 707]]<|/det|>
+- 10mg/kg twice a day for 3 days via i.p. of the NMllii was used for systemic application, but why was this dose/timing/method chosen? This should be listed given the side effects that were observed.  
+
+<|ref|>text<|/ref|><|det|>[[113, 717, 576, 735]]<|/det|>
+We have included this information in the Methods section.  
+
+<|ref|>text<|/ref|><|det|>[[112, 771, 875, 807]]<|/det|>
+Reviewer #2 (Remarks on code availability): I could not access the code when I tried to click on the GitHub link.  
+
+<|ref|>text<|/ref|><|det|>[[113, 816, 415, 834]]<|/det|>
+The code is now available on GitHub.  
+
+<|ref|>text<|/ref|><|det|>[[112, 843, 608, 862]]<|/det|>
+https://github.com/selwynjayakar/Multi- image- neurite- analysis
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[65, 90, 228, 108]]<|/det|>
+## 412 Reviewer #3:  
+
+<|ref|>text<|/ref|><|det|>[[65, 117, 865, 188]]<|/det|>
+413 The manuscript entitled "Non- muscle myosin II inhibition at the site of axon injury increases 414 axon regeneration" is a study in which the authors produce a novel myosin II inhibitor (NMIli2; 415 analog of blebbistatin) and show it's efficacy in induce axon growth in tissue culture and after 416 peripheral nerve injury in adult mice.  
+
+<|ref|>text<|/ref|><|det|>[[65, 199, 88, 213]]<|/det|>
+417  
+
+<|ref|>text<|/ref|><|det|>[[65, 223, 880, 345]]<|/det|>
+418 Following a phenotypic screen in human iPSC derived motor neurons, the authors identified 10 419 compounds that induced neurite outgrowth on a CSPG extracellular matrix environment. 420 Blebbistatin was found to be the strongest inducer of growth, with and without laser cut injury, 421 however given it's issues with photosensitivity and bioavailability, the authors produced a more 422 selective non- muscle myosin II inhibitor, termed NMIli2. As non- muscle myosin II is a motor 423 protein that normally stabilises microtubules, use of NMIli2 would be hypothesised to destabilise 424 MTs thus inducing axon elongation/extension.  
+
+<|ref|>text<|/ref|><|det|>[[65, 380, 880, 555]]<|/det|>
+427 To confirm the findings from their screen as well as to confirm the efficacy of NMIli2, the authors 427 grew human iPSC- derived motor neurons, cortical- like neurons, and human cortical organoids 428 on either laminin or CSPGs; as well as adult mouse DRGs and RGCs on laminin only. The 429 authors also compared their findings with NMIli2 with a lentiviral knockdown of MYH9 (encoding 430 the heavy chain of NMIli) in human iPSC- derived motor neurons, demonstrating axon growth 431 with the KD with confirmed KD (~55%) with a western blot. The authors then tested NMIli2 in 432 vivo in a mouse model of peripheral nerve injury, showing that local/direct administration 433 complexed to a collagen hydrogel, compared with intraperitoneal administration or control, was 434 effective at inducing ChAT- labelled axon growth, synaptic reinnervation and better hindpaw 435 plantar placement.  
+
+<|ref|>text<|/ref|><|det|>[[65, 564, 875, 616]]<|/det|>
+436 While the findings of this study are very interesting for the field of axon regeneration specifically 437 in the PNS, there are some areas of the study that require clarification and/or further 438 information.  
+
+<|ref|>text<|/ref|><|det|>[[65, 626, 249, 643]]<|/det|>
+439 Major questions:  
+
+<|ref|>text<|/ref|><|det|>[[65, 652, 864, 687]]<|/det|>
+- Why were the mouse cellular experiments using adult RGCs and adult DRGs not performed on a CSPG matrix?  
+
+<|ref|>text<|/ref|><|det|>[[65, 697, 872, 784]]<|/det|>
+442 We have conducted additional experiments using CSPG-coated surfaces with cells isolated 443 from 6 additional mice to evaluate RGC growth. While neurite outgrowth was more restricted in 444 this inhibitory environment compared to laminin-coated conditions, we consistently observed 445 that NMIli2 treatment significantly increased total branch length compared to control conditions 446 (see response to Reviewer #1: line number: 77).  
+
+<|ref|>text<|/ref|><|det|>[[65, 793, 860, 845]]<|/det|>
+447 Regarding DRG neurons it is more physiologically relevant to evaluate their regeneration in a 448 peripheral- like environment (laminin- coated surfaces) rather than a CNS- like environment 449 (CSPG- coated surfaces).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[113, 88, 880, 228]]<|/det|>
+CSPG substrates were used in the motor and cortical- like neuron experiments with success. All of these cell types have been shown to respond to CSPGs with retraction and/or overall stunted growth, especially in adult ages. Some studies have shown that the lack of growth on these substrates is due to integrin are downregulated and or subsequent inactivation, leaving adult cells unable to respond to the CSPG matrix. Given the mechanism of action of destabilising MTs, do the authors think that NMIIi2 would induce the growth in the mouse cells on CSPG regardless of integrin regulation? If not, can the authors discuss the mechanism of action in the CNS neurons that differs in the PNS neurons?  
+
+<|ref|>text<|/ref|><|det|>[[113, 237, 880, 377]]<|/det|>
+We speculate that NMIIi2 may promote growth in central neurons on CSPG substrates independently of integrin regulation. However, in the PNS, integrins likely play a more prominent role in regeneration, as described by the Roca- Cusachs molecular clutch model. This model highlights two key contributions of integrins to growth cone dynamics: the regulation of actin retrograde flow and the generation of mechanical forces through integrin- ECM interactions at the growth cone (6, 7). In contrast, CNS neurons are surrounded by a perineuronal net (PNN) composed of CSPGs (8). These differences may explain why integrins have a more substantial impact on promoting regeneration in the PNS compared to the CNS.  
+
+<|ref|>text<|/ref|><|det|>[[113, 385, 808, 420]]<|/det|>
+We have expanded the discussion in the revised manuscript to address these potential differences of NMII inhibition across different neuronal subtypes.  
+
+<|ref|>text<|/ref|><|det|>[[113, 457, 850, 527]]<|/det|>
+- The axon growth in vitro suggests that blebbistatin and NMIIi2 both induce neurite/axon elongation as well as branching. The branching effect is indeed an interesting finding, which would in some cases compete with elongation. Can the authors discuss the mechanisms of action by which the compound can drive both an elongation and a branching phenotype?  
+
+<|ref|>text<|/ref|><|det|>[[112, 536, 855, 555]]<|/det|>
+We propose two possible mechanisms to explain the role of NMIIi in supporting regeneration.  
+
+<|ref|>text<|/ref|><|det|>[[112, 563, 880, 719]]<|/det|>
+First, NMIIi treatment appears to facilitate two different growth modes following nerve injury. Under non- injured conditions, neurons on CSPG- coated surfaces generally cannot extend neurites (Fig. 2E). However, with NMIIi treatment, axonal projection occurs, suggesting that changes in actomyosin force are a key factor driving microtubule destabilization and enabling neuronal elongation. In contrast, on laminin- coated surfaces, where axons are already extending, NMIIi treatment enhances both branch protrusion and elongation of newly formed branches (Fig. 2F), emphasizing the critical role of actin polymerization and integrin- ECM interactions at the growth cone. Under injured conditions, we hypothesize that both mechanisms are simultaneously activated.  
+
+<|ref|>text<|/ref|><|det|>[[112, 728, 875, 902]]<|/det|>
+Second, NMII inhibition may accelerate neuronal growth after injury while potentially modulating neural activity. During nerve regeneration, two distinct growth modes are observed: (1) an elongation mode characterized by typical axonal morphology, and (2) a branching mode distinguished by thin, numerous projections (9). During branch arborization, neuronal activity diminishes as it is distributed across multiple branches (10). Once branches form, their fate is determined during the maturation phase, where they either stabilize and grow or undergo retraction (10). Surviving neurons eventually restore neuronal activity, which is essential for functional recovery (10). Based on these previous findings, we suspect that neuronal activity changes in response to NMIIi treatment may contribute to the recovery, but this will need to be shown directly in future studies.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 117, 883, 170]]<|/det|>
+- In the MYH9 experiments, a western blot was performed to indicate knockdown of MYH9. Have the authors looked at changes in MYH9 in their NMIIi2 treated cells? Would they expect a greater effect, or a less specific effect?  
+
+<|ref|>text<|/ref|><|det|>[[115, 179, 877, 231]]<|/det|>
+To address this, we compared MYH9 levels in cells treated with NMIIi2 for 24 hours with control cells using a Western blot analysis. Unlike MYH9 knockdown, the relative intensity of MYH9 remained unchanged in the NMIIi2-treated cells.  
+
+<|ref|>image<|/ref|><|det|>[[115, 240, 377, 415]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[115, 426, 752, 444]]<|/det|>
+<center>Figure (not included in MS). NMIIi treatment does not decrease MYH9 level </center>  
+
+<|ref|>text<|/ref|><|det|>[[115, 444, 870, 510]]<|/det|>
+Western blot of motor neurons treated with either DMSO or NMIIi2 labeled with anti- MYH9 and anti- GAPDH, and quantification of MYH9 intensity normalized to GAPDH. Experiments conducted in two replicates, statistical analysis unpaired two- tailed Student's \(t\) - test and SEM error bars.  
+
+<|ref|>text<|/ref|><|det|>[[115, 534, 880, 586]]<|/det|>
+This suggests that the effects of NMIIi2 are not mediated through changes in MYH9 expression but rather through an inhibition of NMII activity, which likely reduces actomyosin contractility and facilitates cytoskeletal remodeling, enhancing axonal growth.  
+
+<|ref|>text<|/ref|><|det|>[[115, 595, 880, 700]]<|/det|>
+This outcome aligns with the known mechanism of action of blebbistatin as a myosin II ATPase inhibitor, which targets the enzymatic activity of non- muscle myosin II without altering its expression (11, 12). We decided not to include this data in the revised manuscript, as a more comprehensive analysis is necessary to fully elucidate how NMII activity, rather than expression, contributes to the observed effects. We have incorporated a discussion of this in the revised manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[115, 737, 830, 771]]<|/det|>
+Can the authors perform immunofluorescence with the MYH9 antibody to confirm that the knockdown cells are the cells with increased axon growth?  
+
+<|ref|>text<|/ref|><|det|>[[115, 781, 848, 868]]<|/det|>
+While we did not perform immunofluorescence for MYH9 in this study, we confirmed the efficiency of MYH9 knockdown through a Western blot analysis, which showed a significant reduction in MYH9 protein levels after knockdown compared to controls (Fig. S6A and S6B). The increased axonal growth observed in these cells aligns with the known functional consequences of MYH9 suppression, as previously described.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[113, 88, 870, 211]]<|/det|>
+We recognize that immunofluorescence could validate the link between MYH9 knockdown and enhanced axonal growth. However, our attempts to use commercially available MYH9 antibodies were unfortunately unsuccessful due to their nonspecific binding, which prevented reliable detection. Additionally, the technical challenges of distinguishing axonal growth at single- cell resolution in dense cultures further complicated this approach. We acknowledge the importance of directly visualizing the spatial correlation in MYH9 knockdown axons and have included this in the Discussion of the revised manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[113, 247, 879, 300]]<|/det|>
+- Hydrogels were used in the peripheral nerve injury experiments for drug delivery. What are the biokinetics of these hydrogels? Are they absorbable, and do they have a regenerative effect alone, when used long term for nerve repair?  
+
+<|ref|>text<|/ref|><|det|>[[113, 308, 857, 396]]<|/det|>
+For the peripheral nerve injury experiments we used bovine type I collagen as a drug delivery system for NMIIi2 given its biocompatibility. Based on our observations and the existing literature, we estimate that the hydrogel remains effective for approximately 3- 7 days post- application, coinciding with the critical early phase of nerve regeneration (see response to Reviewer #2: line number: 383).  
+
+<|ref|>text<|/ref|><|det|>[[113, 404, 864, 544]]<|/det|>
+Regarding regenerative effects of the hydrogel itself, previous studies have suggested that collagen- based hydrogels provide structural support and a favorable extracellular matrix environment, potentially aiding in nerve repair when used long- term. To verify this, we performed additional experiments comparing conditions with and without the hydrogel in the absence of NMIIi2. However, we observed minimal differences between these conditions (Fig. S8D), indicating that the hydrogel alone does not have a significant regenerative effect. These findings highlight the utility of the hydrogel as a delivery vehicle for NMIIi2 rather than as an independent regenerative agent and have included this in the revised manuscript.  
+
+<|ref|>image<|/ref|><|det|>[[120, 551, 610, 730]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[113, 742, 870, 810]]<|/det|>
+<center>Figure S8. Evaluation of the hydrogel's intrinsic regenerative effects after nerve injury (D) Quantification of luminance ratio of left/right paws post sciatic nerve crush with and without hydrogel. Data collected from 8 mice. Stats: two-way ANOVA with Tukey's post hoc test, SEM error bars </center>  
+
+<|ref|>text<|/ref|><|det|>[[113, 847, 869, 882]]<|/det|>
+There was recovery of function with vehicle at 28 days equal to NMIIi2 treated animals, did the authors expect this?
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[57, 90, 605, 108]]<|/det|>
+557 Please see our responses to Reviewer #2 (line number: 338).  
+
+<|ref|>text<|/ref|><|det|>[[57, 122, 88, 135]]<|/det|>
+558  
+
+<|ref|>text<|/ref|><|det|>[[115, 145, 877, 230]]<|/det|>
+- ChAT immunohistochemistry was used to examine motor axon regeneration in the sciatic nerve. Is there any possibility of tracing axons, rather than relying on antibody staining for regenerative growth? How was the injury confirmed, as the methods only state that the sural nerve was crushed for 20 sec with a hemostat? The sural nerve is primarily a sensory nerve, so I query why this was used to assess motor axon regeneration.  
+
+<|ref|>text<|/ref|><|det|>[[115, 240, 877, 345]]<|/det|>
+We selected ChAT staining to specifically identify regenerating motor axons, as this provides robust and reliable labeling of cholinergic neurons. Dye- based tracing approaches, such as Dil, would face challenges in differentiating motor neurons from sensory cell types. This is particularly relevant given our new data that NMII2 also enhances sensory neuron regeneration (Fig. 6). Therefore, we do not anticipate that dye tracing would provide information beyond what can be achieved using immunostaining with specific motor and sensory markers.  
+
+<|ref|>text<|/ref|><|det|>[[115, 354, 872, 424]]<|/det|>
+Regarding injury confirmation, the injury was visually confirmed during the surgery and further validated through immunohistochemical analysis (example image of Fig. 6A) and the functional recovery assays (Fig. 5F and Fig. 6D). The precise injury site was clearly identifiable as an enlarged swollen area caused by the hemostat crush (example image in Fig. 6A)  
+
+<|ref|>text<|/ref|><|det|>[[115, 432, 856, 501]]<|/det|>
+We apologize for the confusion in the Methods section regarding the location of the injury. To clarify, the injury was targeted at the sciatic nerve trunk, which includes the sural, common peroneal, and tibial nerves (refer to schematic in Fig. 5A). We have clarified this point in the revised text.  
+
+<|ref|>text<|/ref|><|det|>[[115, 540, 878, 625]]<|/det|>
+- Likewise, to assess motor recovery, hindpaw placement was assessed. Plantar placement is mainly used to assess sensory recovery, or possibly sensorimotor feedback following injury and repair. While the results found are promising, how do the authors discount the potential for sensory axon changes inducing this effect? Have the authors looked at sensory axon repair in this context?  
+
+<|ref|>text<|/ref|><|det|>[[115, 635, 875, 687]]<|/det|>
+We agree that sensory neuron involvement could contribute to the observed effects. To address this, we have now systematically studied the effect of NMII2 on sensory neurons (Fig. 6); further details are provided in our response to Reviewer #1 (line number: 119).  
+
+<|ref|>text<|/ref|><|det|>[[115, 697, 866, 784]]<|/det|>
+Additionally, our machine- learning approach assesses hindpaw placement by quantifying luminance intensity, which reflects weight- bearing and gait changes, which primarily captures loss and recovery of motor function, as supported by the positive correlation between paw luminance ratio and lumbrical muscle reinnervation (Fig. 5G) but not the sensory reinnervation (Fig. 6E).
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[130, 99, 395, 280]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[113, 297, 878, 333]]<|/det|>
+<center>Figure 6E. Pearson correlation analysis of the paw luminance ratio and pinprick behavior scores on day 14 post-injury. Data collected from 10 male mice per group. </center>  
+
+<|ref|>text<|/ref|><|det|>[[115, 345, 866, 381]]<|/det|>
+Very little detail is included in the methods section for this test. Can further details be included (length of walking tract, timing of test, etc)  
+
+<|ref|>text<|/ref|><|det|>[[115, 389, 875, 460]]<|/det|>
+We now include additional information, such as the length of the walking track (10 minutes) and the specific equipment and methods used to calculate the luminance ratio. The timing of the tests is detailed on the x- axis of each time- course plot. All recordings were conducted between 1:00 PM and 3:00 PM for all animals.  
+
+<|ref|>text<|/ref|><|det|>[[115, 496, 785, 515]]<|/det|>
+- did the authors test for changes in Rho/ROCK activity in their in vitro experiments?  
+
+<|ref|>text<|/ref|><|det|>[[113, 523, 881, 731]]<|/det|>
+To address this, we analyzed growth cone morphology under ROCK and NMII inhibited conditions, as this is indicative of alterations in Rho/ROCK and/or myosin II activity (13- 15). Cytoskeletal components were labeled with \(\alpha\) - tubulin and F- actin to visualize growth cone structure. In the control condition, growth cones displayed the presence of both lamellipodia and filopodia. However, treatment with NMIIi2 resulted in a thin, elongated growth cone morphology without a lamellopodium. We also observed similar growth cone phenotypes in response to treatment with ROCK inhibitors, Y- 27632 or fasudil, as those observed with NMII inhibitors. These results suggest that myosin II activity in growth cones may be regulated through a ROCK- dependent pathway. This observation aligns with previous studies indicating that the regulation of myosin II activity is ROCK dependent. However, we decided not to include this data in the revised MS, since this paper primarily focused on evaluating the functional outcomes of the new NMIIi analogue as we have mentioned above (line number 32- 35).  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 768, 857, 799]]<|/det|>
+## Figure (not included in MS). Assessment of growth cone morphology under ROCK and NMII inhibited conditions  
+
+<|ref|>text<|/ref|><|det|>[[115, 800, 874, 864]]<|/det|>
+(left) Representative images of neurons labeled with anti- TUBB3, F- actin, and DAPI at 24h NMIIi2, Y- 27632, Fasudil post- treatment, and controls. (right) Quantification of F- actin intensity of growth cone. Data normalized to DMSO, and experiments done in three biological replicates. Statistics: one- way ANOVA with Sidak's multiple comparison test, SEM error bars.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[115, 100, 876, 260]]<|/det|>  
+
+<|ref|>text<|/ref|><|det|>[[57, 272, 90, 287]]<|/det|>
+623  
+
+<|ref|>text<|/ref|><|det|>[[57, 297, 223, 313]]<|/det|>
+624 Minor issues:  
+
+<|ref|>text<|/ref|><|det|>[[115, 322, 867, 357]]<|/det|>
+- Page 9, paragraph 2: why is the term 'frustrated' used for total internal reflection? 'Uninjured' misspelled as 'uninured'.  
+
+<|ref|>text<|/ref|><|det|>[[115, 365, 870, 416]]<|/det|>
+We use the term 'frustrated total internal reflection' rather than 'total internal reflection' because it describes a modification or disruption of total internal reflection caused by surface contact. The technique is detailed in our original paper (16).  
+
+<|ref|>text<|/ref|><|det|>[[115, 425, 867, 461]]<|/det|>
+We apologize for the typographical error where 'uninjured' was misspelled as 'uninured.' This has been corrected in the revised manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[115, 499, 625, 516]]<|/det|>
+- Page 17, Methods, Western Blot: 'Proteins' should be 'Protein'  
+
+<|ref|>text<|/ref|><|det|>[[115, 525, 712, 544]]<|/det|>
+We have corrected the text. It is now on page 19 of the revised manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[115, 580, 857, 633]]<|/det|>
+- Page 17, Methods, Mouse surgical procedures and behavioural test: There is an unfinished sentence 'Sciatic nerve crush surgery and subsequent local application were performed aseptically under 2% isoflurane anesthesia followed by'?  
+
+<|ref|>text<|/ref|><|det|>[[115, 642, 675, 660]]<|/det|>
+We have corrected this. It's now on page 20 of the revised manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[115, 697, 847, 749]]<|/det|>
+- Page 18, Methods, Mouse surgical procedures and behavioural test: 'To assess functional recovery, lumbrical muscles were dissected form the paws' This need more detail for a functional recovery assessment.  
+
+<|ref|>text<|/ref|><|det|>[[115, 758, 666, 776]]<|/det|>
+- Page 19, Methods, Statistics: Please add what post-hoc tests used.  
+
+<|ref|>text<|/ref|><|det|>[[115, 785, 816, 803]]<|/det|>
+We have added more information to the Method section (page 20- 22 of the revised MS).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[58, 90, 880, 770]]<|/det|>
+REFERENCESS1. Qian T, Wang P, Chen Q, Yi S, Liu Q, Wang H, et al. The dynamic changes of main cell types in the microenvironment of sciatic nerves following sciatic nerve injury and the influence of let- 7 on their distribution. RSC Adv. 2018;8(72):41181- 91.2. Yang Y, Wen D, Lin F, Song X, Pang R, Sun W, et al. Suppression of non- muscle myosin II boosts T cell cytotoxicity against tumors. Sci Adv. 2024;10(44):eadp0631.3. Ma CH, Omura T, Cobos EJ, Latremoliere A, Ghasemlou N, Brenner GJ, et al. Accelerating axonal growth promotes motor recovery after peripheral nerve injury in mice. J Clin Invest. 2011;121(11):4332- 47.4. Dravid A, Raos B, Svirskis D, O'Carroll SJ. Optimised techniques for high- throughput screening of differentiated SH- SY5Y cells and application for neurite outgrowth assays. Sci Rep. 2021;11(1):23935.5. Elliott H, Fischer RS, Myers KA, Desai RA, Gao L, Chen CS, et al. Myosin II controls cellular branching morphogenesis and migration in three dimensions by minimizing cell- surface curvature. Nat Cell Biol. 2015;17(2):137- 47.6. Swaminathan V, Waterman CM. The molecular clutch model for mechanotransduction evolves. Nat Cell Biol. 2016;18(5):459- 61.7. Vicente- Manzanares M, Ma X, Adelstein RS, Horwitz AR. Non- muscle myosin II takes centre stage in cell adhesion and migration. Nat Rev Mol Cell Biol. 2009;10(11):778- 90.8. Soleman S, Filippov MA, Dityatev A, Fawcett JW. Targeting the neural extracellular matrix in neurological disorders. Neuroscience. 2013;253:194- 213.9. Kerschensteiner M, Schwab ME, Lichtman JW, Misgeld T. In vivo imaging of axonal degeneration and regeneration in the injured spinal cord. Nat Med. 2005;11(5):572- 7.10. Kalil K, Dent EW. Branch management: mechanisms of axon branching in the developing vertebrate CNS. Nat Rev Neurosci. 2014;15(1):7- 18.11. Kovacs M, Toth J, Hetenyi C, Malnasi- Csizmadia A, Sellers JR. Mechanism of blebbistatin inhibition of myosin II. J Biol Chem. 2004;279(34):35557- 63.12. Straight AF, Cheung A, Limouze J, Chen I, Westwood NJ, Sellers JR, et al. Dissecting temporal and spatial control of cytokinesis with a myosin II Inhibitor. Science. 2003;299(5613):1743- 7.13. Dent EW, Gupton SL, Gertler FB. The growth cone cytoskeleton in axon outgrowth and guidance. Cold Spring Harb Perspect Biol. 2011;3(3).14. Turney SG, Bridgman PC. Laminin stimulates and guides axonal outgrowth via growth cone myosin II activity. Nat Neurosci. 2005;8(6):717- 9.15. Woo S, Gomez TM. Rac1 and RhoA promote neurite outgrowth through formation and stabilization of growth cone point contacts. J Neurosci. 2006;26(5):1418- 28.16. Zhang Z, Roberson DP, Kotoda M, Boivin B, Bohnslav JP, Gonzalez- Cano R, et al. Automated preclinical detection of mechanical pain hypersensitivity and analgesia. Pain. 2022;163(12):2326- 36.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[69, 88, 459, 107]]<|/det|>
+## RESPONSE TO REVIEWER COMMENTS  
+
+<|ref|>text<|/ref|><|det|>[[115, 116, 803, 169]]<|/det|>
+We thank the reviewers and have addressed the key points raised by Reviewer 2 by incorporating them into the Results and Discussion sections, and updating a figure, as summarized here. The revised text in the MS is highlighted in yellow.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 206, 466, 223]]<|/det|>
+## Revised figure panel including new data:  
+
+<|ref|>text<|/ref|><|det|>[[115, 232, 850, 251]]<|/det|>
+Fig. S7A now includes an additional panel displaying NF- positive axons on day 7 post- injury  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 288, 228, 305]]<|/det|>
+## Reviewer #2:  
+
+<|ref|>text<|/ref|><|det|>[[115, 322, 881, 477]]<|/det|>
+The authors have thoroughly addressed my questions and those of the other reviewers. My only suggestion is that the authors be careful in their mention of sensory neuroregeneration, which does not lead to functional recovery. The majority of sensory neurons are indeed made up of nociceptors, which were examined here, but when it comes to motor recovery, it is more the mechanoreceptors (TrkB) and proprioceptors (TrKC) that would be of interest that was indicated by reviewer #3. Therefore, please add this caveat. NF, while bringing up myelinated neurons, would not be specific to myelinated sensory neurons alone, which is what the comment on 281 of page 11 makes confusing. These slight clarification changes are all that are required for acceptance.  
+
+<|ref|>text<|/ref|><|det|>[[112, 486, 881, 682]]<|/det|>
+We thank the reviewer for raising this important point. Mechanoreceptor and proprioceptor sensory neurons play a crucial role in sensorimotor feedback and motor function recovery after nerve injury (1- 3). Our paw luminance data likely reflects motor behavior primarily driven by motor axon regeneration but may also capture contributions from the regeneration of those proprioceptor and mechanoreceptor sensory neurons. An additional NF panel in Fig. S7A now shows that NF- positive but ChAT- negative axons are present in the regenerating nerve after local NMII treatment, suggesting that proprioceptor or mechanoreceptor sensory neurons also regrow faster. These findings imply that the motor function recovery observed after NMII could be the result of an accelerated regeneration of both motor and sensory neurons. Identification of the specific actions of NMII on those different sensory subtypes which contribute to motor function recovery in future studies, would provide valuable insights into the utility of this approach for promoting a full recovery of motor function.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[115, 100, 880, 435]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[115, 440, 845, 519]]<|/det|>
+<center>Fig. S7A. Regenerating axons in distal nerve 7 days post-sciatic nerve crush injury (A) Representative images of mouse sciatic nerves labeled with an antibody against Neurofilament (NF) and Choline Acetyltransferase (ChAT) on day 7 post-sciatic nerve crush injury. White dashed lines indicate the boundary of the crush injury site, and arrowhead indicates tips of regenerated axons. </center>  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 567, 241, 584]]<|/det|>
+## REFERENCES  
+
+<|ref|>text<|/ref|><|det|>[[112, 599, 875, 718]]<|/det|>
+1. de Nooij JC, Zampieri N. The making of a proprioceptor: a tale of two identities. Trends Neurosci. 2023;46(12):1083-94.  
+2. Oliver KM, Florez-Paz DM, Badea TC, Mentis GZ, Menon V, de Nooij JC. Molecular correlates of muscle spindle and Golgi tendon organ afferents. Nat Commun. 2021;12(1):1451.  
+3. Wu H, Petitpre C, Fontanet P, Sharma A, Bellardita C, Quadros RM, et al. Distinct subtypes of proprioceptive dorsal root ganglion neurons regulate adaptive proprioception in mice. Nat Commun. 2021;12(1):1026.
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__6d3e1cacdbe49ba268511c76e4b8721b6de96bf0af7ea44a323fd2ae3c615143/images_list.json

@@ -0,0 +1,17 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Fig. 3. ...e, Second law efficiency of the \\(\\mathrm{Ti}_{78}\\mathrm{Nb}_{22}\\) and phase-transitional materials calculated under the same working condition of \\(T_{\\mathrm{c}} = 288 \\mathrm{~K}\\) and \\(T_{\\mathrm{h}} = 298 \\mathrm{~K}\\) ; the calculated values for phase-transition materials are taken from Ref. \\(^{42}\\) while that for the \\(\\mathrm{Ti}_{78}\\mathrm{Nb}_{22}\\) is detailed in Note 2 of Supplementary Information.",
+    "footnote": [],
+    "bbox": [
+      [
+        177,
+        231,
+        850,
+        624
+      ]
+    ],
+    "page_idx": 4
+  }
+]

peer_reviews/supplementary_0_Transparent Peer Review file__6d3e1cacdbe49ba268511c76e4b8721b6de96bf0af7ea44a323fd2ae3c615143/supplementary_0_Transparent Peer Review file__6d3e1cacdbe49ba268511c76e4b8721b6de96bf0af7ea44a323fd2ae3c615143.mmd

@@ -0,0 +1,112 @@
+
+# nature portfolio  
+
+# Peer Review File  
+
+# Large thermoelastic effect in martensitic phase of ferroelastic alloys for high efficiency heat pumping  
+
+Corresponding Author: Dr Qiao Li  
+
+This manuscript has been previously reviewed at another journal. This document only contains information relating to versions considered at Nature Communications.  
+
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+Version 1:  
+
+Reviewer comments:  
+
+Reviewer #2  
+
+(Remarks to the Author)  
+
+I find that there are still issues with the presentation of the concept of the efficiency throughout the manuscript. These need to be changed, otherwise, it does not help and only adds to the confusion. At very worst, it does not do justice to this interesting phenomenon, which is different from caloric effects, thus diminishing the true significance of the work.  
+
+The authors claim that they have struck out the COPmat without mentioning the second law efficiency in the abstract, yet it is still clearly there in the new version of the abstract without 2nd law value.  
+
+In some of the places, there is a marginal improvement in this presentation issue, in that there is mentioning of the 2nd law efficiency together with COPmat in some places. But there are still erroneous/incorrect/misleading sentences in the manuscript.  
+
+On p.2, "Nevertheless, the energy efficiency, or system coefficient of performance (COPsys), of the heat pumps is capped with a relatively low coefficient of performance (COPmat) of the materials"  
+
+This is not true and misleading. COPmat (for a fixed Th and Tc) is sufficiently high, it is the engineering ability to convert to the system COP which is lacking in current technologies.  
+
+"Due to phase- transition hysteresis, most ferroic materials — except for certain magnetocaloric materials undergoing second- order phase transition — possess a COPmat of about 4 – 23. These values are equivalent to a second law efficiency...."  
+
+This is misleading since the most common magnetocaloric material is Gd, the 2nd order material, which has a high 2nd law efficiency.  
+
+"These values are equivalent to..." is also very misleading. It is the 2nd law efficiency which is the intrinsic property, and only with defining of Tc and Th, COP value make sense. COP is not an intrinsic comparison- value without fixing both Tc and Th. So the sentence structure should be the other way around: first introduce 2nd law efficiency, and then "this is equivalent to COPmat of XX with standard Th and Tc".  
+
+In fact, I suggest that the authors change all mention of COPmat to 2nd law efficiency. And in some places, with fixing Th and Tc to some standard values, COP can be mentioned after first mentioning the 2nd law efficiency.  
+
+Along the same line, Fig. 4a must be a comparison of 2nd law efficiency, as done in Fig, 4b. (Table S3 does not contain Th and Tc; it only lists delta T)  
+
+Other aspects (typos, etc.) and supplying more details, as suggested by the other reviewers, seem to be largely resolved.  
+
+Version 2:  
+
+Reviewer comments:  
+
+Reviewer #2  
+
+(Remarks to the Author) The authors seemed to have resolved the issue of the 2nd law efficiency vs COP (with defined Th and Tc) in the manuscript.
+
+<--- Page Split --->
+
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source.  
+
+The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.  
+
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+
+## Reply to comments from Reviewer #2  
+
+1. I find that there are still issues with the presentation of the concept of the efficiency throughout the manuscript. These need to be changed, otherwise, it does not help and only adds to the confusion. At very worst, it does not do justice to this interesting phenomenon, which is different from caloric effects, thus diminishing the true significance of the work.  
+
+Reply: Thank you for your valuable feedback on our presentation issue related to energy efficiency. Following your suggestion, we have revised our draft to eliminate potential confusion. Kindly review our detailed revisions explained below.  
+
+2. The authors claim that they have struck out the COPmat without mentioning the second law efficiency in the abstract, yet it is still clearly there in the new version of the abstract without 2nd law value.  
+
+Reply: Thank you for pointing out this issue. Following your advice in Comment 5, all mention of COPmat in the latest abstract has been changed to the 2nd law value (i.e. the ratio of COPmat to the Carnot theoretical limit), as shown below:  
+
+## Revised abstract:  
+
+"Solid- state heat- pumping using latent heat from first- order ferroic phase transitions is a promising green alternative to traditional vapor- compression technology. However, the intrinsic phase- transition hysteresis poses a limitation on heat- pumping energy efficiency. Here, we propose heat- pumping using heat from anhysteretic elastic deformation in martensitic phase of ferroelastic alloys. Conventionally, this thermoelastic effect (TeE) is considered too weak to be practical. But we find that in [100]- textured \(\mathrm{Ti_{78}Nb_{22}}\) martensitic polycrystals, the TeE can produce a large adiabatic temperature change \((\Delta T_{ad})\) of \(4 - 5\mathrm{K}\) at \(413 - 473\mathrm{K}\) due to macroscopic large linear thermal expansion \((\alpha_{l} = 10^{- 4} / \mathrm{K})\) . This large TeE not only far exceeds those of ordinary metals \((\Delta T_{ad}\approx 0.2K)\) but also brings a material- level high energy efficiency that reaches about \(90\%\) of the Carnot theoretical limit. In other ferroelastic martensitic alloys with larger intrinsic \(\alpha_{l}\) (up to \(5.4\times 10^{- 4} / \mathrm{K})\) , the TeE is predicted to bring an even larger \(\Delta T_{ad}\) (up to \(22\mathrm{K}\) ) while maintaining relatively high efficiency. Our findings offer a non- phase- transition- based way for high- efficiency solid- state heat- pumping."  
+
+3. In some of the places, there is a marginal improvement in this presentation issue, in that there is mentioning of the 2nd law efficiency together with COPmat in some places. But there are still erroneous/incorrect/misleading sentences in the manuscript. On p.2, "Nevertheless, the energy efficiency, or system coefficient of performance (COPsys), of the heat pumps is capped with a relatively low coefficient of performance (COPmat) of the materials". This is not true and misleading. COPmat (for a fixed Th and Tc) is sufficiently high, it is the engineering ability to convert to the system COP which is lacking in current technologies.  
+
+Reply: Thank you for your comment. The above- mentioned sentence has now been removed. Please review the updated text provided in reply to your Comment 4.  
+
+4. "Due to phase-transition hysteresis, most ferroic materials — except for certain magnetocaloric materials undergoing second-order phase transition — possess a COPmat of about \(4 - 23\) . These values are equivalent to a second law efficiency...."This is misleading since the most common magnetocaloric material
+
+<--- Page Split --->
+
+is Gd, the 2nd order material, which has a high 2nd law efficiency.  
+
+Reply: Thank you for your feedback. We have removed magnetocaloric materials from the Introduction section, which now only mentions heat- pumping using first- order ferroelastic phase transition of shape memory alloys. You will see that this revision eliminates the above- mentioned misleading point without diminishing the significance of our study (i.e. introduction of a non- phase- transitional heat- pumping approach). Please review the updated text below:  
+
+The second paragraph of the Introduction section:  
+
+"Among green alternatives, solid- state heat- pumping using first- order ferroelastic phase transition of shape memory alloys (SMAs) has attracted much attention \(^{8 - 15}\) . Various high- performance heat- pumping devices have been developed using SMAs capable of exhibiting adiabatic temperature changes ( \(\Delta T_{ad}\) ) up to \(30 \mathrm{K}^{6,9 - 11}\) . Nevertheless, restricted by phase- transition hysteresis, the energy efficiency of SMAs reaches only about \(50\% - 70\%\) of the Carnot theoretical limit, significantly lower than those of VC- based refrigerants \((- 90\%)\) . Higher material- level efficiency is critical since it stands for the upper limit of the device- level energy efficiency. However, this goal is difficult to achieve via the typical first- order phase transition of ferroelastic alloys \(^{15}\) ; exploring anhysteretic non- phase- transition pathways may open a new way out."  
+
+5. "These values are equivalent to..." is also very misleading. It is the 2nd law efficiency which is the intrinsic property, and only with defining of Tc and Th, COP value make sense. COP is not an intrinsic comparison-value without fixing both Tc and Th. So the sentence structure should be the other way around: first introduce 2nd law efficiency, and then "this is equivalent to COPmat of XX with standard Th and Tc". In fact, I suggest that the authors change all mention of COPmat to 2nd law efficiency. And in some places, with fixing Th and Tc to some standard values, COP can be mentioned after first mentioning the 2nd law efficiency.  
+
+Reply: Thank you for your constructive suggestion. We have changed all mention of COPmat to the second law efficiency in the abstract (see reply to Comment 2) and introduction (see reply to Comment 4) of the draft. Detailed COPmat under standard Th and Tc is only provided in the main text, as shown below:  
+
+Lines 21- 32 at Page 5:  
+
+"We calculated the material- level coefficient of performance (COPmat) of the [100]- textured \(\mathrm{Ti}_{78}\mathrm{Nb}_{22}\) in a Stirling heat- pumping cycle by using a thermodynamic model developed by Qian et al. \(^{42,43}\) (see Note 2 of Supplementary information) along with the measured properties of the \(\mathrm{Ti}_{78}\mathrm{Nb}_{22}\) (see Table S3). Under \(T_{\mathrm{c}} = 288 \mathrm{K}\) (heat source temperature) and \(T_{\mathrm{h}} = 298 \mathrm{K}\) (heat sink temperature), the \(\mathrm{Ti}_{78}\mathrm{Nb}_{22}\) exhibits a COPmat of 25.3, equivalent to \(88\%\) of the Carnot COP (COPcarnot = \(T_{\mathrm{c}} / (T_{\mathrm{h}} - T_{\mathrm{c}}) = 28.8\) ). As shown in Fig. 3(e), this ratio, known as second- law efficiency \(^{42,43}\) , well surpasses those from first- order phase transitions in electrocaloric materials (31% - 41%, calculated values from Ref. 42) and elastocaloric SMAs (55% - 71% \(^{42}\) ) and rivals those from second- order phase transition in magnetocaloric materials (79% - 91% \(^{42}\) ) and liquid- vapor phase transition in commercial VC refrigerants (86% - 91% \(^{42}\) ). Near its optimal working temperature (i.e. under \(T_{\mathrm{c}} = 468 \mathrm{K}\) and \(T_{\mathrm{h}} = 478 \mathrm{K}\) ), the \(\mathrm{Ti}_{78}\mathrm{Nb}_{22}\) has an even higher second- law efficiency of 94%.
+
+<--- Page Split --->
+
+6. Along the same line, Fig. 4a must be a comparison of 2nd law efficiency, as done in Fig. 4b. (Table S3 does not contain Th and Tc; it only lists delta T). Other aspects (typos, etc.) and supplying more details, as suggested by the other reviewers, seem to be largely resolved.  
+
+Reply: Thank you for your comment. We have deleted Fig. 4a and its related figures (i.e. original Fig. 3(e) and Fig. S12) and table (i.e. Table S3). For compactness, the original Fig. 4(b) (i.e. comparison of second law efficiency) has been integrated into Fig. 3 as subfigure (e), as shown below:  
+
+![](images/Figure_3.jpg)
+
+<center>Fig. 3. ...e, Second law efficiency of the \(\mathrm{Ti}_{78}\mathrm{Nb}_{22}\) and phase-transitional materials calculated under the same working condition of \(T_{\mathrm{c}} = 288 \mathrm{~K}\) and \(T_{\mathrm{h}} = 298 \mathrm{~K}\) ; the calculated values for phase-transition materials are taken from Ref. \(^{42}\) while that for the \(\mathrm{Ti}_{78}\mathrm{Nb}_{22}\) is detailed in Note 2 of Supplementary Information. </center>
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__6d3e1cacdbe49ba268511c76e4b8721b6de96bf0af7ea44a323fd2ae3c615143/supplementary_0_Transparent Peer Review file__6d3e1cacdbe49ba268511c76e4b8721b6de96bf0af7ea44a323fd2ae3c615143_det.mmd

@@ -0,0 +1,156 @@
+<|ref|>title<|/ref|><|det|>[[72, 50, 296, 79]]<|/det|>
+# nature portfolio  
+
+<|ref|>title<|/ref|><|det|>[[75, 96, 296, 119]]<|/det|>
+# Peer Review File  
+
+<|ref|>title<|/ref|><|det|>[[74, 161, 777, 211]]<|/det|>
+# Large thermoelastic effect in martensitic phase of ferroelastic alloys for high efficiency heat pumping  
+
+<|ref|>text<|/ref|><|det|>[[75, 224, 364, 241]]<|/det|>
+Corresponding Author: Dr Qiao Li  
+
+<|ref|>text<|/ref|><|det|>[[72, 275, 875, 303]]<|/det|>
+This manuscript has been previously reviewed at another journal. This document only contains information relating to versions considered at Nature Communications.  
+
+<|ref|>text<|/ref|><|det|>[[70, 314, 866, 330]]<|/det|>
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+<|ref|>text<|/ref|><|det|>[[73, 366, 145, 380]]<|/det|>
+Version 1:  
+
+<|ref|>text<|/ref|><|det|>[[73, 392, 219, 406]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[73, 418, 162, 432]]<|/det|>
+Reviewer #2  
+
+<|ref|>text<|/ref|><|det|>[[73, 445, 238, 458]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 458, 921, 497]]<|/det|>
+I find that there are still issues with the presentation of the concept of the efficiency throughout the manuscript. These need to be changed, otherwise, it does not help and only adds to the confusion. At very worst, it does not do justice to this interesting phenomenon, which is different from caloric effects, thus diminishing the true significance of the work.  
+
+<|ref|>text<|/ref|><|det|>[[72, 496, 920, 523]]<|/det|>
+The authors claim that they have struck out the COPmat without mentioning the second law efficiency in the abstract, yet it is still clearly there in the new version of the abstract without 2nd law value.  
+
+<|ref|>text<|/ref|><|det|>[[72, 523, 911, 562]]<|/det|>
+In some of the places, there is a marginal improvement in this presentation issue, in that there is mentioning of the 2nd law efficiency together with COPmat in some places. But there are still erroneous/incorrect/misleading sentences in the manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[72, 561, 904, 589]]<|/det|>
+On p.2, "Nevertheless, the energy efficiency, or system coefficient of performance (COPsys), of the heat pumps is capped with a relatively low coefficient of performance (COPmat) of the materials"  
+
+<|ref|>text<|/ref|><|det|>[[72, 588, 907, 614]]<|/det|>
+This is not true and misleading. COPmat (for a fixed Th and Tc) is sufficiently high, it is the engineering ability to convert to the system COP which is lacking in current technologies.  
+
+<|ref|>text<|/ref|><|det|>[[72, 614, 870, 653]]<|/det|>
+"Due to phase- transition hysteresis, most ferroic materials — except for certain magnetocaloric materials undergoing second- order phase transition — possess a COPmat of about 4 – 23. These values are equivalent to a second law efficiency...."  
+
+<|ref|>text<|/ref|><|det|>[[72, 652, 912, 679]]<|/det|>
+This is misleading since the most common magnetocaloric material is Gd, the 2nd order material, which has a high 2nd law efficiency.  
+
+<|ref|>text<|/ref|><|det|>[[72, 679, 921, 731]]<|/det|>
+"These values are equivalent to..." is also very misleading. It is the 2nd law efficiency which is the intrinsic property, and only with defining of Tc and Th, COP value make sense. COP is not an intrinsic comparison- value without fixing both Tc and Th. So the sentence structure should be the other way around: first introduce 2nd law efficiency, and then "this is equivalent to COPmat of XX with standard Th and Tc".  
+
+<|ref|>text<|/ref|><|det|>[[72, 731, 910, 758]]<|/det|>
+In fact, I suggest that the authors change all mention of COPmat to 2nd law efficiency. And in some places, with fixing Th and Tc to some standard values, COP can be mentioned after first mentioning the 2nd law efficiency.  
+
+<|ref|>text<|/ref|><|det|>[[72, 758, 911, 783]]<|/det|>
+Along the same line, Fig. 4a must be a comparison of 2nd law efficiency, as done in Fig, 4b. (Table S3 does not contain Th and Tc; it only lists delta T)  
+
+<|ref|>text<|/ref|><|det|>[[70, 783, 900, 797]]<|/det|>
+Other aspects (typos, etc.) and supplying more details, as suggested by the other reviewers, seem to be largely resolved.  
+
+<|ref|>text<|/ref|><|det|>[[73, 821, 144, 835]]<|/det|>
+Version 2:  
+
+<|ref|>text<|/ref|><|det|>[[73, 847, 219, 860]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[73, 873, 161, 886]]<|/det|>
+Reviewer #2  
+
+<|ref|>text<|/ref|><|det|>[[73, 899, 922, 927]]<|/det|>
+(Remarks to the Author) The authors seemed to have resolved the issue of the 2nd law efficiency vs COP (with defined Th and Tc) in the manuscript.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 45, 916, 99]]<|/det|>
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+<|ref|>text<|/ref|><|det|>[[72, 100, 797, 113]]<|/det|>
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source.  
+
+<|ref|>text<|/ref|><|det|>[[72, 113, 911, 166]]<|/det|>
+The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.  
+
+<|ref|>text<|/ref|><|det|>[[72, 166, 618, 180]]<|/det|>
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[149, 84, 472, 101]]<|/det|>
+## Reply to comments from Reviewer #2  
+
+<|ref|>text<|/ref|><|det|>[[150, 112, 850, 195]]<|/det|>
+1. I find that there are still issues with the presentation of the concept of the efficiency throughout the manuscript. These need to be changed, otherwise, it does not help and only adds to the confusion. At very worst, it does not do justice to this interesting phenomenon, which is different from caloric effects, thus diminishing the true significance of the work.  
+
+<|ref|>text<|/ref|><|det|>[[178, 204, 849, 254]]<|/det|>
+Reply: Thank you for your valuable feedback on our presentation issue related to energy efficiency. Following your suggestion, we have revised our draft to eliminate potential confusion. Kindly review our detailed revisions explained below.  
+
+<|ref|>text<|/ref|><|det|>[[149, 262, 849, 312]]<|/det|>
+2. The authors claim that they have struck out the COPmat without mentioning the second law efficiency in the abstract, yet it is still clearly there in the new version of the abstract without 2nd law value.  
+
+<|ref|>text<|/ref|><|det|>[[179, 320, 849, 371]]<|/det|>
+Reply: Thank you for pointing out this issue. Following your advice in Comment 5, all mention of COPmat in the latest abstract has been changed to the 2nd law value (i.e. the ratio of COPmat to the Carnot theoretical limit), as shown below:  
+
+<|ref|>sub_title<|/ref|><|det|>[[179, 380, 320, 396]]<|/det|>
+## Revised abstract:  
+
+<|ref|>text<|/ref|><|det|>[[177, 401, 850, 650]]<|/det|>
+"Solid- state heat- pumping using latent heat from first- order ferroic phase transitions is a promising green alternative to traditional vapor- compression technology. However, the intrinsic phase- transition hysteresis poses a limitation on heat- pumping energy efficiency. Here, we propose heat- pumping using heat from anhysteretic elastic deformation in martensitic phase of ferroelastic alloys. Conventionally, this thermoelastic effect (TeE) is considered too weak to be practical. But we find that in [100]- textured \(\mathrm{Ti_{78}Nb_{22}}\) martensitic polycrystals, the TeE can produce a large adiabatic temperature change \((\Delta T_{ad})\) of \(4 - 5\mathrm{K}\) at \(413 - 473\mathrm{K}\) due to macroscopic large linear thermal expansion \((\alpha_{l} = 10^{- 4} / \mathrm{K})\) . This large TeE not only far exceeds those of ordinary metals \((\Delta T_{ad}\approx 0.2K)\) but also brings a material- level high energy efficiency that reaches about \(90\%\) of the Carnot theoretical limit. In other ferroelastic martensitic alloys with larger intrinsic \(\alpha_{l}\) (up to \(5.4\times 10^{- 4} / \mathrm{K})\) , the TeE is predicted to bring an even larger \(\Delta T_{ad}\) (up to \(22\mathrm{K}\) ) while maintaining relatively high efficiency. Our findings offer a non- phase- transition- based way for high- efficiency solid- state heat- pumping."  
+
+<|ref|>text<|/ref|><|det|>[[149, 658, 850, 789]]<|/det|>
+3. In some of the places, there is a marginal improvement in this presentation issue, in that there is mentioning of the 2nd law efficiency together with COPmat in some places. But there are still erroneous/incorrect/misleading sentences in the manuscript. On p.2, "Nevertheless, the energy efficiency, or system coefficient of performance (COPsys), of the heat pumps is capped with a relatively low coefficient of performance (COPmat) of the materials". This is not true and misleading. COPmat (for a fixed Th and Tc) is sufficiently high, it is the engineering ability to convert to the system COP which is lacking in current technologies.  
+
+<|ref|>text<|/ref|><|det|>[[177, 798, 848, 832]]<|/det|>
+Reply: Thank you for your comment. The above- mentioned sentence has now been removed. Please review the updated text provided in reply to your Comment 4.  
+
+<|ref|>text<|/ref|><|det|>[[149, 841, 849, 909]]<|/det|>
+4. "Due to phase-transition hysteresis, most ferroic materials — except for certain magnetocaloric materials undergoing second-order phase transition — possess a COPmat of about \(4 - 23\) . These values are equivalent to a second law efficiency...."This is misleading since the most common magnetocaloric material
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[179, 84, 715, 100]]<|/det|>
+is Gd, the 2nd order material, which has a high 2nd law efficiency.  
+
+<|ref|>text<|/ref|><|det|>[[179, 110, 850, 209]]<|/det|>
+Reply: Thank you for your feedback. We have removed magnetocaloric materials from the Introduction section, which now only mentions heat- pumping using first- order ferroelastic phase transition of shape memory alloys. You will see that this revision eliminates the above- mentioned misleading point without diminishing the significance of our study (i.e. introduction of a non- phase- transitional heat- pumping approach). Please review the updated text below:  
+
+<|ref|>text<|/ref|><|det|>[[180, 218, 586, 234]]<|/det|>
+The second paragraph of the Introduction section:  
+
+<|ref|>text<|/ref|><|det|>[[179, 243, 850, 425]]<|/det|>
+"Among green alternatives, solid- state heat- pumping using first- order ferroelastic phase transition of shape memory alloys (SMAs) has attracted much attention \(^{8 - 15}\) . Various high- performance heat- pumping devices have been developed using SMAs capable of exhibiting adiabatic temperature changes ( \(\Delta T_{ad}\) ) up to \(30 \mathrm{K}^{6,9 - 11}\) . Nevertheless, restricted by phase- transition hysteresis, the energy efficiency of SMAs reaches only about \(50\% - 70\%\) of the Carnot theoretical limit, significantly lower than those of VC- based refrigerants \((- 90\%)\) . Higher material- level efficiency is critical since it stands for the upper limit of the device- level energy efficiency. However, this goal is difficult to achieve via the typical first- order phase transition of ferroelastic alloys \(^{15}\) ; exploring anhysteretic non- phase- transition pathways may open a new way out."  
+
+<|ref|>text<|/ref|><|det|>[[150, 440, 850, 575]]<|/det|>
+5. "These values are equivalent to..." is also very misleading. It is the 2nd law efficiency which is the intrinsic property, and only with defining of Tc and Th, COP value make sense. COP is not an intrinsic comparison-value without fixing both Tc and Th. So the sentence structure should be the other way around: first introduce 2nd law efficiency, and then "this is equivalent to COPmat of XX with standard Th and Tc". In fact, I suggest that the authors change all mention of COPmat to 2nd law efficiency. And in some places, with fixing Th and Tc to some standard values, COP can be mentioned after first mentioning the 2nd law efficiency.  
+
+<|ref|>text<|/ref|><|det|>[[179, 584, 850, 650]]<|/det|>
+Reply: Thank you for your constructive suggestion. We have changed all mention of COPmat to the second law efficiency in the abstract (see reply to Comment 2) and introduction (see reply to Comment 4) of the draft. Detailed COPmat under standard Th and Tc is only provided in the main text, as shown below:  
+
+<|ref|>text<|/ref|><|det|>[[180, 658, 373, 675]]<|/det|>
+Lines 21- 32 at Page 5:  
+
+<|ref|>text<|/ref|><|det|>[[179, 683, 864, 899]]<|/det|>
+"We calculated the material- level coefficient of performance (COPmat) of the [100]- textured \(\mathrm{Ti}_{78}\mathrm{Nb}_{22}\) in a Stirling heat- pumping cycle by using a thermodynamic model developed by Qian et al. \(^{42,43}\) (see Note 2 of Supplementary information) along with the measured properties of the \(\mathrm{Ti}_{78}\mathrm{Nb}_{22}\) (see Table S3). Under \(T_{\mathrm{c}} = 288 \mathrm{K}\) (heat source temperature) and \(T_{\mathrm{h}} = 298 \mathrm{K}\) (heat sink temperature), the \(\mathrm{Ti}_{78}\mathrm{Nb}_{22}\) exhibits a COPmat of 25.3, equivalent to \(88\%\) of the Carnot COP (COPcarnot = \(T_{\mathrm{c}} / (T_{\mathrm{h}} - T_{\mathrm{c}}) = 28.8\) ). As shown in Fig. 3(e), this ratio, known as second- law efficiency \(^{42,43}\) , well surpasses those from first- order phase transitions in electrocaloric materials (31% - 41%, calculated values from Ref. 42) and elastocaloric SMAs (55% - 71% \(^{42}\) ) and rivals those from second- order phase transition in magnetocaloric materials (79% - 91% \(^{42}\) ) and liquid- vapor phase transition in commercial VC refrigerants (86% - 91% \(^{42}\) ). Near its optimal working temperature (i.e. under \(T_{\mathrm{c}} = 468 \mathrm{K}\) and \(T_{\mathrm{h}} = 478 \mathrm{K}\) ), the \(\mathrm{Ti}_{78}\mathrm{Nb}_{22}\) has an even higher second- law efficiency of 94%.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[148, 83, 850, 150]]<|/det|>
+6. Along the same line, Fig. 4a must be a comparison of 2nd law efficiency, as done in Fig. 4b. (Table S3 does not contain Th and Tc; it only lists delta T). Other aspects (typos, etc.) and supplying more details, as suggested by the other reviewers, seem to be largely resolved.  
+
+<|ref|>text<|/ref|><|det|>[[177, 158, 850, 226]]<|/det|>
+Reply: Thank you for your comment. We have deleted Fig. 4a and its related figures (i.e. original Fig. 3(e) and Fig. S12) and table (i.e. Table S3). For compactness, the original Fig. 4(b) (i.e. comparison of second law efficiency) has been integrated into Fig. 3 as subfigure (e), as shown below:  
+
+<|ref|>image<|/ref|><|det|>[[177, 231, 850, 624]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[177, 631, 850, 699]]<|/det|>
+<center>Fig. 3. ...e, Second law efficiency of the \(\mathrm{Ti}_{78}\mathrm{Nb}_{22}\) and phase-transitional materials calculated under the same working condition of \(T_{\mathrm{c}} = 288 \mathrm{~K}\) and \(T_{\mathrm{h}} = 298 \mathrm{~K}\) ; the calculated values for phase-transition materials are taken from Ref. \(^{42}\) while that for the \(\mathrm{Ti}_{78}\mathrm{Nb}_{22}\) is detailed in Note 2 of Supplementary Information. </center>
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__6e15618c9c8095f82d3987af8240d75577adac9e0623880cd07bd35f8e1fe9e9/images_list.json

@@ -0,0 +1,63 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_0.jpg",
+    "caption": "Figure R1. Robustness of predictive alignment against noise. (A) Networks with different strengths of noise were trained with the FORCE (red) and the predictive alignment (green) with the patterned target signal. Error bars stand for s.d.s over 20 independent simulations. (B) Example readout activities during the late phase of training are shown. Colors are the same as in A. The gray traces represent the target signal.",
+    "footnote": [],
+    "bbox": [
+      [
+        150,
+        125,
+        848,
+        318
+      ]
+    ],
+    "page_idx": 5
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_1.jpg",
+    "caption": "Figure R3. Learning fixed point attractors. (A) Example three inputs (black) and outputs (blue, orange, and green) are shown. (B) Low dimensional network dynamics are shown. Network dynamics perturbed with the varying strength of inputs shows that a saddle point mediates the transitions between attractors.",
+    "footnote": [],
+    "bbox": [],
+    "page_idx": 5
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_2.jpg",
+    "caption": "Figure R4. Robustness to hyperparameter choices. (A) The networks were trained with different levels of learning rates. (B) Same as A, but trained with different degrees of the initial strength of the plastic recurrent weights. (C) Same as A, but trained with different",
+    "footnote": [],
+    "bbox": [
+      [
+        144,
+        623,
+        844,
+        782
+      ]
+    ],
+    "page_idx": 7
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_3.jpg",
+    "caption": "Figure R5. Robustness of predictive alignment against noise. (A) Networks with different strengths of noise were trained with the FORCE (red) and the predictive alignment (green) with the patterned target signal. Error bars stand for s.d.s over 20 independent simulations. (B)",
+    "footnote": [],
+    "bbox": [
+      [
+        147,
+        629,
+        848,
+        824
+      ]
+    ],
+    "page_idx": 8
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_4.jpg",
+    "caption": "Figure R7. Neural activities for performing Ready-Set-Go task. (A) In each trial, two input units (blue and orange) send pulses to the network with a random delay \\(T_{\\text{delay}}\\) between pulses. The target network output (green) should be a pulse delayed by \\(T_{\\text{delay}}\\) relative to the second input pulse. (B) Networks were trained on a set of samples (colored squares) and tested for generalization to novel inputs, including interpolation within the training range (shaded region) as well as extrapolation beyond the training range. (C) The network after training failed to extrapolate beyond the training range. (D) Principal component analysis (PCA) of the trained network revealed that as the delay \\(T_{\\text{delay}}\\) increased, the network states corresponding to the output peak shifted linearly along a manifold in state space.",
+    "footnote": [],
+    "bbox": [],
+    "page_idx": 9
+  }
+]

peer_reviews/supplementary_0_Transparent Peer Review file__6e15618c9c8095f82d3987af8240d75577adac9e0623880cd07bd35f8e1fe9e9/supplementary_0_Transparent Peer Review file__6e15618c9c8095f82d3987af8240d75577adac9e0623880cd07bd35f8e1fe9e9.mmd

@@ -0,0 +1,350 @@
+
+# Taming the chaos gently: a Predictive Alignment learning rule in recurrent neural networks  
+
+Corresponding Author: Dr Toshitake Asabuki  
+
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+A version of this paper was originally rejected for publication by Nature Communications, however that decision was reconsidered after appeal by the authors.  
+
+Version 0:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author)  
+
+In this manuscript, the authors present a learning method for firing rate- based recurrent neural networks (RNNs), referred to as predictive alignment, and demonstrate that the method can learn and generate a variety of patterned activities, ranging from periodic signals to chaotic Lorenz attractors and high- dimensional movie data.  
+
+Strengths  
+
+As neural circuits exhibit complex chaotic spontaneous activity, understanding how such chaotic activity can be efficiently reorganized to give rise to coherent activity patterns that generate behaviour is indeed an important question. The study attempts to address this question, particularly from the angle that prediction, to some extent, can guide learning.  
+
+Weaknesses  
+
+This study, however, has some serious weaknesses, indicating that the results reported in this manuscript do not necessarily represent a significant advance in the field. More on these in a point- by- point fashion below:  
+
+(1) From a machine learning point of view, the tasks that can be performed by training RNNs using the predictive alignment method can all be implemented by the FORCE method and its variants, often with better performance; some of these papers are cited in the manuscript. From a biological perspective, the methods developed in this study lack the biological plausibility of training RNNs with spiking neurons. Regarding this latter point, the e-prop method (Bellec et al., Nature Communications, 2020), which is also local and leverages error signals, offers more biological plausibility. As the field stands now, a powerful new method that can effectively train recurrent spiking neural networks would represent significant progress.  
+
+(2) It is unclear why the learning method works. In other words, what changes occur during the learning process that enable the RNNs to perform the intended computations? Are these changes related to the reorganization of the underlying chaotic attractor into a rich combination of fixed points or continuous attractors of the corresponding dynamical systems? I would like to see a deeper treatment of this question to open the "black box" of this process. The learning performance is best at the "edge of chaos," which is a very interesting observation, but why do such critical dynamics facilitate learning?  
+
+(3) What is the biological relevance of assuming the recurrent connectivity is a summation of M and G?  
+
+(4) What are the effects of certain hyperparameters, such as learning rates, on learning performance? Without thorough exploration of this aspect, it is difficult to assess whether the learning method is robust and effective.  
+
+(Remarks on code availability)
+
+<--- Page Split --->
+
+(Remarks to the Author)  
+
+Here the authors propose a novel learning rule to "tame" high- dimensional chaotic patterns in recurrent neural networks. The learning rule, "predictive alignment", adjusts recurrent weights in order to best predict the feedback in an initially chaotic RNN network. After training, the chaotic activity resembles an RNN in which the network is suppressed by an external or feedback signal except that the suppression is driven by the recurrent network (although by an independent recurrent weight matrix M). The authors show that RNNs trained with predictive alignment can learn a number of easy and difficult time- varying tasks. An important contribution of this paper is that the learning rule is indeed much more biologically plausible in terms of the learning rule itself. However, there are other aspects of the implementation that are not normally present in RNN models and might be considered implausible by some (e.g., the "silent feedback" and the two weight matrices). But these are not necessarily implausible, and may be considered to be predictions. Overall the presentation of the novel and potentially biologically plausible model is a significant contribution, however, the methods are very brief and the task implementations are not well described, making it somewhat difficult to evaluate the work in detail.  
+
+My primary concern relates to the apparent absence of noise in the RNN. As far as I could tell by looking at the Methods and sample code (perhaps I'm wrong), in contrast to most RNN models, there was no noise term during training or testing. When proposing a biologically plausible learning rule that relies on feedback from a chaotic RNN the absence of noise is potentially a serious issue, because training may not work at all when noise is present as the feedback signal is a moving target. It is thus necessary to include noise and parametrically vary noise to understand the influence of noise in predictive alignment.  
+
+The use of the Measure- Wait- Go task is interesting, and an important demonstration of the power of the learning rule. However, the time scale used was less than an order of magnitude longer than that of tau. Does the approach work for more biologically relevant time scales such as those used in the Jazayeri and Shadlen paper: hundreds of milliseconds? Also, this task is normally referred to as the ready- set- go task (Jazayeri and Shadlen, 2010). It would also be helpful to show a sample of the RNN activity during the task rather than just the PC's, because linear PCs can emerge from RNNs without much linear activity.  
+
+Replay task. It was not clear what time \(= 0\) was for the movie generation? Does the produced movie run in a limit cycle?  
+
+Pg. 2. To support the statement "Cortical circuits often exhibit chaotic spontaneous activity" a number of papers are cited, but they all seem to be models, not experimental papers (and I don't think Rajan et al, 2016 even address chaos). RNNs are not cortical circuits so these references do not seem appropriate. London, ..., Latham, 2010 provides some evidence for chaos in cortical circuits.  
+
+The presentation and discussion omitted many relevant studies that laid the foundation for the current work.  
+
+Jaeger and Haas 2004, essentially published what later came to known as FORCE in their "online" learning rule.  
+
+The paper has commonalities with Laje and Buonomano, which also took the approach of taming chaos by changing the recurrent weights and using high dimensional structure of a local stable trajectories as a reservoir. That work was not cited in the text. And Buonomano and Merzenich 1995, was the first description of what later came to be called a reservoir network.  
+
+Supplemental Figure 4. A schematic of the network would be helpful.  
+
+Pg. 14 "while failed AT extrapolation"  
+
+Pg. 15 "Here, we train recurrent neural" ... typo.  
+
+It is best to refer to the units and weights as such, rather than neurons and synapses.  
+
+(Remarks on code availability)  
+
+Version 1:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author)  
+
+I appreciate the authors' detailed responses and the corresponding revisions. However, my concerns regarding the training of RNNs with spiking neurons and the changes that occur during the learning process to enable these networks to perform the intended computations remain unclear. Specifically:
+
+<--- Page Split --->
+
+(1) For the proposed learning rule to be effectively evaluated in a spiking neural network, it is essential to compare its performance with existing methods for training SNNs. Additionally, applying the proposed method to balanced spiking neural networks with both excitatory and inhibitory neurons would be more appropriate for two reasons: (i) it aligns better with biological plausibility, and (ii) balanced networks are known to exhibit chaotic dynamics, which are more consistent with the problem addressed in this study.  
+
+(2) While the revision presents some results on learned fixed points, it remains unclear how and why the underlying chaotic dynamics are reorganized during the training process.  
+
+(3) The added explanation for why the model performs best at the edge of chaos is largely qualitative and descriptive (see Line 310). In machine learning, it is well established that artificial neural networks achieve optimal training performance near the edge of chaos [Schoenholz, et al., Deep information propagation, ICLR 2017], supported by quantitative analysis. Similarly, I believe a quantitative investigation is necessary to strengthen this argument.  
+
+(Remarks on code availability)  
+
+Reviewer #2  
+
+(Remarks to the Author) The authors have adequately addressed my concerns, and the inclusion of the spiking novel further improves the paper?  
+
+"matrixes" should be "matrices"  
+
+(Remarks on code availability)  
+
+Version 2:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author)  
+
+In the revised manuscript, the authors have included comparisons between the proposed learning rule and two other supervised learning methods, and have added some quantitative explanations concerning why learning performance is maximised at the edge of chaos. My main concerns have now been addressed in the revisions.  
+
+(Remarks on code availability)
+
+<--- Page Split --->
+
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source.  
+
+The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.  
+
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+
+# Response to the reviewers  
+
+Toshitake Asabuki and Claudia Clopath  
+
+We thank the reviewers for their thorough and critical evaluation of our work. We read the reviewers' comments carefully and we think we have addressed reviewers' concerns fully. Below, we explain how we addressed each of their concerns point- by- point. In particular, our additional results show that we can implement our rule in a spiking network and that the rule is robust to noise, two majors concerns of the reviewers. Thank you very much in advance for your time.  
+
+## Reviewer #1  
+
+Reviewer #1Reviewer Point P 1.1 —From a machine learning point of view, the tasks that can be performed by training RNNs using the predictive alignment method can all be implemented by the FORCE method and its variants, often with better performance; some of these papers are cited in the manuscript. From a biological perspective, the methods developed in this study lack the biological plausibility of training RNNs with spiking neurons. Regarding this latter point, the e- prop method (Bellec et al., Nature Communications, 2020), which is also local and leverages error signals, offers more biological plausibility. As the field stands now, a powerful new method that can effectively train recurrent spiking neural networks would represent significant progress.  
+
+Reply: Regarding the first point, while it is true that the tasks used in our study can all be performed by FORCE learning in most cases, we would like to emphasize that we have shown the proposed model performs even better than FORCE in the presence of noise (Supplementary Figure 2, shown again below). This happens because FORCE aims to make its output match the target signal as quickly as possible, which can be disrupted by strong noise. In contrast, our model learns more slowly, making the learning process less affected by noise. We have explained these results in II. 227- 231 in the revised manuscript.
+
+<--- Page Split --->
+![](images/Figure_unknown_0.jpg)
+
+<center>Figure R1. Robustness of predictive alignment against noise. (A) Networks with different strengths of noise were trained with the FORCE (red) and the predictive alignment (green) with the patterned target signal. Error bars stand for s.d.s over 20 independent simulations. (B) Example readout activities during the late phase of training are shown. Colors are the same as in A. The gray traces represent the target signal. </center>  
+
+To address the second point, we asked whether our learning rule can be implemented in a spiking recurrent network. We implemented a recurrent network composed of 1,000 leaky integrate- and- fire (LIF) neurons, coupled with a single linear readout unit as the output. We ran additional simulations and found that our rule can be applied in a network of spiking neurons (Fig. R2 shown below). We have include the new results as a new Supplementary Figure 8 in the revised manuscript. The details of simulation results and implementations are shown in II. 494- 500 and II.756- 783 in the revised manuscript, respectively.  
+
+![](images/Figure_unknown_1.jpg)
+
+
+<--- Page Split --->
+
+Figure R2. The spiking recurrent network was trained to learn periodic target signal. (A) The blue trace represents the target signal, and the green line represents the output. (B) Activities of spiking network neurons before training are shown. (C) Activities of trained spiking network neurons are shown. The modified recurrent connections generate the sequential network activity.  
+
+Reviewer Point P 1.2 — It is unclear why the learning method works. In other words, what changes occur during the learning process that enable the RNNs to perform the intended computations? Are these changes related to the reorganization of the underlying chaotic attractor into a rich combination of fixed points or continuous attractors of the corresponding dynamical systems? I would like to see a deeper treatment of this question to open the "black box" of this process. The learning performance is best at the "edge of chaos," which is a very interesting observation, but why do such critical dynamics facilitate learning?  
+
+Reply: The reviewer's intuition is correct. Our rule suppresses the chaotic spontaneous activity and reorganize it to multiple attractors by modifying the recurrent connections. To address this, we adopted the approach of Sussillo and Barak (2013), where three output units were specifically trained to read out the corresponding fixed points of a network dynamics. The state of each output is determined by transient pulses from the corresponding input units. These pulses affect the corresponding output unit, causing it to switch or maintain a value of either \(+1\) or \(- 1\) . Once the output value is set, it remains fixed until the arrival of the next pulse from the input unit. We found that the trained network is capable of producing outputs that show transition between states in response to input (Figure R3A).  
+
+To observe how the trained network switches between memory states, we perturbed the state of the trained network with input pulses to observe transitions between two fixed points over six trials. In these trials, the strength of the input was gradually increased. When the input was weak, we found that the network activity briefly deviated from its fixed point and returned to the original fixed point when the input was removed. In striking contrast, when the input was strong enough, the network switched to a different fixed point, suggesting the existence of saddle points between the two stable fixed points, demonstrating its mechanistic role in the corresponding transition (Figure R3B). We have included these additional simulation results as Supplementary Figure 5 and explained the
+
+<--- Page Split --->
+
+results in II.325- 358 and implementation details in II.747- 754 in the revised manuscript.  
+
+Furthermore, we have included an additional sentence to explain why the edge of chaos is important in the proposed model in II. 310- 312: "The model performs best at the edge of chaos because it supports richness of network dynamics, which in turn provides rich basis functions from which the output units can readout the appropriate dynamics.".  
+
+![](images/Figure_unknown_2.jpg)
+
+<center>Figure R3. Learning fixed point attractors. (A) Example three inputs (black) and outputs (blue, orange, and green) are shown. (B) Low dimensional network dynamics are shown. Network dynamics perturbed with the varying strength of inputs shows that a saddle point mediates the transitions between attractors. </center>  
+
+Reviewer Point P 1.3 — What is the biological relevance of assuming the recurrent connectivity is a summation of M and G?  
+
+Reply: The reviewer raised an important point. Our plasticity rule assumes two types of recurrent connections (i.e., M and G), one is plastic and the other is static. While the biological relevance of this is still an open question, it is known experimentally that the degree of synaptic plasticity varies across different compartments of dendrites (Gordon et al., J Neurosci. 2006). Based on this experimental evidence, we can speculate that the plastic connections are synapses that project onto proximal basal dendrites and the static connections
+
+<--- Page Split --->
+
+project onto distal basal dendrites. We have included this point in II.631- 637 in the revised manuscript.  
+
+Reviewer Point P 1.4 — What are the effects of certain hyperparameters, such as learning rates, on learning performance? Without thorough exploration of this aspect, it is difficult to assess whether the learning method is robust and effective.  
+
+Reply: We have investigated the robustness of the model with respect to its hyperparameters. First, we examined the effect of the learning rate on performance. We found that as the learning rate increased, so did the output error; however, if the learning rate was sufficiently small, the error remained small even when the learning rate was increased tenfold (Figure R4A). Next, we examined the initial strength of the plastic recurrent connections, where the connections were fully connected. We found that as long as the initial connection strength was not extremely large (e.g., when it was larger than that of a strong connection G), the error remained sufficiently small (Figure R4B). Finally, we examined against the connection probability of M and found that as long as the connections were not extremely sparse, the error remained sufficiently small (Figure R4C). These results suggest that the learning performance was not susceptible to certain hyperparameter choices. We have included these new results as a new Supplementary Figure 3 and explained in II. 233- 237 in the revised manuscript.  
+
+![](images/Figure_unknown_3.jpg)
+
+<center>Figure R4. Robustness to hyperparameter choices. (A) The networks were trained with different levels of learning rates. (B) Same as A, but trained with different degrees of the initial strength of the plastic recurrent weights. (C) Same as A, but trained with different </center>
+
+<--- Page Split --->
+
+degrees of connection probabilities of the plastic recurrent weights. Error bars represent SEs across five independent simulations.  
+
+## Reviewer #2  
+
+Reviewer #2Reviewer Point P 2.1 —My primary concern relates to the apparent absence of noise in the RNN. As far as I could tell by looking at the Methods and sample code (perhaps I’m wrong), in contrast to most RNN models, there was no noise term during training or testing. When proposing a biologically plausible learning rule that relies on feedback from a chaotic RNN the absence of noise is potentially a serious issue, because training may not work at all when noise is present as the feedback signal is a moving target. It is thus necessary to include noise and parametrically vary noise to understand the influence of noise in predictive alignment.  
+
+Reply: We apologize to the reviewer for any confusion caused by our manuscript. We had already included external noise in the network during learning. In the Supplementary Figure 2 (shown again below), we have shown that the proposed model performs even better than FORCE in the presence of noise. These results indicate that our model is robust to external noise in the network. We have explained these results in II. 227- 231 in the revised manuscript.  
+
+In the revised manuscript, we clarified this point by including noise term explicitly in the Equation 7, and mentioned "We will consider network dynamics with external drives and noise." in II. 84- 85 in the Results section to avoid confusion.  
+
+![](images/Figure_unknown_4.jpg)
+
+<center>Figure R5. Robustness of predictive alignment against noise. (A) Networks with different strengths of noise were trained with the FORCE (red) and the predictive alignment (green) with the patterned target signal. Error bars stand for s.d.s over 20 independent simulations. (B) </center>
+
+<--- Page Split --->
+
+Example readout activities during the late phase of training are shown. Colors are the same as in A. The gray traces represent the target signal.  
+
+Reviewer Point P 2.2 —The use of the Measure- Wait- Go task is interesting, and an important demonstration of the power of the learning rule. However, the time scale used was less than an order of magnitude longer than that of tau. Does the approach work for more biologically relevant time scales such as those used in the Jazayeri and Shadlen paper: hundreds of milliseconds? Also, this task is normally referred to as the ready- set- go task (Jazayeri and Shadlen, 2010). It would also be helpful to show a sample of the RNN activity during the task rather than just the PC's, because linear PCs can emerge from RNNs without much linear activity.  
+
+Reply: We increased the delay period to be remembered in the task to more than 100 ms. We found that although the delay period was more than 10 times the time scale of the neuron's dynamics, the model still showed the same behavior as in the previous setting. In the revised manuscript, previous Figure 5 has been replaced by the new simulation results. In a supplementary Figure 7, we have shown ten samples of the RNN activity in the revised manuscript (shown as Fig. R7 below). Furthermore, in the revised manuscript, we have changed the term "Measure- Wait- Go task" to "Ready- Set- Go task".  
+
+Learning over much longer timescales (i.e., hundreds of milliseconds) can in principle be achieved by extending network timescales through larger networks, adaptation, and short- term plasticity. We want to leave this as a future work.  
+
+![PLACEHOLDER_10_0]
+
+
+<--- Page Split --->
+
+Figure R6. Learning delay- matching task. (A) In each trial, two input units (blue and orange) send pulses to the network with a random delay \(T_{\text{delay}}\) between pulses. The target network output (green) should be a pulse delayed by \(T_{\text{delay}}\) relative to the second input pulse. (B) Networks were trained on a set of samples (colored squares) and tested for generalization to novel inputs, including interpolation within the training range (shaded region) as well as extrapolation beyond the training range. (C) The network after training failed to extrapolate beyond the training range. (D) Principal component analysis (PCA) of the trained network revealed that as the delay \(T_{\text{delay}}\) increased, the network states corresponding to the output peak shifted linearly along a manifold in state space.  
+
+![PLACEHOLDER_11_0]
+
+<center>Figure R7. Neural activities for performing Ready-Set-Go task. (A) In each trial, two input units (blue and orange) send pulses to the network with a random delay \(T_{\text{delay}}\) between pulses. The target network output (green) should be a pulse delayed by \(T_{\text{delay}}\) relative to the second input pulse. (B) Networks were trained on a set of samples (colored squares) and tested for generalization to novel inputs, including interpolation within the training range (shaded region) as well as extrapolation beyond the training range. (C) The network after training failed to extrapolate beyond the training range. (D) Principal component analysis (PCA) of the trained network revealed that as the delay \(T_{\text{delay}}\) increased, the network states corresponding to the output peak shifted linearly along a manifold in state space. </center>
+
+<--- Page Split --->
+
+**Reviewer Point P 2.3** —Replay task. It was not clear what time \(= 0\) was for the movie generation? Does the produced movie run in a limit cycle?  
+
+Reply: We thank the reviewer for pointing this out. At time zero, we set the network activity to the predefined initial states. The same initial state was used in both the learning and testing phases. We have included these explanations in II. 515- 518 in the revised manuscript. While we reset the initial state over trials in the results shown in Figure 6, in principle the movie can run in a limit cycle and restart when the movie ends.  
+
+Reviewer Point P 2.4 —Pg. 2. To support the statement "Cortical circuits often exhibit chaotic spontaneous activity" a number of papers are cited, but they all seem to be models, not experimental papers (and I don't think Rajan et al, 2016 even address chaos). RNNs are not cortical circuits so these references do not seem appropriate. London, ..., Latham, 2010 provides some evidence for chaos in cortical circuits.  
+
+Reply: The reviewer is correct, we cited only computational papers in our previous manuscript. We have included the suggested reference and changed the sentence as: "Cortical circuits often exhibit chaotic spontaneous activity (London et al., 2010), and RNNs can generate such dynamics through feedback loops (van Vreeswijk and Sompolinsky, 1996; Amit and Brunel, 1997; Brunel, 2000; Toyoizumi and Abbott, 2011; Rajan et al., 2016).", in II. 42- 45 in the revised manuscript. Here, we have kept the Rajan et al. paper since they used RNNs showing chaotic behavior before training.  
+
+Reviewer Points P 2.5 —The presentation and discussion omitted many relevant studies that laid the foundation for the current work.  
+
+Reply: We hope we have included the appropriate references in the revised manuscript. If there are any missing, could the reviewer let us know?  
+
+Reviewer Points P 2.6 —Jaeger and Haas 2004, essentially published what later came to known as FORCE in their "online" learning rule.  
+
+Reply: We have included the reference in the revised manuscript.
+
+<--- Page Split --->
+
+Reviewer Points P 2.7 — The paper has commonalities with Laje and Buonomano, which also took the approach of taming chaos by changing the recurrent weights and using high dimensional structure of a local stable trajectories as a reservoir. That work was not cited in the text. And Buonomano and Merzenich 1995, was the first description of what later came to be called a reservoir network.  
+
+Reply: We have included the references in the revised manuscript.  
+
+Reviewer Points P 2.8 — Supplemental Figure 4. A schematic of the network would be helpful.  
+
+Reply: We have included new schematics to help readers. We thank the reviewer for pointing this out.  
+
+Reviewer Points P 2.9 — Pg. 14 “while failed AT extrapolation”; Pg. 15 “Here, we train recurrent neural” ... typo.  
+
+Reply: We have corrected the typo. We thank the reviewer for their careful review.  
+
+Reviewer Points P 2.10 — It is best to refer to the units and weights as such, rather than neurons and synapses.  
+
+Reply: We agree with the reviewer to change our terminology. We have changed the terminology in the revised manuscript.  
+
+We look forward to hearing from you soon. Thank you very much for your kind consideration in advance.  
+
+Sincerely yours,  
+
+![PLACEHOLDER_13_0]
+  
+
+Toshitake Asabuki  RIKEN Center for Brain Science, Japan
+
+<--- Page Split --->
+
+toshitake.asabuki@riken.jp 
+
+![PLACEHOLDER_14_0]
+
+Prof. Claudia Clopath  Department of Bioengineering, Imperial College London, London, UK.  c.clopath@ imperial.ac.uk; +44 (0)20 7594 1435
+
+<--- Page Split --->
+
+# Response to the reviewers  
+
+Toshitake Asabuki and Claudia Clopath  
+
+We thank the reviewers for their thorough and critical evaluation of our work. We read the reviewers' comments carefully and we think we have addressed reviewers' concerns fully. Below, we explain how we addressed each of their concerns point- by- point. Thank you very much in advance for your time.  
+
+## Reviewer #1  
+
+Reviewer #1Reviewer Point P 1.1 —I appreciate the authors' detailed responses and the corresponding revisions. However, my concerns regarding the training of RNNs with spiking neurons and the changes that occur during the learning process to enable these networks to perform the intended computations remain unclear. Specifically:  
+
+(1) For the proposed learning rule to be effectively evaluated in a spiking neural network, it is essential to compare its performance with existing methods for training SNNs. Additionally, applying the proposed method to balanced spiking neural networks with both excitatory and inhibitory neurons would be more appropriate for two reasons: (i) it aligns better with biological plausibility, and (ii) balanced networks are known to exhibit chaotic dynamics, which are more consistent with the problem addressed in this study.  
+
+Reply: That's a great suggestion. We have compared the performance with two types of stae- of- the- art spiking recurrent network for supervised learning: FORCE and e- prop. We have included the new results as Supplementary Fig. 10C and mentioned in II.524- 525 in the revised manuscript.  
+
+Furthermore, we showed that our proposed rule enables learning even in a recurrent spiking neural network composed of two populations (i.e., excitatory and inhibitory). We have included the new result as Supplementary Figure 11 and mentioned in II.525- 528 in the revised manuscript.  
+
+Reviewer Point P 1.2 —(2) While the revision presents some results on learned fixed points, it remains unclear how and why the underlying chaotic dynamics are reorganized during the training process.
+
+<--- Page Split --->
+
+Reply: The revierer raised an important point. While we have shown that aligning the predictive recurrent and chaotic dynamics is important for stable learning in Figure 3, we did not explain how it changes the recurrent dynamics. To see this further and understand how the alignment enables stable learning, we monitored the network's Lyapunov exponent during training. We found that in the aligned case, the network's Lyapunov exponent was shifted further toward the negative side compared to the control case, indicating more effective suppression of chaos. We have included the new simulation result as Supplementary Figure 4 and explained in II. 270- 272 in the revised manuscript.  
+
+Reviewer Point P 1.3 —(3) The added explanation for why the model performs best at the edge of chaos is largely qualitative and descriptive (see Line 310). In machine learning, it is well established that artificial neural networks achieve optimal training performance near the edge of chaos [Schoenholz, et al., Deep information propagation, ICLR 2017], supported by quantitative analysis. Similarly, I believe a quantitative investigation is necessary to strengthen this argument.  
+
+Reply: To investigate why learning performance is maximized at the edge of chaos, we analyzed the eigenvalue distribution of the correlation matrix of the network activity. Optimal learning is thought to require a balance between representational diversity and dimensionality. High diversity allows the network to encode rich information, while low dimensionality ensures compact representations. We quantified the diversity of neural representations using the entropy of the eigenvalues of the correlation matrix \((H_{\lambda})\) , and measured the balance between information spread and localization using the square root of the participation ratio \((\sqrt{PR})\) . We computed the ratio of these two measures,  
+
+\[\mathrm{Efficiency} = \frac{H_{\lambda}}{\sqrt{PR}}\]  
+
+and found that this balance index was maximized specifically at the edge of chaos. In the subcritical regime, the participation ratio was low, indicating that activity was concentrated in a low- dimensional subspace, reducing the network's ability to learn diverse representations. In contrast, in the chaotic regime, the participation ratio was high, but the eigenvalue entropy decreased, suggesting
+
+<--- Page Split --->
+
+excessive dispersion of information, leading to loss of meaningful structure. At the edge of chaos, both measures were optimally balanced, enabling the network to achieve the highest learning performance. These findings suggest that learning is most efficient when the network exhibits rich yet structured dynamics, balancing representational diversity and stability. We have included the new simulation results as Supplementary Figure 6 in the revised manuscript and explained in II. 327- 347 in the revised manuscript.  
+
+Reviewer #2 (Remarks to the Author):  
+
+The authors have adequately addressed my concerns, and the inclusion of the spiking novel further improves the paper?  
+
+"matrixes" should be "matrices"  
+
+Reply: We thank the reviewer for the careful review. We have fixed the typo.  
+
+We look forward to hearing from you soon. Thank you very much for your kind consideration in advance.  
+
+Sincerely yours,  
+
+![PLACEHOLDER_17_0]
+  
+
+Toshitake Asabuki  RIKEN Center for Brain Science, Japan  toshitake.asabuki@riken.jp  
+
+![PLACEHOLDER_17_1]
+  
+
+Prof. Claudia Clopath  Department of Bioengineering, Imperial College London, London, UK.  c.clopath@ imperial.ac.uk; +44 (0)20 7594 1435
+
+<--- Page Split --->
+
+# Response to the reviewersToshitake Asabuki and Claudia Clopath  
+
+## Reviewer #1  
+
+Reviewer Point P 1.1 —In the revised manuscript, the authors have included comparisons between the proposed learning rule and two other supervised learning methods, and have added some quantitative explanations concerning why learning performance is maximised at the edge of chaos. My main concerns have now been addressed in the revisions.  
+
+Reply: We thank the reviewer for their fruitful feedback.  
+
+We look forward to hearing from you soon. Thank you very much for your kind consideration in advance.  
+
+Sincerely yours,  
+
+![PLACEHOLDER_18_0]
+  
+
+Toshitake Asabuki RIKEN Center for Brain Science, Japan toshitake.asabuki@riken.jp  
+
+![PLACEHOLDER_18_1]
+  
+
+Prof. Claudia Clopath  Department of Bioengineering, Imperial College London, London, UK.  c.clopath@ imperial.ac.uk; +44 (0)20 7594 1435
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__6e15618c9c8095f82d3987af8240d75577adac9e0623880cd07bd35f8e1fe9e9/supplementary_0_Transparent Peer Review file__6e15618c9c8095f82d3987af8240d75577adac9e0623880cd07bd35f8e1fe9e9_det.mmd

@@ -0,0 +1,461 @@
+<|ref|>title<|/ref|><|det|>[[73, 161, 864, 209]]<|/det|>
+# Taming the chaos gently: a Predictive Alignment learning rule in recurrent neural networks  
+
+<|ref|>text<|/ref|><|det|>[[73, 223, 460, 240]]<|/det|>
+Corresponding Author: Dr Toshitake Asabuki  
+
+<|ref|>text<|/ref|><|det|>[[70, 273, 865, 289]]<|/det|>
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+<|ref|>text<|/ref|><|det|>[[70, 300, 867, 328]]<|/det|>
+A version of this paper was originally rejected for publication by Nature Communications, however that decision was reconsidered after appeal by the authors.  
+
+<|ref|>text<|/ref|><|det|>[[73, 365, 145, 379]]<|/det|>
+Version 0:  
+
+<|ref|>text<|/ref|><|det|>[[73, 391, 219, 405]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[73, 417, 160, 431]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[73, 444, 238, 457]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[73, 456, 911, 497]]<|/det|>
+In this manuscript, the authors present a learning method for firing rate- based recurrent neural networks (RNNs), referred to as predictive alignment, and demonstrate that the method can learn and generate a variety of patterned activities, ranging from periodic signals to chaotic Lorenz attractors and high- dimensional movie data.  
+
+<|ref|>text<|/ref|><|det|>[[73, 509, 140, 523]]<|/det|>
+Strengths  
+
+<|ref|>text<|/ref|><|det|>[[73, 534, 905, 575]]<|/det|>
+As neural circuits exhibit complex chaotic spontaneous activity, understanding how such chaotic activity can be efficiently reorganized to give rise to coherent activity patterns that generate behaviour is indeed an important question. The study attempts to address this question, particularly from the angle that prediction, to some extent, can guide learning.  
+
+<|ref|>text<|/ref|><|det|>[[73, 587, 163, 600]]<|/det|>
+Weaknesses  
+
+<|ref|>text<|/ref|><|det|>[[70, 612, 921, 640]]<|/det|>
+This study, however, has some serious weaknesses, indicating that the results reported in this manuscript do not necessarily represent a significant advance in the field. More on these in a point- by- point fashion below:  
+
+<|ref|>text<|/ref|><|det|>[[72, 652, 919, 744]]<|/det|>
+(1) From a machine learning point of view, the tasks that can be performed by training RNNs using the predictive alignment method can all be implemented by the FORCE method and its variants, often with better performance; some of these papers are cited in the manuscript. From a biological perspective, the methods developed in this study lack the biological plausibility of training RNNs with spiking neurons. Regarding this latter point, the e-prop method (Bellec et al., Nature Communications, 2020), which is also local and leverages error signals, offers more biological plausibility. As the field stands now, a powerful new method that can effectively train recurrent spiking neural networks would represent significant progress.  
+
+<|ref|>text<|/ref|><|det|>[[72, 755, 920, 822]]<|/det|>
+(2) It is unclear why the learning method works. In other words, what changes occur during the learning process that enable the RNNs to perform the intended computations? Are these changes related to the reorganization of the underlying chaotic attractor into a rich combination of fixed points or continuous attractors of the corresponding dynamical systems? I would like to see a deeper treatment of this question to open the "black box" of this process. The learning performance is best at the "edge of chaos," which is a very interesting observation, but why do such critical dynamics facilitate learning?  
+
+<|ref|>text<|/ref|><|det|>[[70, 833, 773, 848]]<|/det|>
+(3) What is the biological relevance of assuming the recurrent connectivity is a summation of M and G?  
+
+<|ref|>text<|/ref|><|det|>[[70, 860, 884, 887]]<|/det|>
+(4) What are the effects of certain hyperparameters, such as learning rates, on learning performance? Without thorough exploration of this aspect, it is difficult to assess whether the learning method is robust and effective.  
+
+<|ref|>text<|/ref|><|det|>[[73, 911, 282, 925]]<|/det|>
+(Remarks on code availability)
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[73, 87, 238, 100]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 100, 923, 243]]<|/det|>
+Here the authors propose a novel learning rule to "tame" high- dimensional chaotic patterns in recurrent neural networks. The learning rule, "predictive alignment", adjusts recurrent weights in order to best predict the feedback in an initially chaotic RNN network. After training, the chaotic activity resembles an RNN in which the network is suppressed by an external or feedback signal except that the suppression is driven by the recurrent network (although by an independent recurrent weight matrix M). The authors show that RNNs trained with predictive alignment can learn a number of easy and difficult time- varying tasks. An important contribution of this paper is that the learning rule is indeed much more biologically plausible in terms of the learning rule itself. However, there are other aspects of the implementation that are not normally present in RNN models and might be considered implausible by some (e.g., the "silent feedback" and the two weight matrices). But these are not necessarily implausible, and may be considered to be predictions. Overall the presentation of the novel and potentially biologically plausible model is a significant contribution, however, the methods are very brief and the task implementations are not well described, making it somewhat difficult to evaluate the work in detail.  
+
+<|ref|>text<|/ref|><|det|>[[72, 255, 920, 335]]<|/det|>
+My primary concern relates to the apparent absence of noise in the RNN. As far as I could tell by looking at the Methods and sample code (perhaps I'm wrong), in contrast to most RNN models, there was no noise term during training or testing. When proposing a biologically plausible learning rule that relies on feedback from a chaotic RNN the absence of noise is potentially a serious issue, because training may not work at all when noise is present as the feedback signal is a moving target. It is thus necessary to include noise and parametrically vary noise to understand the influence of noise in predictive alignment.  
+
+<|ref|>text<|/ref|><|det|>[[72, 346, 919, 425]]<|/det|>
+The use of the Measure- Wait- Go task is interesting, and an important demonstration of the power of the learning rule. However, the time scale used was less than an order of magnitude longer than that of tau. Does the approach work for more biologically relevant time scales such as those used in the Jazayeri and Shadlen paper: hundreds of milliseconds? Also, this task is normally referred to as the ready- set- go task (Jazayeri and Shadlen, 2010). It would also be helpful to show a sample of the RNN activity during the task rather than just the PC's, because linear PCs can emerge from RNNs without much linear activity.  
+
+<|ref|>text<|/ref|><|det|>[[72, 437, 888, 452]]<|/det|>
+Replay task. It was not clear what time \(= 0\) was for the movie generation? Does the produced movie run in a limit cycle?  
+
+<|ref|>text<|/ref|><|det|>[[72, 464, 921, 517]]<|/det|>
+Pg. 2. To support the statement "Cortical circuits often exhibit chaotic spontaneous activity" a number of papers are cited, but they all seem to be models, not experimental papers (and I don't think Rajan et al, 2016 even address chaos). RNNs are not cortical circuits so these references do not seem appropriate. London, ..., Latham, 2010 provides some evidence for chaos in cortical circuits.  
+
+<|ref|>text<|/ref|><|det|>[[72, 529, 803, 543]]<|/det|>
+The presentation and discussion omitted many relevant studies that laid the foundation for the current work.  
+
+<|ref|>text<|/ref|><|det|>[[70, 555, 841, 569]]<|/det|>
+Jaeger and Haas 2004, essentially published what later came to known as FORCE in their "online" learning rule.  
+
+<|ref|>text<|/ref|><|det|>[[72, 581, 924, 621]]<|/det|>
+The paper has commonalities with Laje and Buonomano, which also took the approach of taming chaos by changing the recurrent weights and using high dimensional structure of a local stable trajectories as a reservoir. That work was not cited in the text. And Buonomano and Merzenich 1995, was the first description of what later came to be called a reservoir network.  
+
+<|ref|>text<|/ref|><|det|>[[72, 633, 545, 647]]<|/det|>
+Supplemental Figure 4. A schematic of the network would be helpful.  
+
+<|ref|>text<|/ref|><|det|>[[73, 659, 325, 673]]<|/det|>
+Pg. 14 "while failed AT extrapolation"  
+
+<|ref|>text<|/ref|><|det|>[[73, 686, 396, 700]]<|/det|>
+Pg. 15 "Here, we train recurrent neural" ... typo.  
+
+<|ref|>text<|/ref|><|det|>[[72, 711, 650, 725]]<|/det|>
+It is best to refer to the units and weights as such, rather than neurons and synapses.  
+
+<|ref|>text<|/ref|><|det|>[[73, 777, 282, 790]]<|/det|>
+(Remarks on code availability)  
+
+<|ref|>text<|/ref|><|det|>[[73, 816, 144, 828]]<|/det|>
+Version 1:  
+
+<|ref|>text<|/ref|><|det|>[[73, 842, 219, 854]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[73, 867, 160, 880]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[73, 894, 238, 906]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 907, 912, 946]]<|/det|>
+I appreciate the authors' detailed responses and the corresponding revisions. However, my concerns regarding the training of RNNs with spiking neurons and the changes that occur during the learning process to enable these networks to perform the intended computations remain unclear. Specifically:
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 58, 923, 127]]<|/det|>
+(1) For the proposed learning rule to be effectively evaluated in a spiking neural network, it is essential to compare its performance with existing methods for training SNNs. Additionally, applying the proposed method to balanced spiking neural networks with both excitatory and inhibitory neurons would be more appropriate for two reasons: (i) it aligns better with biological plausibility, and (ii) balanced networks are known to exhibit chaotic dynamics, which are more consistent with the problem addressed in this study.  
+
+<|ref|>text<|/ref|><|det|>[[72, 138, 916, 167]]<|/det|>
+(2) While the revision presents some results on learned fixed points, it remains unclear how and why the underlying chaotic dynamics are reorganized during the training process.  
+
+<|ref|>text<|/ref|><|det|>[[72, 177, 921, 231]]<|/det|>
+(3) The added explanation for why the model performs best at the edge of chaos is largely qualitative and descriptive (see Line 310). In machine learning, it is well established that artificial neural networks achieve optimal training performance near the edge of chaos [Schoenholz, et al., Deep information propagation, ICLR 2017], supported by quantitative analysis. Similarly, I believe a quantitative investigation is necessary to strengthen this argument.  
+
+<|ref|>text<|/ref|><|det|>[[73, 255, 283, 270]]<|/det|>
+(Remarks on code availability)  
+
+<|ref|>text<|/ref|><|det|>[[73, 295, 162, 308]]<|/det|>
+Reviewer #2  
+
+<|ref|>text<|/ref|><|det|>[[73, 320, 899, 349]]<|/det|>
+(Remarks to the Author) The authors have adequately addressed my concerns, and the inclusion of the spiking novel further improves the paper?  
+
+<|ref|>text<|/ref|><|det|>[[73, 359, 290, 373]]<|/det|>
+"matrixes" should be "matrices"  
+
+<|ref|>text<|/ref|><|det|>[[73, 386, 283, 400]]<|/det|>
+(Remarks on code availability)  
+
+<|ref|>text<|/ref|><|det|>[[73, 425, 145, 439]]<|/det|>
+Version 2:  
+
+<|ref|>text<|/ref|><|det|>[[73, 451, 219, 464]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[73, 476, 160, 490]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[73, 503, 238, 516]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[73, 515, 889, 556]]<|/det|>
+In the revised manuscript, the authors have included comparisons between the proposed learning rule and two other supervised learning methods, and have added some quantitative explanations concerning why learning performance is maximised at the edge of chaos. My main concerns have now been addressed in the revisions.  
+
+<|ref|>text<|/ref|><|det|>[[73, 568, 283, 581]]<|/det|>
+(Remarks on code availability)
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 45, 916, 99]]<|/det|>
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+<|ref|>text<|/ref|><|det|>[[72, 100, 797, 113]]<|/det|>
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source.  
+
+<|ref|>text<|/ref|><|det|>[[72, 113, 911, 166]]<|/det|>
+The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.  
+
+<|ref|>text<|/ref|><|det|>[[72, 166, 618, 180]]<|/det|>
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[370, 116, 629, 135]]<|/det|>
+# Response to the reviewers  
+
+<|ref|>text<|/ref|><|det|>[[336, 137, 660, 153]]<|/det|>
+Toshitake Asabuki and Claudia Clopath  
+
+<|ref|>text<|/ref|><|det|>[[140, 166, 858, 275]]<|/det|>
+We thank the reviewers for their thorough and critical evaluation of our work. We read the reviewers' comments carefully and we think we have addressed reviewers' concerns fully. Below, we explain how we addressed each of their concerns point- by- point. In particular, our additional results show that we can implement our rule in a spiking network and that the rule is robust to noise, two majors concerns of the reviewers. Thank you very much in advance for your time.  
+
+<|ref|>sub_title<|/ref|><|det|>[[140, 336, 247, 353]]<|/det|>
+## Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[140, 377, 859, 567]]<|/det|>
+Reviewer #1Reviewer Point P 1.1 —From a machine learning point of view, the tasks that can be performed by training RNNs using the predictive alignment method can all be implemented by the FORCE method and its variants, often with better performance; some of these papers are cited in the manuscript. From a biological perspective, the methods developed in this study lack the biological plausibility of training RNNs with spiking neurons. Regarding this latter point, the e- prop method (Bellec et al., Nature Communications, 2020), which is also local and leverages error signals, offers more biological plausibility. As the field stands now, a powerful new method that can effectively train recurrent spiking neural networks would represent significant progress.  
+
+<|ref|>text<|/ref|><|det|>[[140, 590, 859, 759]]<|/det|>
+Reply: Regarding the first point, while it is true that the tasks used in our study can all be performed by FORCE learning in most cases, we would like to emphasize that we have shown the proposed model performs even better than FORCE in the presence of noise (Supplementary Figure 2, shown again below). This happens because FORCE aims to make its output match the target signal as quickly as possible, which can be disrupted by strong noise. In contrast, our model learns more slowly, making the learning process less affected by noise. We have explained these results in II. 227- 231 in the revised manuscript.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[150, 125, 848, 318]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[139, 330, 858, 404]]<|/det|>
+<center>Figure R1. Robustness of predictive alignment against noise. (A) Networks with different strengths of noise were trained with the FORCE (red) and the predictive alignment (green) with the patterned target signal. Error bars stand for s.d.s over 20 independent simulations. (B) Example readout activities during the late phase of training are shown. Colors are the same as in A. The gray traces represent the target signal. </center>  
+
+<|ref|>text<|/ref|><|det|>[[139, 445, 860, 614]]<|/det|>
+To address the second point, we asked whether our learning rule can be implemented in a spiking recurrent network. We implemented a recurrent network composed of 1,000 leaky integrate- and- fire (LIF) neurons, coupled with a single linear readout unit as the output. We ran additional simulations and found that our rule can be applied in a network of spiking neurons (Fig. R2 shown below). We have include the new results as a new Supplementary Figure 8 in the revised manuscript. The details of simulation results and implementations are shown in II. 494- 500 and II.756- 783 in the revised manuscript, respectively.  
+
+<|ref|>image<|/ref|><|det|>[[150, 650, 840, 840]]<|/det|>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[140, 115, 858, 175]]<|/det|>
+Figure R2. The spiking recurrent network was trained to learn periodic target signal. (A) The blue trace represents the target signal, and the green line represents the output. (B) Activities of spiking network neurons before training are shown. (C) Activities of trained spiking network neurons are shown. The modified recurrent connections generate the sequential network activity.  
+
+<|ref|>text<|/ref|><|det|>[[139, 197, 860, 365]]<|/det|>
+Reviewer Point P 1.2 — It is unclear why the learning method works. In other words, what changes occur during the learning process that enable the RNNs to perform the intended computations? Are these changes related to the reorganization of the underlying chaotic attractor into a rich combination of fixed points or continuous attractors of the corresponding dynamical systems? I would like to see a deeper treatment of this question to open the "black box" of this process. The learning performance is best at the "edge of chaos," which is a very interesting observation, but why do such critical dynamics facilitate learning?  
+
+<|ref|>text<|/ref|><|det|>[[139, 404, 860, 638]]<|/det|>
+Reply: The reviewer's intuition is correct. Our rule suppresses the chaotic spontaneous activity and reorganize it to multiple attractors by modifying the recurrent connections. To address this, we adopted the approach of Sussillo and Barak (2013), where three output units were specifically trained to read out the corresponding fixed points of a network dynamics. The state of each output is determined by transient pulses from the corresponding input units. These pulses affect the corresponding output unit, causing it to switch or maintain a value of either \(+1\) or \(- 1\) . Once the output value is set, it remains fixed until the arrival of the next pulse from the input unit. We found that the trained network is capable of producing outputs that show transition between states in response to input (Figure R3A).  
+
+<|ref|>text<|/ref|><|det|>[[139, 661, 860, 872]]<|/det|>
+To observe how the trained network switches between memory states, we perturbed the state of the trained network with input pulses to observe transitions between two fixed points over six trials. In these trials, the strength of the input was gradually increased. When the input was weak, we found that the network activity briefly deviated from its fixed point and returned to the original fixed point when the input was removed. In striking contrast, when the input was strong enough, the network switched to a different fixed point, suggesting the existence of saddle points between the two stable fixed points, demonstrating its mechanistic role in the corresponding transition (Figure R3B). We have included these additional simulation results as Supplementary Figure 5 and explained the
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[140, 118, 857, 159]]<|/det|>
+results in II.325- 358 and implementation details in II.747- 754 in the revised manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[139, 182, 860, 288]]<|/det|>
+Furthermore, we have included an additional sentence to explain why the edge of chaos is important in the proposed model in II. 310- 312: "The model performs best at the edge of chaos because it supports richness of network dynamics, which in turn provides rich basis functions from which the output units can readout the appropriate dynamics.".  
+
+<|ref|>image<|/ref|><|det|>[[172, 332, 825, 514]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[139, 520, 860, 601]]<|/det|>
+<center>Figure R3. Learning fixed point attractors. (A) Example three inputs (black) and outputs (blue, orange, and green) are shown. (B) Low dimensional network dynamics are shown. Network dynamics perturbed with the varying strength of inputs shows that a saddle point mediates the transitions between attractors. </center>  
+
+<|ref|>text<|/ref|><|det|>[[140, 648, 857, 688]]<|/det|>
+Reviewer Point P 1.3 — What is the biological relevance of assuming the recurrent connectivity is a summation of M and G?  
+
+<|ref|>text<|/ref|><|det|>[[139, 711, 860, 858]]<|/det|>
+Reply: The reviewer raised an important point. Our plasticity rule assumes two types of recurrent connections (i.e., M and G), one is plastic and the other is static. While the biological relevance of this is still an open question, it is known experimentally that the degree of synaptic plasticity varies across different compartments of dendrites (Gordon et al., J Neurosci. 2006). Based on this experimental evidence, we can speculate that the plastic connections are synapses that project onto proximal basal dendrites and the static connections
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[140, 118, 858, 158]]<|/det|>
+project onto distal basal dendrites. We have included this point in II.631- 637 in the revised manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[140, 183, 859, 244]]<|/det|>
+Reviewer Point P 1.4 — What are the effects of certain hyperparameters, such as learning rates, on learning performance? Without thorough exploration of this aspect, it is difficult to assess whether the learning method is robust and effective.  
+
+<|ref|>text<|/ref|><|det|>[[139, 268, 860, 586]]<|/det|>
+Reply: We have investigated the robustness of the model with respect to its hyperparameters. First, we examined the effect of the learning rate on performance. We found that as the learning rate increased, so did the output error; however, if the learning rate was sufficiently small, the error remained small even when the learning rate was increased tenfold (Figure R4A). Next, we examined the initial strength of the plastic recurrent connections, where the connections were fully connected. We found that as long as the initial connection strength was not extremely large (e.g., when it was larger than that of a strong connection G), the error remained sufficiently small (Figure R4B). Finally, we examined against the connection probability of M and found that as long as the connections were not extremely sparse, the error remained sufficiently small (Figure R4C). These results suggest that the learning performance was not susceptible to certain hyperparameter choices. We have included these new results as a new Supplementary Figure 3 and explained in II. 233- 237 in the revised manuscript.  
+
+<|ref|>image<|/ref|><|det|>[[144, 623, 844, 782]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[139, 803, 859, 863]]<|/det|>
+<center>Figure R4. Robustness to hyperparameter choices. (A) The networks were trained with different levels of learning rates. (B) Same as A, but trained with different degrees of the initial strength of the plastic recurrent weights. (C) Same as A, but trained with different </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[140, 118, 857, 158]]<|/det|>
+degrees of connection probabilities of the plastic recurrent weights. Error bars represent SEs across five independent simulations.  
+
+<|ref|>sub_title<|/ref|><|det|>[[140, 177, 249, 194]]<|/det|>
+## Reviewer #2  
+
+<|ref|>text<|/ref|><|det|>[[139, 198, 860, 366]]<|/det|>
+Reviewer #2Reviewer Point P 2.1 —My primary concern relates to the apparent absence of noise in the RNN. As far as I could tell by looking at the Methods and sample code (perhaps I’m wrong), in contrast to most RNN models, there was no noise term during training or testing. When proposing a biologically plausible learning rule that relies on feedback from a chaotic RNN the absence of noise is potentially a serious issue, because training may not work at all when noise is present as the feedback signal is a moving target. It is thus necessary to include noise and parametrically vary noise to understand the influence of noise in predictive alignment.  
+
+<|ref|>text<|/ref|><|det|>[[139, 388, 860, 515]]<|/det|>
+Reply: We apologize to the reviewer for any confusion caused by our manuscript. We had already included external noise in the network during learning. In the Supplementary Figure 2 (shown again below), we have shown that the proposed model performs even better than FORCE in the presence of noise. These results indicate that our model is robust to external noise in the network. We have explained these results in II. 227- 231 in the revised manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[140, 538, 859, 600]]<|/det|>
+In the revised manuscript, we clarified this point by including noise term explicitly in the Equation 7, and mentioned "We will consider network dynamics with external drives and noise." in II. 84- 85 in the Results section to avoid confusion.  
+
+<|ref|>image<|/ref|><|det|>[[147, 629, 848, 824]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[139, 836, 860, 881]]<|/det|>
+<center>Figure R5. Robustness of predictive alignment against noise. (A) Networks with different strengths of noise were trained with the FORCE (red) and the predictive alignment (green) with the patterned target signal. Error bars stand for s.d.s over 20 independent simulations. (B) </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[140, 115, 857, 146]]<|/det|>
+Example readout activities during the late phase of training are shown. Colors are the same as in A. The gray traces represent the target signal.  
+
+<|ref|>text<|/ref|><|det|>[[139, 168, 860, 337]]<|/det|>
+Reviewer Point P 2.2 —The use of the Measure- Wait- Go task is interesting, and an important demonstration of the power of the learning rule. However, the time scale used was less than an order of magnitude longer than that of tau. Does the approach work for more biologically relevant time scales such as those used in the Jazayeri and Shadlen paper: hundreds of milliseconds? Also, this task is normally referred to as the ready- set- go task (Jazayeri and Shadlen, 2010). It would also be helpful to show a sample of the RNN activity during the task rather than just the PC's, because linear PCs can emerge from RNNs without much linear activity.  
+
+<|ref|>text<|/ref|><|det|>[[139, 360, 860, 529]]<|/det|>
+Reply: We increased the delay period to be remembered in the task to more than 100 ms. We found that although the delay period was more than 10 times the time scale of the neuron's dynamics, the model still showed the same behavior as in the previous setting. In the revised manuscript, previous Figure 5 has been replaced by the new simulation results. In a supplementary Figure 7, we have shown ten samples of the RNN activity in the revised manuscript (shown as Fig. R7 below). Furthermore, in the revised manuscript, we have changed the term "Measure- Wait- Go task" to "Ready- Set- Go task".  
+
+<|ref|>text<|/ref|><|det|>[[140, 552, 859, 614]]<|/det|>
+Learning over much longer timescales (i.e., hundreds of milliseconds) can in principle be achieved by extending network timescales through larger networks, adaptation, and short- term plasticity. We want to leave this as a future work.  
+
+<|ref|>image<|/ref|><|det|>[[144, 641, 860, 880]]<|/det|>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[137, 117, 861, 308]]<|/det|>
+Figure R6. Learning delay- matching task. (A) In each trial, two input units (blue and orange) send pulses to the network with a random delay \(T_{\text{delay}}\) between pulses. The target network output (green) should be a pulse delayed by \(T_{\text{delay}}\) relative to the second input pulse. (B) Networks were trained on a set of samples (colored squares) and tested for generalization to novel inputs, including interpolation within the training range (shaded region) as well as extrapolation beyond the training range. (C) The network after training failed to extrapolate beyond the training range. (D) Principal component analysis (PCA) of the trained network revealed that as the delay \(T_{\text{delay}}\) increased, the network states corresponding to the output peak shifted linearly along a manifold in state space.  
+
+<|ref|>image<|/ref|><|det|>[[160, 344, 833, 645]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[137, 665, 861, 852]]<|/det|>
+<center>Figure R7. Neural activities for performing Ready-Set-Go task. (A) In each trial, two input units (blue and orange) send pulses to the network with a random delay \(T_{\text{delay}}\) between pulses. The target network output (green) should be a pulse delayed by \(T_{\text{delay}}\) relative to the second input pulse. (B) Networks were trained on a set of samples (colored squares) and tested for generalization to novel inputs, including interpolation within the training range (shaded region) as well as extrapolation beyond the training range. (C) The network after training failed to extrapolate beyond the training range. (D) Principal component analysis (PCA) of the trained network revealed that as the delay \(T_{\text{delay}}\) increased, the network states corresponding to the output peak shifted linearly along a manifold in state space. </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[140, 118, 857, 159]]<|/det|>
+**Reviewer Point P 2.3** —Replay task. It was not clear what time \(= 0\) was for the movie generation? Does the produced movie run in a limit cycle?  
+
+<|ref|>text<|/ref|><|det|>[[139, 182, 859, 309]]<|/det|>
+Reply: We thank the reviewer for pointing this out. At time zero, we set the network activity to the predefined initial states. The same initial state was used in both the learning and testing phases. We have included these explanations in II. 515- 518 in the revised manuscript. While we reset the initial state over trials in the results shown in Figure 6, in principle the movie can run in a limit cycle and restart when the movie ends.  
+
+<|ref|>text<|/ref|><|det|>[[139, 332, 859, 437]]<|/det|>
+Reviewer Point P 2.4 —Pg. 2. To support the statement "Cortical circuits often exhibit chaotic spontaneous activity" a number of papers are cited, but they all seem to be models, not experimental papers (and I don't think Rajan et al, 2016 even address chaos). RNNs are not cortical circuits so these references do not seem appropriate. London, ..., Latham, 2010 provides some evidence for chaos in cortical circuits.  
+
+<|ref|>text<|/ref|><|det|>[[139, 459, 860, 629]]<|/det|>
+Reply: The reviewer is correct, we cited only computational papers in our previous manuscript. We have included the suggested reference and changed the sentence as: "Cortical circuits often exhibit chaotic spontaneous activity (London et al., 2010), and RNNs can generate such dynamics through feedback loops (van Vreeswijk and Sompolinsky, 1996; Amit and Brunel, 1997; Brunel, 2000; Toyoizumi and Abbott, 2011; Rajan et al., 2016).", in II. 42- 45 in the revised manuscript. Here, we have kept the Rajan et al. paper since they used RNNs showing chaotic behavior before training.  
+
+<|ref|>text<|/ref|><|det|>[[140, 653, 857, 693]]<|/det|>
+Reviewer Points P 2.5 —The presentation and discussion omitted many relevant studies that laid the foundation for the current work.  
+
+<|ref|>text<|/ref|><|det|>[[140, 716, 857, 757]]<|/det|>
+Reply: We hope we have included the appropriate references in the revised manuscript. If there are any missing, could the reviewer let us know?  
+
+<|ref|>text<|/ref|><|det|>[[140, 780, 857, 821]]<|/det|>
+Reviewer Points P 2.6 —Jaeger and Haas 2004, essentially published what later came to known as FORCE in their "online" learning rule.  
+
+<|ref|>text<|/ref|><|det|>[[140, 844, 728, 864]]<|/det|>
+Reply: We have included the reference in the revised manuscript.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[139, 140, 860, 245]]<|/det|>
+Reviewer Points P 2.7 — The paper has commonalities with Laje and Buonomano, which also took the approach of taming chaos by changing the recurrent weights and using high dimensional structure of a local stable trajectories as a reservoir. That work was not cited in the text. And Buonomano and Merzenich 1995, was the first description of what later came to be called a reservoir network.  
+
+<|ref|>text<|/ref|><|det|>[[140, 268, 737, 287]]<|/det|>
+Reply: We have included the references in the revised manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[140, 311, 858, 351]]<|/det|>
+Reviewer Points P 2.8 — Supplemental Figure 4. A schematic of the network would be helpful.  
+
+<|ref|>text<|/ref|><|det|>[[140, 375, 858, 415]]<|/det|>
+Reply: We have included new schematics to help readers. We thank the reviewer for pointing this out.  
+
+<|ref|>text<|/ref|><|det|>[[140, 439, 858, 480]]<|/det|>
+Reviewer Points P 2.9 — Pg. 14 “while failed AT extrapolation”; Pg. 15 “Here, we train recurrent neural” ... typo.  
+
+<|ref|>text<|/ref|><|det|>[[139, 503, 860, 522]]<|/det|>
+Reply: We have corrected the typo. We thank the reviewer for their careful review.  
+
+<|ref|>text<|/ref|><|det|>[[140, 546, 858, 586]]<|/det|>
+Reviewer Points P 2.10 — It is best to refer to the units and weights as such, rather than neurons and synapses.  
+
+<|ref|>text<|/ref|><|det|>[[140, 610, 858, 650]]<|/det|>
+Reply: We agree with the reviewer to change our terminology. We have changed the terminology in the revised manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[140, 674, 858, 714]]<|/det|>
+We look forward to hearing from you soon. Thank you very much for your kind consideration in advance.  
+
+<|ref|>text<|/ref|><|det|>[[140, 739, 285, 757]]<|/det|>
+Sincerely yours,  
+
+<|ref|>image<|/ref|><|det|>[[140, 783, 284, 833]]<|/det|>  
+
+<|ref|>text<|/ref|><|det|>[[139, 846, 496, 885]]<|/det|>
+Toshitake Asabuki  RIKEN Center for Brain Science, Japan
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[139, 120, 383, 136]]<|/det|>
+toshitake.asabuki@riken.jp 
+
+<|ref|>image<|/ref|><|det|>[[142, 168, 245, 215]]<|/det|>
+
+
+<|ref|>text<|/ref|><|det|>[[139, 227, 761, 285]]<|/det|>
+Prof. Claudia Clopath  Department of Bioengineering, Imperial College London, London, UK.  c.clopath@ imperial.ac.uk; +44 (0)20 7594 1435
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[370, 116, 630, 135]]<|/det|>
+# Response to the reviewers  
+
+<|ref|>text<|/ref|><|det|>[[336, 137, 660, 154]]<|/det|>
+Toshitake Asabuki and Claudia Clopath  
+
+<|ref|>text<|/ref|><|det|>[[140, 166, 858, 239]]<|/det|>
+We thank the reviewers for their thorough and critical evaluation of our work. We read the reviewers' comments carefully and we think we have addressed reviewers' concerns fully. Below, we explain how we addressed each of their concerns point- by- point. Thank you very much in advance for your time.  
+
+<|ref|>sub_title<|/ref|><|det|>[[140, 280, 247, 297]]<|/det|>
+## Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[140, 322, 858, 404]]<|/det|>
+Reviewer #1Reviewer Point P 1.1 —I appreciate the authors' detailed responses and the corresponding revisions. However, my concerns regarding the training of RNNs with spiking neurons and the changes that occur during the learning process to enable these networks to perform the intended computations remain unclear. Specifically:  
+
+<|ref|>text<|/ref|><|det|>[[139, 428, 859, 554]]<|/det|>
+(1) For the proposed learning rule to be effectively evaluated in a spiking neural network, it is essential to compare its performance with existing methods for training SNNs. Additionally, applying the proposed method to balanced spiking neural networks with both excitatory and inhibitory neurons would be more appropriate for two reasons: (i) it aligns better with biological plausibility, and (ii) balanced networks are known to exhibit chaotic dynamics, which are more consistent with the problem addressed in this study.  
+
+<|ref|>text<|/ref|><|det|>[[140, 577, 858, 660]]<|/det|>
+Reply: That's a great suggestion. We have compared the performance with two types of stae- of- the- art spiking recurrent network for supervised learning: FORCE and e- prop. We have included the new results as Supplementary Fig. 10C and mentioned in II.524- 525 in the revised manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[140, 684, 858, 766]]<|/det|>
+Furthermore, we showed that our proposed rule enables learning even in a recurrent spiking neural network composed of two populations (i.e., excitatory and inhibitory). We have included the new result as Supplementary Figure 11 and mentioned in II.525- 528 in the revised manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[140, 808, 858, 870]]<|/det|>
+Reviewer Point P 1.2 —(2) While the revision presents some results on learned fixed points, it remains unclear how and why the underlying chaotic dynamics are reorganized during the training process.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[139, 156, 862, 345]]<|/det|>
+Reply: The revierer raised an important point. While we have shown that aligning the predictive recurrent and chaotic dynamics is important for stable learning in Figure 3, we did not explain how it changes the recurrent dynamics. To see this further and understand how the alignment enables stable learning, we monitored the network's Lyapunov exponent during training. We found that in the aligned case, the network's Lyapunov exponent was shifted further toward the negative side compared to the control case, indicating more effective suppression of chaos. We have included the new simulation result as Supplementary Figure 4 and explained in II. 270- 272 in the revised manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[139, 386, 862, 512]]<|/det|>
+Reviewer Point P 1.3 —(3) The added explanation for why the model performs best at the edge of chaos is largely qualitative and descriptive (see Line 310). In machine learning, it is well established that artificial neural networks achieve optimal training performance near the edge of chaos [Schoenholz, et al., Deep information propagation, ICLR 2017], supported by quantitative analysis. Similarly, I believe a quantitative investigation is necessary to strengthen this argument.  
+
+<|ref|>text<|/ref|><|det|>[[139, 535, 860, 727]]<|/det|>
+Reply: To investigate why learning performance is maximized at the edge of chaos, we analyzed the eigenvalue distribution of the correlation matrix of the network activity. Optimal learning is thought to require a balance between representational diversity and dimensionality. High diversity allows the network to encode rich information, while low dimensionality ensures compact representations. We quantified the diversity of neural representations using the entropy of the eigenvalues of the correlation matrix \((H_{\lambda})\) , and measured the balance between information spread and localization using the square root of the participation ratio \((\sqrt{PR})\) . We computed the ratio of these two measures,  
+
+<|ref|>equation<|/ref|><|det|>[[420, 730, 577, 768]]<|/det|>
+\[\mathrm{Efficiency} = \frac{H_{\lambda}}{\sqrt{PR}}\]  
+
+<|ref|>text<|/ref|><|det|>[[139, 787, 861, 892]]<|/det|>
+and found that this balance index was maximized specifically at the edge of chaos. In the subcritical regime, the participation ratio was low, indicating that activity was concentrated in a low- dimensional subspace, reducing the network's ability to learn diverse representations. In contrast, in the chaotic regime, the participation ratio was high, but the eigenvalue entropy decreased, suggesting
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[139, 118, 860, 266]]<|/det|>
+excessive dispersion of information, leading to loss of meaningful structure. At the edge of chaos, both measures were optimally balanced, enabling the network to achieve the highest learning performance. These findings suggest that learning is most efficient when the network exhibits rich yet structured dynamics, balancing representational diversity and stability. We have included the new simulation results as Supplementary Figure 6 in the revised manuscript and explained in II. 327- 347 in the revised manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[140, 327, 480, 345]]<|/det|>
+Reviewer #2 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[140, 370, 858, 410]]<|/det|>
+The authors have adequately addressed my concerns, and the inclusion of the spiking novel further improves the paper?  
+
+<|ref|>text<|/ref|><|det|>[[140, 434, 422, 452]]<|/det|>
+"matrixes" should be "matrices"  
+
+<|ref|>text<|/ref|><|det|>[[140, 477, 824, 496]]<|/det|>
+Reply: We thank the reviewer for the careful review. We have fixed the typo.  
+
+<|ref|>text<|/ref|><|det|>[[140, 520, 858, 559]]<|/det|>
+We look forward to hearing from you soon. Thank you very much for your kind consideration in advance.  
+
+<|ref|>text<|/ref|><|det|>[[140, 584, 285, 602]]<|/det|>
+Sincerely yours,  
+
+<|ref|>image<|/ref|><|det|>[[147, 621, 288, 670]]<|/det|>  
+
+<|ref|>text<|/ref|><|det|>[[139, 672, 496, 732]]<|/det|>
+Toshitake Asabuki  RIKEN Center for Brain Science, Japan  toshitake.asabuki@riken.jp  
+
+<|ref|>image<|/ref|><|det|>[[140, 761, 247, 810]]<|/det|>  
+
+<|ref|>text<|/ref|><|det|>[[139, 818, 763, 880]]<|/det|>
+Prof. Claudia Clopath  Department of Bioengineering, Imperial College London, London, UK.  c.clopath@ imperial.ac.uk; +44 (0)20 7594 1435
+
+<--- Page Split --->
+<|ref|>title<|/ref|><|det|>[[333, 116, 661, 155]]<|/det|>
+# Response to the reviewersToshitake Asabuki and Claudia Clopath  
+
+<|ref|>sub_title<|/ref|><|det|>[[140, 180, 247, 197]]<|/det|>
+## Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[140, 222, 859, 326]]<|/det|>
+Reviewer Point P 1.1 —In the revised manuscript, the authors have included comparisons between the proposed learning rule and two other supervised learning methods, and have added some quantitative explanations concerning why learning performance is maximised at the edge of chaos. My main concerns have now been addressed in the revisions.  
+
+<|ref|>text<|/ref|><|det|>[[140, 350, 635, 368]]<|/det|>
+Reply: We thank the reviewer for their fruitful feedback.  
+
+<|ref|>text<|/ref|><|det|>[[140, 412, 858, 453]]<|/det|>
+We look forward to hearing from you soon. Thank you very much for your kind consideration in advance.  
+
+<|ref|>text<|/ref|><|det|>[[140, 478, 285, 496]]<|/det|>
+Sincerely yours,  
+
+<|ref|>image<|/ref|><|det|>[[145, 518, 286, 560]]<|/det|>  
+
+<|ref|>text<|/ref|><|det|>[[139, 564, 496, 625]]<|/det|>
+Toshitake Asabuki RIKEN Center for Brain Science, Japan toshitake.asabuki@riken.jp  
+
+<|ref|>image<|/ref|><|det|>[[140, 655, 247, 704]]<|/det|>  
+
+<|ref|>text<|/ref|><|det|>[[139, 713, 763, 775]]<|/det|>
+Prof. Claudia Clopath  Department of Bioengineering, Imperial College London, London, UK.  c.clopath@ imperial.ac.uk; +44 (0)20 7594 1435
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__6e3070f47b530d68e5f8d912919727215aa0f9efcd341ef546885015d497a4fc/images_list.json

@@ -0,0 +1 @@
+[]

peer_reviews/supplementary_0_Transparent Peer Review file__6e3070f47b530d68e5f8d912919727215aa0f9efcd341ef546885015d497a4fc/supplementary_0_Transparent Peer Review file__6e3070f47b530d68e5f8d912919727215aa0f9efcd341ef546885015d497a4fc.mmd

@@ -0,0 +1,503 @@
+
+# nature portfolio  
+
+# Peer Review File  
+
+# Pesticides have negative effects on non-target organisms  
+
+Corresponding Author: Professor Nian- Feng Wan  
+
+This manuscript has been previously reviewed at another journal. This document only contains information relating to versions considered at Nature Communications. Mentions of the other journal have been redacted.  
+
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+Version 1:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author)  
+
+The authors have put significant effort into revising their work, which is substantially improved now. However, there are some remaining concerns.  
+
+1. The GitHub report now contains the code but not the data to re-run the code. Also, there are multiple R files and no description how to use them to recreate study results – this should be described in detail in the README file. Meta-data should be also provided for all used variables.  
+2. Poor coverage of grey and non-English literature should be acknowledged as limitations in the main manuscript.  
+3. Trim-and-fill is a publication bias test not sensitivity analysis.  
+4. The issue of percentage and proportion data is not in the normality of the residuals, but in the fact that they are bounded between 0 and 1 (100). As such they should be arcsin-transformed to make them unbounded. All transformations can be documented in R code, so there is no issue of transparency.  
+5. The issue with folded normal distribution when using absolute values has not been resolved. See Nakagawa, S. & Lagisz, M. (2016) Visualizing unbiased and biased unweighted meta-analyses. 334. Journal of Evolutionary Biology, 29, 1914-1916, for a visual clarification of the issue.  
+6. A synthetic phylogenetic tree can be retrieved from Open Tree of Life, which is accessible via R package rotl. As, such phylogenetic analyses can and should be conducted beyond plants.  
+7. Publication year should be included in the analyses (meta-regression) – it is not confounded with study identity, and it is potentially important and interesting moderator of the effect.  
+8. Funding and conflict of interest statements can be easily extracted from the included papers, even if they are not consistently included. Then, presence of potential links with the relevant industry should be analysed in a meta-regression. This is a too important issue to dismiss.  
+
+## Reviewer #3  
+
+(Remarks to the Author)  
+
+The current ms derives from a previous version submitted to [Redacted] that has been commented by three reviewers. I now have checked for the implementation of the comments of Reviewers 2 and 3 in the revised version.  
+
+To my opinion, almost all comments of Reviewer 2 have been adequately addressed, except for  
+
+(1) Comment: 216-217 were vertebrates and invertebrates equally affected, or one more than the other? Does that depend on the intended target of an pesticide, e.g. distinguishing between pesticides aimed at invertebrates or otherwise?
+
+<--- Page Split --->
+
+Response: We have removed the sentence due to space limited, but from the statistical values in Supplementary Table 4, we can see that vertebrates and invertebrates were equally affected (see Supplementary Table 4).  
+
+and  
+
+(2) Comment: 223 'affected different taxonomic groups' how?  
+
+Response: We have removed the sentence due to space limited, but from the statistical values in Supplementary Tables 32- 36, we can see that "insecticides, fungicides and herbicides affected different taxonomic groups with decreased growth, reproduction or behavior and with perturbed biomarkers" (see Supplementary Tables 32-36).  
+
+I feel that both informations are important and should be kept within the main body of text rather than being implied to be taken from the supplementary material.  
+
+All comments of Reviewer 3 have been perfectly addressed in this version of the ms.  
+
+In the Competing Interests paragraph, "B.W." should be replaced by "B.A.W.".  
+
+Version 2:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author)  
+
+Reviewer comments:  
+
+The authors addressed my concerns partially. However some new concerns emerged in the process.  
+
+Response to the referee comments  
+
+Replies to Reviewer #1:  
+
+Comment: The authors have put significant effort into revising their work, which is substantially improved now.  
+
+Response: We thank the reviewer for this positive evaluation (see this revision).  
+
+Comment: However, there are some remaining concerns. 1. The GitHub report now contains the code but not the data to re- run the code.  
+
+Response: We have uploaded both code and data to GitHub (https://github.com/Liwan- Fu/Impact- of- pesticides).  
+
+New reviewer comment - Round3: Thank you, for providing a link to the files on GitHub. I note that now the data and code are also provided on Zenodo and the contents of these two repositories do not match. Specifically, some code files that are archived on Zenodo are absent from GitHub version. Please remove any special (Chinese) characters from your R code.  
+
+Comment: Also, there are multiple R files and no description how to use them to recreate study results- this should be described in detail in the README file. Meta- data should be also provided for all used variables.  
+
+Response: We have provided a README file for all used variables and uploaded the file to GitHub (https://github.com/Liwan- Fu/Impact- of- pesticides).  
+
+New reviewer comment - Round3: Thank you for providing a README file. Unfortunately, the meta- data provided applies to only one of the data files, "Meta combined data.csv" (there seems to be 16 data files in total). The meta- data for this one key data table used to run the key models is still too rudimentary to allow cross- checking of the data or replication of the data extraction process. For example, the variable described as "Control (value): the value in control group;" does not explain which value should extracted be there, apart that it should relate to the control group. Another, example - "Control- n: the number in control group;", which is supposed to be the sample size, it also needs explaining the units of counting sample size, as these can be different for different taxa and experiment types - for animals it might be number of individuals, for plants could be individuals or plots, for microorganisms it could be a colony in a tube/plate, or number of natural plots, etc. One column in the data table called "LnR" is not described - it seems to be an effect size that was not be used in the analyses, as the main effect size used was SMD. However, I cannot find the code the authors used for calculating SMD. Also, I cannot find the code for Figure 1 (the map), and instead there is code for Figure 1b, a histogram which is not in the main manuscript. Further, the study codes in the main data file "Meta combined data.csv" do not match the study numbers in the table characterizing included studies (with their references) - "447069_2_data_set_8320286_s47vd1.xls". Please include full study references as a column in the data file "Meta combined data.csv", so that each data point can be linked to its original source. Overall, lack of consistent and comprehensive documentation is still a major limitation to transparency and reproducibility of this meta-analysis, which should be addressed before this work is accepted.  
+
+Comment: 2. Poor coverage of grey and non- English literature should be acknowledged as limitations in the main manuscript.  
+
+Response: We have now made it explicit that our study did not include grey or non- English literature (lines 420- 421). New reviewer comment - Round3: I can see it in the methods section now. I suggest adding this also to the header of Figure 1 (the global) map, because it is highly relevant to the pattern presented there.
+
+<--- Page Split --->
+
+Comment: 3. Trim- and- fill is a publication bias test not sensitivity analysis. Response: We have removed text referring to this as a sensitivity analysis (see lines 631- 632 in the main text; line 326 in the Supplementary Results). New reviewer comment - Round3: thank you.  
+
+Comment: 4. The issue of percentage and proportion data is not in the normality of the residuals, but in the fact that they are bounded between 0 and 1 (100). As such they should be arcsin- transformed to make them unbounded. All transformations can be documented in R code, so there is no issue of transparency.  
+
+Response: We have consulted with our institutional statistician (Dr Pete Henrys - CStat, Royal Statistical Society Chartered statistician, CSci, The Science Council Chartered Scientist) who states that a transformation is there to ensure model assumptions are met. Transformation purely on the basis that data is of a particular type is not appropriate, and rather transformation should only be applied on a case by case basis to address underlying issues that may affect model assumptions. He confirms that focusing on model assumption checks (as we have done) including assessing residuals represents the critical stage in the process for assessing the need for underlying processing of input data. We strongly argue that percentage and proportion data in the context of this analysis does not warrant to be arcsin- transformed without a reason relating to model distributional fit. In the context of this analysis this is not required.  
+
+New c reviewer comment - Round3: it is great that you consulted the statistician and conducted the model assumptions checks. Please provide the results of the model assumptions test in the supplementary materials. Conducting sensitivity analysis for the main model that show that applying the recommended transformation does not change the results should be also presented in the supplementary materials to support your claims.  
+
+Comment: 5. The issue with folded normal distribution when using absolute values has not been resolved. See Nakagawa, S. & Lagisz, M. (2016) Visualizing unbiased and biased unweighted meta- analyses. 334. Journal of Evolutionary Biology, 29, 1914- 1916, for a visual clarification of the issue.  
+
+Response: This relates to our standardizing of biomarkers (e.g. protein regulation, gene expression) which may respond positively or negatively to pesticides. Our approach has been to use absolute values and focus on a deviation from zero in the meta- analysis, essentially quantifying departure from the norm. Again, we emphasize that checks of model assumptions suggest that the current use of a normal distribution was robust. From a practical perspective it is not possible within our modeling framework to specify a folded distribution (which we do not think is required). However, we would be prepared to simply no longer use the absolute values thus negating the suggested need for a folded distribution. We do feel that this would potentially lose important information about a general impact of pesticides on either up- or down- regulation of biomarkers - which we think is hugely significant from the context of the impacts of pesticides on non- obvious species level effects. The author team has discussed this issue extensively and we still believe that taking absolute values biologically makes more sense here.  
+
+New reviewer comment - Round3: Please provide the results of the model assumptions test in the supplementary materials. Conducting sensitivity analysis for the key model that show that applying the recommended transformation does not change the results for biomarkers data subset should be also presented in the supplementary materials to support your claim. Also, please flip the sign on the absolute values to make it easier to compare magnitude of estimated effects in all forest plots.  
+
+Comment: 6. A synthetic phylogenetic tree can be retrieved from Open Tree of Life, which is accessible via R package rotl. As, such phylogenetic analyses can and should be conducted beyond plants.  
+
+Response: We have used rotl, a freely available package for R designed to takes advantage of the Open Tree of Life's Application Programming Interfaces (APIs) to access subtrees from the synthetic Open Tree, as well as the published source trees that contribute to the synthesis to generate comprehensive phylogenies for animals and microorganisms (see lines 296- 321 in the Supplementary Methods). In this revision, we have now conducted all the phylogenetic analyses for plants, animals and microorganisms (see detailed results in Supplementary Tables 1.3- 1.5, 2.3- 2.5, 45.3- 45.5 and 46.3- 46.5). We still note that an overall agreed- upon phylogeny for the tree of life will likely never arrive, but we did our best to address the referee's comment.  
+
+New reviewer comment - Round3: Thank you and I agree regarding the conditionality of any phylogeny used - but we will never arrive there.  
+
+Comment: 7. Publication year should be included in the analyses (meta- regression)- it is not confounded with study identity, and it is potentially important and interesting moderator of the effect.  
+
+Response: We have clarified this now as "Publication year: a continuous metric according to the year when the articles were published" (lines 528- 529 in the main text). We have included publication year in the meta- regression model (see detailed results in Supplementary Tables 1.1- 1.5, 2.1- 2.5, 45.1- 45.5 and 46.1- 46.5; Extended Data Figs. 2- 7 and other associated supplementary tables).  
+
+New reviewer comment - Round3: thank you.  
+
+Comment: 8. Funding and conflict of interest statements can be easily extracted from the included papers, even if they are not consistently included. Then, presence of potential links with the relevant industry should be analysed in a meta- regression. This is a too important issue to dismiss.  
+
+Response: We now extracted conflict- of- interest statements where possible. Among the 1705 papers, we could confirm that 1,411 ones did not have conflict of interest while 25 self- identified as having a conflict of interest. The remainder made no statement about conflicts of interest (lines 244- 248 in the main text; Supplementary Table 58). In the meta- regression analysis, we consider "Conflict of interest status" as one binary variable (i.e., "1" denotes that a certain article has conflict of interest and "0" represents no conflict) (lines 526- 528 in the main text). Detailed results for "Conflict of interest status" were presented in Supplementary Tables 1.1- 1.5, 2.1- 2.5, 45.1- 45.5, 46.1- 46.5 and 58, Extended Data Figs. 2- 7 and other associated supplementary tables).
+
+<--- Page Split --->
+
+New comment - Round3: thank you for doing this.  
+
+New comment - Round3:  
+
+By looking at the raw data, "Meta combined data.csv", I can see "Life expectancy" and "Longevity" are in "insecticide- animal reproduction" category. It seems like all (and there are many) measures related to survival are classified as "insecticide- animal reproduction" (similar issue is also present for plants and microorganisms in the data set). Please provide justification for this choice. Also animal feeding rates could represent both growth and behavior - having them solely as a measure of growth may require additional sensitivity analyses and clarification in the methods section.  
+
+New reviewer comment - Round3:  
+
+In the "Meta combined data.csv", there are some clear mistakes too, e.g. the column for describing measurements is named "Animal growth indicator", and sometimes is empty, "number of eggs/female" is classified as "insecticide- animal behavior", this raises concerns for the consistency of data extractions and the quality control procedures in this work.  
+
+New reviewer comment - Round3:  
+
+Lines 463- 466: "When a study included different levels of pesticide application rates, measurements for the control groups without pesticides versus different pesticide application rates were considered as independent paired observations" - This is a very bold assumption given "26,096 estimates of pesticide effects reported from 1,705". Now looking at the actual data, I can see that although sometimes multiple pesticides or species were used in a single study, in most cases multiple doses were used, which means data non- independence related to the repeated use of the same control group to compare it with different dose levels is a serious issue. Specifically, by not controlling for repeated use of the same individuals, the precision of the overall estimate is very likely inflated. One very simple way is to divide sample size of the control group by the number of comparisons it is used in, thus avoiding "double- counting" of the same replicated units.  
+
+New reviewer comment - Round3:  
+
+In the main text, please acknowledge study limitations due to sample sizes in the included experiments (from the raw data I can see that the median sample size is likely 3) - having a histogram of the distribution of sample sizes by type of organism and/or treatment would be very useful too.  
+
+Version 3:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author)  
+
+Reviewer comments:  
+
+The authors addressed some of my concerns but there are still outstanding issues related to transparency and reproducibility of this work, as well as some methodological concerns.  
+
+My previous comments and authors responses are quoted below when they are linked to new issues identified. My new comments are provided as "New reviewer comment - Round4:  
+
+"New reviewer comment - Round3: Thank you, for providing a link to the files on GitHub. I note that now the data and code are also provided on Zenodo and the contents of these two repositories do not match. Specifically, some code files that are archived on Zenodo are absent from GitHub version. Please remove any special (Chinese) characters from your R code. Response: We thank the referee for pointing us at this. To avoid confusion and duplication we have now uploaded all raw data and code to Zenodo and removed them from GitHub (see this revision). We have checked that all relevant files are present. We have now also removed all Chinese characters as suggested. All data used in this analysis are available on Zenodo  
+
+(https://zenodo.org/records/10495263). All code for this analysis is available on both Zenodo  
+
+(https://zenodo.org/records/10509420).  
+
+New reviewer comment - Round4: Thank you for providing your data and code on Zenodo. It seems like putting the contents of these two repositories together still does not allow reproducing all the analyses and they are really hard to peer- review. There are three reasons: first the code is very poorly annotated, and thus it is hard to figure out what different sections of code are supposed to produce. Second, it seems like pieces of code are missing, e.g. file "Figures 2- 3. R" starts from loading the file 'Meta combined data.csv' and then immediately runs a meta analytic model on the data from it, however the file does not contain effect sizes and there is no line of code that calculates effect sizes before running the model. Third, the code refers to loading many additional data files, that look like pre- processed data, but they are not included in the Zenodo repository, e.g. "Supplementary Tables 1(1- 3). R" calls "Old_New_pesticide <-  
+
+read_excel(paste("/lustre/huyueqing/ssy/Wan2023/DataPre/", "Supplementary Data 2- Classification of old and new pesticides- 2023- 0825. xls", sep = ""), sheet = 1)", but there is no such file or no explanation where it originates from. Overall, I think all the analytic pipeline should be made fully transparent reproducible and then subjected to independent peer- review.  
+
+"New reviewer comment - Round3: Thank you for providing a README file. Unfortunately, the meta- data provided applies to only one of the data files, "Meta combined data.csv" (there seems to be 16 data files in total). The meta- data for this one key data table used to run the key models is still too rudimentary to allow cross- checking of the data or replication of the data extraction process. For example, the variable described as "Control (value): the value in control group;" does not explain
+
+<--- Page Split --->
+
+which value should extracted be there, apart that it should relate to the control group. Another, example - "Control- n: the number in control group;", which is supposed to be the sample size, it also needs explaining the units of counting sample size, as these can be different for different taxa and experiment types - for animals it might be number of individuals, for plants could be individuals or plots, for microorganisms it could be a colony in a tube/plate, or number of natural plots, etc. Response: We have now revised the README file to improve clarity. We hope this addresses the problems."  
+
+New reviewer comment - Round 4: Thank you. The README file only describes data in the main data file, but not in any of the additional data files and does not provide any information about the project in general and how the code and files should be used to reproduce the analyses. Please provide more comprehensive description of all your files, starting from overview and then providing meta- data for all data that is needed to reproduce your work. Consider using Quarto with R to produce computationally reproducible documents that contain the R code, verbal descriptions, and outputs.  
+
+"Comment: One column in the data table called "LnR" is not described - it seems to be an effect size that was not be used in the analyses, as the main effect size used was SMD. However, I cannot find the code the authors used for calculating SMD.  
+
+Response: The column in the data table called "LnR" had not been used in the analyses, and we have now deleted it. The code used for calculating SMD has been added in the README file." New reviewer comment - Round 4: The code used for calculating SMD should not be placed in the README file, but in the code where it is needed to calculate effect sizes to reproduce the results.  
+
+"Comment: Further, the study codes in the main data file "Meta combined data.csv" do not match the study numbers in the table characterizing included studies (with their references) - "447069_2_data_set_8320286_s47vd1.xls". Please include full study references as a column in the data file "Meta combined data.csv", so that each data point can be linked to its original source. Overall, lack of consistent and comprehensive documentation is still a major limitation to transparency and reproducibility of this meta-analysis, which should be addressed before this work is accepted. Response: We thank the referee for pointing us at this. We have now made sure that all study codes match up. We have listed all the 1705 study references in the file "Meta combined data" with a new sheet (Full study references). The paper code order in "Meta combined data" is based on the time order that we extracted data, and the paper order in Supplementary Data 1 is based on the authors' name in alphabetical order (see Meta combined data). In addition, we have added another data file "ExtendData" for other supplementary files (e.g., for Supplementary Tables 59- 61, Supplementary Figures 24- 27)." New reviewer comment - Round 4: Thank you for adding the study references next to study codes in the data file. I was now able to compare the extracted data with the information reported in the original papers. I randomly selected 4 papers I have full- text access to and found the following issues:  
+
+- Paper 16. (Colin et al. 2004. A Method to Quantify and Analyze the Foraging Activity of Honey Bees: Relevance to the Sublethal Effects Induced by Systemic Insecticides): rows 16-21: the concentration of Fipronil is wrong - the study reports using 2 micrograms per kilogram while the extracted data is 6 g/kg. For Imidacloprid the unit is also wrong - it should be microgram per kilogram not gram per kilogram. For the outcome "number of daily attendance of active bees-nuclei" which was extracted for nuclei A, B, C, it is and somehow from control nuclei, but I cannot find the matching exact numbers in the text, and it is not possible to extract them from the figures in the text. Given the complexity of the experimental design, raw data would need to be used to calculate such values. There are no comments in the data file how the numbers were obtained, and thus they are not reproducible.  
+
+- Paper 1628. (Zhao X, Li YL, Zhang LX, Dorna H, 2003. Effects of priming and fungicide treatment on germination of china aster (Callistephus chinensis L.) seeds. Seed Science & Technology, 32: 451-457.): rows 1943-1944: it looks like only germination capacity data from Sample 1 and only from 30C temperature treatment were extracted and the standard deviations were manually imputed by assuming they are 10% of the measured value. This is only a small fraction of available data in this paper (multiple control and exposure groups, all samples should be extracted and used for calculating mean and SD) and is incorrectly extracted.  
+
+- Paper 1008. (Gabaston J, Khawand TE, Wafo-Teguo P, Decendit A, Richard T, Merillon JM, Pavel A, 2018. Stilbenes from grapevine root: a promising natural insecticide against Leptinotarsa decimlineata. Journal of Pest Science, 91: 897-906.) rows 5808-5809: looks like data was extracted from Table 6 for 3rd day of exposure only and two lower doses out of 3 used (this is not reported on the data extraction sheet). The mean mortality % was converted to survival %, but the reported standard deviation values were somehow replaced with 10% of the men value. The sample size is claimed to be 4, but I cannot find such information in the article, and it seems improbable given the type of organism, and reported values.  
+
+Overall, this small exercise clearly shows that 3/3 papers had some issues with data extractions. This indicates a very high rate of errors in the whole data set. The data lacks information on the exact sources of the extracted values for the outcomes at least (table/figure number, text page, raw data, etc.), any extractions decisions made, calculations, imputations, etc. - this makes assessing trustworthiness of the data set very hard and makes data extractions irreproducible. Because of this I request that all extracted data is annotated in detail. Further it has to be cross-checked by a researcher who was not involved in original extractions, re- extracted where issues are found and clearly described to make it fully transparent and trustworthy.  
+
+New reviewer comment - Round 4: When checking the extracted data, I noticed that many of the extracted standard deviation values are exactly 10% of the mean value. This cannot be coincidence and the manuscript does not mention doing manual imputations for missing standard deviation values (or any other missing values). Please explain how these values originated and the detailed procedures used to deal with any missing data.
+
+<--- Page Split --->
+
+"New reviewer comment - Round 3: it is great that you consulted the statistician and conducted the model assumptions checks. Please provide the results of the model assumptions test in the supplementary materials. Conducting sensitivity analysis for the main model that show that applying the recommended transformation does not change the results should be also presented in the supplementary materials to support your claims. Response: As suggested, we now include a sensitivity analysis where we apply arcsine transformation for percentage and proportion data (as you suggest) and compare the results to our approach (no transformation or proportion/percentage data) (see Supplementary Table 59). This shows no qualitative difference in the results. We also provide a supplementary figure in the supporting information for the model checks of the original analysis (see Supplementary Fig. 24). Original comment for context: 5. The issue with folded normal distribution when using absolute values has not been resolved. See Nakagawa, S. & Lagisz, M. (2016) Visualizing unbiased and biased unweighted meta-analyses. 334. Journal of Evolutionary Biology, 29, 1914- 1916, for a visual clarification of the issue."  
+
+New reviewer comment - Round 4:  
+
+The code for reproducing this should be in the file "Supplementary Tables 59- 61. R", and as such cannot be rerun. In the provided code I cannot find any line that would execute arcsine transformation for proportion data. And the code starts from loading a mysterious data file that is not archived on Zenodo, so can't tell what is there nor how it got there: AllData <- read.table(paste(DataPre_path,"SuppTab59Data.txt",sep=""),header = T). This seems to be ongoing issues across the provided code which is compounded by almost complete lack of code comments and descriptions.  
+
+"New reviewer comment - Round 3: Lines 463- 466: "When a study included different levels of pesticide application rates, measurements for the control groups without pesticides versus different pesticide application rates were considered as independent paired observations" - This is a very bold assumption given "26,096 estimates of pesticide effects reported from 1,705". Now looking at the actual data, I can see that although sometimes multiple pesticides or species were used in a single study, in most cases multiple doses were used, which means data non- independence related to the repeated use of the same control group to compare it with different dose levels is a serious issue. Specifically, by not controlling for repeated use of the same individuals, the precision of the overall estimate is very likely inflated. One very simple way is to divide sample size of the control group by the number of comparisons it is used in, thus avoiding "double- counting" of the same replicated units.  
+
+Response: Observations with no pesticide had been considered as control groups, while those with different pesticide application rates were considered as the treatment groups. When an article included different levels of pesticide application rates, measurements for the control groups (no pesticide) were compared to all other treatment levels of pesticide application rates and treated as independent paired observations. Within the context of these analysis, this reflects typical experimental designs for studies considered in ecotoxicology that have a single control which is then compared to multiple test treatment residue levels. The random allocation of treatments to replicates, combined with study as a random effect in the meta- analytical models, means that these comparisons are appropriate and do not represent double accounting as suggested. In response to the referee's comment, we have now re- run our analyses with the correction suggested, i.e. dividing the sample size of the control group by the number of comparisons it had been used in.  
+
+The results of this additional analysis can now be found in a new Supplementary Table 62. When comparing these results to our former approach (see Supplementary Table 3), there were no qualitative differences in the results (see Supplementary Table 3 vs. Supplementary Table 62). In particular, all negative effect sizes remained negative (and significant), and all positive effect sizes remained positive and significant.  
+
+New reviewer comment - Round 4: The authors are wrong when stating that "The random allocation of treatments to replicates, combined with study as a random effect in the meta- analytical models, means that these comparisons are appropriate and do not represent double accounting as suggested." The measures from the same control individual are used repeatedly in comparisons of effects of different doses of the same pesticides and they are such effect sizes are not independent and are a form of "double- counting". The analyses that account for such non- independence should be used as main models not as supplementary. The author used sample- size corrections their sensitivity analyses for a few models. I suggest testing their all models using cluster- robust approach which is already implemented in metafor package (https://wviechtb.github.io/metafor/reference/robust.html) as their main approach. Please also run and report an overall (global) meta- analytic model on all data with organism, trait and pesticide type as random effects.  
+
+"New reviewer comment - Round 3: In the main text, please acknowledge study limitations due to sample sizes in the included experiments (from the raw data I can see that the median sample size is likely 3) - having a histogram of the distribution of sample sizes by type of organism and/or treatment would be very useful too.  
+
+Response: The median value of sample size was 4 and 4 for the control and treatment, respectively (see lines 122- 123 in the main text). In addition, we have added a histogram to show the distribution of sample sizes by types of organisms and separately by control and treatment (see Supplementary Fig. 27)."  
+
+New reviewer comment - Round 4: Thank you for adding information about the median sample sizes. However, this raises new issue since Line 562 states "A large- sample approximation was used to compute the sampling variances46. " This approximation is not suitable for such small sample sizes, and a small- size- corrected version of the SMD (Hedges' g) should be used as an effect size in this meta- analysis.
+
+<--- Page Split --->
+
+(Remarks on code availability) code incomplete and poorly annotated  
+
+Version 4:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author)  
+
+Thank you for your patience and hard work on this manuscript. I am pleased to note that it is much more transparent and robust now (it seems to have some new interesting findings as well). I really hope that in your next meta- analytic project you will take all steps from its inception to make sure it is trustworthy and well- documented, so I dont need to do 5 rounds of review again... good luck  
+
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source.  
+
+The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.  
+
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+
+## Response to the referee comments  
+
+## Replies to Reviewer #1:  
+
+Comment: The authors have put significant effort into revising their work, which is substantially improved now.  
+
+Response: We thank the reviewer for this positive evaluation (see this revision).  
+
+Comment: However, there are some remaining concerns. 1. The GitHub report now contains the code but not the data to re- run the code.  
+
+Response: We have uploaded both code and data to GitHub (https://github.com/Liwan- Fu/Impact- of- pesticides).  
+
+Comment: Also, there are multiple R files and no description how to use them to recreate study results- this should be described in detail in the README file. Meta- data should be also provided for all used variables.  
+
+Response: We have provided a README file for all used variables and uploaded the file to GitHub (https://github.com/Liwan- Fu/Impact- of- pesticides).  
+
+Comment: 2. Poor coverage of grey and non- English literature should be acknowledged as limitations in the main manuscript.  
+
+Response: We have now made it explicit that our study did not include grey or non- English literature (lines 420- 421).  
+
+Comment: 3. Trim- and- fill is a publication bias test not sensitivity analysis.  
+
+Response: We have removed text referring to this as a sensitivity analysis (see lines 631- 632 in the main text; line 326 in the Supplementary Results).  
+
+Comment: 4. The issue of percentage and proportion data is not in the normality of the residuals, but in the fact that they are bounded between 0 and 1 (100). As such they should be acris- transformed to make them unbounded. All transformations can be documented in R code, so there is no issue of transparency.  
+
+Response: We have consulted with our institutional statistician (Dr Pete Henrys - CStat, Royal Statistical Society Chartered statistician, CSci, The Science Council Chartered Scientist) who states that a transformation is there to ensure model assumptions are met. Transformation purely on the basis that data is of a particular type is not appropriate, and rather transformation should only be applied on a case by case basis to address underlying issues that may affect model assumptions. He confirms that focusing on model assumption checks (as we have done) including assessing residuals represents the critical stage in the process for assessing the need for underlying processing of input data. We strongly argue that percentage and proportion data in the context of this analysis does not
+
+<--- Page Split --->
+
+warrant to be arcsin- transformed without a reason relating to model distributional fit. In the context of this analysis this is not required.  
+
+Comment: 5. The issue with folded normal distribution when using absolute values has not been resolved. See Nakagawa, S. & Lagisz, M. (2016) Visualizing unbiased and biased unweighted meta- analyses. 334. Journal of Evolutionary Biology, 29, 1914- 1916, for a visual clarification of the issue. Response: This relates to our standardizing of biomarkers (e.g. protein regulation, gene expression) which may respond positively or negatively to pesticides. Our approach has been to use absolute values and focus on a deviation from zero in the meta- analysis, essentially quantifying departure from the norm. Again, we emphasize that checks of model assumptions suggest that the current use of a normal distribution was robust. From a practical perspective it is not possible within our modeling framework to specify a folded distribution (which we do not think is required). However, we would be prepared to simply no longer use the absolute values thus negating the suggested need for a folded distribution. We do feel that this would potentially lose important information about a general impact of pesticides on either up- or down- regulation of biomarkers - which we think is hugely significant from the context of the impacts of pesticides on non- obvious species level effects. The author team has discussed this issue extensively and we still believe that taking absolute values biologically makes more sense here.  
+
+Comment: 6. A synthetic phylogenetic tree can be retrieved from Open Tree of Life, which is accessible via R package rotl. As, such phylogenetic analyses can and should be conducted beyond plants.  
+
+Response: We have used rotl, a freely available package for R designed to takes advantage of the Open Tree of Life's Application Programming Interfaces (APIs) to access subtrees from the synthetic Open Tree, as well as the published source trees that contribute to the synthesis to generate comprehensive phylogenies for animals and microorganisms (see lines 296- 321 in the Supplementary Methods). In this revision, we have now conducted all the phylogenetic analyses for plants, animals and microorganisms (see detailed results in Supplementary Tables 1.3- 1.5, 2.3- 2.5, 45.3- 45.5 and 46.3- 46.5). We still note that an overall agreed- upon phylogeny for the tree of life will likely never arrive, but we did our best to address the referee's comment.  
+
+Comment: 7. Publication year should be included in the analyses (meta- regression)- it is not confounded with study identity, and it is potentially important and interesting moderator of the effect. Response: We have clarified this now as "Publication year: a continuous metric according to the year when the articles were published" (lines 528- 529 in the main text). We have included publication year in the meta- regression model (see detailed results in Supplementary Tables 1.1- 1.5, 2.1- 2.5, 45.1- 45.5 and 46.1- 46.5; Extended Data Figs. 2- 7 and other associated supplementary tables).
+
+<--- Page Split --->
+
+Comment: 8. Funding and conflict of interest statements can be easily extracted from the included papers, even if they are not consistently included. Then, presence of potential links with the relevant industry should be analysed in a meta- regression. This is a too important issue to dismiss.  
+
+Comment: 8. Funding and conflict of interest statements can be easily extracted from the included papers, even if they are not consistently included. Then, presence of potential links with the relevant industry should be analysed in a meta- regression. This is a too important issue to dismiss.Response: We now extracted conflict- of- interest statements where possible. Among the 1705 papers, we could confirm that 1,411 ones did not have conflict of interest while 25 self- identified as having a conflict of interest. The remainder made no statement about conflicts of interest (lines 244- 248 in the main text; Supplementary Table 58). In the meta- regression analysis, we consider "Conflict of interest status" as one binary variable (i.e., "1" denotes that a certain article has conflict of interest and "0" represents no conflict) (lines 526- 528 in the main text). Detailed results for "Conflict of interest status" were presented in Supplementary Tables 1.1- 1.5, 2.1- 2.5, 45.1- 45.5, 46.1- 46.5 and 58, Extended Data Figs. 2- 7 and other associated supplementary tables).  
+
+## Replies to Reviewer #3:  
+
+Comment: The current ms derives from a previous version submitted to [Redacted] that has been commented by three reviewers. I now have checked for the implementation of the comments of Reviewers 2 and 3 in the revised version.  
+
+Response: We thank the reviewer for this positive evaluation. We carefully addressed all comments in revising the manuscript. We are grateful for the suggestions, which have improved the quality of our study (see this revision).  
+
+Comment: To my opinion, almost all comments of Reviewer 2 have been adequately addressed, except for "(1) Comment: 216- 217 were vertebrates and invertebrates equally affected, or one more than the other? Does that depend on the intended target of an pesticide, e.g. distinguishing between pesticides aimed at invertebrates or otherwise?  
+
+Response: We have removed the sentence due to space limited, but from the statistical values in Supplementary Table 4, we can see that vertebrates and invertebrates were equally affected (see Supplementary Table 4).  
+
+Response: We have added this information in the main text, and we have clarified it as "vertebrates and invertebrates were equally affected by pesticides" (lines 258- 259 in the main text).  
+
+Comment: and "(2) Comment: 223 'affected different taxonomic groups' how?  
+
+Response: We have removed the sentence due to space limited, but from the statistical values in Supplementary Tables 32- 36, we can see that "insecticides, fungicides and herbicides affected different taxonomic groups with decreased growth, reproduction or behavior and with perturbed biomarkers" (see Supplementary Tables 32- 36)."  
+
+Response: We have added this information in the main text, and we have clarified it as
+
+<--- Page Split --->
+
+“Insecticides, fungicides and herbicides affected different taxonomic groups though decreased growth, reproduction or behavioural responses, and biomarkers being perturbed from baseline conditions (see Supplementary Tables 32–36)” (lines 216- 218 in the main text).  
+
+Comment: I feel that both informations are important and should be kept within the main body of text rather than being implied to be taken from the supplementary material.  
+
+Response: Please see above two responses.  
+
+Comment: All comments of Reviewer 3 have been perfectly addressed in this version of the ms. Response: Thanks for your positive comments.  
+
+Comment: In the Competing Interests paragraph, "B.W." should be replaced by "B.A.W.". Response: Done as suggested (lines 717, 720 and 721).
+
+<--- Page Split --->
+
+## Response to the referee comments  
+
+## Replies to Reviewer #1:  
+
+Note we only show responses here for outstanding issues you identified in your previous review. Issues you identified as being addressed to your satisfaction we do not refer to here.  
+
+New reviewer comment - Round3: Thank you, for providing a link to the files on GitHub. I note that now the data and code are also provided on Zenodo and the contents of these two repositories do not match. Specifically, some code files that are archived on Zenodo are absent from GitHub version. Please remove any special (Chinese) characters from your R code.  
+
+Response: We thank the referee for pointing us at this. To avoid confusion and duplication we have now uploaded all raw data and code to Zenodo and removed them from GitHub (see this revision). We have checked that all relevant files are present. We have now also removed all Chinese characters as suggested. All data used in this analysis are available on Zenodo (https://zenodo.org/records/10495263). All code for this analysis is available on both Zenodo (https://zenodo.org/records/10509420).  
+
+New reviewer comment - Round3: Thank you for providing a README file. Unfortunately, the meta- data provided applies to only one of the data files, "Meta combined data.csv" (there seems to be 16 data files in total). The meta- data for this one key data table used to run the key models is still too rudimentary to allow cross- checking of the data or replication of the data extraction process. For example, the variable described as "Control (value): the value in control group;" does not explain which value should extracted be there, apart that it should relate to the control group. Another, example - "Control- n: the number in control group;", which is supposed to be the sample size, it also needs explaining the units of counting sample size, as these can be different for different taxa and experiment types - for animals it might be number of individuals, for plants could be individuals or plots, for microorganisms it could be a colony in a tube/plate, or number of natural plots, etc.  
+
+Response: We have now revised the README file to improve clarity. We hope this addresses the problems.  
+
+Comment: One column in the data table called "LnR" is not described - it seems to be an effect size that was not be used in the analyses, as the main effect size used was SMD. However, I cannot find the code the authors used for calculating SMD.  
+
+Response: The column in the data table called "LnR" had not been used in the analyses, and we have now deleted it. The code used for calculating SMD has been added in the README file.  
+
+Comment: Also, I cannot find the code for Figure 1 (the map), and instead there is code for Figure 1b, a histogram which is not in the main manuscript.  
+
+Response: Figure 1 had been prepared in ArcGIS, so no R code can be provided for that. We have
+
+<--- Page Split --->
+
+now deleted code for “Figure 1b” which was part of a prior version of the manuscript.  
+
+Comment: Further, the study codes in the main data file “Meta combined data.csv” do not match the study numbers in the table characterizing included studies (with their references) – “447069_2_data_set_8320286_s47vd1.xls”. Please include full study references as a column in the data file “Meta combined data.csv”, so that each data point can be linked to its original source. Overall, lack of consistent and comprehensive documentation is still a major limitation to transparency and reproducibility of this meta-analysis, which should be addressed before this work is accepted.  
+
+Response: We thank the referee for pointing us at this. We have now made sure that all study codes match up. We have listed all the 1705 study references in the file “Meta combined data” with a new sheet (Full study references). The paper code order in “Meta combined data” is based on the time order that we extracted data, and the paper order in Supplementary Data 1 is based on the authors' name in alphabetical order (see Meta combined data). In addition, we have added another data file “ExtendData” for other supplementary files (e.g., for Supplementary Tables 59- 61, Supplementary Figures 24- 27).  
+
+New reviewer comment - Round 3: I can see it in the methods section now. I suggest adding this [grey and non- English literature] also to the header of Figure 1 (the global) map, because it is highly relevant to the pattern presented there.  
+
+Response: We have added “Grey and non- English literature was not included in this meta- analysis” in the caption of Figure 1.  
+
+New reviewer comment - Round 3: it is great that you consulted the statistician and conducted the model assumptions checks. Please provide the results of the model assumptions test in the supplementary materials. Conducting sensitivity analysis for the main model that show that applying the recommended transformation does not change the results should be also presented in the supplementary materials to support your claims.  
+
+Response: As suggested, we now include a sensitivity analysis where we apply arcsine transformation for percentage and proportion data (as you suggest) and compare the results to our approach (no transformation or proportion/percentage data) (see Supplementary Table 59). This shows no qualitative difference in the results. We also provide a supplementary figure in the supporting information for the model checks of the original analysis (see Supplementary Fig. 24).  
+
+Original comment for context: 5. The issue with folded normal distribution when using absolute values has not been resolved. See Nakagawa, S. & Lagisz, M. (2016) Visualizing unbiased and biased unweighted meta- analyses. 334. Journal of Evolutionary Biology, 29, 1914- 1916, for a visual clarification of the issue. New reviewer comment - Round 3: Please provide the results of the
+
+<--- Page Split --->
+
+model assumptions test in the supplementary materials. Conducting sensitivity analysis for the key model that show that applying the recommended transformation does not change the results for biomarkers data subset should be also presented in the supplementary materials to support your claim. Also, please flip the sign on the absolute values to make it easier to compare magnitude of estimated effects in all forest plots.  
+
+Response: As suggested, we now present additiopal model checks in Supplementary Fig. 25 that show that the distribution used was appropriate (see Supplementary Fig. 25). However, we acknowledge the referee's wider concern about the use of absolute values for biomarker responses. As already pointed out, biomarkers include metrics that may potentially be both be up- and downregulated as a response to exposure to pesticides. In both cases, this would represent biologically meaningful responses that may have longer term fitness consequences for individual populations. To account for the potential of both up- and down- regulation of biomarkers, we have treated these as absolute values within the main analyses. To address the referee's comment, and to support a broader interpretation of the data, we now provide a comparative sensitivity analysis for the response of biomarkers to overall pesticide exposure for animals, plants, and microorganisms using absolute values of 'biomarker' (as described in the methods) and the original raw values of 'biomarker', i.e. with directionality of the effect (see Supplementary Table 60). As can be seen, even when the biomarkers are not treated as absolute values, they are still on average characterized by negative effects in response to pesticide exposure so that the results remain qualitatively equivalent to those presented in the main paper. This would suggest that while biomarkers may in principal be both up and down regulated following pesticide exposure, across the studies considered this effect remains overwhelmingly down regulation. We hope that this addresses your wider concerns about the biomarker responses.  
+
+New comment - Round 3: By looking at the raw data, "Meta combined data.csv", I can see "Life expectancy" and "Longevity" are in "insecticide- animal reproduction" category. It seems like all (and there are many) measures related to survival are classified as "insecticide- animal reproduction" (similar issue is also present for plants and microorganisms in the data set). Please provide justification for this choice.  
+
+Response: We acknowledge that to some extent the categories we had used for reproduction, growth, behaviour and biomarkers are synthetic, but we believe they still allow the most meaningful interpretation of the plethora of metrics measured across the 1705 studies as responses to pesticide exposure. While further sub- divisions of these categories could have been undertaken, this would not have added to the clarity of the manuscript. As indicated, certain responses to pesticides were allocated to these four categories based on the authors' combined opinion, as well as though solicitation of opinion for colleagues. As such, "Life expectancy" and "Longevity" were considered as "animal reproduction" as premature death may limit lifetime potential reproductive output. Similarly, we think that animal survival determines population reproduction of animals, and thus
+
+<--- Page Split --->
+
+should be considered as "animal reproduction". There are no "Life expectancy" and "Longevity" measures for plants or microorganisms. The overwhelming negative effects (or positive in the case of the use of absolute values for biomarkers) across reproduction, growth, behavior and biomarkers suggests that, while further sub- categorization of these responses could have been undertaken, this would not have changed the results.  
+
+New comment - Round 3: Also animal feeding rates could represent both growth and behavior - having them solely as a measure of growth may require additional sensitivity analyses and clarification in the methods section.  
+
+Response: The suggestion of the reviewer is correct as animal feeding rates could represent both growth and feeding behavior. Here, we assumed that animal feeding is primarily for animal growth, and thus we put "animal feeding" indicator into "animal growth" in this meta analysis. We do however follow the referee's suggestion and apply a sensitivity analysis to assess the effect of allocation of feeding to either growth or behavior in our meta analysis. We have added a supplementary table and a supplementary figure to measure the sensitivity analyses of animal growth and animal behavior when "animal feeding" indicator was included in animal growth or in animal behaviour (see Supplementary Table 61 and Supplementary Fig. 26).  
+
+New reviewer comment - Round 3: In the "Meta combined data.csv", there are some clear mistakes too, e.g. the column for describing measurements is named "Animal growth indicator", and sometimes is empty, "number of eggs/female" is classified as "insecticide- animal behavior", this raises concerns for the consistency of data extractions and the quality control procedures in this work.  
+
+Response: We have addressed these minor errors and checked that this has not affected data extraction. In "Meta combined data", we have replaced the column "Animal growth indicator" with a new column ("Non- target organism indicator"), added the missing experimental year (i.e., 2017) in rows 7849- 7853, and have added the missing unit (i.e., Individuals/g) in row 14201. In "Meta combined data.csv", we have replaced "number of eggs/female" with "Walking distance (cm)" in row 8051, and have replaced "number of eggs/female" with "Walking speed (10- 2 cm s- 1)" in row 8053. Both "Walking distance" and "Walking speed" belong to "animal behavior".  
+
+New reviewer comment - Round 3: Lines 463- 466: "When a study included different levels of pesticide application rates, measurements for the control groups without pesticides versus different pesticide application rates were considered as independent paired observations" - This is a very bold assumption given "26,096 estimates of pesticide effects reported from 1,705". Now looking at the actual data, I can see that although sometimes multiple pesticides or species were used in a single study, in most cases multiple doses were used, which means data non- independence related to the repeated use of the same control group to compare it with different dose levels is a serious issue.
+
+<--- Page Split --->
+
+Specifically, by not controlling for repeated use of the same individuals, the precision of the overall estimate is very likely inflated. One very simple way is to divide sample size of the control group by the number of comparisons it is used in, thus avoiding "double- counting" of the same replicated units.  
+
+Response: Observations with no pesticide had been considered as control groups, while those with different pesticide application rates were considered as the treatment groups. When an article included different levels of pesticide application rates, measurements for the control groups (no pesticide) were compared to all other treatment levels of pesticide application rates and treated as independent paired observations. Within the context of these analysis, this reflects typical experimental designs for studies considered in ecotoxicology that have a single control which is then compared to multiple test treatment residue levels. The random allocation of treatments to replicates, combined with study as a random effect in the meta- analytical models, means that these comparisons are appropriate and do not represent double accounting as suggested. In response to the referee's comment, we have now re- run our analyses with the correction suggested, i.e. dividing the sample size of the control group by the number of comparisons it had been used in.  
+
+The results of this additional analysis can now be found in a new Supplementary Table 62. When comparing these results to our former approach (see Supplementary Table 3), there were no qualitative differences in the results (see Supplementary Table 3 vs. Supplementary Table 62). In particular, all negative effect sizes remained negative (and significant), and all positive effect sizes remained positive and significant.  
+
+New reviewer comment - Round 3: In the main text, please acknowledge study limitations due to sample sizes in the included experiments (from the raw data I can see that the median sample size is likely 3) - having a histogram of the distribution of sample sizes by type of organism and/or treatment would be very useful too.  
+
+Response: The median value of sample size was 4 and 4 for the control and treatment, respectively (see lines 122- 123 in the main text). In addition, we have added a histogram to show the distribution of sample sizes by types of organisms and separately by control and treatment (see Supplementary Fig. 27).
+
+<--- Page Split --->
+
+## Response to the referee comments  
+
+## Replies to Reviewer #1:  
+
+The authors addressed some of my concerns but there are still outstanding issues related to transparency and reproducibility of this work, as well as some methodological concerns. My previous comments and authors responses are quoted below when they are linked to new issues identified. My new comments are provided as "New reviewer comment - Round4:  
+
+Response: We thank you for these suggestions to improve transparency and reproducibility of our manuscript. We have now revised the manuscript according to these suggestions. The responses to the below comments describe the changes that have been implemented in greater detail. In particular, we have made sure that the work is transparent and reproducible. We have also re- done the data extraction from all original studies, which has greatly improved data quality.  
+
+New reviewer comment - Round4: Thank you for providing your data and code on Zenodo. It seems like putting the contents of these two repositories together still does not allow reproducing all the analyses and they are really hard to peer- review. There are three reasons: first the code is very poorly annotated, and thus it is hard to figure out what different sections of code are supposed to produce. Second, it seems like pieces of code are missing, e.g. file "Figures 2- 3. R" starts from loading the file 'Meta combined data.csv' and then immediately rands a meta analytic model on the data from it, however the file does not contain effect sizes and there is no line of code that calculates effect sizes before running the model. Third, the code refers to loading many additional data files, that look like pre- processed data, but they are not included in the Zenodo repository, e.g. "Supplementary Tables 1(1- 3). R" calls "Old_New_pesticide <- read_excel(paste("/lustre/huyueqing/ssy/Wan2023/DataPre/","Supplementary Data 2- Classification of old and new pesticides- 2023- 0825. xls",sep = ""),sheet = 1)", but there is no such file or no explanation where it originates from. Overall, I think all the analytic pipeline should be made fully transparent reproducible and then subjected to independent peer- review.  
+
+Response: Thank you for your comment. We have now completely overhauled the analytical pipeline in response to your comments. 1) We have now gone through the code and improved the annotation so that it allows a full interpretation of what has been done. 2) We have ensured that all pieces of code are present, including those relating to the production of all figures, tables and supplementary materials; 3) All files required for the analysis are included.  
+
+New reviewer comment - Round 4: Thank you. The README file only describes data in the main data file, but not in any of the additional data files and does not provide any information about the project in general and how the code and files should be used to reproduce the analyses. Please provide more comprehensive description of all your files, starting from overview and then providing meta- data for all data that is needed to reproduce your work. Consider using Quarto with R to
+
+<--- Page Split --->
+
+produce computationally reproducible documents that contain the R code, verbal descriptions, and outputs.  
+
+Response: We acknowledge that the previous version lacked reproducibility and had been hard to follow. Addressing this recommendation, we have now re- written the README file using Quarto with R, and extended this to include information on additional data files, as well as our analysis approach in general. We hope that this has now improved transparency of dataset descriptions and relevant code information, allowing clear interpretation of the approach used for all stages of the analysis. We have double- checked all code and made sure it runs on machines across continents.  
+
+New reviewer comment - Round 4: The code used for calculating SMD should not be placed in the README file, but in the code where it is needed to calculate effect sizes to reproduce the results. Response: Thank you for your comment. Following your wider recommendation, all R code and a README file with R annotation developed using Quarto have now been uploaded to Zenodo. The specific code you are referring to for calculating SMD has been also been revised again in this version to avoid problems stemming from our initial use of SMD rather than log response ratios. In our revised version, we now use the log response ratio (InRR) to calculate the effect sizes (the range of effect sizes in InRR is smaller than when using SMD). We have contacted Shinichi Nakagawa, an expert in meta- analysis and statistics, whose recommendations we follow in our revisions. Briefly, we now use log response ratios rather than SMD, because these do not require standard deviations and are more robust for use in meta- analysis, as demonstrated by Nakagawa et al. (2023; see reference below). For missing value imputation, we have now employed the so- called 'all cases' method, for sensitivity analysis and main analysis. Reporting follows the PRISMA- EcoEvo checklist to enhance quality. Additionally, we used the latest publication bias tests and submitted our fully annotated R code to Zenodo for transparency. We believe that these extensive revisions have considerably increased the robustness of our analysis.  
+
+Reference: Nakagawa et al., 2023. A robust and readily implementable method for the meta- analysis of response ratios with and without missing standard deviations. Ecol. Lett. 26: 232- 244.  
+
+New reviewer comment - Round 4: Thank you for adding the study references next to study codes in the data file. I was now able to compare the extracted data with the information reported in the original papers. I randomly selected 4 papers I have full- text access to and found the following issues:  
+
+Response: Following your comments about potential errors in the data (and as an explanation for the significant delay in producing this review) we have now re- checked all the original data and studies used in this paper. This involved going back to the original papers and confirming values derived to produce effect sizes as used in the analysis (note this is the reason for the considerable delay in this revision as the complete recheck took a lot of time). Where on reassessment papers were not considered to meet requirements for data derivation, they have now been excluded. In
+
+<--- Page Split --->
+
+some cases, new papers were identified from citations within those that were rejected for the above reasons but not identified in the initial scanning process. The process for including these additional papers is detailed within the PRISMA process and diagram. Below, we reply to the specific further comments on individual papers.  
+
+- Paper 16. (Colin et al. 2004. A Method to Quantify and Analyze the Foraging Activity of Honey Bees: Relevance to the Sublethal Effects Induced by Systemic Insecticides): rows 16-21: the concentration of Fipronil is wrong - the study reports using 2 micrograms per kilogram while the extracted data is \(6 \mathrm{g / kg}\) . For Imidacloprid the unit is also wrong-it should be microgram per kilogram not gram per kilogram. For the outcome "number of daily attendance of active bees-nuclei" which was extracted for nuclei A, B, C, it is and somehow from control nuclei, but I cannot find the matching exact numbers in the text, and it is not possible to extract them from the figures in the text. Given the complexity of the experimental design, raw data would need to be used to calculate such values. There are no comments in the data file how the numbers were obtained, and thus they are not reproducible.  
+
+Response: In the context of the paper you refer to, the means, SD and sample size \((n)\) of the indicator (i.e., "number of daily attendance of active bees-nuclei A") were calculated on days 0, 1, and 4. Because this was not a strictly repeated trial, we have removed this paper.  
+
+- Paper 1628. (Zhao X, Li YL, Zhang LX, Dorna H, 2003. Effects of priming and fungicide treatment on germination of china aster (Callistephus chinensis L.) seeds. Seed Science & Technology, 32: 451-457.): rows 1943-1944: it looks like only germination capacity data from Sample 1 and only from 30C temperature treatment were extracted and the standard deviations were manually imputed by assuming they are \(10\%\) of the measured value. This is only a small fraction of available data in this paper (multiple control and exposure groups, all samples should be extracted and used for calculating mean and SD) and is incorrectly extracted.  
+
+Response: We thank you for pointing us at this. To exclude the interactive effects of pesticide doses and temperature levels on non- target organisms, and only to measure the individual effect of pesticides on non- target organisms, our principle is that we extract the data at the highest temperature to conduct our analysis to exclude the interactive effects of pesticides and temperatures. The authors mentioned that "The germination tests were conducted at \(30^{\circ} \mathrm{C}, 20^{\circ} \mathrm{C}\) and \(10^{\circ} \mathrm{C}\) in darkness on 8 replicates of 50 seeds for each treatment." This is why the sample size was 8, and the SD of the value "Germination capacity (%)" at different temperatures is missing in Table 1. Because this is a comparatively old paper, we did not receive a response from the original corresponding author. To exclude your worries about this paper, we have now also removed this paper.
+
+<--- Page Split --->
+
+## Seed germination tests  
+
+The germination tests were conducted at \(30^{\circ}\mathrm{C}\) \(20^{\circ}\mathrm{C}\) and \(10^{\circ}\mathrm{C}\) in darkness on 8 replicates of 50 seeds for each treatment. Control seeds were untreated. Fifty seeds were placed in each \(9\mathrm{cm}\) diameter Petri dish containing 6 layers blotting paper wetted with \(5\mathrm{ml}\) of distilled water. Seeds were considered as germinating when there was a visible protrusion of the radical through the seed coat and pericarp.  
+
+For germination index and mean germination time were recorded at 24h intervals for 14 days. Germination index \(\mathrm{(G~I) = (Gt/Tt))}\) Where G is the number of seeds germinated on day t and t is the number of days. Mean germination time (MGT) \(=\) TiNi/ Ni, where Ni is the number of newly germinated seed at time Ti.  
+
+Figure: Screenshot proof showing that these were 8 replicates  
+
+(from Zhao et al., Seed Science & Technology, 32: 451- 457)  
+
+- Paper 1008. (Gabaston J, Khawand TE, Wafo-Teguo P, Decendit A, Richard T, Merillon JM, Pavel A, 2018. Stilbenes from grapevine root: a promising natural insecticide against Leptinotarsa decemlineata. Journal of Pest Science, 91: 897-906.) rows 5808-5809: looks like data was extracted from Table 6 for 3rd day of exposure only and two lower doses out of 3 used (this is not reported on the data extraction sheet). The mean mortality \(\%\) was converted to survival \(\%\) , but the reported standard deviation values were somehow replaced with \(10\%\) of the men value. The sample size is claimed to be 4, but I cannot find such information in the article, and it seems improbable given the type of the organism, and reported values.  
+
+Response: In the data extraction sheet (i.e., Meta combined data- 2024- 0814), in rows 131- 132 in column "N" (i.e., Non- target organism indicator), we have presented it as "survival percent- transformed by mortality- 3rd day- Table 6". In the Methods section (i.e., Earthworm acute toxicity test), the authors have clearly shown that the sample size was 4: "Ten adult earthworms were placed in glass containers (1 L) filled with the test substrate (650 g), and the test containers were enclosed with a polythene sheet with integrated gauze to prevent the worms from escaping and to ensure optimal ventilation. After 3, 7 and 14 days of incubation, living worms were sorted by hand; the test endpoint was mortality. The tests were repeated four times. The glass containers were placed in a growth chamber in an artificial climate (L16:D8, \(20 \pm 1^{\circ}\mathrm{C}\) ).  
+
+Ten adult earthworms were placed in glass containers (1 L) filled with the test substrate (650 g), and the test containers were enclosed with a polythene sheet with integrated gauze to prevent the worms from escaping and to ensure optimal ventilation. After 3, 7 and 14 days of incubation, living worms were sorted by hand; the test endpoint was mortality. The tests were repeated four times. The glass containers were placed in a growth chamber in an artificial climate (L16:D8, \(20 \pm 1^{\circ}\mathrm{C}\) ).  
+
+Fig. Screenshot proof that this paper had 4 replicates (Gabaston et al., 2018. Journal of Pest
+
+<--- Page Split --->
+
+Question: Overall, this small exercise clearly shows that 3/3 papers had some issues with data extractions. This indicates a very high rate of errors in the whole data set. The data lacks information on the exact sources of the extracted values for the outcomes at least (table/figure number, text page, raw data, etc.), any extractions decisions made, calculations, imputations, etc. - this makes assessing trustworthiness of the data set very hard and makes data extractions irreproducible. Because of this I request that all extracted data is annotated in detail. Further it has to be cross-checked by a researcher who was not involved in original extractions, re-extracted where issues are found and clearly described to make it fully transparent and trustworthy.  
+
+Response: As an expansion to our initial statement about a complete re- check of the data, we have now added the column "Y" (i.e., Exact sources of extracted values), the column "Z" (i.e., Extraction decisions made) and listed the extracted data information in column "N" (i.e., Non- target organism indicator). In this revision, Dr. Siyuan Shen re- extracted and cross- checked the data. After Siyuan's re- extracting and checking, Dr. Nian- Feng Wan and Dr. Ben A. Woodcock re- checked all the extracted information again, respectively, with additional checks being done by Christoph Scherber. During this re- extraction process, papers considered to have been unclear in their definition of measures of variance or described sample sizes were subjected to a group discussion (mainly composed of Nian- Feng Wan, Ben A. Woodcock, Dave Goulson, Adam J. Vanbergen, David J. Spurgeon and Siyuan Shen) to produce consensus on these derived metrics. At least three group members participated in each discussion. If our understanding was the same, we retained the data. If not, we contacted the authors directly to ask for the data. If the authors did not respond or could not be contacted, we removed the dataset originating from the respective paper. However, to collect more data to test pesticide effects on non- target organisms, we replaced papers from which no valid data could be extracted with a new, matching paper by the same authors or institutions (see lines 16- 29 in Supplementary methods). This process has (as we state above) now been clarified within the revised PRISMA diagram and associated processing description for data set derivation.  
+
+New reviewer comment - Round 4: When checking the extracted data, I noticed that many of the extracted standard deviation values are exactly \(10\%\) of the mean value. This cannot be coincidence and the manuscript does not mention doing manual imputations for missing standard deviation values (or any other missing values). Please explain how these values originated and the detailed procedures used to deal with any missing data.  
+
+Response: You are correct that where the values were \(10\%\) , this was due to an imputation for missing values (reffering to: Luo et al., 2006. Elevated \(\mathrm{CO_2}\) stimulates net accumulations of carbon and nitrogen in land ecosystems: a meta- analysis. Ecology, 87: 53- 63). In consultation with Professor Shinichi Nakagawa we have now revised this approach to deal with missing SDs in meta analyses. According to Dr. Nakagawa's suggestions, we now followed the approach described in his
+
+<--- Page Split --->
+
+paper (Nakagawa et al., 2023. A robust and readily implementable method for the meta-analysis of response ratios with and without missing standard deviations. Ecol. Lett. 26: 232- 244) and used Equations 6 and 7 (see methods section in our manuscript) to calculate effect sizes and sampling variances when SDs were missing by simply imputing the pooled coefficient variation (CV) from the subset of studies that do report SDs. This is now fully described in the new revised version of our manuscript (see lines 30- 34 in Supplementary methods).  
+
+New reviewer comment - Round 4: The code for reproducing this should be in the file "Supplementary Tables 59- 61. R", and as such cannot be rerun. In the provided code I cannot find any line that would execute arcsine transformation for proportion data. And the code starts from loading a mysterious data file that is not archived on Zenodo, so cant tell what is there nor how it got there: AllData <- read.table(paste(DataPre_path,"SuppTab59Data.txt",sep=""),header = T). This seems to be ongoing issues across the provided code which is compounded by almost complete lack of code comments and descriptions.  
+
+Response: In this revision, and in response to your other above comments, we have re- written the code for executing the arcsine transformation for proportion data. Code for the arcsine transformation for proportion data is now included and annotated in the README file and all operational R files. We also had an internal discussion on using arcsine vs. logit transformation by hand, but stic ked with the arcsine as it is more established in a meta analysis context. The datasets are described in README file and can now all be fully accessed. Annotation has been improved for the code in general using as you suggested using Quarto with R.  
+
+New reviewer comment - Round 4: The authors are wrong when stating that "The random allocation of treatments to replicates, combined with study as a random effect in the meta- analytical models, means that these comparisons are appropriate and do not represent double accounting as suggested." The measures from the same control individual are used repeatedly in comparisons of effects of different doses of the same pesticides and they are such effect sizes are not independent and are a form of "double- counting". The analyses that account for such non- independence should be used as main models not as supplementary. The author used sample- size corrections their sensitivity analyses for a few models. I suggest testing their all models using cluster- rubust approach which is already implemented in metafor package (https://wviechtb.github.io/metafor/reference/robust.html) as their main approach.  
+
+Response: Thanks for your comment. We have addressed this important issue following the approach described by Shinichi Nakagawa in his paper on "Quantitative evidence synthesis: a practical guide on meta- analysis, meta- regression, and publication bias tests for environmental sciences." (Nakagawa et al., 2023. Environ. Evid. 12: 8). In order to adjust for repeated measurements of control values, we assigned the argument "V" in the "rma.mv()" function with the sampling variance- covariance matrix estimated by function "vcalc()" (lines 584- 592).
+
+<--- Page Split --->
+
+New reviewer comment - Round 4: Please also run and report an overall (global) meta-analytic model on all data with organism, trait and pesticide type as random effects.  
+
+Response: We derived responses for 830 different species (560 animals, 192 plants, 78 microorganisms) and 129 non- species- level groups (i.e. 79 non- species- level animals, 8 non- species- level plants and 42 non- species- level microorganisms) to 471 different pesticide active ingredients (243 insecticides, 104 fungicides and 124 herbicides) (Supplementary Tables 14- 28; Supplementary Data 1- 3) (lines 118- 122). In the last version of this paper, pesticide identity (active ingredients pesticide name) and study identity were used as random effects that were recognized by you. In our revision, we have now added organisms and traits as random effects, according to your suggestions. Namely, we run/reported an overall (global) meta- analytic model on all data with organisms (unique species name and unique non- level species groups), traits (growth, reproduction, biomarker and behavior), pesticide identity (active ingredients pesticide names) and study identity as random effects (see Supplementary Table 1.1, Supplementary Table 1.2, Supplementary Table 2.1 and Supplementary Table 2.2), run/reported an overall meta- analytic model on all data with plant species (excluding non- level species groups), traits (growth, reproduction and biomarker), pesticide identity (active ingredients pesticide names), and study identity for plants (see Supplementary Table 1.3 and Supplementary Table 2.3), run/reported an overall meta- analytic model on all data with animal species (excluding non- level species groups), traits (growth, reproduction, biomarker and behavior), pesticide identity (active ingredients pesticide names), and study identity as random effects for animals (see Supplementary Table 1.4 and Supplementary Table 2.4), and run/reported an overall meta- analytic model on all data with microorganism species (excluding non- level species groups), traits (growth, reproduction and biomarker), pesticide identity (active ingredients pesticide names), and study identity as random effects for microorganisms (see Supplementary Table 1.5 and Supplementary Table 2.5).  
+
+New reviewer comment - Round 4: Thank you for adding information about the median sample sizes. However, this raises new issue since Line 562 states "A large- sample approximation was used to compute the sampling variances46." This approximation is not suitable for such small sample sizes, and a small- size- corrected version of the SMD (Hedges' g) should be used as an effect size in this meta- analysis.  
+
+Response: Thank you for your comment. According to suggestions from Professor Shinichi Nakagawa, we have replaced SMD with InRR to calculate the effect size and sample variance. Professor Shinichi Nakagawa reported four new methods to calculate effect size when missing SDs exist, which was reported in his paper entitled "A robust and readily implementable method for the meta- analysis of response ratios with and without missing standard deviations" (Nakagawa et al., 2023. Ecol. Lett. 26: 232- 244). He suggested that using the average CV to estimate sampling variances for all observations, regardless of missingness, performs with minimal
+
+<--- Page Split --->
+
+bias. We have exactly followed his suggestions.
+
+<--- Page Split --->
+
+## Response to the referee comments  
+
+## Replies to Reviewer #1:  
+
+Comment: Thank you for your patience and hard work on this manuscript. I am pleased to note that it is much more transparent and robust now (it seems to have some new interesting findings as well). I really hope that in your next meta- analytic project you will take all steps from its inception to make sure it is trustworthy and well- documented, so I dont need to do 5 rounds of review again... good luck.  
+
+Response: We thank you for your positive evaluation. We have checked the manuscript again.
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__6e3070f47b530d68e5f8d912919727215aa0f9efcd341ef546885015d497a4fc/supplementary_0_Transparent Peer Review file__6e3070f47b530d68e5f8d912919727215aa0f9efcd341ef546885015d497a4fc_det.mmd

@@ -0,0 +1,701 @@
+<|ref|>title<|/ref|><|det|>[[72, 50, 295, 80]]<|/det|>
+# nature portfolio  
+
+<|ref|>title<|/ref|><|det|>[[74, 96, 296, 120]]<|/det|>
+# Peer Review File  
+
+<|ref|>title<|/ref|><|det|>[[72, 161, 870, 187]]<|/det|>
+# Pesticides have negative effects on non-target organisms  
+
+<|ref|>text<|/ref|><|det|>[[72, 200, 500, 218]]<|/det|>
+Corresponding Author: Professor Nian- Feng Wan  
+
+<|ref|>text<|/ref|><|det|>[[72, 250, 875, 278]]<|/det|>
+This manuscript has been previously reviewed at another journal. This document only contains information relating to versions considered at Nature Communications. Mentions of the other journal have been redacted.  
+
+<|ref|>text<|/ref|><|det|>[[72, 289, 866, 304]]<|/det|>
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+<|ref|>text<|/ref|><|det|>[[72, 341, 144, 355]]<|/det|>
+Version 1:  
+
+<|ref|>text<|/ref|><|det|>[[72, 368, 220, 381]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[72, 394, 160, 407]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[72, 420, 238, 434]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 445, 884, 473]]<|/det|>
+The authors have put significant effort into revising their work, which is substantially improved now. However, there are some remaining concerns.  
+
+<|ref|>text<|/ref|><|det|>[[70, 497, 900, 733]]<|/det|>
+1. The GitHub report now contains the code but not the data to re-run the code. Also, there are multiple R files and no description how to use them to recreate study results – this should be described in detail in the README file. Meta-data should be also provided for all used variables.  
+2. Poor coverage of grey and non-English literature should be acknowledged as limitations in the main manuscript.  
+3. Trim-and-fill is a publication bias test not sensitivity analysis.  
+4. The issue of percentage and proportion data is not in the normality of the residuals, but in the fact that they are bounded between 0 and 1 (100). As such they should be arcsin-transformed to make them unbounded. All transformations can be documented in R code, so there is no issue of transparency.  
+5. The issue with folded normal distribution when using absolute values has not been resolved. See Nakagawa, S. & Lagisz, M. (2016) Visualizing unbiased and biased unweighted meta-analyses. 334. Journal of Evolutionary Biology, 29, 1914-1916, for a visual clarification of the issue.  
+6. A synthetic phylogenetic tree can be retrieved from Open Tree of Life, which is accessible via R package rotl. As, such phylogenetic analyses can and should be conducted beyond plants.  
+7. Publication year should be included in the analyses (meta-regression) – it is not confounded with study identity, and it is potentially important and interesting moderator of the effect.  
+8. Funding and conflict of interest statements can be easily extracted from the included papers, even if they are not consistently included. Then, presence of potential links with the relevant industry should be analysed in a meta-regression. This is a too important issue to dismiss.  
+
+<|ref|>sub_title<|/ref|><|det|>[[72, 810, 161, 823]]<|/det|>
+## Reviewer #3  
+
+<|ref|>text<|/ref|><|det|>[[72, 835, 238, 848]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[70, 848, 915, 875]]<|/det|>
+The current ms derives from a previous version submitted to [Redacted] that has been commented by three reviewers. I now have checked for the implementation of the comments of Reviewers 2 and 3 in the revised version.  
+
+<|ref|>text<|/ref|><|det|>[[70, 887, 728, 902]]<|/det|>
+To my opinion, almost all comments of Reviewer 2 have been adequately addressed, except for  
+
+<|ref|>text<|/ref|><|det|>[[70, 913, 905, 941]]<|/det|>
+(1) Comment: 216-217 were vertebrates and invertebrates equally affected, or one more than the other? Does that depend on the intended target of an pesticide, e.g. distinguishing between pesticides aimed at invertebrates or otherwise?
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 47, 904, 75]]<|/det|>
+Response: We have removed the sentence due to space limited, but from the statistical values in Supplementary Table 4, we can see that vertebrates and invertebrates were equally affected (see Supplementary Table 4).  
+
+<|ref|>text<|/ref|><|det|>[[73, 87, 101, 99]]<|/det|>
+and  
+
+<|ref|>text<|/ref|><|det|>[[73, 112, 490, 126]]<|/det|>
+(2) Comment: 223 'affected different taxonomic groups' how?  
+
+<|ref|>text<|/ref|><|det|>[[73, 126, 921, 179]]<|/det|>
+Response: We have removed the sentence due to space limited, but from the statistical values in Supplementary Tables 32- 36, we can see that "insecticides, fungicides and herbicides affected different taxonomic groups with decreased growth, reproduction or behavior and with perturbed biomarkers" (see Supplementary Tables 32-36).  
+
+<|ref|>text<|/ref|><|det|>[[70, 189, 891, 217]]<|/det|>
+I feel that both informations are important and should be kept within the main body of text rather than being implied to be taken from the supplementary material.  
+
+<|ref|>text<|/ref|><|det|>[[73, 229, 650, 243]]<|/det|>
+All comments of Reviewer 3 have been perfectly addressed in this version of the ms.  
+
+<|ref|>text<|/ref|><|det|>[[73, 255, 602, 269]]<|/det|>
+In the Competing Interests paragraph, "B.W." should be replaced by "B.A.W.".  
+
+<|ref|>text<|/ref|><|det|>[[73, 294, 144, 307]]<|/det|>
+Version 2:  
+
+<|ref|>text<|/ref|><|det|>[[73, 320, 219, 333]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[73, 346, 160, 359]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[73, 372, 238, 384]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[73, 386, 219, 398]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[73, 399, 761, 412]]<|/det|>
+The authors addressed my concerns partially. However some new concerns emerged in the process.  
+
+<|ref|>text<|/ref|><|det|>[[73, 425, 313, 438]]<|/det|>
+Response to the referee comments  
+
+<|ref|>text<|/ref|><|det|>[[73, 439, 234, 451]]<|/det|>
+Replies to Reviewer #1:  
+
+<|ref|>text<|/ref|><|det|>[[73, 451, 822, 465]]<|/det|>
+Comment: The authors have put significant effort into revising their work, which is substantially improved now.  
+
+<|ref|>text<|/ref|><|det|>[[73, 465, 616, 478]]<|/det|>
+Response: We thank the reviewer for this positive evaluation (see this revision).  
+
+<|ref|>text<|/ref|><|det|>[[70, 490, 912, 515]]<|/det|>
+Comment: However, there are some remaining concerns. 1. The GitHub report now contains the code but not the data to re- run the code.  
+
+<|ref|>text<|/ref|><|det|>[[73, 515, 833, 528]]<|/det|>
+Response: We have uploaded both code and data to GitHub (https://github.com/Liwan- Fu/Impact- of- pesticides).  
+
+<|ref|>text<|/ref|><|det|>[[73, 528, 911, 569]]<|/det|>
+New reviewer comment - Round3: Thank you, for providing a link to the files on GitHub. I note that now the data and code are also provided on Zenodo and the contents of these two repositories do not match. Specifically, some code files that are archived on Zenodo are absent from GitHub version. Please remove any special (Chinese) characters from your R code.  
+
+<|ref|>text<|/ref|><|det|>[[72, 581, 870, 608]]<|/det|>
+Comment: Also, there are multiple R files and no description how to use them to recreate study results- this should be described in detail in the README file. Meta- data should be also provided for all used variables.  
+
+<|ref|>text<|/ref|><|det|>[[72, 608, 750, 634]]<|/det|>
+Response: We have provided a README file for all used variables and uploaded the file to GitHub (https://github.com/Liwan- Fu/Impact- of- pesticides).  
+
+<|ref|>text<|/ref|><|det|>[[72, 634, 918, 840]]<|/det|>
+New reviewer comment - Round3: Thank you for providing a README file. Unfortunately, the meta- data provided applies to only one of the data files, "Meta combined data.csv" (there seems to be 16 data files in total). The meta- data for this one key data table used to run the key models is still too rudimentary to allow cross- checking of the data or replication of the data extraction process. For example, the variable described as "Control (value): the value in control group;" does not explain which value should extracted be there, apart that it should relate to the control group. Another, example - "Control- n: the number in control group;", which is supposed to be the sample size, it also needs explaining the units of counting sample size, as these can be different for different taxa and experiment types - for animals it might be number of individuals, for plants could be individuals or plots, for microorganisms it could be a colony in a tube/plate, or number of natural plots, etc. One column in the data table called "LnR" is not described - it seems to be an effect size that was not be used in the analyses, as the main effect size used was SMD. However, I cannot find the code the authors used for calculating SMD. Also, I cannot find the code for Figure 1 (the map), and instead there is code for Figure 1b, a histogram which is not in the main manuscript. Further, the study codes in the main data file "Meta combined data.csv" do not match the study numbers in the table characterizing included studies (with their references) - "447069_2_data_set_8320286_s47vd1.xls". Please include full study references as a column in the data file "Meta combined data.csv", so that each data point can be linked to its original source. Overall, lack of consistent and comprehensive documentation is still a major limitation to transparency and reproducibility of this meta-analysis, which should be addressed before this work is accepted.  
+
+<|ref|>text<|/ref|><|det|>[[70, 866, 848, 893]]<|/det|>
+Comment: 2. Poor coverage of grey and non- English literature should be acknowledged as limitations in the main manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[70, 893, 872, 933]]<|/det|>
+Response: We have now made it explicit that our study did not include grey or non- English literature (lines 420- 421). New reviewer comment - Round3: I can see it in the methods section now. I suggest adding this also to the header of Figure 1 (the global) map, because it is highly relevant to the pattern presented there.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 47, 928, 100]]<|/det|>
+Comment: 3. Trim- and- fill is a publication bias test not sensitivity analysis. Response: We have removed text referring to this as a sensitivity analysis (see lines 631- 632 in the main text; line 326 in the Supplementary Results). New reviewer comment - Round3: thank you.  
+
+<|ref|>text<|/ref|><|det|>[[72, 111, 920, 155]]<|/det|>
+Comment: 4. The issue of percentage and proportion data is not in the normality of the residuals, but in the fact that they are bounded between 0 and 1 (100). As such they should be arcsin- transformed to make them unbounded. All transformations can be documented in R code, so there is no issue of transparency.  
+
+<|ref|>text<|/ref|><|det|>[[72, 153, 920, 256]]<|/det|>
+Response: We have consulted with our institutional statistician (Dr Pete Henrys - CStat, Royal Statistical Society Chartered statistician, CSci, The Science Council Chartered Scientist) who states that a transformation is there to ensure model assumptions are met. Transformation purely on the basis that data is of a particular type is not appropriate, and rather transformation should only be applied on a case by case basis to address underlying issues that may affect model assumptions. He confirms that focusing on model assumption checks (as we have done) including assessing residuals represents the critical stage in the process for assessing the need for underlying processing of input data. We strongly argue that percentage and proportion data in the context of this analysis does not warrant to be arcsin- transformed without a reason relating to model distributional fit. In the context of this analysis this is not required.  
+
+<|ref|>text<|/ref|><|det|>[[72, 254, 920, 310]]<|/det|>
+New c reviewer comment - Round3: it is great that you consulted the statistician and conducted the model assumptions checks. Please provide the results of the model assumptions test in the supplementary materials. Conducting sensitivity analysis for the main model that show that applying the recommended transformation does not change the results should be also presented in the supplementary materials to support your claims.  
+
+<|ref|>text<|/ref|><|det|>[[72, 320, 916, 362]]<|/det|>
+Comment: 5. The issue with folded normal distribution when using absolute values has not been resolved. See Nakagawa, S. & Lagisz, M. (2016) Visualizing unbiased and biased unweighted meta- analyses. 334. Journal of Evolutionary Biology, 29, 1914- 1916, for a visual clarification of the issue.  
+
+<|ref|>text<|/ref|><|det|>[[72, 360, 918, 490]]<|/det|>
+Response: This relates to our standardizing of biomarkers (e.g. protein regulation, gene expression) which may respond positively or negatively to pesticides. Our approach has been to use absolute values and focus on a deviation from zero in the meta- analysis, essentially quantifying departure from the norm. Again, we emphasize that checks of model assumptions suggest that the current use of a normal distribution was robust. From a practical perspective it is not possible within our modeling framework to specify a folded distribution (which we do not think is required). However, we would be prepared to simply no longer use the absolute values thus negating the suggested need for a folded distribution. We do feel that this would potentially lose important information about a general impact of pesticides on either up- or down- regulation of biomarkers - which we think is hugely significant from the context of the impacts of pesticides on non- obvious species level effects. The author team has discussed this issue extensively and we still believe that taking absolute values biologically makes more sense here.  
+
+<|ref|>text<|/ref|><|det|>[[72, 489, 918, 544]]<|/det|>
+New reviewer comment - Round3: Please provide the results of the model assumptions test in the supplementary materials. Conducting sensitivity analysis for the key model that show that applying the recommended transformation does not change the results for biomarkers data subset should be also presented in the supplementary materials to support your claim. Also, please flip the sign on the absolute values to make it easier to compare magnitude of estimated effects in all forest plots.  
+
+<|ref|>text<|/ref|><|det|>[[72, 554, 911, 581]]<|/det|>
+Comment: 6. A synthetic phylogenetic tree can be retrieved from Open Tree of Life, which is accessible via R package rotl. As, such phylogenetic analyses can and should be conducted beyond plants.  
+
+<|ref|>text<|/ref|><|det|>[[72, 580, 925, 672]]<|/det|>
+Response: We have used rotl, a freely available package for R designed to takes advantage of the Open Tree of Life's Application Programming Interfaces (APIs) to access subtrees from the synthetic Open Tree, as well as the published source trees that contribute to the synthesis to generate comprehensive phylogenies for animals and microorganisms (see lines 296- 321 in the Supplementary Methods). In this revision, we have now conducted all the phylogenetic analyses for plants, animals and microorganisms (see detailed results in Supplementary Tables 1.3- 1.5, 2.3- 2.5, 45.3- 45.5 and 46.3- 46.5). We still note that an overall agreed- upon phylogeny for the tree of life will likely never arrive, but we did our best to address the referee's comment.  
+
+<|ref|>text<|/ref|><|det|>[[72, 671, 905, 699]]<|/det|>
+New reviewer comment - Round3: Thank you and I agree regarding the conditionality of any phylogeny used - but we will never arrive there.  
+
+<|ref|>text<|/ref|><|det|>[[72, 710, 915, 738]]<|/det|>
+Comment: 7. Publication year should be included in the analyses (meta- regression)- it is not confounded with study identity, and it is potentially important and interesting moderator of the effect.  
+
+<|ref|>text<|/ref|><|det|>[[72, 737, 920, 792]]<|/det|>
+Response: We have clarified this now as "Publication year: a continuous metric according to the year when the articles were published" (lines 528- 529 in the main text). We have included publication year in the meta- regression model (see detailed results in Supplementary Tables 1.1- 1.5, 2.1- 2.5, 45.1- 45.5 and 46.1- 46.5; Extended Data Figs. 2- 7 and other associated supplementary tables).  
+
+<|ref|>text<|/ref|><|det|>[[72, 792, 381, 805]]<|/det|>
+New reviewer comment - Round3: thank you.  
+
+<|ref|>text<|/ref|><|det|>[[72, 816, 905, 857]]<|/det|>
+Comment: 8. Funding and conflict of interest statements can be easily extracted from the included papers, even if they are not consistently included. Then, presence of potential links with the relevant industry should be analysed in a meta- regression. This is a too important issue to dismiss.  
+
+<|ref|>text<|/ref|><|det|>[[72, 856, 915, 945]]<|/det|>
+Response: We now extracted conflict- of- interest statements where possible. Among the 1705 papers, we could confirm that 1,411 ones did not have conflict of interest while 25 self- identified as having a conflict of interest. The remainder made no statement about conflicts of interest (lines 244- 248 in the main text; Supplementary Table 58). In the meta- regression analysis, we consider "Conflict of interest status" as one binary variable (i.e., "1" denotes that a certain article has conflict of interest and "0" represents no conflict) (lines 526- 528 in the main text). Detailed results for "Conflict of interest status" were presented in Supplementary Tables 1.1- 1.5, 2.1- 2.5, 45.1- 45.5, 46.1- 46.5 and 58, Extended Data Figs. 2- 7 and other associated supplementary tables).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[73, 47, 410, 61]]<|/det|>
+New comment - Round3: thank you for doing this.  
+
+<|ref|>text<|/ref|><|det|>[[73, 73, 245, 85]]<|/det|>
+New comment - Round3:  
+
+<|ref|>text<|/ref|><|det|>[[72, 86, 923, 152]]<|/det|>
+By looking at the raw data, "Meta combined data.csv", I can see "Life expectancy" and "Longevity" are in "insecticide- animal reproduction" category. It seems like all (and there are many) measures related to survival are classified as "insecticide- animal reproduction" (similar issue is also present for plants and microorganisms in the data set). Please provide justification for this choice. Also animal feeding rates could represent both growth and behavior - having them solely as a measure of growth may require additional sensitivity analyses and clarification in the methods section.  
+
+<|ref|>text<|/ref|><|det|>[[73, 165, 308, 178]]<|/det|>
+New reviewer comment - Round3:  
+
+<|ref|>text<|/ref|><|det|>[[73, 178, 920, 218]]<|/det|>
+In the "Meta combined data.csv", there are some clear mistakes too, e.g. the column for describing measurements is named "Animal growth indicator", and sometimes is empty, "number of eggs/female" is classified as "insecticide- animal behavior", this raises concerns for the consistency of data extractions and the quality control procedures in this work.  
+
+<|ref|>text<|/ref|><|det|>[[73, 231, 308, 244]]<|/det|>
+New reviewer comment - Round3:  
+
+<|ref|>text<|/ref|><|det|>[[72, 244, 921, 348]]<|/det|>
+Lines 463- 466: "When a study included different levels of pesticide application rates, measurements for the control groups without pesticides versus different pesticide application rates were considered as independent paired observations" - This is a very bold assumption given "26,096 estimates of pesticide effects reported from 1,705". Now looking at the actual data, I can see that although sometimes multiple pesticides or species were used in a single study, in most cases multiple doses were used, which means data non- independence related to the repeated use of the same control group to compare it with different dose levels is a serious issue. Specifically, by not controlling for repeated use of the same individuals, the precision of the overall estimate is very likely inflated. One very simple way is to divide sample size of the control group by the number of comparisons it is used in, thus avoiding "double- counting" of the same replicated units.  
+
+<|ref|>text<|/ref|><|det|>[[73, 360, 308, 373]]<|/det|>
+New reviewer comment - Round3:  
+
+<|ref|>text<|/ref|><|det|>[[72, 373, 916, 413]]<|/det|>
+In the main text, please acknowledge study limitations due to sample sizes in the included experiments (from the raw data I can see that the median sample size is likely 3) - having a histogram of the distribution of sample sizes by type of organism and/or treatment would be very useful too.  
+
+<|ref|>text<|/ref|><|det|>[[73, 437, 145, 450]]<|/det|>
+Version 3:  
+
+<|ref|>text<|/ref|><|det|>[[73, 464, 219, 477]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[73, 489, 160, 502]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[73, 515, 238, 528]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[73, 529, 220, 541]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[72, 541, 825, 568]]<|/det|>
+The authors addressed some of my concerns but there are still outstanding issues related to transparency and reproducibility of this work, as well as some methodological concerns.  
+
+<|ref|>text<|/ref|><|det|>[[72, 580, 890, 607]]<|/det|>
+My previous comments and authors responses are quoted below when they are linked to new issues identified. My new comments are provided as "New reviewer comment - Round4:  
+
+<|ref|>text<|/ref|><|det|>[[72, 619, 911, 715]]<|/det|>
+"New reviewer comment - Round3: Thank you, for providing a link to the files on GitHub. I note that now the data and code are also provided on Zenodo and the contents of these two repositories do not match. Specifically, some code files that are archived on Zenodo are absent from GitHub version. Please remove any special (Chinese) characters from your R code. Response: We thank the referee for pointing us at this. To avoid confusion and duplication we have now uploaded all raw data and code to Zenodo and removed them from GitHub (see this revision). We have checked that all relevant files are present. We have now also removed all Chinese characters as suggested. All data used in this analysis are available on Zenodo  
+
+<|ref|>text<|/ref|><|det|>[[72, 714, 707, 728]]<|/det|>
+(https://zenodo.org/records/10495263). All code for this analysis is available on both Zenodo  
+
+<|ref|>text<|/ref|><|det|>[[72, 728, 345, 740]]<|/det|>
+(https://zenodo.org/records/10509420).  
+
+<|ref|>text<|/ref|><|det|>[[72, 740, 921, 808]]<|/det|>
+New reviewer comment - Round4: Thank you for providing your data and code on Zenodo. It seems like putting the contents of these two repositories together still does not allow reproducing all the analyses and they are really hard to peer- review. There are three reasons: first the code is very poorly annotated, and thus it is hard to figure out what different sections of code are supposed to produce. Second, it seems like pieces of code are missing, e.g. file "Figures 2- 3. R" starts from loading the file 'Meta combined data.csv' and then immediately runs a meta analytic model on the data from it, however the file does not contain effect sizes and there is no line of code that calculates effect sizes before running the model. Third, the code refers to loading many additional data files, that look like pre- processed data, but they are not included in the Zenodo repository, e.g. "Supplementary Tables 1(1- 3). R" calls "Old_New_pesticide <-  
+
+<|ref|>text<|/ref|><|det|>[[72, 808, 916, 848]]<|/det|>
+read_excel(paste("/lustre/huyueqing/ssy/Wan2023/DataPre/", "Supplementary Data 2- Classification of old and new pesticides- 2023- 0825. xls", sep = ""), sheet = 1)", but there is no such file or no explanation where it originates from. Overall, I think all the analytic pipeline should be made fully transparent reproducible and then subjected to independent peer- review.  
+
+<|ref|>text<|/ref|><|det|>[[72, 848, 920, 888]]<|/det|>
+"New reviewer comment - Round3: Thank you for providing a README file. Unfortunately, the meta- data provided applies to only one of the data files, "Meta combined data.csv" (there seems to be 16 data files in total). The meta- data for this one key data table used to run the key models is still too rudimentary to allow cross- checking of the data or replication of the data extraction process. For example, the variable described as "Control (value): the value in control group;" does not explain
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 46, 905, 125]]<|/det|>
+which value should extracted be there, apart that it should relate to the control group. Another, example - "Control- n: the number in control group;", which is supposed to be the sample size, it also needs explaining the units of counting sample size, as these can be different for different taxa and experiment types - for animals it might be number of individuals, for plants could be individuals or plots, for microorganisms it could be a colony in a tube/plate, or number of natural plots, etc. Response: We have now revised the README file to improve clarity. We hope this addresses the problems."  
+
+<|ref|>text<|/ref|><|det|>[[72, 124, 920, 192]]<|/det|>
+New reviewer comment - Round 4: Thank you. The README file only describes data in the main data file, but not in any of the additional data files and does not provide any information about the project in general and how the code and files should be used to reproduce the analyses. Please provide more comprehensive description of all your files, starting from overview and then providing meta- data for all data that is needed to reproduce your work. Consider using Quarto with R to produce computationally reproducible documents that contain the R code, verbal descriptions, and outputs.  
+
+<|ref|>text<|/ref|><|det|>[[72, 202, 880, 243]]<|/det|>
+"Comment: One column in the data table called "LnR" is not described - it seems to be an effect size that was not be used in the analyses, as the main effect size used was SMD. However, I cannot find the code the authors used for calculating SMD.  
+
+<|ref|>text<|/ref|><|det|>[[72, 242, 915, 297]]<|/det|>
+Response: The column in the data table called "LnR" had not been used in the analyses, and we have now deleted it. The code used for calculating SMD has been added in the README file." New reviewer comment - Round 4: The code used for calculating SMD should not be placed in the README file, but in the code where it is needed to calculate effect sizes to reproduce the results.  
+
+<|ref|>text<|/ref|><|det|>[[72, 319, 911, 504]]<|/det|>
+"Comment: Further, the study codes in the main data file "Meta combined data.csv" do not match the study numbers in the table characterizing included studies (with their references) - "447069_2_data_set_8320286_s47vd1.xls". Please include full study references as a column in the data file "Meta combined data.csv", so that each data point can be linked to its original source. Overall, lack of consistent and comprehensive documentation is still a major limitation to transparency and reproducibility of this meta-analysis, which should be addressed before this work is accepted. Response: We thank the referee for pointing us at this. We have now made sure that all study codes match up. We have listed all the 1705 study references in the file "Meta combined data" with a new sheet (Full study references). The paper code order in "Meta combined data" is based on the time order that we extracted data, and the paper order in Supplementary Data 1 is based on the authors' name in alphabetical order (see Meta combined data). In addition, we have added another data file "ExtendData" for other supplementary files (e.g., for Supplementary Tables 59- 61, Supplementary Figures 24- 27)." New reviewer comment - Round 4: Thank you for adding the study references next to study codes in the data file. I was now able to compare the extracted data with the information reported in the original papers. I randomly selected 4 papers I have full- text access to and found the following issues:  
+
+<|ref|>text<|/ref|><|det|>[[72, 514, 911, 620]]<|/det|>
+- Paper 16. (Colin et al. 2004. A Method to Quantify and Analyze the Foraging Activity of Honey Bees: Relevance to the Sublethal Effects Induced by Systemic Insecticides): rows 16-21: the concentration of Fipronil is wrong - the study reports using 2 micrograms per kilogram while the extracted data is 6 g/kg. For Imidacloprid the unit is also wrong - it should be microgram per kilogram not gram per kilogram. For the outcome "number of daily attendance of active bees-nuclei" which was extracted for nuclei A, B, C, it is and somehow from control nuclei, but I cannot find the matching exact numbers in the text, and it is not possible to extract them from the figures in the text. Given the complexity of the experimental design, raw data would need to be used to calculate such values. There are no comments in the data file how the numbers were obtained, and thus they are not reproducible.  
+
+<|ref|>text<|/ref|><|det|>[[72, 618, 915, 698]]<|/det|>
+- Paper 1628. (Zhao X, Li YL, Zhang LX, Dorna H, 2003. Effects of priming and fungicide treatment on germination of china aster (Callistephus chinensis L.) seeds. Seed Science & Technology, 32: 451-457.): rows 1943-1944: it looks like only germination capacity data from Sample 1 and only from 30C temperature treatment were extracted and the standard deviations were manually imputed by assuming they are 10% of the measured value. This is only a small fraction of available data in this paper (multiple control and exposure groups, all samples should be extracted and used for calculating mean and SD) and is incorrectly extracted.  
+
+<|ref|>text<|/ref|><|det|>[[72, 696, 915, 777]]<|/det|>
+- Paper 1008. (Gabaston J, Khawand TE, Wafo-Teguo P, Decendit A, Richard T, Merillon JM, Pavel A, 2018. Stilbenes from grapevine root: a promising natural insecticide against Leptinotarsa decimlineata. Journal of Pest Science, 91: 897-906.) rows 5808-5809: looks like data was extracted from Table 6 for 3rd day of exposure only and two lower doses out of 3 used (this is not reported on the data extraction sheet). The mean mortality % was converted to survival %, but the reported standard deviation values were somehow replaced with 10% of the men value. The sample size is claimed to be 4, but I cannot find such information in the article, and it seems improbable given the type of organism, and reported values.  
+
+<|ref|>text<|/ref|><|det|>[[72, 787, 915, 880]]<|/det|>
+Overall, this small exercise clearly shows that 3/3 papers had some issues with data extractions. This indicates a very high rate of errors in the whole data set. The data lacks information on the exact sources of the extracted values for the outcomes at least (table/figure number, text page, raw data, etc.), any extractions decisions made, calculations, imputations, etc. - this makes assessing trustworthiness of the data set very hard and makes data extractions irreproducible. Because of this I request that all extracted data is annotated in detail. Further it has to be cross-checked by a researcher who was not involved in original extractions, re- extracted where issues are found and clearly described to make it fully transparent and trustworthy.  
+
+<|ref|>text<|/ref|><|det|>[[72, 891, 922, 945]]<|/det|>
+New reviewer comment - Round 4: When checking the extracted data, I noticed that many of the extracted standard deviation values are exactly 10% of the mean value. This cannot be coincidence and the manuscript does not mention doing manual imputations for missing standard deviation values (or any other missing values). Please explain how these values originated and the detailed procedures used to deal with any missing data.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 73, 916, 245]]<|/det|>
+"New reviewer comment - Round 3: it is great that you consulted the statistician and conducted the model assumptions checks. Please provide the results of the model assumptions test in the supplementary materials. Conducting sensitivity analysis for the main model that show that applying the recommended transformation does not change the results should be also presented in the supplementary materials to support your claims. Response: As suggested, we now include a sensitivity analysis where we apply arcsine transformation for percentage and proportion data (as you suggest) and compare the results to our approach (no transformation or proportion/percentage data) (see Supplementary Table 59). This shows no qualitative difference in the results. We also provide a supplementary figure in the supporting information for the model checks of the original analysis (see Supplementary Fig. 24). Original comment for context: 5. The issue with folded normal distribution when using absolute values has not been resolved. See Nakagawa, S. & Lagisz, M. (2016) Visualizing unbiased and biased unweighted meta-analyses. 334. Journal of Evolutionary Biology, 29, 1914- 1916, for a visual clarification of the issue."  
+
+<|ref|>text<|/ref|><|det|>[[73, 245, 312, 258]]<|/det|>
+New reviewer comment - Round 4:  
+
+<|ref|>text<|/ref|><|det|>[[73, 256, 911, 322]]<|/det|>
+The code for reproducing this should be in the file "Supplementary Tables 59- 61. R", and as such cannot be rerun. In the provided code I cannot find any line that would execute arcsine transformation for proportion data. And the code starts from loading a mysterious data file that is not archived on Zenodo, so can't tell what is there nor how it got there: AllData <- read.table(paste(DataPre_path,"SuppTab59Data.txt",sep=""),header = T). This seems to be ongoing issues across the provided code which is compounded by almost complete lack of code comments and descriptions.  
+
+<|ref|>text<|/ref|><|det|>[[72, 346, 920, 460]]<|/det|>
+"New reviewer comment - Round 3: Lines 463- 466: "When a study included different levels of pesticide application rates, measurements for the control groups without pesticides versus different pesticide application rates were considered as independent paired observations" - This is a very bold assumption given "26,096 estimates of pesticide effects reported from 1,705". Now looking at the actual data, I can see that although sometimes multiple pesticides or species were used in a single study, in most cases multiple doses were used, which means data non- independence related to the repeated use of the same control group to compare it with different dose levels is a serious issue. Specifically, by not controlling for repeated use of the same individuals, the precision of the overall estimate is very likely inflated. One very simple way is to divide sample size of the control group by the number of comparisons it is used in, thus avoiding "double- counting" of the same replicated units.  
+
+<|ref|>text<|/ref|><|det|>[[72, 460, 884, 603]]<|/det|>
+Response: Observations with no pesticide had been considered as control groups, while those with different pesticide application rates were considered as the treatment groups. When an article included different levels of pesticide application rates, measurements for the control groups (no pesticide) were compared to all other treatment levels of pesticide application rates and treated as independent paired observations. Within the context of these analysis, this reflects typical experimental designs for studies considered in ecotoxicology that have a single control which is then compared to multiple test treatment residue levels. The random allocation of treatments to replicates, combined with study as a random effect in the meta- analytical models, means that these comparisons are appropriate and do not represent double accounting as suggested. In response to the referee's comment, we have now re- run our analyses with the correction suggested, i.e. dividing the sample size of the control group by the number of comparisons it had been used in.  
+
+<|ref|>text<|/ref|><|det|>[[72, 603, 746, 668]]<|/det|>
+The results of this additional analysis can now be found in a new Supplementary Table 62. When comparing these results to our former approach (see Supplementary Table 3), there were no qualitative differences in the results (see Supplementary Table 3 vs. Supplementary Table 62). In particular, all negative effect sizes remained negative (and significant), and all positive effect sizes remained positive and significant.  
+
+<|ref|>text<|/ref|><|det|>[[72, 669, 923, 790]]<|/det|>
+New reviewer comment - Round 4: The authors are wrong when stating that "The random allocation of treatments to replicates, combined with study as a random effect in the meta- analytical models, means that these comparisons are appropriate and do not represent double accounting as suggested." The measures from the same control individual are used repeatedly in comparisons of effects of different doses of the same pesticides and they are such effect sizes are not independent and are a form of "double- counting". The analyses that account for such non- independence should be used as main models not as supplementary. The author used sample- size corrections their sensitivity analyses for a few models. I suggest testing their all models using cluster- robust approach which is already implemented in metafor package (https://wviechtb.github.io/metafor/reference/robust.html) as their main approach. Please also run and report an overall (global) meta- analytic model on all data with organism, trait and pesticide type as random effects.  
+
+<|ref|>text<|/ref|><|det|>[[72, 814, 900, 853]]<|/det|>
+"New reviewer comment - Round 3: In the main text, please acknowledge study limitations due to sample sizes in the included experiments (from the raw data I can see that the median sample size is likely 3) - having a histogram of the distribution of sample sizes by type of organism and/or treatment would be very useful too.  
+
+<|ref|>text<|/ref|><|det|>[[72, 852, 905, 891]]<|/det|>
+Response: The median value of sample size was 4 and 4 for the control and treatment, respectively (see lines 122- 123 in the main text). In addition, we have added a histogram to show the distribution of sample sizes by types of organisms and separately by control and treatment (see Supplementary Fig. 27)."  
+
+<|ref|>text<|/ref|><|det|>[[72, 891, 920, 944]]<|/det|>
+New reviewer comment - Round 4: Thank you for adding information about the median sample sizes. However, this raises new issue since Line 562 states "A large- sample approximation was used to compute the sampling variances46. " This approximation is not suitable for such small sample sizes, and a small- size- corrected version of the SMD (Hedges' g) should be used as an effect size in this meta- analysis.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 60, 338, 87]]<|/det|>
+(Remarks on code availability) code incomplete and poorly annotated  
+
+<|ref|>text<|/ref|><|det|>[[73, 100, 144, 112]]<|/det|>
+Version 4:  
+
+<|ref|>text<|/ref|><|det|>[[73, 125, 220, 139]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[73, 152, 161, 164]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[73, 177, 238, 190]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 191, 916, 244]]<|/det|>
+Thank you for your patience and hard work on this manuscript. I am pleased to note that it is much more transparent and robust now (it seems to have some new interesting findings as well). I really hope that in your next meta- analytic project you will take all steps from its inception to make sure it is trustworthy and well- documented, so I dont need to do 5 rounds of review again... good luck  
+
+<|ref|>text<|/ref|><|det|>[[72, 597, 916, 650]]<|/det|>
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+<|ref|>text<|/ref|><|det|>[[72, 650, 800, 664]]<|/det|>
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source.  
+
+<|ref|>text<|/ref|><|det|>[[72, 664, 911, 716]]<|/det|>
+The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.  
+
+<|ref|>text<|/ref|><|det|>[[72, 716, 618, 729]]<|/det|>
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[150, 85, 408, 99]]<|/det|>
+## Response to the referee comments  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 106, 327, 121]]<|/det|>
+## Replies to Reviewer #1:  
+
+<|ref|>text<|/ref|><|det|>[[148, 127, 849, 163]]<|/det|>
+Comment: The authors have put significant effort into revising their work, which is substantially improved now.  
+
+<|ref|>text<|/ref|><|det|>[[148, 170, 719, 186]]<|/det|>
+Response: We thank the reviewer for this positive evaluation (see this revision).  
+
+<|ref|>text<|/ref|><|det|>[[148, 213, 850, 250]]<|/det|>
+Comment: However, there are some remaining concerns. 1. The GitHub report now contains the code but not the data to re- run the code.  
+
+<|ref|>text<|/ref|><|det|>[[148, 256, 771, 293]]<|/det|>
+Response: We have uploaded both code and data to GitHub (https://github.com/Liwan- Fu/Impact- of- pesticides).  
+
+<|ref|>text<|/ref|><|det|>[[148, 320, 850, 379]]<|/det|>
+Comment: Also, there are multiple R files and no description how to use them to recreate study results- this should be described in detail in the README file. Meta- data should be also provided for all used variables.  
+
+<|ref|>text<|/ref|><|det|>[[148, 385, 850, 422]]<|/det|>
+Response: We have provided a README file for all used variables and uploaded the file to GitHub (https://github.com/Liwan- Fu/Impact- of- pesticides).  
+
+<|ref|>text<|/ref|><|det|>[[148, 449, 850, 486]]<|/det|>
+Comment: 2. Poor coverage of grey and non- English literature should be acknowledged as limitations in the main manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[148, 492, 850, 529]]<|/det|>
+Response: We have now made it explicit that our study did not include grey or non- English literature (lines 420- 421).  
+
+<|ref|>text<|/ref|><|det|>[[148, 556, 690, 572]]<|/det|>
+Comment: 3. Trim- and- fill is a publication bias test not sensitivity analysis.  
+
+<|ref|>text<|/ref|><|det|>[[148, 578, 850, 615]]<|/det|>
+Response: We have removed text referring to this as a sensitivity analysis (see lines 631- 632 in the main text; line 326 in the Supplementary Results).  
+
+<|ref|>text<|/ref|><|det|>[[148, 642, 850, 723]]<|/det|>
+Comment: 4. The issue of percentage and proportion data is not in the normality of the residuals, but in the fact that they are bounded between 0 and 1 (100). As such they should be acris- transformed to make them unbounded. All transformations can be documented in R code, so there is no issue of transparency.  
+
+<|ref|>text<|/ref|><|det|>[[147, 729, 852, 896]]<|/det|>
+Response: We have consulted with our institutional statistician (Dr Pete Henrys - CStat, Royal Statistical Society Chartered statistician, CSci, The Science Council Chartered Scientist) who states that a transformation is there to ensure model assumptions are met. Transformation purely on the basis that data is of a particular type is not appropriate, and rather transformation should only be applied on a case by case basis to address underlying issues that may affect model assumptions. He confirms that focusing on model assumption checks (as we have done) including assessing residuals represents the critical stage in the process for assessing the need for underlying processing of input data. We strongly argue that percentage and proportion data in the context of this analysis does not
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[148, 84, 850, 121]]<|/det|>
+warrant to be arcsin- transformed without a reason relating to model distributional fit. In the context of this analysis this is not required.  
+
+<|ref|>text<|/ref|><|det|>[[147, 147, 852, 466]]<|/det|>
+Comment: 5. The issue with folded normal distribution when using absolute values has not been resolved. See Nakagawa, S. & Lagisz, M. (2016) Visualizing unbiased and biased unweighted meta- analyses. 334. Journal of Evolutionary Biology, 29, 1914- 1916, for a visual clarification of the issue. Response: This relates to our standardizing of biomarkers (e.g. protein regulation, gene expression) which may respond positively or negatively to pesticides. Our approach has been to use absolute values and focus on a deviation from zero in the meta- analysis, essentially quantifying departure from the norm. Again, we emphasize that checks of model assumptions suggest that the current use of a normal distribution was robust. From a practical perspective it is not possible within our modeling framework to specify a folded distribution (which we do not think is required). However, we would be prepared to simply no longer use the absolute values thus negating the suggested need for a folded distribution. We do feel that this would potentially lose important information about a general impact of pesticides on either up- or down- regulation of biomarkers - which we think is hugely significant from the context of the impacts of pesticides on non- obvious species level effects. The author team has discussed this issue extensively and we still believe that taking absolute values biologically makes more sense here.  
+
+<|ref|>text<|/ref|><|det|>[[148, 492, 851, 551]]<|/det|>
+Comment: 6. A synthetic phylogenetic tree can be retrieved from Open Tree of Life, which is accessible via R package rotl. As, such phylogenetic analyses can and should be conducted beyond plants.  
+
+<|ref|>text<|/ref|><|det|>[[147, 556, 852, 724]]<|/det|>
+Response: We have used rotl, a freely available package for R designed to takes advantage of the Open Tree of Life's Application Programming Interfaces (APIs) to access subtrees from the synthetic Open Tree, as well as the published source trees that contribute to the synthesis to generate comprehensive phylogenies for animals and microorganisms (see lines 296- 321 in the Supplementary Methods). In this revision, we have now conducted all the phylogenetic analyses for plants, animals and microorganisms (see detailed results in Supplementary Tables 1.3- 1.5, 2.3- 2.5, 45.3- 45.5 and 46.3- 46.5). We still note that an overall agreed- upon phylogeny for the tree of life will likely never arrive, but we did our best to address the referee's comment.  
+
+<|ref|>text<|/ref|><|det|>[[147, 750, 852, 895]]<|/det|>
+Comment: 7. Publication year should be included in the analyses (meta- regression)- it is not confounded with study identity, and it is potentially important and interesting moderator of the effect. Response: We have clarified this now as "Publication year: a continuous metric according to the year when the articles were published" (lines 528- 529 in the main text). We have included publication year in the meta- regression model (see detailed results in Supplementary Tables 1.1- 1.5, 2.1- 2.5, 45.1- 45.5 and 46.1- 46.5; Extended Data Figs. 2- 7 and other associated supplementary tables).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 104, 852, 163]]<|/det|>
+Comment: 8. Funding and conflict of interest statements can be easily extracted from the included papers, even if they are not consistently included. Then, presence of potential links with the relevant industry should be analysed in a meta- regression. This is a too important issue to dismiss.  
+
+<|ref|>text<|/ref|><|det|>[[147, 168, 850, 338]]<|/det|>
+Comment: 8. Funding and conflict of interest statements can be easily extracted from the included papers, even if they are not consistently included. Then, presence of potential links with the relevant industry should be analysed in a meta- regression. This is a too important issue to dismiss.Response: We now extracted conflict- of- interest statements where possible. Among the 1705 papers, we could confirm that 1,411 ones did not have conflict of interest while 25 self- identified as having a conflict of interest. The remainder made no statement about conflicts of interest (lines 244- 248 in the main text; Supplementary Table 58). In the meta- regression analysis, we consider "Conflict of interest status" as one binary variable (i.e., "1" denotes that a certain article has conflict of interest and "0" represents no conflict) (lines 526- 528 in the main text). Detailed results for "Conflict of interest status" were presented in Supplementary Tables 1.1- 1.5, 2.1- 2.5, 45.1- 45.5, 46.1- 46.5 and 58, Extended Data Figs. 2- 7 and other associated supplementary tables).  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 385, 327, 401]]<|/det|>
+## Replies to Reviewer #3:  
+
+<|ref|>text<|/ref|><|det|>[[148, 406, 851, 465]]<|/det|>
+Comment: The current ms derives from a previous version submitted to [Redacted] that has been commented by three reviewers. I now have checked for the implementation of the comments of Reviewers 2 and 3 in the revised version.  
+
+<|ref|>text<|/ref|><|det|>[[148, 470, 851, 530]]<|/det|>
+Response: We thank the reviewer for this positive evaluation. We carefully addressed all comments in revising the manuscript. We are grateful for the suggestions, which have improved the quality of our study (see this revision).  
+
+<|ref|>text<|/ref|><|det|>[[147, 555, 851, 636]]<|/det|>
+Comment: To my opinion, almost all comments of Reviewer 2 have been adequately addressed, except for "(1) Comment: 216- 217 were vertebrates and invertebrates equally affected, or one more than the other? Does that depend on the intended target of an pesticide, e.g. distinguishing between pesticides aimed at invertebrates or otherwise?  
+
+<|ref|>text<|/ref|><|det|>[[148, 642, 851, 702]]<|/det|>
+Response: We have removed the sentence due to space limited, but from the statistical values in Supplementary Table 4, we can see that vertebrates and invertebrates were equally affected (see Supplementary Table 4).  
+
+<|ref|>text<|/ref|><|det|>[[148, 707, 850, 745]]<|/det|>
+Response: We have added this information in the main text, and we have clarified it as "vertebrates and invertebrates were equally affected by pesticides" (lines 258- 259 in the main text).  
+
+<|ref|>text<|/ref|><|det|>[[148, 771, 712, 788]]<|/det|>
+Comment: and "(2) Comment: 223 'affected different taxonomic groups' how?  
+
+<|ref|>text<|/ref|><|det|>[[148, 793, 851, 874]]<|/det|>
+Response: We have removed the sentence due to space limited, but from the statistical values in Supplementary Tables 32- 36, we can see that "insecticides, fungicides and herbicides affected different taxonomic groups with decreased growth, reproduction or behavior and with perturbed biomarkers" (see Supplementary Tables 32- 36)."  
+
+<|ref|>text<|/ref|><|det|>[[147, 879, 850, 896]]<|/det|>
+Response: We have added this information in the main text, and we have clarified it as
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[148, 84, 850, 143]]<|/det|>
+“Insecticides, fungicides and herbicides affected different taxonomic groups though decreased growth, reproduction or behavioural responses, and biomarkers being perturbed from baseline conditions (see Supplementary Tables 32–36)” (lines 216- 218 in the main text).  
+
+<|ref|>text<|/ref|><|det|>[[148, 170, 850, 208]]<|/det|>
+Comment: I feel that both informations are important and should be kept within the main body of text rather than being implied to be taken from the supplementary material.  
+
+<|ref|>text<|/ref|><|det|>[[149, 214, 459, 229]]<|/det|>
+Response: Please see above two responses.  
+
+<|ref|>text<|/ref|><|det|>[[148, 255, 835, 293]]<|/det|>
+Comment: All comments of Reviewer 3 have been perfectly addressed in this version of the ms. Response: Thanks for your positive comments.  
+
+<|ref|>text<|/ref|><|det|>[[148, 320, 794, 358]]<|/det|>
+Comment: In the Competing Interests paragraph, "B.W." should be replaced by "B.A.W.". Response: Done as suggested (lines 717, 720 and 721).
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[149, 105, 407, 120]]<|/det|>
+## Response to the referee comments  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 127, 327, 142]]<|/det|>
+## Replies to Reviewer #1:  
+
+<|ref|>text<|/ref|><|det|>[[149, 148, 849, 187]]<|/det|>
+Note we only show responses here for outstanding issues you identified in your previous review. Issues you identified as being addressed to your satisfaction we do not refer to here.  
+
+<|ref|>text<|/ref|><|det|>[[148, 191, 851, 272]]<|/det|>
+New reviewer comment - Round3: Thank you, for providing a link to the files on GitHub. I note that now the data and code are also provided on Zenodo and the contents of these two repositories do not match. Specifically, some code files that are archived on Zenodo are absent from GitHub version. Please remove any special (Chinese) characters from your R code.  
+
+<|ref|>text<|/ref|><|det|>[[147, 277, 851, 401]]<|/det|>
+Response: We thank the referee for pointing us at this. To avoid confusion and duplication we have now uploaded all raw data and code to Zenodo and removed them from GitHub (see this revision). We have checked that all relevant files are present. We have now also removed all Chinese characters as suggested. All data used in this analysis are available on Zenodo (https://zenodo.org/records/10495263). All code for this analysis is available on both Zenodo (https://zenodo.org/records/10509420).  
+
+<|ref|>text<|/ref|><|det|>[[147, 426, 852, 637]]<|/det|>
+New reviewer comment - Round3: Thank you for providing a README file. Unfortunately, the meta- data provided applies to only one of the data files, "Meta combined data.csv" (there seems to be 16 data files in total). The meta- data for this one key data table used to run the key models is still too rudimentary to allow cross- checking of the data or replication of the data extraction process. For example, the variable described as "Control (value): the value in control group;" does not explain which value should extracted be there, apart that it should relate to the control group. Another, example - "Control- n: the number in control group;", which is supposed to be the sample size, it also needs explaining the units of counting sample size, as these can be different for different taxa and experiment types - for animals it might be number of individuals, for plants could be individuals or plots, for microorganisms it could be a colony in a tube/plate, or number of natural plots, etc.  
+
+<|ref|>text<|/ref|><|det|>[[148, 642, 850, 680]]<|/det|>
+Response: We have now revised the README file to improve clarity. We hope this addresses the problems.  
+
+<|ref|>text<|/ref|><|det|>[[148, 706, 851, 766]]<|/det|>
+Comment: One column in the data table called "LnR" is not described - it seems to be an effect size that was not be used in the analyses, as the main effect size used was SMD. However, I cannot find the code the authors used for calculating SMD.  
+
+<|ref|>text<|/ref|><|det|>[[148, 771, 850, 809]]<|/det|>
+Response: The column in the data table called "LnR" had not been used in the analyses, and we have now deleted it. The code used for calculating SMD has been added in the README file.  
+
+<|ref|>text<|/ref|><|det|>[[148, 836, 850, 875]]<|/det|>
+Comment: Also, I cannot find the code for Figure 1 (the map), and instead there is code for Figure 1b, a histogram which is not in the main manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[145, 879, 850, 896]]<|/det|>
+Response: Figure 1 had been prepared in ArcGIS, so no R code can be provided for that. We have
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[148, 84, 757, 100]]<|/det|>
+now deleted code for “Figure 1b” which was part of a prior version of the manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[147, 125, 852, 272]]<|/det|>
+Comment: Further, the study codes in the main data file “Meta combined data.csv” do not match the study numbers in the table characterizing included studies (with their references) – “447069_2_data_set_8320286_s47vd1.xls”. Please include full study references as a column in the data file “Meta combined data.csv”, so that each data point can be linked to its original source. Overall, lack of consistent and comprehensive documentation is still a major limitation to transparency and reproducibility of this meta-analysis, which should be addressed before this work is accepted.  
+
+<|ref|>text<|/ref|><|det|>[[147, 277, 852, 423]]<|/det|>
+Response: We thank the referee for pointing us at this. We have now made sure that all study codes match up. We have listed all the 1705 study references in the file “Meta combined data” with a new sheet (Full study references). The paper code order in “Meta combined data” is based on the time order that we extracted data, and the paper order in Supplementary Data 1 is based on the authors' name in alphabetical order (see Meta combined data). In addition, we have added another data file “ExtendData” for other supplementary files (e.g., for Supplementary Tables 59- 61, Supplementary Figures 24- 27).  
+
+<|ref|>text<|/ref|><|det|>[[148, 449, 850, 508]]<|/det|>
+New reviewer comment - Round 3: I can see it in the methods section now. I suggest adding this [grey and non- English literature] also to the header of Figure 1 (the global) map, because it is highly relevant to the pattern presented there.  
+
+<|ref|>text<|/ref|><|det|>[[148, 514, 850, 551]]<|/det|>
+Response: We have added “Grey and non- English literature was not included in this meta- analysis” in the caption of Figure 1.  
+
+<|ref|>text<|/ref|><|det|>[[147, 577, 850, 680]]<|/det|>
+New reviewer comment - Round 3: it is great that you consulted the statistician and conducted the model assumptions checks. Please provide the results of the model assumptions test in the supplementary materials. Conducting sensitivity analysis for the main model that show that applying the recommended transformation does not change the results should be also presented in the supplementary materials to support your claims.  
+
+<|ref|>text<|/ref|><|det|>[[147, 686, 850, 787]]<|/det|>
+Response: As suggested, we now include a sensitivity analysis where we apply arcsine transformation for percentage and proportion data (as you suggest) and compare the results to our approach (no transformation or proportion/percentage data) (see Supplementary Table 59). This shows no qualitative difference in the results. We also provide a supplementary figure in the supporting information for the model checks of the original analysis (see Supplementary Fig. 24).  
+
+<|ref|>text<|/ref|><|det|>[[147, 813, 850, 895]]<|/det|>
+Original comment for context: 5. The issue with folded normal distribution when using absolute values has not been resolved. See Nakagawa, S. & Lagisz, M. (2016) Visualizing unbiased and biased unweighted meta- analyses. 334. Journal of Evolutionary Biology, 29, 1914- 1916, for a visual clarification of the issue. New reviewer comment - Round 3: Please provide the results of the
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 84, 852, 186]]<|/det|>
+model assumptions test in the supplementary materials. Conducting sensitivity analysis for the key model that show that applying the recommended transformation does not change the results for biomarkers data subset should be also presented in the supplementary materials to support your claim. Also, please flip the sign on the absolute values to make it easier to compare magnitude of estimated effects in all forest plots.  
+
+<|ref|>text<|/ref|><|det|>[[147, 192, 852, 574]]<|/det|>
+Response: As suggested, we now present additiopal model checks in Supplementary Fig. 25 that show that the distribution used was appropriate (see Supplementary Fig. 25). However, we acknowledge the referee's wider concern about the use of absolute values for biomarker responses. As already pointed out, biomarkers include metrics that may potentially be both be up- and downregulated as a response to exposure to pesticides. In both cases, this would represent biologically meaningful responses that may have longer term fitness consequences for individual populations. To account for the potential of both up- and down- regulation of biomarkers, we have treated these as absolute values within the main analyses. To address the referee's comment, and to support a broader interpretation of the data, we now provide a comparative sensitivity analysis for the response of biomarkers to overall pesticide exposure for animals, plants, and microorganisms using absolute values of 'biomarker' (as described in the methods) and the original raw values of 'biomarker', i.e. with directionality of the effect (see Supplementary Table 60). As can be seen, even when the biomarkers are not treated as absolute values, they are still on average characterized by negative effects in response to pesticide exposure so that the results remain qualitatively equivalent to those presented in the main paper. This would suggest that while biomarkers may in principal be both up and down regulated following pesticide exposure, across the studies considered this effect remains overwhelmingly down regulation. We hope that this addresses your wider concerns about the biomarker responses.  
+
+<|ref|>text<|/ref|><|det|>[[147, 600, 852, 702]]<|/det|>
+New comment - Round 3: By looking at the raw data, "Meta combined data.csv", I can see "Life expectancy" and "Longevity" are in "insecticide- animal reproduction" category. It seems like all (and there are many) measures related to survival are classified as "insecticide- animal reproduction" (similar issue is also present for plants and microorganisms in the data set). Please provide justification for this choice.  
+
+<|ref|>text<|/ref|><|det|>[[147, 708, 852, 896]]<|/det|>
+Response: We acknowledge that to some extent the categories we had used for reproduction, growth, behaviour and biomarkers are synthetic, but we believe they still allow the most meaningful interpretation of the plethora of metrics measured across the 1705 studies as responses to pesticide exposure. While further sub- divisions of these categories could have been undertaken, this would not have added to the clarity of the manuscript. As indicated, certain responses to pesticides were allocated to these four categories based on the authors' combined opinion, as well as though solicitation of opinion for colleagues. As such, "Life expectancy" and "Longevity" were considered as "animal reproduction" as premature death may limit lifetime potential reproductive output. Similarly, we think that animal survival determines population reproduction of animals, and thus
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[148, 83, 852, 185]]<|/det|>
+should be considered as "animal reproduction". There are no "Life expectancy" and "Longevity" measures for plants or microorganisms. The overwhelming negative effects (or positive in the case of the use of absolute values for biomarkers) across reproduction, growth, behavior and biomarkers suggests that, while further sub- categorization of these responses could have been undertaken, this would not have changed the results.  
+
+<|ref|>text<|/ref|><|det|>[[148, 212, 851, 270]]<|/det|>
+New comment - Round 3: Also animal feeding rates could represent both growth and behavior - having them solely as a measure of growth may require additional sensitivity analyses and clarification in the methods section.  
+
+<|ref|>text<|/ref|><|det|>[[147, 277, 852, 444]]<|/det|>
+Response: The suggestion of the reviewer is correct as animal feeding rates could represent both growth and feeding behavior. Here, we assumed that animal feeding is primarily for animal growth, and thus we put "animal feeding" indicator into "animal growth" in this meta analysis. We do however follow the referee's suggestion and apply a sensitivity analysis to assess the effect of allocation of feeding to either growth or behavior in our meta analysis. We have added a supplementary table and a supplementary figure to measure the sensitivity analyses of animal growth and animal behavior when "animal feeding" indicator was included in animal growth or in animal behaviour (see Supplementary Table 61 and Supplementary Fig. 26).  
+
+<|ref|>text<|/ref|><|det|>[[147, 470, 852, 572]]<|/det|>
+New reviewer comment - Round 3: In the "Meta combined data.csv", there are some clear mistakes too, e.g. the column for describing measurements is named "Animal growth indicator", and sometimes is empty, "number of eggs/female" is classified as "insecticide- animal behavior", this raises concerns for the consistency of data extractions and the quality control procedures in this work.  
+
+<|ref|>text<|/ref|><|det|>[[147, 578, 852, 722]]<|/det|>
+Response: We have addressed these minor errors and checked that this has not affected data extraction. In "Meta combined data", we have replaced the column "Animal growth indicator" with a new column ("Non- target organism indicator"), added the missing experimental year (i.e., 2017) in rows 7849- 7853, and have added the missing unit (i.e., Individuals/g) in row 14201. In "Meta combined data.csv", we have replaced "number of eggs/female" with "Walking distance (cm)" in row 8051, and have replaced "number of eggs/female" with "Walking speed (10- 2 cm s- 1)" in row 8053. Both "Walking distance" and "Walking speed" belong to "animal behavior".  
+
+<|ref|>text<|/ref|><|det|>[[147, 750, 852, 896]]<|/det|>
+New reviewer comment - Round 3: Lines 463- 466: "When a study included different levels of pesticide application rates, measurements for the control groups without pesticides versus different pesticide application rates were considered as independent paired observations" - This is a very bold assumption given "26,096 estimates of pesticide effects reported from 1,705". Now looking at the actual data, I can see that although sometimes multiple pesticides or species were used in a single study, in most cases multiple doses were used, which means data non- independence related to the repeated use of the same control group to compare it with different dose levels is a serious issue.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[148, 84, 852, 164]]<|/det|>
+Specifically, by not controlling for repeated use of the same individuals, the precision of the overall estimate is very likely inflated. One very simple way is to divide sample size of the control group by the number of comparisons it is used in, thus avoiding "double- counting" of the same replicated units.  
+
+<|ref|>text<|/ref|><|det|>[[147, 170, 852, 401]]<|/det|>
+Response: Observations with no pesticide had been considered as control groups, while those with different pesticide application rates were considered as the treatment groups. When an article included different levels of pesticide application rates, measurements for the control groups (no pesticide) were compared to all other treatment levels of pesticide application rates and treated as independent paired observations. Within the context of these analysis, this reflects typical experimental designs for studies considered in ecotoxicology that have a single control which is then compared to multiple test treatment residue levels. The random allocation of treatments to replicates, combined with study as a random effect in the meta- analytical models, means that these comparisons are appropriate and do not represent double accounting as suggested. In response to the referee's comment, we have now re- run our analyses with the correction suggested, i.e. dividing the sample size of the control group by the number of comparisons it had been used in.  
+
+<|ref|>text<|/ref|><|det|>[[147, 406, 852, 509]]<|/det|>
+The results of this additional analysis can now be found in a new Supplementary Table 62. When comparing these results to our former approach (see Supplementary Table 3), there were no qualitative differences in the results (see Supplementary Table 3 vs. Supplementary Table 62). In particular, all negative effect sizes remained negative (and significant), and all positive effect sizes remained positive and significant.  
+
+<|ref|>text<|/ref|><|det|>[[147, 535, 851, 616]]<|/det|>
+New reviewer comment - Round 3: In the main text, please acknowledge study limitations due to sample sizes in the included experiments (from the raw data I can see that the median sample size is likely 3) - having a histogram of the distribution of sample sizes by type of organism and/or treatment would be very useful too.  
+
+<|ref|>text<|/ref|><|det|>[[147, 621, 851, 702]]<|/det|>
+Response: The median value of sample size was 4 and 4 for the control and treatment, respectively (see lines 122- 123 in the main text). In addition, we have added a histogram to show the distribution of sample sizes by types of organisms and separately by control and treatment (see Supplementary Fig. 27).
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[150, 105, 406, 120]]<|/det|>
+## Response to the referee comments  
+
+<|ref|>sub_title<|/ref|><|det|>[[150, 127, 327, 142]]<|/det|>
+## Replies to Reviewer #1:  
+
+<|ref|>text<|/ref|><|det|>[[148, 148, 852, 230]]<|/det|>
+The authors addressed some of my concerns but there are still outstanding issues related to transparency and reproducibility of this work, as well as some methodological concerns. My previous comments and authors responses are quoted below when they are linked to new issues identified. My new comments are provided as "New reviewer comment - Round4:  
+
+<|ref|>text<|/ref|><|det|>[[148, 234, 854, 337]]<|/det|>
+Response: We thank you for these suggestions to improve transparency and reproducibility of our manuscript. We have now revised the manuscript according to these suggestions. The responses to the below comments describe the changes that have been implemented in greater detail. In particular, we have made sure that the work is transparent and reproducible. We have also re- done the data extraction from all original studies, which has greatly improved data quality.  
+
+<|ref|>text<|/ref|><|det|>[[147, 363, 844, 660]]<|/det|>
+New reviewer comment - Round4: Thank you for providing your data and code on Zenodo. It seems like putting the contents of these two repositories together still does not allow reproducing all the analyses and they are really hard to peer- review. There are three reasons: first the code is very poorly annotated, and thus it is hard to figure out what different sections of code are supposed to produce. Second, it seems like pieces of code are missing, e.g. file "Figures 2- 3. R" starts from loading the file 'Meta combined data.csv' and then immediately rands a meta analytic model on the data from it, however the file does not contain effect sizes and there is no line of code that calculates effect sizes before running the model. Third, the code refers to loading many additional data files, that look like pre- processed data, but they are not included in the Zenodo repository, e.g. "Supplementary Tables 1(1- 3). R" calls "Old_New_pesticide <- read_excel(paste("/lustre/huyueqing/ssy/Wan2023/DataPre/","Supplementary Data 2- Classification of old and new pesticides- 2023- 0825. xls",sep = ""),sheet = 1)", but there is no such file or no explanation where it originates from. Overall, I think all the analytic pipeline should be made fully transparent reproducible and then subjected to independent peer- review.  
+
+<|ref|>text<|/ref|><|det|>[[148, 666, 850, 770]]<|/det|>
+Response: Thank you for your comment. We have now completely overhauled the analytical pipeline in response to your comments. 1) We have now gone through the code and improved the annotation so that it allows a full interpretation of what has been done. 2) We have ensured that all pieces of code are present, including those relating to the production of all figures, tables and supplementary materials; 3) All files required for the analysis are included.  
+
+<|ref|>text<|/ref|><|det|>[[148, 795, 851, 898]]<|/det|>
+New reviewer comment - Round 4: Thank you. The README file only describes data in the main data file, but not in any of the additional data files and does not provide any information about the project in general and how the code and files should be used to reproduce the analyses. Please provide more comprehensive description of all your files, starting from overview and then providing meta- data for all data that is needed to reproduce your work. Consider using Quarto with R to
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[148, 84, 850, 121]]<|/det|>
+produce computationally reproducible documents that contain the R code, verbal descriptions, and outputs.  
+
+<|ref|>text<|/ref|><|det|>[[147, 127, 852, 250]]<|/det|>
+Response: We acknowledge that the previous version lacked reproducibility and had been hard to follow. Addressing this recommendation, we have now re- written the README file using Quarto with R, and extended this to include information on additional data files, as well as our analysis approach in general. We hope that this has now improved transparency of dataset descriptions and relevant code information, allowing clear interpretation of the approach used for all stages of the analysis. We have double- checked all code and made sure it runs on machines across continents.  
+
+<|ref|>text<|/ref|><|det|>[[147, 277, 852, 645]]<|/det|>
+New reviewer comment - Round 4: The code used for calculating SMD should not be placed in the README file, but in the code where it is needed to calculate effect sizes to reproduce the results. Response: Thank you for your comment. Following your wider recommendation, all R code and a README file with R annotation developed using Quarto have now been uploaded to Zenodo. The specific code you are referring to for calculating SMD has been also been revised again in this version to avoid problems stemming from our initial use of SMD rather than log response ratios. In our revised version, we now use the log response ratio (InRR) to calculate the effect sizes (the range of effect sizes in InRR is smaller than when using SMD). We have contacted Shinichi Nakagawa, an expert in meta- analysis and statistics, whose recommendations we follow in our revisions. Briefly, we now use log response ratios rather than SMD, because these do not require standard deviations and are more robust for use in meta- analysis, as demonstrated by Nakagawa et al. (2023; see reference below). For missing value imputation, we have now employed the so- called 'all cases' method, for sensitivity analysis and main analysis. Reporting follows the PRISMA- EcoEvo checklist to enhance quality. Additionally, we used the latest publication bias tests and submitted our fully annotated R code to Zenodo for transparency. We believe that these extensive revisions have considerably increased the robustness of our analysis.  
+
+<|ref|>text<|/ref|><|det|>[[148, 622, 850, 660]]<|/det|>
+Reference: Nakagawa et al., 2023. A robust and readily implementable method for the meta- analysis of response ratios with and without missing standard deviations. Ecol. Lett. 26: 232- 244.  
+
+<|ref|>text<|/ref|><|det|>[[148, 686, 851, 765]]<|/det|>
+New reviewer comment - Round 4: Thank you for adding the study references next to study codes in the data file. I was now able to compare the extracted data with the information reported in the original papers. I randomly selected 4 papers I have full- text access to and found the following issues:  
+
+<|ref|>text<|/ref|><|det|>[[148, 771, 852, 896]]<|/det|>
+Response: Following your comments about potential errors in the data (and as an explanation for the significant delay in producing this review) we have now re- checked all the original data and studies used in this paper. This involved going back to the original papers and confirming values derived to produce effect sizes as used in the analysis (note this is the reason for the considerable delay in this revision as the complete recheck took a lot of time). Where on reassessment papers were not considered to meet requirements for data derivation, they have now been excluded. In
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[148, 83, 852, 165]]<|/det|>
+some cases, new papers were identified from citations within those that were rejected for the above reasons but not identified in the initial scanning process. The process for including these additional papers is detailed within the PRISMA process and diagram. Below, we reply to the specific further comments on individual papers.  
+
+<|ref|>text<|/ref|><|det|>[[147, 190, 857, 400]]<|/det|>
+- Paper 16. (Colin et al. 2004. A Method to Quantify and Analyze the Foraging Activity of Honey Bees: Relevance to the Sublethal Effects Induced by Systemic Insecticides): rows 16-21: the concentration of Fipronil is wrong - the study reports using 2 micrograms per kilogram while the extracted data is \(6 \mathrm{g / kg}\) . For Imidacloprid the unit is also wrong-it should be microgram per kilogram not gram per kilogram. For the outcome "number of daily attendance of active bees-nuclei" which was extracted for nuclei A, B, C, it is and somehow from control nuclei, but I cannot find the matching exact numbers in the text, and it is not possible to extract them from the figures in the text. Given the complexity of the experimental design, raw data would need to be used to calculate such values. There are no comments in the data file how the numbers were obtained, and thus they are not reproducible.  
+
+<|ref|>text<|/ref|><|det|>[[148, 406, 847, 465]]<|/det|>
+Response: In the context of the paper you refer to, the means, SD and sample size \((n)\) of the indicator (i.e., "number of daily attendance of active bees-nuclei A") were calculated on days 0, 1, and 4. Because this was not a strictly repeated trial, we have removed this paper.  
+
+<|ref|>text<|/ref|><|det|>[[147, 491, 852, 638]]<|/det|>
+- Paper 1628. (Zhao X, Li YL, Zhang LX, Dorna H, 2003. Effects of priming and fungicide treatment on germination of china aster (Callistephus chinensis L.) seeds. Seed Science & Technology, 32: 451-457.): rows 1943-1944: it looks like only germination capacity data from Sample 1 and only from 30C temperature treatment were extracted and the standard deviations were manually imputed by assuming they are \(10\%\) of the measured value. This is only a small fraction of available data in this paper (multiple control and exposure groups, all samples should be extracted and used for calculating mean and SD) and is incorrectly extracted.  
+
+<|ref|>text<|/ref|><|det|>[[147, 643, 849, 858]]<|/det|>
+Response: We thank you for pointing us at this. To exclude the interactive effects of pesticide doses and temperature levels on non- target organisms, and only to measure the individual effect of pesticides on non- target organisms, our principle is that we extract the data at the highest temperature to conduct our analysis to exclude the interactive effects of pesticides and temperatures. The authors mentioned that "The germination tests were conducted at \(30^{\circ} \mathrm{C}, 20^{\circ} \mathrm{C}\) and \(10^{\circ} \mathrm{C}\) in darkness on 8 replicates of 50 seeds for each treatment." This is why the sample size was 8, and the SD of the value "Germination capacity (%)" at different temperatures is missing in Table 1. Because this is a comparatively old paper, we did not receive a response from the original corresponding author. To exclude your worries about this paper, we have now also removed this paper.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[179, 91, 344, 104]]<|/det|>
+## Seed germination tests  
+
+<|ref|>text<|/ref|><|det|>[[178, 105, 826, 181]]<|/det|>
+The germination tests were conducted at \(30^{\circ}\mathrm{C}\) \(20^{\circ}\mathrm{C}\) and \(10^{\circ}\mathrm{C}\) in darkness on 8 replicates of 50 seeds for each treatment. Control seeds were untreated. Fifty seeds were placed in each \(9\mathrm{cm}\) diameter Petri dish containing 6 layers blotting paper wetted with \(5\mathrm{ml}\) of distilled water. Seeds were considered as germinating when there was a visible protrusion of the radical through the seed coat and pericarp.  
+
+<|ref|>text<|/ref|><|det|>[[178, 181, 826, 241]]<|/det|>
+For germination index and mean germination time were recorded at 24h intervals for 14 days. Germination index \(\mathrm{(G~I) = (Gt/Tt))}\) Where G is the number of seeds germinated on day t and t is the number of days. Mean germination time (MGT) \(=\) TiNi/ Ni, where Ni is the number of newly germinated seed at time Ti.  
+
+<|ref|>text<|/ref|><|det|>[[264, 254, 730, 270]]<|/det|>
+Figure: Screenshot proof showing that these were 8 replicates  
+
+<|ref|>text<|/ref|><|det|>[[275, 276, 720, 291]]<|/det|>
+(from Zhao et al., Seed Science & Technology, 32: 451- 457)  
+
+<|ref|>text<|/ref|><|det|>[[147, 318, 852, 487]]<|/det|>
+- Paper 1008. (Gabaston J, Khawand TE, Wafo-Teguo P, Decendit A, Richard T, Merillon JM, Pavel A, 2018. Stilbenes from grapevine root: a promising natural insecticide against Leptinotarsa decemlineata. Journal of Pest Science, 91: 897-906.) rows 5808-5809: looks like data was extracted from Table 6 for 3rd day of exposure only and two lower doses out of 3 used (this is not reported on the data extraction sheet). The mean mortality \(\%\) was converted to survival \(\%\) , but the reported standard deviation values were somehow replaced with \(10\%\) of the men value. The sample size is claimed to be 4, but I cannot find such information in the article, and it seems improbable given the type of the organism, and reported values.  
+
+<|ref|>text<|/ref|><|det|>[[147, 492, 852, 681]]<|/det|>
+Response: In the data extraction sheet (i.e., Meta combined data- 2024- 0814), in rows 131- 132 in column "N" (i.e., Non- target organism indicator), we have presented it as "survival percent- transformed by mortality- 3rd day- Table 6". In the Methods section (i.e., Earthworm acute toxicity test), the authors have clearly shown that the sample size was 4: "Ten adult earthworms were placed in glass containers (1 L) filled with the test substrate (650 g), and the test containers were enclosed with a polythene sheet with integrated gauze to prevent the worms from escaping and to ensure optimal ventilation. After 3, 7 and 14 days of incubation, living worms were sorted by hand; the test endpoint was mortality. The tests were repeated four times. The glass containers were placed in a growth chamber in an artificial climate (L16:D8, \(20 \pm 1^{\circ}\mathrm{C}\) ).  
+
+<|ref|>text<|/ref|><|det|>[[207, 683, 787, 873]]<|/det|>
+Ten adult earthworms were placed in glass containers (1 L) filled with the test substrate (650 g), and the test containers were enclosed with a polythene sheet with integrated gauze to prevent the worms from escaping and to ensure optimal ventilation. After 3, 7 and 14 days of incubation, living worms were sorted by hand; the test endpoint was mortality. The tests were repeated four times. The glass containers were placed in a growth chamber in an artificial climate (L16:D8, \(20 \pm 1^{\circ}\mathrm{C}\) ).  
+
+<|ref|>text<|/ref|><|det|>[[150, 880, 843, 897]]<|/det|>
+Fig. Screenshot proof that this paper had 4 replicates (Gabaston et al., 2018. Journal of Pest
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 125, 852, 296]]<|/det|>
+Question: Overall, this small exercise clearly shows that 3/3 papers had some issues with data extractions. This indicates a very high rate of errors in the whole data set. The data lacks information on the exact sources of the extracted values for the outcomes at least (table/figure number, text page, raw data, etc.), any extractions decisions made, calculations, imputations, etc. - this makes assessing trustworthiness of the data set very hard and makes data extractions irreproducible. Because of this I request that all extracted data is annotated in detail. Further it has to be cross-checked by a researcher who was not involved in original extractions, re-extracted where issues are found and clearly described to make it fully transparent and trustworthy.  
+
+<|ref|>text<|/ref|><|det|>[[147, 300, 852, 662]]<|/det|>
+Response: As an expansion to our initial statement about a complete re- check of the data, we have now added the column "Y" (i.e., Exact sources of extracted values), the column "Z" (i.e., Extraction decisions made) and listed the extracted data information in column "N" (i.e., Non- target organism indicator). In this revision, Dr. Siyuan Shen re- extracted and cross- checked the data. After Siyuan's re- extracting and checking, Dr. Nian- Feng Wan and Dr. Ben A. Woodcock re- checked all the extracted information again, respectively, with additional checks being done by Christoph Scherber. During this re- extraction process, papers considered to have been unclear in their definition of measures of variance or described sample sizes were subjected to a group discussion (mainly composed of Nian- Feng Wan, Ben A. Woodcock, Dave Goulson, Adam J. Vanbergen, David J. Spurgeon and Siyuan Shen) to produce consensus on these derived metrics. At least three group members participated in each discussion. If our understanding was the same, we retained the data. If not, we contacted the authors directly to ask for the data. If the authors did not respond or could not be contacted, we removed the dataset originating from the respective paper. However, to collect more data to test pesticide effects on non- target organisms, we replaced papers from which no valid data could be extracted with a new, matching paper by the same authors or institutions (see lines 16- 29 in Supplementary methods). This process has (as we state above) now been clarified within the revised PRISMA diagram and associated processing description for data set derivation.  
+
+<|ref|>text<|/ref|><|det|>[[148, 688, 852, 790]]<|/det|>
+New reviewer comment - Round 4: When checking the extracted data, I noticed that many of the extracted standard deviation values are exactly \(10\%\) of the mean value. This cannot be coincidence and the manuscript does not mention doing manual imputations for missing standard deviation values (or any other missing values). Please explain how these values originated and the detailed procedures used to deal with any missing data.  
+
+<|ref|>text<|/ref|><|det|>[[148, 795, 852, 898]]<|/det|>
+Response: You are correct that where the values were \(10\%\) , this was due to an imputation for missing values (reffering to: Luo et al., 2006. Elevated \(\mathrm{CO_2}\) stimulates net accumulations of carbon and nitrogen in land ecosystems: a meta- analysis. Ecology, 87: 53- 63). In consultation with Professor Shinichi Nakagawa we have now revised this approach to deal with missing SDs in meta analyses. According to Dr. Nakagawa's suggestions, we now followed the approach described in his
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 83, 852, 207]]<|/det|>
+paper (Nakagawa et al., 2023. A robust and readily implementable method for the meta-analysis of response ratios with and without missing standard deviations. Ecol. Lett. 26: 232- 244) and used Equations 6 and 7 (see methods section in our manuscript) to calculate effect sizes and sampling variances when SDs were missing by simply imputing the pooled coefficient variation (CV) from the subset of studies that do report SDs. This is now fully described in the new revised version of our manuscript (see lines 30- 34 in Supplementary methods).  
+
+<|ref|>text<|/ref|><|det|>[[147, 233, 852, 378]]<|/det|>
+New reviewer comment - Round 4: The code for reproducing this should be in the file "Supplementary Tables 59- 61. R", and as such cannot be rerun. In the provided code I cannot find any line that would execute arcsine transformation for proportion data. And the code starts from loading a mysterious data file that is not archived on Zenodo, so cant tell what is there nor how it got there: AllData <- read.table(paste(DataPre_path,"SuppTab59Data.txt",sep=""),header = T). This seems to be ongoing issues across the provided code which is compounded by almost complete lack of code comments and descriptions.  
+
+<|ref|>text<|/ref|><|det|>[[147, 383, 852, 530]]<|/det|>
+Response: In this revision, and in response to your other above comments, we have re- written the code for executing the arcsine transformation for proportion data. Code for the arcsine transformation for proportion data is now included and annotated in the README file and all operational R files. We also had an internal discussion on using arcsine vs. logit transformation by hand, but stic ked with the arcsine as it is more established in a meta analysis context. The datasets are described in README file and can now all be fully accessed. Annotation has been improved for the code in general using as you suggested using Quarto with R.  
+
+<|ref|>text<|/ref|><|det|>[[147, 556, 852, 766]]<|/det|>
+New reviewer comment - Round 4: The authors are wrong when stating that "The random allocation of treatments to replicates, combined with study as a random effect in the meta- analytical models, means that these comparisons are appropriate and do not represent double accounting as suggested." The measures from the same control individual are used repeatedly in comparisons of effects of different doses of the same pesticides and they are such effect sizes are not independent and are a form of "double- counting". The analyses that account for such non- independence should be used as main models not as supplementary. The author used sample- size corrections their sensitivity analyses for a few models. I suggest testing their all models using cluster- rubust approach which is already implemented in metafor package (https://wviechtb.github.io/metafor/reference/robust.html) as their main approach.  
+
+<|ref|>text<|/ref|><|det|>[[147, 771, 852, 895]]<|/det|>
+Response: Thanks for your comment. We have addressed this important issue following the approach described by Shinichi Nakagawa in his paper on "Quantitative evidence synthesis: a practical guide on meta- analysis, meta- regression, and publication bias tests for environmental sciences." (Nakagawa et al., 2023. Environ. Evid. 12: 8). In order to adjust for repeated measurements of control values, we assigned the argument "V" in the "rma.mv()" function with the sampling variance- covariance matrix estimated by function "vcalc()" (lines 584- 592).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[148, 104, 850, 143]]<|/det|>
+New reviewer comment - Round 4: Please also run and report an overall (global) meta-analytic model on all data with organism, trait and pesticide type as random effects.  
+
+<|ref|>text<|/ref|><|det|>[[147, 147, 852, 616]]<|/det|>
+Response: We derived responses for 830 different species (560 animals, 192 plants, 78 microorganisms) and 129 non- species- level groups (i.e. 79 non- species- level animals, 8 non- species- level plants and 42 non- species- level microorganisms) to 471 different pesticide active ingredients (243 insecticides, 104 fungicides and 124 herbicides) (Supplementary Tables 14- 28; Supplementary Data 1- 3) (lines 118- 122). In the last version of this paper, pesticide identity (active ingredients pesticide name) and study identity were used as random effects that were recognized by you. In our revision, we have now added organisms and traits as random effects, according to your suggestions. Namely, we run/reported an overall (global) meta- analytic model on all data with organisms (unique species name and unique non- level species groups), traits (growth, reproduction, biomarker and behavior), pesticide identity (active ingredients pesticide names) and study identity as random effects (see Supplementary Table 1.1, Supplementary Table 1.2, Supplementary Table 2.1 and Supplementary Table 2.2), run/reported an overall meta- analytic model on all data with plant species (excluding non- level species groups), traits (growth, reproduction and biomarker), pesticide identity (active ingredients pesticide names), and study identity for plants (see Supplementary Table 1.3 and Supplementary Table 2.3), run/reported an overall meta- analytic model on all data with animal species (excluding non- level species groups), traits (growth, reproduction, biomarker and behavior), pesticide identity (active ingredients pesticide names), and study identity as random effects for animals (see Supplementary Table 1.4 and Supplementary Table 2.4), and run/reported an overall meta- analytic model on all data with microorganism species (excluding non- level species groups), traits (growth, reproduction and biomarker), pesticide identity (active ingredients pesticide names), and study identity as random effects for microorganisms (see Supplementary Table 1.5 and Supplementary Table 2.5).  
+
+<|ref|>text<|/ref|><|det|>[[148, 641, 852, 744]]<|/det|>
+New reviewer comment - Round 4: Thank you for adding information about the median sample sizes. However, this raises new issue since Line 562 states "A large- sample approximation was used to compute the sampling variances46." This approximation is not suitable for such small sample sizes, and a small- size- corrected version of the SMD (Hedges' g) should be used as an effect size in this meta- analysis.  
+
+<|ref|>text<|/ref|><|det|>[[148, 749, 852, 896]]<|/det|>
+Response: Thank you for your comment. According to suggestions from Professor Shinichi Nakagawa, we have replaced SMD with InRR to calculate the effect size and sample variance. Professor Shinichi Nakagawa reported four new methods to calculate effect size when missing SDs exist, which was reported in his paper entitled "A robust and readily implementable method for the meta- analysis of response ratios with and without missing standard deviations" (Nakagawa et al., 2023. Ecol. Lett. 26: 232- 244). He suggested that using the average CV to estimate sampling variances for all observations, regardless of missingness, performs with minimal
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 84, 486, 99]]<|/det|>
+bias. We have exactly followed his suggestions.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[149, 105, 406, 120]]<|/det|>
+## Response to the referee comments  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 127, 327, 141]]<|/det|>
+## Replies to Reviewer #1:  
+
+<|ref|>text<|/ref|><|det|>[[147, 148, 852, 250]]<|/det|>
+Comment: Thank you for your patience and hard work on this manuscript. I am pleased to note that it is much more transparent and robust now (it seems to have some new interesting findings as well). I really hope that in your next meta- analytic project you will take all steps from its inception to make sure it is trustworthy and well- documented, so I dont need to do 5 rounds of review again... good luck.  
+
+<|ref|>text<|/ref|><|det|>[[147, 255, 815, 272]]<|/det|>
+Response: We thank you for your positive evaluation. We have checked the manuscript again.
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__6e810f38300811758ab7eab9b6c03a87ff1dde2eb852213e3e151c27fbcdf8e5/images_list.json

@@ -0,0 +1 @@
+[]

peer_reviews/supplementary_0_Transparent Peer Review file__6e810f38300811758ab7eab9b6c03a87ff1dde2eb852213e3e151c27fbcdf8e5/supplementary_0_Transparent Peer Review file__6e810f38300811758ab7eab9b6c03a87ff1dde2eb852213e3e151c27fbcdf8e5.mmd

@@ -0,0 +1,52 @@
+
+# nature portfolio  
+
+Peer Review File  
+
+# Thermophobic diffusion becomes dominant in ultra-dilute alkali halide aqueous solutions  
+
+Corresponding Author: Professor Juan Felipe Torres  
+
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+Version 0:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author)  
+
+I have read all available versions of the manuscript and the extensive discussions between the authors and reviewer #2. From a technical perspective, the manuscript is carefully executed, and the methodology appears sound. All questionable points were removed during the discussion with the referees. Nevertheless, I agree with reviewer #2 that the work does not represent a sufficiently significant advance in the field of thermal diffusion to merit publication in Nature Communications. The contribution is rather incremental in scope and impact.  
+
+However, I believe the manuscript has merit for the scientific community and would be well suited for Communications Chemistry, which is more specialized and appropriate for this level of advance. I believe it will find an appropriate readership in Communications Chemistry.  
+
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were
+
+<--- Page Split --->
+
+made.  
+
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source. The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.  
+
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+
+# Response to Reviewer's Comment  
+
+September 2, 2025  
+
+## Reviewer 1 (Adjudicator)  
+
+## R1 General Comment  
+
+I have read all available versions of the manuscript and the extensive discussions between the authors and reviewer #2. From a technical perspective, the manuscript is carefully executed, and the methodology appears sound. All questionable points were removed during the discussion with the referees. Nevertheless, I agree with reviewer #2 that the work does not represent a sufficiently significant advance in the field of thermal diffusion to merit publication in Nature Communications. The contribution is rather incremental in scope and impact.  
+
+However, I believe the manuscript has merit for the scientific community and would be well suited for Communications Chemistry, which is more specialized and appropriate for this level of advance. I believe it will find an appropriate readership in Communications Chemistry.  
+
+## Response:  
+
+We sincerely thank the adjudicator for their careful and thorough evaluation of our manuscript, especially given the substantial volume of material reviewed throughout this process. We greatly appreciate the recognition of the technical soundness of our methodology and the merit of our contribution to the field. We also thank the adjudicator for their recommendation. We agree that Communications Chemistry is a highly suitable platform for our work, with the right scope and readership to maximize its impact.
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__6e810f38300811758ab7eab9b6c03a87ff1dde2eb852213e3e151c27fbcdf8e5/supplementary_0_Transparent Peer Review file__6e810f38300811758ab7eab9b6c03a87ff1dde2eb852213e3e151c27fbcdf8e5_det.mmd

@@ -0,0 +1,72 @@
+<|ref|>title<|/ref|><|det|>[[72, 53, 295, 80]]<|/det|>
+# nature portfolio  
+
+<|ref|>text<|/ref|><|det|>[[73, 96, 296, 119]]<|/det|>
+Peer Review File  
+
+<|ref|>title<|/ref|><|det|>[[72, 161, 866, 210]]<|/det|>
+# Thermophobic diffusion becomes dominant in ultra-dilute alkali halide aqueous solutions  
+
+<|ref|>text<|/ref|><|det|>[[72, 224, 525, 240]]<|/det|>
+Corresponding Author: Professor Juan Felipe Torres  
+
+<|ref|>text<|/ref|><|det|>[[72, 274, 864, 289]]<|/det|>
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+<|ref|>text<|/ref|><|det|>[[72, 327, 144, 340]]<|/det|>
+Version 0:  
+
+<|ref|>text<|/ref|><|det|>[[72, 354, 219, 367]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[72, 379, 160, 393]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[72, 404, 238, 418]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 430, 913, 496]]<|/det|>
+I have read all available versions of the manuscript and the extensive discussions between the authors and reviewer #2. From a technical perspective, the manuscript is carefully executed, and the methodology appears sound. All questionable points were removed during the discussion with the referees. Nevertheless, I agree with reviewer #2 that the work does not represent a sufficiently significant advance in the field of thermal diffusion to merit publication in Nature Communications. The contribution is rather incremental in scope and impact.  
+
+<|ref|>text<|/ref|><|det|>[[72, 508, 921, 550]]<|/det|>
+However, I believe the manuscript has merit for the scientific community and would be well suited for Communications Chemistry, which is more specialized and appropriate for this level of advance. I believe it will find an appropriate readership in Communications Chemistry.  
+
+<|ref|>text<|/ref|><|det|>[[72, 903, 916, 944]]<|/det|>
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 48, 117, 60]]<|/det|>
+made.  
+
+<|ref|>text<|/ref|><|det|>[[72, 60, 910, 125]]<|/det|>
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source. The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.  
+
+<|ref|>text<|/ref|><|det|>[[72, 125, 618, 139]]<|/det|>
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[68, 95, 523, 118]]<|/det|>
+# Response to Reviewer's Comment  
+
+<|ref|>text<|/ref|><|det|>[[68, 140, 218, 155]]<|/det|>
+September 2, 2025  
+
+<|ref|>sub_title<|/ref|><|det|>[[68, 174, 352, 195]]<|/det|>
+## Reviewer 1 (Adjudicator)  
+
+<|ref|>sub_title<|/ref|><|det|>[[95, 220, 265, 233]]<|/det|>
+## R1 General Comment  
+
+<|ref|>text<|/ref|><|det|>[[95, 234, 904, 320]]<|/det|>
+I have read all available versions of the manuscript and the extensive discussions between the authors and reviewer #2. From a technical perspective, the manuscript is carefully executed, and the methodology appears sound. All questionable points were removed during the discussion with the referees. Nevertheless, I agree with reviewer #2 that the work does not represent a sufficiently significant advance in the field of thermal diffusion to merit publication in Nature Communications. The contribution is rather incremental in scope and impact.  
+
+<|ref|>text<|/ref|><|det|>[[95, 328, 904, 372]]<|/det|>
+However, I believe the manuscript has merit for the scientific community and would be well suited for Communications Chemistry, which is more specialized and appropriate for this level of advance. I believe it will find an appropriate readership in Communications Chemistry.  
+
+<|ref|>sub_title<|/ref|><|det|>[[95, 389, 175, 402]]<|/det|>
+## Response:  
+
+<|ref|>text<|/ref|><|det|>[[95, 405, 904, 477]]<|/det|>
+We sincerely thank the adjudicator for their careful and thorough evaluation of our manuscript, especially given the substantial volume of material reviewed throughout this process. We greatly appreciate the recognition of the technical soundness of our methodology and the merit of our contribution to the field. We also thank the adjudicator for their recommendation. We agree that Communications Chemistry is a highly suitable platform for our work, with the right scope and readership to maximize its impact.
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__7018d1e716724eb35ce43dd941be54feddd3574f1d0d1d131b6825d727f7db8c/images_list.json

@@ -0,0 +1 @@
+[]

peer_reviews/supplementary_0_Transparent Peer Review file__7018d1e716724eb35ce43dd941be54feddd3574f1d0d1d131b6825d727f7db8c/supplementary_0_Transparent Peer Review file__7018d1e716724eb35ce43dd941be54feddd3574f1d0d1d131b6825d727f7db8c.mmd

@@ -0,0 +1,408 @@
+
+# nature portfolio  
+
+Peer Review File  
+
+# Polyketide synthase-derived sphingolipids mediate microbiota protection against a bacterial pathogen in C. elegans  
+
+Corresponding Author: Dr Katja Dierking  
+
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+Version 0:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author)  
+
+This is an important study characterizing a polyketide synthase from Pseudomonas fluorescens MYb115, a microbe associated with C. elegans in nature that has been shown to offer C. elegans protection from pathogenic Bt infections. The researchers identified the polyketide synthase (PKS) activity, deleted it, and found that the microbe no longer offers protection against Bt infection in C. elegans. They used LC- MS to identify the product of the polyketide synthase, and found that the enzyme produces long- chain sphinganine molecules and then used isotopic labeling to show that serine is incorporated into the sphinganine molecules allowing them to propose a pathway for the activities of the PKS. This is very solid work and is important and interesting because the researchers also identified genes encoding PKS activity in many other microbes, especially those who interact with host organisms. This is a significant finding because the PKS were previously believed to be found in fungi and rarely in bacteria.  
+
+The investigators then sought to determine how the PKS products protect C. elegans from Bt pathogens. Sphingolipids are important components of the intestinal barrier, so the researchers used transcriptomic and lipidomic analysis to determine changes in the C. elegans sphingolipids in worms eating the wild type bacteria vs the bacteria in which the PKS was deleted. Interestingly, they found that the C. elegans that were colonized by the wild type MYb115 bacteria had lower levels of many spingolipid species compared to worms who were colonized by the mutant bacteria lacking PKS. However, this may be minor because only a few species were significantly changed. Researchers then knocked down sphingolipid metabolism genes in C. elegans and examined susceptibility to Bt. They found that knocking down enzyme in the de novo pathway leading to ceramide protected worms from infection, while knocking down sphingomyelin synthase, which converts ceramide to sphingomyelin, resulted in more susceptible worms. The MYb115 somewhat rescued this susceptibility of the sms- 1 knockout, although this is puzzling, because the researchers showed that colonization by MYb115 leads to reduced sphingomyelins in worms, so it is unclear how this rescue occurred.  
+
+This paper is well written and the data are clearly and thoroughly presented. While the C. elegans experiments did not clarify how the sphinganine products from MYb115 lead to protection from Bt infection in C. elegans, the observations are important starting points for future experiments. The researchers used rather vague terms, for example on line 309 "These data indicate that MYb115 interacts with host sphingolipid metabolism...." While the nature of the interaction isn't clear, at this point I think this is a valid statement, because more definitive mechanism can't be determined from the data. One technical comment. When researchers receive deletion lines from the CGC (ok alleles) or NBRP (tm alleles) they typically perform outcrosses against wild type worms since the original deletion strains can contain many other mutations. The researchers should report how many times the strains were outcrossed prior to their experiments, or else say that they were not outcrossed.  
+
+## Reviewer #2  
+
+(Remarks to the Author)  
+
+The manuscript Peters et al. identified an iterative type I polyketide synthase (PKS) in MYb115 and showed that PKS regulates bacterial sphingolipid biogenesis and this interferes with sphingolipid metabolism of the host, thus exerting a
+
+<--- Page Split --->
+
+protective role for the host against pathogen infection. The paper contains a number of interesting observations, especially identification of an SPT- independent mechanism for sphingolipid biogenesis. However, the data did not provide a significant novelty and new insights into mechanistic explanation for microbiota- host interaction in regulation of host immune response, given that the role of sphingolipid in immune response and the interplay between bacterial sphingolipid and host sphingolipid biogenesis/metabolism are known. In addition, some of the claims are also not sufficiently supported by the results. For these reasons, I am not recommending publication in Nature Communications.  
+
+## Major issues:  
+
+1. sgaAB encoded PKS catalyses sphingolipid biosynthesis and improves resistance of animals against Bt247. 
+a) Is the protective effect of MYb115 linked with altered activity of known pathways for worm innate immunity, e.g. p38, MAPK, JUN-1, ELT-2, necrosis pathway or bre genes? 
+b) How does MYb115 affect toxicity of the other Gram-positive pathogen, e.g. Bt679? 
+c) Can the MYb115 ΔsgaAB phenotype be rescued by restoring sgaAB expression in this mutant bacterium? 
+d) Can expression of sgaAB in another bacterium which normally doesn't have a protective role against Bt247 result a protection? 
+e) How is the intestine morphology and integrity upon Bt247 infection affected by MYb115 and sgaAB? 
+f) Line 146: MYb115 lipidomic analysis was performed to test whether the bacteria sphinginates exist as free compounds or are part of lipids and sphinginates 1-3 and PG-sphingolipids 4-6 were identified. Does this result suggest that the bacteria sphinginates exist as both free compounds and part of lipids? A clear conclusion to the question should be provided to make understanding easier, especially for those who ate not familiar with sphingolipid metabolism. In addition, I don't understand why elucidating the existing form of sphinginates is important to know how MYb115 interact with the host. Additional background explanations are required to understand the underlying logic. The author concluded that they are not able to differentiate between the effects of the individual sphingolipid species. I don't think that lipidomic analysis alone would help to answer this question. But supplementation of these sphinginates or PG-sphingolipids might do.  
+
+## 2. Transcriptome study:  
+
+a) it is surprising that Bt247 infection only cause few numbers of DEG in MYb115 and worms on the MYb115 ΔsgaAB mutants, given their significantly different survival rate. As the RNA-seq results was not shown in any main figures or supplemental data, expect an enrichment analysis in Fig. 4A and S4A, it is difficult to make much comments. Does Bt247 infection affect transcriptome significantly? Does it result in activation of innate immunity genes? If Bt247 does not cause a significant alteration in gene expression, it would be hard to find involved factors via comparing transcriptome of animals on MYb115 and on the MYb115 ΔsgaAB mutants. In a such scenario, MYb115 might protect animals via other mechanisms that are independent of influencing transcript level. How does MYb115 affect transcriptome without Bt247 infection? If the Bt247 infectious bacteria remodel the worms' transcriptome much more than MYb115 does, the the potential changed genes that are only influenced by MYb115 or MYb115 ΔsgaAB mutants might submerge among Bt247-caused alteration in gene expression.  
+
+b) The transcriptome data was then integrated into the iCEL1314 genome-scale metabolic model and 24 and 23 significant differences in the presence or absence of Bt247 were received. What does "24 and 23 significant differences" mean? Does it refer to Differences between generated metabolic models? The results must be described in more details so that non-experts in iCEL1314 could understand the data. What does the "Gene Ratio" in the Figure 4A mean? Fig. 4A and S4A only show enrichment of metabolic processes even without notifying whether these processes are activated or inactivated by sgaAB in MYb115 (whether genes involved are up- or down- regulated). A list of genes used for the enrichment analysis would also be important to understand the result.  
+
+c) Line 220: "Here, we also saw an enrichment in valine, leucine, and isoleucine degradation, which are branched-chain amino acids (BCAA). This pathway is directly connected with propanate metabolism that provides components for the synthesis of the C15iso fatty acid, which is the precursor for sphingolipids in C. elegans": are these genes up- or down-regulated in MYb115 vs. worms on the MYb115 ΔsgaAB mutants? Again, provide more RNA-seq information will avoid such a confusion.  
+
+d) The RNA-seq data should be validated via qRT-PCR.  
+
+3. Host sphingolipid metabolism interferes with Bt247 susceptibility  
+
+a) From the host lipidomics analysis shown in Fig. 4b, sgaAB in MYb115 seems to reduce level of Cer, HexCer, DhCer and SM in the host. To test whether accumulation of these molecules reduces survival of animals upon Bt247 infection, the authors used different mutants in the sphingolipid metabolism pathway to test their susceptibility to Bt247. However, cgt-1 and cerk-1 mutant animals fed with OP50 show increased survival, although these animals are supposed to have Cer accumulation. In contrast, asm-3, hyl-1, hyl-2, sptl-1 and sptl-3 mutants which probably have reduced Cer level are also resistant. Do these results argue against an important role of Cer in Bt247 susceptibility? In summary, from these Bt247 survival assay with different mutants in sphingolipid metabolism pathway, the author could not provide a clear answer which classes of host sphingolipid are responsible for the altered susceptibility to the pathogen. 
+b) What is account for Bt247 toxicity? Why should altered sphingolipid metabolism influence toxicity of Bt247? A study from Ruvkun's lab (Liu et al., 2014, Nature) has shown that some nature habitat bacteria cause mitochondria dysfunction and animals respond with mitochondrial surveillance machinery which is ceramide dependent. At the same time some other bacteria could inhibit mitochondrial surveillance to render a more effective virulence. Does Bt247 modulate the mitochondria surveillance? If yes, could sgaAB in MYb115 counteract the Bt247 effect on mitochondria?  
+
+4. Sphingolipid in MYb115 affect host sphingolipid biogenesis and metabolism.  
+
+There is no attempt to address how this could happen. Is uptake of the bacterial sphingolipid into worm intestine necessary? Can feeding worms with the identified sphinginates 1-3 or sphingolipid 4-6 impact Bt247 resistance? If yes, does this supplementation change sphingolipid composition of worms?
+
+<--- Page Split --->
+
+## Minor issues  
+
+1. Zeile 377: Notably, in C. elegans, glucosylceramide deficiency was linked to an increase in autophagy which plays an important role in cellular defence after attack by certain Bt pore-forming toxins (PFTs). Does Myb affect autophagy? If yes, could manipulating autophagy affect survival of animals to Bt247?  
+
+2. Fig. S6: Should A1 be compared with B1, A2 with B2 and so on? If yes, why is the mutant A10 in red line and B10 in yellow? The color used for asah-1 and asah-2 are too similar to be differentiated. Maybe just add the name of the mutant directly to the figures. Why are there several figures for some mutants, e.g. asah-1 while for some others only one e.g. cgt-1 on OP50? It does not seem to be biological replicates as \(n = 4\) for each figure.  
+
+3. Fig. 3B: Does the width of the boxes present the number of bacteria? Yes. Information is shown in the supplemental table. Please indicate this in the main figure to make understanding easier.  
+
+## Reviewer #3  
+
+(Remarks to the Author) In silico analyses revealed three biosynthetic gene clusters in P. fluorescens MYb115. Two were characterized, and the authors demonstrate that a PKS cluster produces sphingolipids, which alters sphingolipid metabolism in the host. An active PKS cluster in MYb115 is required to provide protection against Bacillus thuringiensis (Bt), which is (in part) mediated by host sphingolipid metabolism. Thus, a nice interplay between a natural member of the C. elegans microbiota, C. elegans and microbial pathogens has been uncovered.  
+
+## Major Comments:  
+
+1. Three biosynthetic gene clusters were identified, yet only two were functionally characterized. For completeness, all three BGCs should be included in the analyses for their role in providing protection against Bt247; 
+2. Are the protective effects specific for Bt247? 
+3. one cannot conclude that NRPS does not contribute to the protection against Bt247 with an active PKS. The experiments should be repeated in a Asga background to address whether AsgaAnrpA reduces survival compared to Asga and/or AnrpA. The third BGC should also be included. 
+4. while there is a phenotype observed, what is known about leaky expression of the arabinose promoter in this system? 
+5. It is expected that gene deletion mutants are complemented for function  
+
+## Minor Comments:  
+
+6. L82: it should read 'intraspecies' and not 'interspecies'; 
+7. L109: this reads odd; please rephrase; 
+8. Fig. 1: since these are separate infection studies, and not consecutive sampling, data should be shown as bars instead of continuous lines. 
+9. Fig 5A: please show actual data rather than the abstract interpretation  
+
+## Version 1:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author)  
+
+This is an interesting, important study. I do not have any further concerns with the manuscript.  
+
+## Reviewer #2  
+
+(Remarks to the Author)  
+
+The authors have revised their manuscript with great diligence and have addressed the majority of the concerns raised, particularly those related to the identification of bacterial SLs in the first section. I support the public communication of this work.  
+
+However, I am still not fully satisfied with the organization of the figures. The order in which the figures are presented does not always correspond to the sequence in which they are referenced in the Results section. For example, Figure 1C is first mentioned in the second paragraph, while Figures 1D- 1F are already discussed in the first paragraph.  
+
+Additionally, the main figures present only a limited portion of the results, whereas the supplementary material contains 22 figures, some of which include highly relevant data. Given that many readers may not consult the supplementary information in detail, I strongly recommend that the authors incorporate some of these key findings into the main figures to enhance the accessibility and impact of the manuscript.  
+
+Additional minor remark
+
+<--- Page Split --->
+
+line 212: two dots after "Figure S4)"line 279: This may indicate that MYb115- derived SLs do not 280 strongly affect C. elegans on the transcript level, but more strongly influence the host on the proteome or metabolome level.  
+
+line 305: "we \(\Delta \Delta\) integrated" what does \(\Delta \Delta\) integrated mean?Similarly, line 303: colonisation \(\Delta\) propanate  
+
+## Reviewer #3  
+
+(Remarks to the Author)  
+
+The authors have performed a comprehensive revision of their original submission, and I believe that all of my original concerns have been fully addressed. The (new) data strongly support the authors' conclusions. Their findings present a major advance. The authors must be lauded for their efforts!  
+
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source. The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.  
+
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+
+Reviewer's Responses to Questions in grey  
+
+## Our responses in blue  
+
+We thank all reviewers for thoughtful reading of the manuscript and the very constructive criticism.  
+
+Reviewer #1 (Remarks to the Author):  
+
+This is an important study characterizing a polyketide synthase from Pseudomonas fluorescens MYb115, a microbe associated with C. elegans in nature that has been shown to offer C. elegans protection from pathogenic Bt infections. The researchers identified the polyketide synthase (PKS) activity, deleted it, and found that the microbe no longer offers protection against Bt infection in C. elegans. They used LC- MS to identify the product of the polyketide synthase, and found that the enzyme produces long- chain sphinganine molecules and then used isotopic labeling to show that serine is incorporated into the sphinganine molecules allowing them to propose a pathway for the activities of the PKS. This is very solid work and is important and interesting because the researchers also identified genes encoding PKS activity in many other microbes, especially those who interact with host organisms. This is a significant finding because the PKS were previously believed to be found in fungi and rarely in bacteria.  
+
+The investigators then sought to determine how the PKS products protect C. elegans from Bt pathogens. Sphingolipids are important components of the intestinal barrier, so the researchers used transcriptomic and lipidomic analysis to determine changes in the C. elegans sphingolipids in worms eating the wild type bacteria vs the bacteria in which the PKS was deleted. Interestingly, they found that the C. elegans that were colonized by the wild type MYb115 bacteria had lower levels of many spingolipid species compared to worms who were colonized by the mutant bacteria lacking PKS. However, this may be minor because only a few species were significantly changed. Researchers then knocked down sphingolipid metabolism genes in C. elegans and examined susceptibility to Bt. They found that knocking down enzyme in the de novo pathway leading to ceramide protected worms from infection, while knocking down sphingomyelin synthase, which converts ceramide to sphingomyelin, resulted in more susceptible worms. The MYb115 somewhat rescued this susceptibility of the sms- 1 knockout, although this is puzzling, because the researchers showed that colonization by MYb115 leads to reduced sphingomyelins in worms, so it is unclear how this rescue occurred.  
+
+Our reply: We thank the reviewer for this valid comment. We agree that we do not understand how exactly MYb115 interacts with C. elegans sphingolipid metabolism. Sphingolipid homeostasis is controlled by a complex network comprising several levels of regulation. The enzymes responsible for sphingolipid production and turnover comprise a metabolic network that gives rise to numerous bioactive molecules, which participate in highly complex and interconnected pathways influencing a multitude of physiological processes (Hannun and Obeid, 2018 https://doi.org/10.1038/nrm.2017.107). Also, sphingolipid metabolism shares common substrates with other metabolic routes and is, for example, highly connected to other lipid metabolic networks. Consequently, imbalances in sphingolipid metabolism in a mutant may have far- reaching consequences for host physiology. Thus, the effect of MYb115 on the phenotype of a sphingolipid metabolism
+
+<--- Page Split --->
+
+mutant (increased survival after Bt infection) may not be directly linked to its effect on wildtype worms (decrease in sphingomyelin).  
+
+Nevertheless, our comprehensive data sets, including several new data sets now added to the revised manuscript, do allow us to draw several important conclusions on commensal- mediated immune protection. Important new insights are for example that sphingolipid- producing MYb115 does cause a decrease of certain host sphingolipid species, including sphingomyelin species, in comparison to non- sphingolipid- producing MYb115. Moreover, MYb115 ameliorates the survival phenotype of C. elegans sphingolipid enzyme mutants following Bt infection in comparison to OP50. We now integrated a critical evaluation of our data into the discussion and further clarify that the current comprehensive data sets do not yet provide a final answer on whether the observed effect of MYb115 on the survival phenotype of the sphingolipid enzyme mutant is direct (on sphingolipid metabolism) or indirect (p. 17 and 18, line 495- 501).  
+
+Also, we have now invested considerable efforts in further improving our understanding of MYb115 sphingolipid biosynthesis and the effects produced by MYb115, including biochemical characterization of the enzymes encoded in the biosynthetic gene cluster (BGC) that catalyze MYb115 sphingolipid biosynthesis (in collaboration with experts in sphingolipid biosynthesis), further metabolic analyses of bacterial cultures (in collaboration with experts in metabolic analyses), and further phenotypic analyses in C. elegans demonstrating – among others – that protection against another Bt strain, Bt679, and protection of the C. elegans intestinal barrier following Bt infection depends on MYb115- produced sphingolipids and that MYb115- mediated protection is independent of two known C. elegans Bt defense pathways (p38 and JNK MAPK pathways), the mitochondrial surveillance response, and of a Bt toxin glycosphingolipid receptor (please also see our answers to the reviewers' comments below). This paper is well written and the data are clearly and thoroughly presented. While the C. elegans experiments did not clarify how the sphinganine products from MYb115 lead to protection from Bt infection in C. elegans, the observations are important starting points for future experiments. The researchers used rather vague terms, for example on line 309 "These data indicate that MYb115 interacts with host sphingolipid metabolism...." While the nature of the interaction isn't clear, at this point I think this is a valid statement, because more definitive mechanism can't be determined from the data.  
+
+Our reply: We thank the reviewer for their comments.  
+
+One technical comment. When researchers receive deletion lines from the CGC (ok alleles) or NBRP (tm alleles) they typically perform outcrosses against wild type worms since the original deletion strains can contain many other mutations. The researchers should report how many times the strains were outcrossed prior to their experiments, or else say that they were not outcrossed.  
+
+Our reply: We genotyped the respective sphingolipid metabolism pathway mutant and thus confirmed the deletion alleles, but we did not outcross the mutants. We now clarify that the mutants were not outcrossed in the materials and methods section and show the detailed genotyping results in a supplementary figure (Figure S26).  
+
+## Reviewer #2 (Remarks to the Author):  
+
+The manuscript Peters et al. identified an iterative type I polyketide synthase (PKS) in MYb115 and showed that PKS regulates bacterial sphingolipid biogenesis and this interferes with sphingolipid metabolism of the host, thus exerting a protective role for the host against pathogen infection. The paper contains a number of interesting observations, especially identification of an SPT- independent mechanism for sphingolipid biogenesis.
+
+<--- Page Split --->
+
+Our reply: We agree with the reviewer that the discovery of a new way of bacterial sphingolipid synthesis significantly advances our knowledge on bacterial sphingolipid metabolism. While the few known bacterial sphingolipid producers (such as Bacteroidetes), like eukaryotes, produce sphingolipid as primary metabolites in a manner that depends on the enzyme serine palmitoyltransferase (SPT), P. fluorescens MYb115 produces sphingolipids as secondary metabolites in a non- canonical way that depends on a BGC, the polyketide synthase (PKS) P/SgaB. To our knowledge this is the first example of a bacterial PKS shown to be involved in sphingolipid biosynthesis and also the first description of a Pseudomonas isolate as sphingolipid producer.  
+
+Moreover, in the revised manuscript we now present and discuss a comprehensive set of new results that considerably expand our knowledge on the P. fluorescens BGC- dependent sphingolipid biosynthesis pathway: We previously could only assume that the enzyme encoded by P/SgaB within the BGC substitutes the function of SPT - the above- mentioned rate- limiting enzyme for sphingolipid biosynthesis in all eukaryotes and the few bacterial sphingolipid producers. In the revised manuscript, we now integrated comprehensive new data, which we generated in collaboration with Dominic Campopiano, Michael Herrera, and Francesca Lubbock (University of Edinburgh, all now co- authors) on the function of P/SgaB. Using heterologous expression in E. coli and subsequent in vitro functional analysis with the purified enzyme we could prove that P/SgaB functions as SPT. Furthermore, we identified a putative short chain dehydrogenase/reductase (SDR) in the MYb115 sphingolipid BGC, which is predicted to share structural homology with 3- ketodihydropsphingine reductase (KDSR). KDSR is known to catalyze the reduction of 3- KDS to dihydropsphinganine (DHS) (Beeler et al., 1998. https://doi.org/10.1074/jbc.273.46.30688; Fornarotto et al., 2006. https://doi.org/10.1016/j.bbalip.2005.11.013); whilst this step is ubiquitous in eukaryotic sphingolipid biosynthesis, it is unusual in bacterial sphingolipid pathways (Stankevicute et al., 2022. https://doi.org/10.1038/s41589-021-00948-7). The inclusion of this eukaryotic- like step further distinguishes the P. fluorescens sphingolipid- producing BGC from canonical bacterial sphingolipid biosynthesis. In the revised manuscript, we now highlight these new exciting findings in a new section and new figure (see lines 204- 241 and Figure 2).  
+
+However, the data did not provide a significant novelty and new insights into mechanistic explanation for microbiota- host interaction in regulation of host immune response, given that the role of sphingolipid in immune response and the interplay between bacterial sphingolipid and host sphingolipid biogenesis/metabolism are known.  
+
+Our reply: Many thanks for this comment. We politely disagree. Our work very clearly and substantially advances our knowledge on the functional significance of a bacterial, sphingolipid- producing PKS in microbiota- mediated protection against pathogens. We found this PKS in many other bacteria that, like Pseudomonas, are not yet known sphingolipid producers. This is a significant novelty since the interplay between bacterial sphingolipids and host sphingolipid metabolism has so far only been reported for one bacterial phylum and one host, the Bacteroidetes in the human gut (e.g. Johnsen et al., 2020. https://doi.org/10.1038/s41467- 020- 16274- w; Le et al., 2022. https://doi.org/10.1016/j.chom.2022.05.002; Brown et al., 2019. https://doi.org/10.1016/j.chom.2019.04.002). Most importantly, to date, the relevance of this interplay has never been reported in the context of microbiota- mediated protection against pathogens. Also, we show that a Pseudomonas species alters sphingolipid metabolism in C. elegans and establish the importance of C. elegans sphingolipid metabolism for survival after Bt infection. We agree with the reviewer that a more detailed study on the mechanism by which MYb115- derived sphingolipids modify host sphingolipid metabolism and/or immune response is an interesting topic for a future study. However, such an analysis is a manuscript of its own and thus goes beyond the scope of the present study. In response to the reviewer's comment, we now more clearly highlight the importance of our numerous novel insights and further point to promising future research directions.  
+
+interplay has never been reported in the context of microbiota- mediated protection against pathogens. Also, we show that a Pseudomonas species alters sphingolipid metabolism in C. elegans and establish the importance of C. elegans sphingolipid metabolism for survival after Bt infection. We agree with the reviewer that a more detailed study on the mechanism by which MYb115- derived sphingolipids modify host sphingolipid metabolism and/or immune response is an interesting topic for a future study. However, such an analysis is a manuscript of its own and thus goes beyond the scope of the present study. In response to the reviewer's comment, we now more clearly highlight the importance of our numerous novel insights and further point to promising future research directions.
+
+<--- Page Split --->
+
+In addition, some of the claims are also not sufficiently supported by the results.  
+
+Our reply: We hope that we could sufficiently address this concern (see our answers below).  
+
+For these reasons, I am not recommending publication in Nature Communications. Major issues:  
+
+1. sgaAB encoded PKS catalyses sphingolipid biosynthesis and improves resistance of animals against Bt247.  
+
+Our reply: We thank the reviewer for bringing to our attention the ambiguities in MYb115- mediated protection listed below. In the revised manuscript, we have added a new paragraph (line 264- 299) summarizing the new insights we have gained on how the host response to Bt247 is affected by MYb115 and MYb115ΔsgaA.  
+
+a) Is the protective effect of MYb115 linked with altered activity of known pathways for worm innate immunity, e.g. p38, MAPK, JUN-1, ELT-2, necrosis pathway or bre genes?  
+
+Our reply: Following the reviewer's suggestion, we tested the involvement of the p38 MAPK and the JNK-like MAPK pathway in MYb115-mediated protection against Bt247. We found that the protective effect is completely independent of the p38 MAPK pathway (MYb115 also protects the p38 MAPK pathway tir-1, nsy-1, sek-1, and pmk-1 mutants) or the JNK-like MAPK KGB-1 (Figure S19). We could previously already exclude an involvement of the bre genes in C. elegans defense against Bt247, given that bre mutants are susceptible to Bt247 infection (see discussion line 511- 518). However, because of the direct link between bre genes and the biogenesis of complex glycosphingolipids, we confirmed these previous findings and included the data in Figure S25B. Please also see our response to the reviewer's comment 3b below.  
+
+b) How does MYb115 affect toxicity of the other Gram-positive pathogen, e.g. Bt679?  
+
+Our reply: We have previously already shown that MYb115 also protects against another Bt strain, Bt679 (Kissoyan and Peters et al. https://doi.org/10.3389/fcimb.2022.775728). We now mention this in the revised manuscript (line 272- 274). The comments by reviewer #2 and #3 prompted us to test if protection against Bt679 is also lost on the MYb115 ΔsgaA mutant and indeed it is. We now included this new additional result in Figure S17. We did not test other Gram-positive bacteria.  
+
+c) Can the MYb115 ΔsgaAB phenotype be rescued by restoring sgaAB expression in this mutant bacterium?  
+
+Our reply: In response to the comments of reviewer #2 and #3 we have now conducted a series of experiments to functionally complement the MYb115 ΔsgaA mutant by inserting the vanillic acid inducible PvanCC promoter in front of sgaA or the complete sgaAB BGC on the plasmid pSEVA631, which was then introduced into the MYb115 mutant. Only the complementation that included the complete sgaAB BGC restored sphingolipid production (detection of compounds 1 (m/z 414.4 [M+H]+), 2 (m/z 386.4 [M+H]+), and 3 (m/z 442.4 [M+H]+) by LC-MS (Figure S27). We observed a clear increase in resistance to Bt247 of worms on this complemented MYb115 ΔsgaA mutant and now present these additional new results in Figure 1F of the revised manuscript.  
+
+Of note: We obtained two complemented MYb115 ΔsgaA mutant strains. While these strains were meant for targeted activation of the BGC, we realized that the vanillic acid inducible promoter was leaky (and more so in one strain than in the other), resulting in sphingolipid production also in the absence of vanillic acid. However, we used this to show that variations in sphingolipid production are reflected in variations in the protective effect, providing further evidence that host protection is dependent on bacterial sphingolipid production. The whole set of obtained results do clearly demonstrate the role of the complete P/SgaAB BGC in sphingolipid production and host protection, which we now emphasize in the revised manuscript (line 179- 192 and Figure 1H).
+
+<--- Page Split --->
+
+d) Can expression of sgaAB in another bacterium which normally doesn't have a protective role against Bt247 result a protection?  
+
+Our reply: This is an interesting point. Yet, it is not essential to demonstrate a causal effect of sgaAB on protection against pathogenic Bt. Such evidence was produced with MYb115, which we now substantiated with the complementation of the MYb115 ΔsgaA mutant. Therefore, we decided to prioritize these additional complementation experiments (requested by reviewer #2 and #3), in order to further advance our understanding, and as a consequence, we did not include the here proposed experiment at this time.  
+
+e) How is the intestine morphology and integrity upon Bt247 infection affected by MYb115 and sgaAB?  
+
+Our reply: MYb115 limits Bt- induced damage to the intestinal epithelium, as we already demonstrated previously (Kissoyan et al., 2019. https://doi.org/10.1016/j.cub.2019.01.050). Prompted by the reviewer's question, we now further tested if MYb115- derived sphingolipids are involved in mitigating Bt- induced damage. Using the C. elegans PGP- 1::GFP reporter strain, we confirmed that MYb115 PKS- derived sphingolipids reduced membrane damage caused by Bt infection, while the MYb115 ΔsgaA mutant did not. We included these new results in the revised manuscript (line 290- 299. Figure S20).  
+
+f) Line 146: MYb115 lipidomic analysis was performed to test whether the bacteria sphingamines exist as free compounds or are part of lipids and sphingamines 1-3 and PG- sphingolipids 4-6 were identified. Does this result suggest that the bacteria sphingamines exist as both free compounds and part of lipids? A clear conclusion to the question should be provided to make understanding easier, especially for those who are not familiar with sphingolipid metabolism. In addition, I don't understand why elucidating the existing form of sphingamines is important to know how MYb115 interact with the host. Additional background explanations are required to understand the underlying logic. The author concluded that they are not able to differentiate between the effects of the individual sphingolipid species. I don't think that lipidomic analysis alone would help to answer this question. But supplementation of these sphingamines or PG- sphingolipids might do.  
+
+Our reply: Many thanks for these valuable comments. The MYb115 lipidomic analysis using high- resolution Liquid Chromatography Tandem Mass Spectrometry (HRES- LC- MS/MS) allowed us to identify the PG- sphingolipids 4- 6. We now clarify this (line 167- 170). The supplementation experiments would require purified sphingolipids from MYb115, which however is not available and thus, could not easily be obtained. Therefore, we decided for a different approach to further elucidate the involvement of different sphingolipids. In collaboration with Manuel Liebeke (Kiel University, now co- author) using MALDI mass spectrometry spot assays (https://doi.org/10.1038/s41596- 023- 00864- 1), we visualized sphingamines 1- 3 and PG- sphingolipid 4- 6 in bacterial cultures, whose protective effect we then tested in survival analyses. These additional experiments demonstrate that protection significantly correlates with the abundance of sphingamines 1- 3 and PG- sphingolipid 4, indicating that host protection is dependent on these sphingolipids. In the revised manuscript, we now added and explain these new data (line 179- 192 and Figure 1H).  
+
+2. Transcriptome study:  
+
+a) it is surprising that Bt247 infection only cause few numbers of DEG in MYb115 and worms on the MYb115 ΔsgaAB mutants, given their significantly different survival rate. As the RNA-seq results was not shown in any main figures or supplemental data, expect an enrichment analysis in Fig. 4A and S4A, it is difficult to make much comments.  
+
+Our reply: The transcriptome response of C. elegans to MYb115 and MYb115 ΔsgaA is indeed very similar under both conditions, infected and non- infected. We agree with reviewer #2 that this result is surprising and unexpected. As discussed also below, the only difference between MYb115 and MYb115 ΔsgaA is the production of sphingolipids. We do think that sphingolipid production may not strongly affect C. elegans on the transcript level, but more strongly influences the host on the proteome/metabolome level. We mention this in the
+
+<--- Page Split --->
+
+revised manuscript in line 279- 281. We thank the reviewer for pointing out that we did not clearly present the RNAseq data in the manuscript. We now added a short description of the results in the context of host pathogen defense (line274- 287) and present the data in Figure S18 and Table S7. As before, raw data and processed data are accessible through GEO Series accession number GSE245296 at NCBI's Gene Expression Omnibus. Does Bt247 infection affect transcriptome significantly? Does it result in activation of innate immunity genes?  
+
+Our reply: Yes, we have previously shown in several independent studies that Bt247 infection has a strong effect on the C. elegans transcriptome and results in activation of C. elegans pathogen- responsive/innate immunity genes (Boehnisch et al., 2011. https://doi.org/10.1371/journal.pone.0024619. Nakad et al., 2016. https://doi.org/10.1186/s12864- 016- 2603- 8. Yang, Dierking et al., 2015. https://doi.org/10.1016/j.dci.2015.02.010 Zarate- Potes et al., 2020. https://doi.org/10.1371/journal.ppat.100882).  
+
+If Bt247 does not cause a significant alteration in gene expression, it would be hard to find involved factors via comparing transcriptome of animals on MYb115 and on the MYb115 AsgaAB mutants. In a such scenario, MYb115 might protect animals via other mechanisms that are independent of influencing transcript level.  
+
+How does MYb115 affect transcriptome without Bt247 infection?  
+
+Our reply: For this study, we specifically compared the C. elegans transcriptome response to MYb115 in comparison to the non- sphingolipid producing MYb115 AsgaA mutant. We did not include the laboratory food bacterium E. coli OP50 as control, which would be necessary to assess the general effect of MYb115 on the C. elegans transcriptome without Bt247 infection. The comparison between the C. elegans response to OP50 and MYb115 was not the subject of the study.  
+
+If the Bt247 infectious bacteria remodel the worms' transcriptome much more than MYb115 does, the the potential changed genes that are only influenced by MYb115 or MYb115 AsgaAB mutants might submerge among Bt247- caused alteration in gene expression.  
+
+Our reply: We thank the reviewer for their comment. This may indeed be the case and, as discussed above, MYb115- derived sphingolipid may affect the host response more strongly on the proteome level. We now address this aspect in the revised manuscript (line 279- 281). b) The transcriptome data was then integrated into the iCEL1314 genome- scale metabolic model and 24 and 23 significant differences in the presence or absence of Bt247 were received. What does "24 and 23 significant differences" mean? Does it refer to Differences between generated metabolic models? The results must be described in more details so that non- experts in iCEL1314 could understand the data.  
+
+Our reply: In the methods section of the revised manuscript we did explain more in detail what the different data types are and how metabolic models were applied. In the revised results, we now point to the methods section for these details (line 307).  
+
+Additionally, we have now simplified the data types that we analysed. Previously, we had the OFD results, lower bound and upper bound - we performed statistical analysis on them separately. However, the interpretation of differences was difficult. In the revised manuscript, we now combined the lower and upper bound values into one score - the center, which represents: (upper bound - lower bound) / 2. Centers are more representative of the solution spaces for each of the reactions, and coefficients are easier to interpret. After doing so, a few of the previous reactions are no longer present among the significant results - instead of 24 (KO vs WT with Bt247) and 23 (KO vs WT without Bt247), we obtained 16 reactions in each contrast. We have updated the manuscript accordingly (line 305- 322).  
+
+What does the "Gene Ratio" in the Figure 4A mean? Fig. 4A and S4A only show enrichment of metabolic processes even without notifying whether these processes are activated or
+
+<--- Page Split --->
+
+inactivated by sgaAB in MYb115 (whether genes involved are up- or down- regulated). A list of genes used for the enrichment analysis would also be important to understand the result. Our reply: We would like to thank the reviewer for drawing our attention to this point. We did not use genes for enrichment, rather the reactions significant after our statistical analyses. The "pathway" universe were the subsystem annotations of all reactions within the model. Since we identified only few reactions as significantly up- or down- regulated an enrichment on them separately would not yield any results. We changed the mislabeled "Gene ratio" to "Reactions". The individual significant reactions are provided in Table S10.  
+
+Importantly, our goal for this analysis was to understand which pathways are affected, not necessarily specify which reactions are affected. Additionally, we do not want to show the coefficients in the main manuscript because they can be misinterpreted in the traditional sense due to the fact that in metabolic models, the positive or negative sign of reactions actually represents the directionality of a reaction. For example, a reaction can be irreversible, e.g. substrate - > product, or reversible, i.e. substrate <-> product. In our hypothetical irreversible reaction, the reaction flux can only have positive values (it's irreversible), so here a negative coefficient would mean that the flux through this reaction in the worms grown on mutant bacteria is lower. However, for our hypothetical irreversible reactions, values can be both positive and negative. So, if all our values were negative, a positive coefficient would actually mean that the worms grown on mutant bacteria would have less flux (e.g. - 0.5 vs - 2) through this reaction. Since we have a large mix of reversible and irreversible reactions within our network, and we have not only looked at flux potentials, but also at mathematical optimal flux distribution (which is the optimal but not unique solution to the linear optimization problem), we cannot reliably infer the significant involvement of individual reactions based on these statistical analysis results. Nevertheless, we can see trends from our analysis on the pathway level, which is why we have described our results as we have. These aspects have now been briefly explained in the methods section (line 701- 725).  
+
+c) Line 220: "Here, we also saw an enrichment in valine, leucine, and isoleucine degradation, which are branched-chain amino acids (BCAA). This pathway is directly connected with propanoate metabolism that provides components for the synthesis of the C15iso fatty acid, which is the precursor for sphingolipids in C. elegans": are these genes up- or down-regulated in MYb115 vs. worms on the MYb115 \(\Delta\) sgaAB mutants? Again, provide more RNA-seq information will avoid such a confusion.  
+
+Our reply: In Table S10 we have added the information about the genes that encode the significant reactions, as well as the logFC values of these genes from the transcriptomic analysis. We would like to point out that the iMAT++ algorithm used to integrate the transcriptomic data into the iCEL1314 metabolic model is done on a sample basis, therefore any statistical comparisons of gene expression is not taken into account during this process. This means that comparing the differences in simulation results of reactions encoded by a certain gene and the logFC values of this gene might not directly match in direction. This is a desirable attribute of metabolic modelling since we can predict metabolic requirements of an organism that are in conflict with the gene expression differences - these might be caused by post-translational modifications or other effects. These aspects have now been briefly explained in the methods section (line 710- 716).  
+
+d) The RNA-seq data should be validated via qRT-PCR.  
+
+Our reply: The metabolic network analysis, which is based on the RNAseq data, revealed that sphingolipid metabolism reactions show differential activity between MYb115 and MYb115 \(\Delta\) sgaA. We focused on this result and validated the transcriptomic data using lipidomic profiling. The other results of the RNAseq study are not the focus of this project. 3. Host sphingolipid metabolism interferes with Bt247 susceptibility  
+
+a) From the host lipidomics analysis shown in Fig. 4b, sgaAB in MYb115 seems to reduce
+
+<--- Page Split --->
+
+level of Cer, HexCer, DhCer and SM in the host. To test whether accumulation of these molecules reduces survival of animals upon Bt247 infection, the authors used different mutants in the sphingolipid metabolism pathway to test their susceptibility to Bt247. However, cgt- 1 and cerk- 1 mutant animals fed with OP50 show increased survival, although these animals are supposed to have Cer accumulation. In contrast, asm- 3, hyl- 1, hyl- 2, splt- 1 and splt- 3 mutants which probably have reduced Cer level are also resistant. Do these results argue against an important role of Cer in Bt247 susceptibility? In summary, from these Bt247 survival assay with different mutants in sphingolipid metabolism pathway, the author could not provide a clear answer which classes of host sphingolipid are responsible for the altered susceptibility to the pathogen.  
+
+Our reply: We thank the reviewer for this valid comment. We tested four ceramide metabolic gene mutants, all with an expected increased ceramide content, but two mutants (cerk- 1 and cgt- 1) show a variable survival phenotype and two mutants (asah- 1,2 and sms- 1) are more susceptible to infection. In response to the comments of reviewer #2 we now further assessed the role of specific sphingolipids in defense against Bt infection and supplemented Bt infected worms, using a set of new experiments with the commercially available sphingolipids ceramide, sphingomyelin, and sphingosine- 1- phosphate. We found that supplementation with C18 and C20 ceramide significantly improved survival rates, while C22 ceramide, C16 sphingomyelin, C18 sphingomyelin, d- sphingosine, and S1P did not affect survival. These new results and the phenotypic differences between the ceramide metabolic gene mutants imply that the inhibition of ceramide metabolism (and the associated increase in ceramide content) is not the only factor determining susceptibility to infection. We present and discuss these additional new results in the revised manuscript (line 390- 396. Figure S13 and line 483- 492).  
+
+b) What is account for Bt247 toxicity? Why should altered sphingolipid metabolism influence toxicity of Bt247?  
+
+Our reply: How altered host sphingolipid metabolism affect defense against Bt and thus Bt toxicity remains mysterious. Since sphingolipids are required for the integrity of cellular membranes, but can also act as bioactive signalling molecules involved in regulation of a myriad of cell activities, there also are a myriad of potential links between sphingolipid metabolism and defense against Bt247. In response to the reviewer's comments above, but also below, we further explored the two most apparent potential links between sphingolipids and C. elegans defense against Bt infection: The involvement of mitochondrial surveillance (see comment by the reviewer below) and the bre genes (see comment 1 a). We did not find any evidence of a link between Bt infection and mitochondrial surveillance or the bre genes. We present and discuss these results in the revised manuscript (line 426- 438. Figure S25 and line 504- 518). Based on our comprehensive data set, we discuss possible alternative routes of how sphingolipid metabolism interacts with Bt247 (line 502- 522).  
+
+A study from Ruvkun's lab (Liu et al., 2014, Nature) has shown that some nature habitat bacteria cause mitochondria dysfunction and animals respond with mitochondrial surveillance machinery which is ceramide dependent. At the same time some other bacteria could inhibit mitochondrial surveillance to render a more effective virulence. Does Bt247 modulate the mitochondria surveillance?  
+
+Our reply: Following the reviewer's suggestion, we tested if Bt247 infection induces the mitochondrial stress- induced hsp- 6p::gfp reporter that was also used in the study by Liu et al. Neither Bt247 infection, nor MYb115 induced expression of hsp- 6p::gfp, indicating that mitochondrial surveillance is not involved in MYb115- mediated protection against Bt247 infection. We added these results to the revised version of our manuscript (line 426- 438. Figure S25A).
+
+<--- Page Split --->
+
+4. Sphingolipid in MYb115 affect host sphingolipid biogenesis and metabolism.  
+
+There is no attempt to address how this could happen. Is uptake of the bacterial sphingolipid into worm intestine necessary? Can feeding worms with the identified sphinganines 1- 3 or sphingolipid 4- 6 impact Bt247 resistance? If yes, does this supplementation change sphingolipid composition of worms?  
+
+Our reply: We agree with the reviewer that the question how exactly MYb115- derived sphingolipids affects host sphingolipid metabolism is highly intriguing, yet its further analysis goes beyond the scope of the current study. Our study did reveal for the first time that P. fluorescens- derived sphingolipids affect C. elegans sphingolipid metabolism and that modulation of host sphingolipid metabolism increases tolerance towards Bt infection - among others. We now provide a comprehensive set of new data to substantiate the findings made. Supplementation experiments, as suggested by the reviewer, would indeed be helpful to differentiate between the effects of the identified sphinganines 1- 3 and PG- sphingolipids 4- 6. As also discussed above (comment 1 f), we unfortunately do not have the MYb115- derived sphingolipids purified, which we would need to do these supplementation experiments. However, in collaboration with Manuel Liebeke (Kiel University, now co- author) using MALDI mass spectrometry spot assays (https://doi.org/10.1038/s41596-023-00864-1), we visualized sphinganines 1- 3 and PG- sphingolipid 4- 6 in bacterial cultures, whose protective effect we then tested in survival analyses. These additional experiments demonstrate that protection significantly correlates with the abundance of sphinganines 1- 3 and PG- sphingolipid 4, indicating that host protection is dependent on these sphingolipids. In the revised manuscript, we now added and explain these new data (line 179- 192 and Figure 1H).  
+
+Minor issues  
+
+1. Zeile 377: Notably, in C. elegans, glucosylceramide deficiency was linked to an increase in autophagy which plays an important role in cellular defence after attack by certain Bt performing toxins (PFTs). Does Myb affect autophagy? If yes, could manipulating autophagy affect survival of animals to Bt247?  
+
+Our reply: We appreciate the reviewer's suggestion regarding the possible involvement of autophagy. This is indeed an interesting idea and could provide valuable insights for future work. However, due to time and resource constraints, we prioritized addressing the most important points in our revisions. Thus, we decided against an additional investigation of autophagy.  
+
+Fig. S6: Should A1 be compared with B1, A2 with B2 and so on? If yes, why is the mutant A10 in red line and B10 in yellow? The color used for asah- 1 and asah- 2 are too similar to be differentiated. Maybe just add the name of the mutant directly to the figures. Why are there several figures for some mutants, e.g. asah- 1 while for some others only one e.g. cgt- 1 on OP50? It does not seem to be biological replicates as \(n = 4\) for each figure.  
+
+Our reply: We appreciate the reviewer's feedback regarding the data representation. In response, we have revised the figure to improve clarity. Specifically, we have updated the color scheme: for all assays where worms were exposed to OP50, the lines are now depicted in grey (panel A), while survival assays with worms exposed to MYb115 are shown in blue (panel B). This adjustment ensures consistency with the color scheme used throughout the manuscript. Additionally, as suggested, we have directly labeled each facet with the name of the corresponding mutant strain, further enhancing the figure's clarity. Moreover, we have now added more survival assays, yielding a larger number of independent experiments for each C. elegans mutant strain, and thus much more robust insights into the role of the respective genes. For some mutants i.e. (cgt- 1 and cerk- 1) the additional experiments revealed that the survival phenotype is variable. We thus included the technical and biological replicates in the heat- map of Figure 5C, so that the reader can see
+
+<--- Page Split --->
+
+at a glance how many experimental runs were done and identify variation across technical and biological replicates. The previous and the additional new data now provide an in- depth as well as statistically sound overview of the involvement of different branches of the entire sphingolipid metabolism.  
+
+3. Fig. 3B: Does the width of the boxes present the number of bacteria? Yes. Information is shown in the supplemental table. Please indicate this in the main figure to make understanding easier.  
+
+Our reply: We thank the reviewer for pointing out this ambiguity. We added the information in the legend of Figure 3.  
+
+Reviewer #3 (Remarks to the Author):  
+
+In silico analyses revealed three biosynthetic gene clusters in P. fluorescens MYb115. Two were characterized, and the authors demonstrate that a PKS cluster produces sphingolipids, which alters sphingolipid metabolism in the host. An active PKS cluster in MYb115 is required to provide protection against Bacillus thuringiensis (Bt), which is (in part) mediated by host sphingolipid metabolism. Thus, a nice interplay between a natural member of the C. elegans microbiota, C. elegans and microbial pathogens has been uncovered.  
+
+Major Comments:  
+
+1. Three biosynthetic gene clusters were identified, yet only two were functionally characterized. For completeness, all three BGCs should be included in the analyses for their role in providing protection against Bt247;  
+
+Our reply: We thank the reviewer for their comments. Although interesting, analysis of the third BSG is not essential to prove involvement of the sgaAB BGC in sphingolipid metabolism and protection against pathogens. We decided to prioritize on other experiments which now yielded a comprehensive data base for obtaining critical additional insights into the role of sgaAB BGC (see also above and below replies), and thus omitted an analysis of the arylpolene pathway.  
+
+2. Are the protective effects specific for Bt247?  
+
+Our reply: We have previously already shown that MYb115 also protects against another Bt strain, Bt679 (Kissoyan and Peters et al. https://doi.org/10.3389/fcimb.2022.775728). We now mention this in the revised manuscript (line 272-274). The comments by reviewer #2 and #3 prompted us to test if protection against Bt679 is also lost on the MYb115 ΔsgaA mutant and indeed it is. We now included this result in Figure S17. We did not test other Gram-positive bacteria.  
+
+3. one cannot conclude that NRPS does not contribute to the protection against Bt247 with an active PKS. The experiments should be repeated in a Δsga background to address whether ΔsgaΔnrpA reduces survival compared to Δsga and/or ΔnrpA. The third BGC should also be included.  
+
+Our reply: We thank the reviewer for pointing out this ambiguity. Due to time and resource constraints, we prioritized addressing the complementation of the MYb115 ΔsgaA and ΔsgaB mutants requested by reviewer #2 and #3. We now refrain from drawing conclusions regarding the NRPS cluster and rephrased the sentence in line 117- 120 "While the PKS gene cluster affects MYb115- mediated protection, we did not observe significant differences in worm survival with or without arabinose supplementation on the MYb115 PBADnrpA strain (Figure 1B, Table S1), indicating that the MYb115 NRPS gene cluster is not involved in MYb115- mediated protection against Bt infection." to "While the PKS gene cluster affects MYb115- mediated protection, we did not observe significant differences in worm survival with or without arabinose supplementation on the MYb115 PBADnrpA strain (Figure 1B, Table S1). We therefore focused on the MYb115 PKS gene cluster in our subsequent analyses."
+
+<--- Page Split --->
+
+Since we observed a complete loss of sphingolipid production and a clear decrease in host protection for the MYb115 \(\Delta sgaA\) and MYb115 \(\Delta sgaB\) single mutants, we think that it is justified to focus on the PKS cluster within the scope of the current project.  
+
+4. while there is a phenotype observed, what is known about leaky expression of the arabinose promoter in this system?  
+
+Our reply: We assessed MYb115 P \(_{BADsga}\) sphingolipid production in an induced (+ arabinose) and non- induced (- arabinose) state by LC- MS and did not detect any sphingolipids in the non- induced state (Figure 1C). In addition, in the revised manuscript, we included MYb115 P \(_{BADsga}\) in an induced (+ arabinose) and non- induced (- arabinose) state in an experiment to visualize production of the different sphingolipids (collaboration with Manuel Liebeke, Kiel University now co- author) and could confirm that P \(_{BAD}\) is not leaky in this system (datasets: MPIMM_514_QE_P https://metaspace2020. org/dataset/2025- 02- 27_13h37m58s).  
+
+5. It is expected that gene deletion mutants are complemented for function  
+
+Our reply (also see our response to the comment 1c of reviewer #2 above): In response to the comments of reviewer #2 and #3 we have conducted a series of additional new experiments to functionally complement the MYb115 \(\Delta sgaA\) mutant by inserting the vanillic acid inducible PvanCC promoter in front of sgaA or the complete sgaAB BGC on the plasmid pSEVA631, which was then introduced into the MYb115 mutants. Only the complementation that included the complete sgaAB BGC restored sphingolipid production (detection of compounds 1 (m/z 414.4 [M+H]+), 2 (m/z 386.4 [M+H]+), and 3 (m/z 442.4 [M+H]+) by LC- MS (Figure S27) and only in the MYb115 \(\Delta sgaA\) mutant. We observed a clear increase in resistance to Bt247 of worms on this complemented MYb115 \(\Delta sgaA\) mutant and present the results in Figure 1F of the revised manuscript.  
+
+Of note: We obtained two complemented MYb115 \(\Delta sgaA\) mutant strains. While these strains were meant for targeted activation of the BGC, we realized that the vanillic acid inducible promoter was leaky, resulting in sphingolipid production also in the absence of vanillic acid (datasets: MPIMM_514_QE_P https://metaspace2020. org/dataset/2025- 02- 27_13h37m58s). However, we used this to show that variations in sphingolipid production are reflected in variations in the protective effect, providing further evidence that host protection is dependent on bacterial sphingolipid production. The whole set of obtained results do clearly demonstrate the role of the complete MYb115 sgaAB BGC in sphingolipid production and host protection, which we now emphasize in the revised manuscript (line 179- 192 and Figure 1H).  
+
+Minor Comments:  
+
+6. L82: it should read 'intraspecies' and not 'interspecies';  
+
+Our reply: We have now replaced "interspecies" with "intraspecies"  
+
+7. L109: this reads odd; please rephrase;  
+
+Our reply: The sentence: "Supplementation of the C. elegans laboratory food Escherichia coli OP50 with arabinose did not affect survival, showing that arabinose itself does not influence C. elegans resistance to Bt (Table S1)." has now been reworded to: "Arabinose supplementation had no effect on resistance of C. elegans to Bt infection on its standard laboratory food Escherichia coli OP50."  
+
+8. Fig. 1: since these are separate infection studies, and not consecutive sampling, data should be shown as bars instead of continuous lines.  
+
+Our reply: The continuous lines in Figure 1 A, B and E represent a dose- response relationship. As it is common practice for dose- response curves, the data points are connected.  
+
+9. Fig 5A: please show actual data rather than the abstract interpretation  
+
+Our reply: We show the survival data in Figure 5C as heatmaps to summarize the results of 32 and 24 survival assays on E. coli OP50 and P. fluorescens MYb115, respectively, and to facilitate the comparison of results between different knockout mutants. We now also
+
+<--- Page Split --->
+
+included technical and biological replicates, so that the reader can see at a glance how many experimental runs were done and to facilitate the identification of variation across technical and biological replicates. For the interested reader, we do provide each individual survival curve in the supplementary Figure S23.
+
+<--- Page Split --->
+
+Reviewer's Remarks to the author in grey  
+
+## Our responses in blue  
+
+We thank all reviewers for thoughtful re- evaluation of the manuscript and their comments.  
+
+Reviewer #1 (Remarks to the Author):  
+
+This is an interesting, important study. I do not have any further concerns with the manuscript.  
+
+Our reply: We thank the reviewer for their comments.  
+
+Reviewer #2 (Remarks to the Author):  
+
+The authors have revised their manuscript with great diligence and have addressed the majority of the concerns raised, particularly those related to the identification of bacterial SLs in the first section. I support the public communication of this work. However, I am still not fully satisfied with the organization of the figures. The order in which the figures are presented does not always correspond to the sequence in which they are referenced in the Results section. For example, Figure 1C is first mentioned in the second paragraph, while Figures 1D- 1F are already discussed in the first paragraph.  
+
+Our reply: We changed the text according to the reviewer's suggestions.  
+
+Additionally, the main figures present only a limited portion of the results, whereas the supplementary material contains 22 figures, some of which include highly relevant data. Given that many readers may not consult the supplementary information in detail, I strongly recommend that the authors incorporate some of these key findings into the main figures to enhance the accessibility and impact of the manuscript.  
+
+Our reply: We thank the reviewer for pointing this out. We now incorporated the data shown in figures S13A into figure 2D, figure S18 B is now (new) figure 4A, figure S19 is now incorporated in figure 4 (new figure 4B and C), and figure S20 is now (new) figure 5. In addition, we combined several supplementary figures, so that we now have a total of 21 (instead of 27) supplementary figures.  
+
+Additional minor remark  
+
+line 212: two dots after "Figure S4)  
+
+Our reply: We removed one dot.  
+
+line 279: This may indicate that MYb115- derived SLs do not 280 strongly affect C. elegans on the transcript level, but more strongly influence the host on the proteome or metabolome level.  
+
+Our reply: We changed the sentence according to the reviewer's suggestion. line 305: "we \(\Delta \Delta\) integrated" what does \(\Delta \Delta\) integrated mean?  
+
+Our reply: We thank the reviewer for pointing out this mistake. We removed the ' \(\Delta \Delta\)  
+
+Similarly, line 303: colonisation \(\Delta\) propanoate  
+
+Our reply: We removed the 'colonisation \(\Delta\) .  
+
+Reviewer #3 (Remarks to the Author):  
+
+The authors have performed a comprehensive revision of their original submission, and I believe that all of my original concerns have been fully addressed. The (new) data strongly
+
+<--- Page Split --->
+
+support the authors' conclusions. Their findings present a major advance. The authors must be lauded for their efforts!  
+
+Our reply: We thank the reviewer for their encouraging comments.
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__7018d1e716724eb35ce43dd941be54feddd3574f1d0d1d131b6825d727f7db8c/supplementary_0_Transparent Peer Review file__7018d1e716724eb35ce43dd941be54feddd3574f1d0d1d131b6825d727f7db8c_det.mmd

@@ -0,0 +1,569 @@
+<|ref|>title<|/ref|><|det|>[[72, 50, 296, 80]]<|/det|>
+# nature portfolio  
+
+<|ref|>text<|/ref|><|det|>[[74, 96, 296, 119]]<|/det|>
+Peer Review File  
+
+<|ref|>title<|/ref|><|det|>[[73, 161, 841, 237]]<|/det|>
+# Polyketide synthase-derived sphingolipids mediate microbiota protection against a bacterial pathogen in C. elegans  
+
+<|ref|>text<|/ref|><|det|>[[73, 249, 421, 267]]<|/det|>
+Corresponding Author: Dr Katja Dierking  
+
+<|ref|>text<|/ref|><|det|>[[72, 299, 864, 314]]<|/det|>
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+<|ref|>text<|/ref|><|det|>[[73, 352, 144, 365]]<|/det|>
+Version 0:  
+
+<|ref|>text<|/ref|><|det|>[[73, 378, 219, 392]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[73, 404, 160, 417]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[73, 431, 238, 444]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 444, 914, 563]]<|/det|>
+This is an important study characterizing a polyketide synthase from Pseudomonas fluorescens MYb115, a microbe associated with C. elegans in nature that has been shown to offer C. elegans protection from pathogenic Bt infections. The researchers identified the polyketide synthase (PKS) activity, deleted it, and found that the microbe no longer offers protection against Bt infection in C. elegans. They used LC- MS to identify the product of the polyketide synthase, and found that the enzyme produces long- chain sphinganine molecules and then used isotopic labeling to show that serine is incorporated into the sphinganine molecules allowing them to propose a pathway for the activities of the PKS. This is very solid work and is important and interesting because the researchers also identified genes encoding PKS activity in many other microbes, especially those who interact with host organisms. This is a significant finding because the PKS were previously believed to be found in fungi and rarely in bacteria.  
+
+<|ref|>text<|/ref|><|det|>[[72, 574, 917, 717]]<|/det|>
+The investigators then sought to determine how the PKS products protect C. elegans from Bt pathogens. Sphingolipids are important components of the intestinal barrier, so the researchers used transcriptomic and lipidomic analysis to determine changes in the C. elegans sphingolipids in worms eating the wild type bacteria vs the bacteria in which the PKS was deleted. Interestingly, they found that the C. elegans that were colonized by the wild type MYb115 bacteria had lower levels of many spingolipid species compared to worms who were colonized by the mutant bacteria lacking PKS. However, this may be minor because only a few species were significantly changed. Researchers then knocked down sphingolipid metabolism genes in C. elegans and examined susceptibility to Bt. They found that knocking down enzyme in the de novo pathway leading to ceramide protected worms from infection, while knocking down sphingomyelin synthase, which converts ceramide to sphingomyelin, resulted in more susceptible worms. The MYb115 somewhat rescued this susceptibility of the sms- 1 knockout, although this is puzzling, because the researchers showed that colonization by MYb115 leads to reduced sphingomyelins in worms, so it is unclear how this rescue occurred.  
+
+<|ref|>text<|/ref|><|det|>[[72, 728, 923, 847]]<|/det|>
+This paper is well written and the data are clearly and thoroughly presented. While the C. elegans experiments did not clarify how the sphinganine products from MYb115 lead to protection from Bt infection in C. elegans, the observations are important starting points for future experiments. The researchers used rather vague terms, for example on line 309 "These data indicate that MYb115 interacts with host sphingolipid metabolism...." While the nature of the interaction isn't clear, at this point I think this is a valid statement, because more definitive mechanism can't be determined from the data. One technical comment. When researchers receive deletion lines from the CGC (ok alleles) or NBRP (tm alleles) they typically perform outcrosses against wild type worms since the original deletion strains can contain many other mutations. The researchers should report how many times the strains were outcrossed prior to their experiments, or else say that they were not outcrossed.  
+
+<|ref|>sub_title<|/ref|><|det|>[[73, 871, 161, 884]]<|/det|>
+## Reviewer #2  
+
+<|ref|>text<|/ref|><|det|>[[73, 898, 237, 911]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 911, 884, 939]]<|/det|>
+The manuscript Peters et al. identified an iterative type I polyketide synthase (PKS) in MYb115 and showed that PKS regulates bacterial sphingolipid biogenesis and this interferes with sphingolipid metabolism of the host, thus exerting a
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 45, 925, 126]]<|/det|>
+protective role for the host against pathogen infection. The paper contains a number of interesting observations, especially identification of an SPT- independent mechanism for sphingolipid biogenesis. However, the data did not provide a significant novelty and new insights into mechanistic explanation for microbiota- host interaction in regulation of host immune response, given that the role of sphingolipid in immune response and the interplay between bacterial sphingolipid and host sphingolipid biogenesis/metabolism are known. In addition, some of the claims are also not sufficiently supported by the results. For these reasons, I am not recommending publication in Nature Communications.  
+
+<|ref|>sub_title<|/ref|><|det|>[[72, 139, 165, 151]]<|/det|>
+## Major issues:  
+
+<|ref|>text<|/ref|><|det|>[[70, 152, 920, 360]]<|/det|>
+1. sgaAB encoded PKS catalyses sphingolipid biosynthesis and improves resistance of animals against Bt247. 
+a) Is the protective effect of MYb115 linked with altered activity of known pathways for worm innate immunity, e.g. p38, MAPK, JUN-1, ELT-2, necrosis pathway or bre genes? 
+b) How does MYb115 affect toxicity of the other Gram-positive pathogen, e.g. Bt679? 
+c) Can the MYb115 ΔsgaAB phenotype be rescued by restoring sgaAB expression in this mutant bacterium? 
+d) Can expression of sgaAB in another bacterium which normally doesn't have a protective role against Bt247 result a protection? 
+e) How is the intestine morphology and integrity upon Bt247 infection affected by MYb115 and sgaAB? 
+f) Line 146: MYb115 lipidomic analysis was performed to test whether the bacteria sphinginates exist as free compounds or are part of lipids and sphinginates 1-3 and PG-sphingolipids 4-6 were identified. Does this result suggest that the bacteria sphinginates exist as both free compounds and part of lipids? A clear conclusion to the question should be provided to make understanding easier, especially for those who ate not familiar with sphingolipid metabolism. In addition, I don't understand why elucidating the existing form of sphinginates is important to know how MYb115 interact with the host. Additional background explanations are required to understand the underlying logic. The author concluded that they are not able to differentiate between the effects of the individual sphingolipid species. I don't think that lipidomic analysis alone would help to answer this question. But supplementation of these sphinginates or PG-sphingolipids might do.  
+
+<|ref|>sub_title<|/ref|><|det|>[[72, 373, 235, 386]]<|/det|>
+## 2. Transcriptome study:  
+
+<|ref|>text<|/ref|><|det|>[[70, 386, 926, 515]]<|/det|>
+a) it is surprising that Bt247 infection only cause few numbers of DEG in MYb115 and worms on the MYb115 ΔsgaAB mutants, given their significantly different survival rate. As the RNA-seq results was not shown in any main figures or supplemental data, expect an enrichment analysis in Fig. 4A and S4A, it is difficult to make much comments. Does Bt247 infection affect transcriptome significantly? Does it result in activation of innate immunity genes? If Bt247 does not cause a significant alteration in gene expression, it would be hard to find involved factors via comparing transcriptome of animals on MYb115 and on the MYb115 ΔsgaAB mutants. In a such scenario, MYb115 might protect animals via other mechanisms that are independent of influencing transcript level. How does MYb115 affect transcriptome without Bt247 infection? If the Bt247 infectious bacteria remodel the worms' transcriptome much more than MYb115 does, the the potential changed genes that are only influenced by MYb115 or MYb115 ΔsgaAB mutants might submerge among Bt247-caused alteration in gene expression.  
+
+<|ref|>text<|/ref|><|det|>[[70, 515, 925, 606]]<|/det|>
+b) The transcriptome data was then integrated into the iCEL1314 genome-scale metabolic model and 24 and 23 significant differences in the presence or absence of Bt247 were received. What does "24 and 23 significant differences" mean? Does it refer to Differences between generated metabolic models? The results must be described in more details so that non-experts in iCEL1314 could understand the data. What does the "Gene Ratio" in the Figure 4A mean? Fig. 4A and S4A only show enrichment of metabolic processes even without notifying whether these processes are activated or inactivated by sgaAB in MYb115 (whether genes involved are up- or down- regulated). A list of genes used for the enrichment analysis would also be important to understand the result.  
+
+<|ref|>text<|/ref|><|det|>[[70, 606, 925, 671]]<|/det|>
+c) Line 220: "Here, we also saw an enrichment in valine, leucine, and isoleucine degradation, which are branched-chain amino acids (BCAA). This pathway is directly connected with propanate metabolism that provides components for the synthesis of the C15iso fatty acid, which is the precursor for sphingolipids in C. elegans": are these genes up- or down-regulated in MYb115 vs. worms on the MYb115 ΔsgaAB mutants? Again, provide more RNA-seq information will avoid such a confusion.  
+
+<|ref|>text<|/ref|><|det|>[[72, 671, 460, 685]]<|/det|>
+d) The RNA-seq data should be validated via qRT-PCR.  
+
+<|ref|>text<|/ref|><|det|>[[70, 697, 532, 710]]<|/det|>
+3. Host sphingolipid metabolism interferes with Bt247 susceptibility  
+
+<|ref|>text<|/ref|><|det|>[[70, 710, 920, 877]]<|/det|>
+a) From the host lipidomics analysis shown in Fig. 4b, sgaAB in MYb115 seems to reduce level of Cer, HexCer, DhCer and SM in the host. To test whether accumulation of these molecules reduces survival of animals upon Bt247 infection, the authors used different mutants in the sphingolipid metabolism pathway to test their susceptibility to Bt247. However, cgt-1 and cerk-1 mutant animals fed with OP50 show increased survival, although these animals are supposed to have Cer accumulation. In contrast, asm-3, hyl-1, hyl-2, sptl-1 and sptl-3 mutants which probably have reduced Cer level are also resistant. Do these results argue against an important role of Cer in Bt247 susceptibility? In summary, from these Bt247 survival assay with different mutants in sphingolipid metabolism pathway, the author could not provide a clear answer which classes of host sphingolipid are responsible for the altered susceptibility to the pathogen. 
+b) What is account for Bt247 toxicity? Why should altered sphingolipid metabolism influence toxicity of Bt247? A study from Ruvkun's lab (Liu et al., 2014, Nature) has shown that some nature habitat bacteria cause mitochondria dysfunction and animals respond with mitochondrial surveillance machinery which is ceramide dependent. At the same time some other bacteria could inhibit mitochondrial surveillance to render a more effective virulence. Does Bt247 modulate the mitochondria surveillance? If yes, could sgaAB in MYb115 counteract the Bt247 effect on mitochondria?  
+
+<|ref|>text<|/ref|><|det|>[[70, 891, 616, 904]]<|/det|>
+4. Sphingolipid in MYb115 affect host sphingolipid biogenesis and metabolism.  
+
+<|ref|>text<|/ref|><|det|>[[70, 904, 920, 944]]<|/det|>
+There is no attempt to address how this could happen. Is uptake of the bacterial sphingolipid into worm intestine necessary? Can feeding worms with the identified sphinginates 1-3 or sphingolipid 4-6 impact Bt247 resistance? If yes, does this supplementation change sphingolipid composition of worms?
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[72, 60, 163, 72]]<|/det|>
+## Minor issues  
+
+<|ref|>text<|/ref|><|det|>[[72, 72, 902, 113]]<|/det|>
+1. Zeile 377: Notably, in C. elegans, glucosylceramide deficiency was linked to an increase in autophagy which plays an important role in cellular defence after attack by certain Bt pore-forming toxins (PFTs). Does Myb affect autophagy? If yes, could manipulating autophagy affect survival of animals to Bt247?  
+
+<|ref|>text<|/ref|><|det|>[[72, 138, 911, 192]]<|/det|>
+2. Fig. S6: Should A1 be compared with B1, A2 with B2 and so on? If yes, why is the mutant A10 in red line and B10 in yellow? The color used for asah-1 and asah-2 are too similar to be differentiated. Maybe just add the name of the mutant directly to the figures. Why are there several figures for some mutants, e.g. asah-1 while for some others only one e.g. cgt-1 on OP50? It does not seem to be biological replicates as \(n = 4\) for each figure.  
+
+<|ref|>text<|/ref|><|det|>[[72, 202, 918, 231]]<|/det|>
+3. Fig. 3B: Does the width of the boxes present the number of bacteria? Yes. Information is shown in the supplemental table. Please indicate this in the main figure to make understanding easier.  
+
+<|ref|>sub_title<|/ref|><|det|>[[72, 269, 162, 282]]<|/det|>
+## Reviewer #3  
+
+<|ref|>text<|/ref|><|det|>[[72, 295, 922, 372]]<|/det|>
+(Remarks to the Author) In silico analyses revealed three biosynthetic gene clusters in P. fluorescens MYb115. Two were characterized, and the authors demonstrate that a PKS cluster produces sphingolipids, which alters sphingolipid metabolism in the host. An active PKS cluster in MYb115 is required to provide protection against Bacillus thuringiensis (Bt), which is (in part) mediated by host sphingolipid metabolism. Thus, a nice interplay between a natural member of the C. elegans microbiota, C. elegans and microbial pathogens has been uncovered.  
+
+<|ref|>sub_title<|/ref|><|det|>[[72, 385, 197, 398]]<|/det|>
+## Major Comments:  
+
+<|ref|>text<|/ref|><|det|>[[70, 398, 920, 504]]<|/det|>
+1. Three biosynthetic gene clusters were identified, yet only two were functionally characterized. For completeness, all three BGCs should be included in the analyses for their role in providing protection against Bt247; 
+2. Are the protective effects specific for Bt247? 
+3. one cannot conclude that NRPS does not contribute to the protection against Bt247 with an active PKS. The experiments should be repeated in a Asga background to address whether AsgaAnrpA reduces survival compared to Asga and/or AnrpA. The third BGC should also be included. 
+4. while there is a phenotype observed, what is known about leaky expression of the arabinose promoter in this system? 
+5. It is expected that gene deletion mutants are complemented for function  
+
+<|ref|>sub_title<|/ref|><|det|>[[72, 541, 196, 553]]<|/det|>
+## Minor Comments:  
+
+<|ref|>text<|/ref|><|det|>[[70, 553, 920, 618]]<|/det|>
+6. L82: it should read 'intraspecies' and not 'interspecies'; 
+7. L109: this reads odd; please rephrase; 
+8. Fig. 1: since these are separate infection studies, and not consecutive sampling, data should be shown as bars instead of continuous lines. 
+9. Fig 5A: please show actual data rather than the abstract interpretation  
+
+<|ref|>sub_title<|/ref|><|det|>[[72, 645, 144, 658]]<|/det|>
+## Version 1:  
+
+<|ref|>text<|/ref|><|det|>[[72, 670, 218, 684]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[72, 696, 160, 710]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[72, 723, 236, 736]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 737, 710, 750]]<|/det|>
+This is an interesting, important study. I do not have any further concerns with the manuscript.  
+
+<|ref|>sub_title<|/ref|><|det|>[[72, 763, 162, 776]]<|/det|>
+## Reviewer #2  
+
+<|ref|>text<|/ref|><|det|>[[72, 789, 236, 802]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 802, 904, 841]]<|/det|>
+The authors have revised their manuscript with great diligence and have addressed the majority of the concerns raised, particularly those related to the identification of bacterial SLs in the first section. I support the public communication of this work.  
+
+<|ref|>text<|/ref|><|det|>[[72, 841, 911, 880]]<|/det|>
+However, I am still not fully satisfied with the organization of the figures. The order in which the figures are presented does not always correspond to the sequence in which they are referenced in the Results section. For example, Figure 1C is first mentioned in the second paragraph, while Figures 1D- 1F are already discussed in the first paragraph.  
+
+<|ref|>text<|/ref|><|det|>[[72, 880, 916, 933]]<|/det|>
+Additionally, the main figures present only a limited portion of the results, whereas the supplementary material contains 22 figures, some of which include highly relevant data. Given that many readers may not consult the supplementary information in detail, I strongly recommend that the authors incorporate some of these key findings into the main figures to enhance the accessibility and impact of the manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[72, 933, 238, 945]]<|/det|>
+Additional minor remark
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 46, 910, 88]]<|/det|>
+line 212: two dots after "Figure S4)"line 279: This may indicate that MYb115- derived SLs do not 280 strongly affect C. elegans on the transcript level, but more strongly influence the host on the proteome or metabolome level.  
+
+<|ref|>text<|/ref|><|det|>[[72, 99, 482, 127]]<|/det|>
+line 305: "we \(\Delta \Delta\) integrated" what does \(\Delta \Delta\) integrated mean?Similarly, line 303: colonisation \(\Delta\) propanate  
+
+<|ref|>sub_title<|/ref|><|det|>[[73, 152, 162, 165]]<|/det|>
+## Reviewer #3  
+
+<|ref|>text<|/ref|><|det|>[[73, 177, 238, 190]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[73, 191, 889, 231]]<|/det|>
+The authors have performed a comprehensive revision of their original submission, and I believe that all of my original concerns have been fully addressed. The (new) data strongly support the authors' conclusions. Their findings present a major advance. The authors must be lauded for their efforts!  
+
+<|ref|>text<|/ref|><|det|>[[72, 583, 916, 638]]<|/det|>
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+<|ref|>text<|/ref|><|det|>[[72, 638, 915, 704]]<|/det|>
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source. The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.  
+
+<|ref|>text<|/ref|><|det|>[[73, 702, 618, 716]]<|/det|>
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 85, 478, 100]]<|/det|>
+Reviewer's Responses to Questions in grey  
+
+<|ref|>sub_title<|/ref|><|det|>[[119, 116, 310, 131]]<|/det|>
+## Our responses in blue  
+
+<|ref|>text<|/ref|><|det|>[[118, 164, 844, 197]]<|/det|>
+We thank all reviewers for thoughtful reading of the manuscript and the very constructive criticism.  
+
+<|ref|>text<|/ref|><|det|>[[119, 221, 430, 237]]<|/det|>
+Reviewer #1 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[117, 253, 881, 450]]<|/det|>
+This is an important study characterizing a polyketide synthase from Pseudomonas fluorescens MYb115, a microbe associated with C. elegans in nature that has been shown to offer C. elegans protection from pathogenic Bt infections. The researchers identified the polyketide synthase (PKS) activity, deleted it, and found that the microbe no longer offers protection against Bt infection in C. elegans. They used LC- MS to identify the product of the polyketide synthase, and found that the enzyme produces long- chain sphinganine molecules and then used isotopic labeling to show that serine is incorporated into the sphinganine molecules allowing them to propose a pathway for the activities of the PKS. This is very solid work and is important and interesting because the researchers also identified genes encoding PKS activity in many other microbes, especially those who interact with host organisms. This is a significant finding because the PKS were previously believed to be found in fungi and rarely in bacteria.  
+
+<|ref|>text<|/ref|><|det|>[[117, 464, 878, 707]]<|/det|>
+The investigators then sought to determine how the PKS products protect C. elegans from Bt pathogens. Sphingolipids are important components of the intestinal barrier, so the researchers used transcriptomic and lipidomic analysis to determine changes in the C. elegans sphingolipids in worms eating the wild type bacteria vs the bacteria in which the PKS was deleted. Interestingly, they found that the C. elegans that were colonized by the wild type MYb115 bacteria had lower levels of many spingolipid species compared to worms who were colonized by the mutant bacteria lacking PKS. However, this may be minor because only a few species were significantly changed. Researchers then knocked down sphingolipid metabolism genes in C. elegans and examined susceptibility to Bt. They found that knocking down enzyme in the de novo pathway leading to ceramide protected worms from infection, while knocking down sphingomyelin synthase, which converts ceramide to sphingomyelin, resulted in more susceptible worms. The MYb115 somewhat rescued this susceptibility of the sms- 1 knockout, although this is puzzling, because the researchers showed that colonization by MYb115 leads to reduced sphingomyelins in worms, so it is unclear how this rescue occurred.  
+
+<|ref|>text<|/ref|><|det|>[[117, 708, 857, 888]]<|/det|>
+Our reply: We thank the reviewer for this valid comment. We agree that we do not understand how exactly MYb115 interacts with C. elegans sphingolipid metabolism. Sphingolipid homeostasis is controlled by a complex network comprising several levels of regulation. The enzymes responsible for sphingolipid production and turnover comprise a metabolic network that gives rise to numerous bioactive molecules, which participate in highly complex and interconnected pathways influencing a multitude of physiological processes (Hannun and Obeid, 2018 https://doi.org/10.1038/nrm.2017.107). Also, sphingolipid metabolism shares common substrates with other metabolic routes and is, for example, highly connected to other lipid metabolic networks. Consequently, imbalances in sphingolipid metabolism in a mutant may have far- reaching consequences for host physiology. Thus, the effect of MYb115 on the phenotype of a sphingolipid metabolism
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 816, 116]]<|/det|>
+mutant (increased survival after Bt infection) may not be directly linked to its effect on wildtype worms (decrease in sphingomyelin).  
+
+<|ref|>text<|/ref|><|det|>[[117, 117, 866, 295]]<|/det|>
+Nevertheless, our comprehensive data sets, including several new data sets now added to the revised manuscript, do allow us to draw several important conclusions on commensal- mediated immune protection. Important new insights are for example that sphingolipid- producing MYb115 does cause a decrease of certain host sphingolipid species, including sphingomyelin species, in comparison to non- sphingolipid- producing MYb115. Moreover, MYb115 ameliorates the survival phenotype of C. elegans sphingolipid enzyme mutants following Bt infection in comparison to OP50. We now integrated a critical evaluation of our data into the discussion and further clarify that the current comprehensive data sets do not yet provide a final answer on whether the observed effect of MYb115 on the survival phenotype of the sphingolipid enzyme mutant is direct (on sphingolipid metabolism) or indirect (p. 17 and 18, line 495- 501).  
+
+<|ref|>text<|/ref|><|det|>[[117, 295, 879, 586]]<|/det|>
+Also, we have now invested considerable efforts in further improving our understanding of MYb115 sphingolipid biosynthesis and the effects produced by MYb115, including biochemical characterization of the enzymes encoded in the biosynthetic gene cluster (BGC) that catalyze MYb115 sphingolipid biosynthesis (in collaboration with experts in sphingolipid biosynthesis), further metabolic analyses of bacterial cultures (in collaboration with experts in metabolic analyses), and further phenotypic analyses in C. elegans demonstrating – among others – that protection against another Bt strain, Bt679, and protection of the C. elegans intestinal barrier following Bt infection depends on MYb115- produced sphingolipids and that MYb115- mediated protection is independent of two known C. elegans Bt defense pathways (p38 and JNK MAPK pathways), the mitochondrial surveillance response, and of a Bt toxin glycosphingolipid receptor (please also see our answers to the reviewers' comments below). This paper is well written and the data are clearly and thoroughly presented. While the C. elegans experiments did not clarify how the sphinganine products from MYb115 lead to protection from Bt infection in C. elegans, the observations are important starting points for future experiments. The researchers used rather vague terms, for example on line 309 "These data indicate that MYb115 interacts with host sphingolipid metabolism...." While the nature of the interaction isn't clear, at this point I think this is a valid statement, because more definitive mechanism can't be determined from the data.  
+
+<|ref|>text<|/ref|><|det|>[[120, 587, 560, 602]]<|/det|>
+Our reply: We thank the reviewer for their comments.  
+
+<|ref|>text<|/ref|><|det|>[[118, 603, 872, 683]]<|/det|>
+One technical comment. When researchers receive deletion lines from the CGC (ok alleles) or NBRP (tm alleles) they typically perform outcrosses against wild type worms since the original deletion strains can contain many other mutations. The researchers should report how many times the strains were outcrossed prior to their experiments, or else say that they were not outcrossed.  
+
+<|ref|>text<|/ref|><|det|>[[119, 684, 869, 749]]<|/det|>
+Our reply: We genotyped the respective sphingolipid metabolism pathway mutant and thus confirmed the deletion alleles, but we did not outcross the mutants. We now clarify that the mutants were not outcrossed in the materials and methods section and show the detailed genotyping results in a supplementary figure (Figure S26).  
+
+<|ref|>sub_title<|/ref|><|det|>[[120, 780, 430, 796]]<|/det|>
+## Reviewer #2 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[118, 813, 877, 894]]<|/det|>
+The manuscript Peters et al. identified an iterative type I polyketide synthase (PKS) in MYb115 and showed that PKS regulates bacterial sphingolipid biogenesis and this interferes with sphingolipid metabolism of the host, thus exerting a protective role for the host against pathogen infection. The paper contains a number of interesting observations, especially identification of an SPT- independent mechanism for sphingolipid biogenesis.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 83, 872, 230]]<|/det|>
+Our reply: We agree with the reviewer that the discovery of a new way of bacterial sphingolipid synthesis significantly advances our knowledge on bacterial sphingolipid metabolism. While the few known bacterial sphingolipid producers (such as Bacteroidetes), like eukaryotes, produce sphingolipid as primary metabolites in a manner that depends on the enzyme serine palmitoyltransferase (SPT), P. fluorescens MYb115 produces sphingolipids as secondary metabolites in a non- canonical way that depends on a BGC, the polyketide synthase (PKS) P/SgaB. To our knowledge this is the first example of a bacterial PKS shown to be involved in sphingolipid biosynthesis and also the first description of a Pseudomonas isolate as sphingolipid producer.  
+
+<|ref|>text<|/ref|><|det|>[[115, 230, 874, 560]]<|/det|>
+Moreover, in the revised manuscript we now present and discuss a comprehensive set of new results that considerably expand our knowledge on the P. fluorescens BGC- dependent sphingolipid biosynthesis pathway: We previously could only assume that the enzyme encoded by P/SgaB within the BGC substitutes the function of SPT - the above- mentioned rate- limiting enzyme for sphingolipid biosynthesis in all eukaryotes and the few bacterial sphingolipid producers. In the revised manuscript, we now integrated comprehensive new data, which we generated in collaboration with Dominic Campopiano, Michael Herrera, and Francesca Lubbock (University of Edinburgh, all now co- authors) on the function of P/SgaB. Using heterologous expression in E. coli and subsequent in vitro functional analysis with the purified enzyme we could prove that P/SgaB functions as SPT. Furthermore, we identified a putative short chain dehydrogenase/reductase (SDR) in the MYb115 sphingolipid BGC, which is predicted to share structural homology with 3- ketodihydropsphingine reductase (KDSR). KDSR is known to catalyze the reduction of 3- KDS to dihydropsphinganine (DHS) (Beeler et al., 1998. https://doi.org/10.1074/jbc.273.46.30688; Fornarotto et al., 2006. https://doi.org/10.1016/j.bbalip.2005.11.013); whilst this step is ubiquitous in eukaryotic sphingolipid biosynthesis, it is unusual in bacterial sphingolipid pathways (Stankevicute et al., 2022. https://doi.org/10.1038/s41589-021-00948-7). The inclusion of this eukaryotic- like step further distinguishes the P. fluorescens sphingolipid- producing BGC from canonical bacterial sphingolipid biosynthesis. In the revised manuscript, we now highlight these new exciting findings in a new section and new figure (see lines 204- 241 and Figure 2).  
+
+<|ref|>text<|/ref|><|det|>[[117, 559, 875, 620]]<|/det|>
+However, the data did not provide a significant novelty and new insights into mechanistic explanation for microbiota- host interaction in regulation of host immune response, given that the role of sphingolipid in immune response and the interplay between bacterial sphingolipid and host sphingolipid biogenesis/metabolism are known.  
+
+<|ref|>text<|/ref|><|det|>[[117, 620, 876, 779]]<|/det|>
+Our reply: Many thanks for this comment. We politely disagree. Our work very clearly and substantially advances our knowledge on the functional significance of a bacterial, sphingolipid- producing PKS in microbiota- mediated protection against pathogens. We found this PKS in many other bacteria that, like Pseudomonas, are not yet known sphingolipid producers. This is a significant novelty since the interplay between bacterial sphingolipids and host sphingolipid metabolism has so far only been reported for one bacterial phylum and one host, the Bacteroidetes in the human gut (e.g. Johnsen et al., 2020. https://doi.org/10.1038/s41467- 020- 16274- w; Le et al., 2022. https://doi.org/10.1016/j.chom.2022.05.002; Brown et al., 2019. https://doi.org/10.1016/j.chom.2019.04.002). Most importantly, to date, the relevance of this interplay has never been reported in the context of microbiota- mediated protection against pathogens. Also, we show that a Pseudomonas species alters sphingolipid metabolism in C. elegans and establish the importance of C. elegans sphingolipid metabolism for survival after Bt infection. We agree with the reviewer that a more detailed study on the mechanism by which MYb115- derived sphingolipids modify host sphingolipid metabolism and/or immune response is an interesting topic for a future study. However, such an analysis is a manuscript of its own and thus goes beyond the scope of the present study. In response to the reviewer's comment, we now more clearly highlight the importance of our numerous novel insights and further point to promising future research directions.  
+
+<|ref|>text<|/ref|><|det|>[[117, 779, 880, 905]]<|/det|>
+interplay has never been reported in the context of microbiota- mediated protection against pathogens. Also, we show that a Pseudomonas species alters sphingolipid metabolism in C. elegans and establish the importance of C. elegans sphingolipid metabolism for survival after Bt infection. We agree with the reviewer that a more detailed study on the mechanism by which MYb115- derived sphingolipids modify host sphingolipid metabolism and/or immune response is an interesting topic for a future study. However, such an analysis is a manuscript of its own and thus goes beyond the scope of the present study. In response to the reviewer's comment, we now more clearly highlight the importance of our numerous novel insights and further point to promising future research directions.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 770, 100]]<|/det|>
+In addition, some of the claims are also not sufficiently supported by the results.  
+
+<|ref|>text<|/ref|><|det|>[[118, 100, 816, 132]]<|/det|>
+Our reply: We hope that we could sufficiently address this concern (see our answers below).  
+
+<|ref|>text<|/ref|><|det|>[[118, 133, 795, 164]]<|/det|>
+For these reasons, I am not recommending publication in Nature Communications. Major issues:  
+
+<|ref|>text<|/ref|><|det|>[[118, 165, 833, 197]]<|/det|>
+1. sgaAB encoded PKS catalyses sphingolipid biosynthesis and improves resistance of animals against Bt247.  
+
+<|ref|>text<|/ref|><|det|>[[118, 198, 856, 262]]<|/det|>
+Our reply: We thank the reviewer for bringing to our attention the ambiguities in MYb115- mediated protection listed below. In the revised manuscript, we have added a new paragraph (line 264- 299) summarizing the new insights we have gained on how the host response to Bt247 is affected by MYb115 and MYb115ΔsgaA.  
+
+<|ref|>text<|/ref|><|det|>[[118, 262, 870, 295]]<|/det|>
+a) Is the protective effect of MYb115 linked with altered activity of known pathways for worm innate immunity, e.g. p38, MAPK, JUN-1, ELT-2, necrosis pathway or bre genes?  
+
+<|ref|>text<|/ref|><|det|>[[117, 295, 876, 444]]<|/det|>
+Our reply: Following the reviewer's suggestion, we tested the involvement of the p38 MAPK and the JNK-like MAPK pathway in MYb115-mediated protection against Bt247. We found that the protective effect is completely independent of the p38 MAPK pathway (MYb115 also protects the p38 MAPK pathway tir-1, nsy-1, sek-1, and pmk-1 mutants) or the JNK-like MAPK KGB-1 (Figure S19). We could previously already exclude an involvement of the bre genes in C. elegans defense against Bt247, given that bre mutants are susceptible to Bt247 infection (see discussion line 511- 518). However, because of the direct link between bre genes and the biogenesis of complex glycosphingolipids, we confirmed these previous findings and included the data in Figure S25B. Please also see our response to the reviewer's comment 3b below.  
+
+<|ref|>text<|/ref|><|det|>[[118, 444, 820, 461]]<|/det|>
+b) How does MYb115 affect toxicity of the other Gram-positive pathogen, e.g. Bt679?  
+
+<|ref|>text<|/ref|><|det|>[[117, 461, 879, 558]]<|/det|>
+Our reply: We have previously already shown that MYb115 also protects against another Bt strain, Bt679 (Kissoyan and Peters et al. https://doi.org/10.3389/fcimb.2022.775728). We now mention this in the revised manuscript (line 272- 274). The comments by reviewer #2 and #3 prompted us to test if protection against Bt679 is also lost on the MYb115 ΔsgaA mutant and indeed it is. We now included this new additional result in Figure S17. We did not test other Gram-positive bacteria.  
+
+<|ref|>text<|/ref|><|det|>[[118, 559, 850, 590]]<|/det|>
+c) Can the MYb115 ΔsgaAB phenotype be rescued by restoring sgaAB expression in this mutant bacterium?  
+
+<|ref|>text<|/ref|><|det|>[[117, 591, 878, 737]]<|/det|>
+Our reply: In response to the comments of reviewer #2 and #3 we have now conducted a series of experiments to functionally complement the MYb115 ΔsgaA mutant by inserting the vanillic acid inducible PvanCC promoter in front of sgaA or the complete sgaAB BGC on the plasmid pSEVA631, which was then introduced into the MYb115 mutant. Only the complementation that included the complete sgaAB BGC restored sphingolipid production (detection of compounds 1 (m/z 414.4 [M+H]+), 2 (m/z 386.4 [M+H]+), and 3 (m/z 442.4 [M+H]+) by LC-MS (Figure S27). We observed a clear increase in resistance to Bt247 of worms on this complemented MYb115 ΔsgaA mutant and now present these additional new results in Figure 1F of the revised manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[117, 737, 878, 883]]<|/det|>
+Of note: We obtained two complemented MYb115 ΔsgaA mutant strains. While these strains were meant for targeted activation of the BGC, we realized that the vanillic acid inducible promoter was leaky (and more so in one strain than in the other), resulting in sphingolipid production also in the absence of vanillic acid. However, we used this to show that variations in sphingolipid production are reflected in variations in the protective effect, providing further evidence that host protection is dependent on bacterial sphingolipid production. The whole set of obtained results do clearly demonstrate the role of the complete P/SgaAB BGC in sphingolipid production and host protection, which we now emphasize in the revised manuscript (line 179- 192 and Figure 1H).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 84, 864, 116]]<|/det|>
+d) Can expression of sgaAB in another bacterium which normally doesn't have a protective role against Bt247 result a protection?  
+
+<|ref|>text<|/ref|><|det|>[[117, 117, 868, 216]]<|/det|>
+Our reply: This is an interesting point. Yet, it is not essential to demonstrate a causal effect of sgaAB on protection against pathogenic Bt. Such evidence was produced with MYb115, which we now substantiated with the complementation of the MYb115 ΔsgaA mutant. Therefore, we decided to prioritize these additional complementation experiments (requested by reviewer #2 and #3), in order to further advance our understanding, and as a consequence, we did not include the here proposed experiment at this time.  
+
+<|ref|>text<|/ref|><|det|>[[118, 216, 860, 246]]<|/det|>
+e) How is the intestine morphology and integrity upon Bt247 infection affected by MYb115 and sgaAB?  
+
+<|ref|>text<|/ref|><|det|>[[117, 246, 874, 358]]<|/det|>
+Our reply: MYb115 limits Bt- induced damage to the intestinal epithelium, as we already demonstrated previously (Kissoyan et al., 2019. https://doi.org/10.1016/j.cub.2019.01.050). Prompted by the reviewer's question, we now further tested if MYb115- derived sphingolipids are involved in mitigating Bt- induced damage. Using the C. elegans PGP- 1::GFP reporter strain, we confirmed that MYb115 PKS- derived sphingolipids reduced membrane damage caused by Bt infection, while the MYb115 ΔsgaA mutant did not. We included these new results in the revised manuscript (line 290- 299. Figure S20).  
+
+<|ref|>text<|/ref|><|det|>[[117, 359, 866, 536]]<|/det|>
+f) Line 146: MYb115 lipidomic analysis was performed to test whether the bacteria sphingamines exist as free compounds or are part of lipids and sphingamines 1-3 and PG- sphingolipids 4-6 were identified. Does this result suggest that the bacteria sphingamines exist as both free compounds and part of lipids? A clear conclusion to the question should be provided to make understanding easier, especially for those who are not familiar with sphingolipid metabolism. In addition, I don't understand why elucidating the existing form of sphingamines is important to know how MYb115 interact with the host. Additional background explanations are required to understand the underlying logic. The author concluded that they are not able to differentiate between the effects of the individual sphingolipid species. I don't think that lipidomic analysis alone would help to answer this question. But supplementation of these sphingamines or PG- sphingolipids might do.  
+
+<|ref|>text<|/ref|><|det|>[[117, 537, 880, 735]]<|/det|>
+Our reply: Many thanks for these valuable comments. The MYb115 lipidomic analysis using high- resolution Liquid Chromatography Tandem Mass Spectrometry (HRES- LC- MS/MS) allowed us to identify the PG- sphingolipids 4- 6. We now clarify this (line 167- 170). The supplementation experiments would require purified sphingolipids from MYb115, which however is not available and thus, could not easily be obtained. Therefore, we decided for a different approach to further elucidate the involvement of different sphingolipids. In collaboration with Manuel Liebeke (Kiel University, now co- author) using MALDI mass spectrometry spot assays (https://doi.org/10.1038/s41596- 023- 00864- 1), we visualized sphingamines 1- 3 and PG- sphingolipid 4- 6 in bacterial cultures, whose protective effect we then tested in survival analyses. These additional experiments demonstrate that protection significantly correlates with the abundance of sphingamines 1- 3 and PG- sphingolipid 4, indicating that host protection is dependent on these sphingolipids. In the revised manuscript, we now added and explain these new data (line 179- 192 and Figure 1H).  
+
+<|ref|>text<|/ref|><|det|>[[118, 735, 313, 749]]<|/det|>
+2. Transcriptome study:  
+
+<|ref|>text<|/ref|><|det|>[[118, 750, 878, 813]]<|/det|>
+a) it is surprising that Bt247 infection only cause few numbers of DEG in MYb115 and worms on the MYb115 ΔsgaAB mutants, given their significantly different survival rate. As the RNA-seq results was not shown in any main figures or supplemental data, expect an enrichment analysis in Fig. 4A and S4A, it is difficult to make much comments.  
+
+<|ref|>text<|/ref|><|det|>[[117, 814, 878, 910]]<|/det|>
+Our reply: The transcriptome response of C. elegans to MYb115 and MYb115 ΔsgaA is indeed very similar under both conditions, infected and non- infected. We agree with reviewer #2 that this result is surprising and unexpected. As discussed also below, the only difference between MYb115 and MYb115 ΔsgaA is the production of sphingolipids. We do think that sphingolipid production may not strongly affect C. elegans on the transcript level, but more strongly influences the host on the proteome/metabolome level. We mention this in the
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 83, 875, 189]]<|/det|>
+revised manuscript in line 279- 281. We thank the reviewer for pointing out that we did not clearly present the RNAseq data in the manuscript. We now added a short description of the results in the context of host pathogen defense (line274- 287) and present the data in Figure S18 and Table S7. As before, raw data and processed data are accessible through GEO Series accession number GSE245296 at NCBI's Gene Expression Omnibus. Does Bt247 infection affect transcriptome significantly? Does it result in activation of innate immunity genes?  
+
+<|ref|>text<|/ref|><|det|>[[117, 198, 860, 310]]<|/det|>
+Our reply: Yes, we have previously shown in several independent studies that Bt247 infection has a strong effect on the C. elegans transcriptome and results in activation of C. elegans pathogen- responsive/innate immunity genes (Boehnisch et al., 2011. https://doi.org/10.1371/journal.pone.0024619. Nakad et al., 2016. https://doi.org/10.1186/s12864- 016- 2603- 8. Yang, Dierking et al., 2015. https://doi.org/10.1016/j.dci.2015.02.010 Zarate- Potes et al., 2020. https://doi.org/10.1371/journal.ppat.100882).  
+
+<|ref|>text<|/ref|><|det|>[[117, 310, 864, 374]]<|/det|>
+If Bt247 does not cause a significant alteration in gene expression, it would be hard to find involved factors via comparing transcriptome of animals on MYb115 and on the MYb115 AsgaAB mutants. In a such scenario, MYb115 might protect animals via other mechanisms that are independent of influencing transcript level.  
+
+<|ref|>text<|/ref|><|det|>[[120, 375, 647, 391]]<|/det|>
+How does MYb115 affect transcriptome without Bt247 infection?  
+
+<|ref|>text<|/ref|><|det|>[[117, 392, 875, 483]]<|/det|>
+Our reply: For this study, we specifically compared the C. elegans transcriptome response to MYb115 in comparison to the non- sphingolipid producing MYb115 AsgaA mutant. We did not include the laboratory food bacterium E. coli OP50 as control, which would be necessary to assess the general effect of MYb115 on the C. elegans transcriptome without Bt247 infection. The comparison between the C. elegans response to OP50 and MYb115 was not the subject of the study.  
+
+<|ref|>text<|/ref|><|det|>[[117, 483, 868, 530]]<|/det|>
+If the Bt247 infectious bacteria remodel the worms' transcriptome much more than MYb115 does, the the potential changed genes that are only influenced by MYb115 or MYb115 AsgaAB mutants might submerge among Bt247- caused alteration in gene expression.  
+
+<|ref|>text<|/ref|><|det|>[[117, 530, 875, 655]]<|/det|>
+Our reply: We thank the reviewer for their comment. This may indeed be the case and, as discussed above, MYb115- derived sphingolipid may affect the host response more strongly on the proteome level. We now address this aspect in the revised manuscript (line 279- 281). b) The transcriptome data was then integrated into the iCEL1314 genome- scale metabolic model and 24 and 23 significant differences in the presence or absence of Bt247 were received. What does "24 and 23 significant differences" mean? Does it refer to Differences between generated metabolic models? The results must be described in more details so that non- experts in iCEL1314 could understand the data.  
+
+<|ref|>text<|/ref|><|det|>[[117, 658, 878, 707]]<|/det|>
+Our reply: In the methods section of the revised manuscript we did explain more in detail what the different data types are and how metabolic models were applied. In the revised results, we now point to the methods section for these details (line 307).  
+
+<|ref|>text<|/ref|><|det|>[[117, 708, 880, 855]]<|/det|>
+Additionally, we have now simplified the data types that we analysed. Previously, we had the OFD results, lower bound and upper bound - we performed statistical analysis on them separately. However, the interpretation of differences was difficult. In the revised manuscript, we now combined the lower and upper bound values into one score - the center, which represents: (upper bound - lower bound) / 2. Centers are more representative of the solution spaces for each of the reactions, and coefficients are easier to interpret. After doing so, a few of the previous reactions are no longer present among the significant results - instead of 24 (KO vs WT with Bt247) and 23 (KO vs WT without Bt247), we obtained 16 reactions in each contrast. We have updated the manuscript accordingly (line 305- 322).  
+
+<|ref|>text<|/ref|><|det|>[[117, 856, 880, 888]]<|/det|>
+What does the "Gene Ratio" in the Figure 4A mean? Fig. 4A and S4A only show enrichment of metabolic processes even without notifying whether these processes are activated or
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 85, 881, 214]]<|/det|>
+inactivated by sgaAB in MYb115 (whether genes involved are up- or down- regulated). A list of genes used for the enrichment analysis would also be important to understand the result. Our reply: We would like to thank the reviewer for drawing our attention to this point. We did not use genes for enrichment, rather the reactions significant after our statistical analyses. The "pathway" universe were the subsystem annotations of all reactions within the model. Since we identified only few reactions as significantly up- or down- regulated an enrichment on them separately would not yield any results. We changed the mislabeled "Gene ratio" to "Reactions". The individual significant reactions are provided in Table S10.  
+
+<|ref|>text<|/ref|><|det|>[[115, 216, 881, 512]]<|/det|>
+Importantly, our goal for this analysis was to understand which pathways are affected, not necessarily specify which reactions are affected. Additionally, we do not want to show the coefficients in the main manuscript because they can be misinterpreted in the traditional sense due to the fact that in metabolic models, the positive or negative sign of reactions actually represents the directionality of a reaction. For example, a reaction can be irreversible, e.g. substrate - > product, or reversible, i.e. substrate <-> product. In our hypothetical irreversible reaction, the reaction flux can only have positive values (it's irreversible), so here a negative coefficient would mean that the flux through this reaction in the worms grown on mutant bacteria is lower. However, for our hypothetical irreversible reactions, values can be both positive and negative. So, if all our values were negative, a positive coefficient would actually mean that the worms grown on mutant bacteria would have less flux (e.g. - 0.5 vs - 2) through this reaction. Since we have a large mix of reversible and irreversible reactions within our network, and we have not only looked at flux potentials, but also at mathematical optimal flux distribution (which is the optimal but not unique solution to the linear optimization problem), we cannot reliably infer the significant involvement of individual reactions based on these statistical analysis results. Nevertheless, we can see trends from our analysis on the pathway level, which is why we have described our results as we have. These aspects have now been briefly explained in the methods section (line 701- 725).  
+
+<|ref|>text<|/ref|><|det|>[[117, 512, 852, 609]]<|/det|>
+c) Line 220: "Here, we also saw an enrichment in valine, leucine, and isoleucine degradation, which are branched-chain amino acids (BCAA). This pathway is directly connected with propanoate metabolism that provides components for the synthesis of the C15iso fatty acid, which is the precursor for sphingolipids in C. elegans": are these genes up- or down-regulated in MYb115 vs. worms on the MYb115 \(\Delta\) sgaAB mutants? Again, provide more RNA-seq information will avoid such a confusion.  
+
+<|ref|>text<|/ref|><|det|>[[117, 609, 881, 790]]<|/det|>
+Our reply: In Table S10 we have added the information about the genes that encode the significant reactions, as well as the logFC values of these genes from the transcriptomic analysis. We would like to point out that the iMAT++ algorithm used to integrate the transcriptomic data into the iCEL1314 metabolic model is done on a sample basis, therefore any statistical comparisons of gene expression is not taken into account during this process. This means that comparing the differences in simulation results of reactions encoded by a certain gene and the logFC values of this gene might not directly match in direction. This is a desirable attribute of metabolic modelling since we can predict metabolic requirements of an organism that are in conflict with the gene expression differences - these might be caused by post-translational modifications or other effects. These aspects have now been briefly explained in the methods section (line 710- 716).  
+
+<|ref|>text<|/ref|><|det|>[[117, 790, 579, 806]]<|/det|>
+d) The RNA-seq data should be validated via qRT-PCR.  
+
+<|ref|>text<|/ref|><|det|>[[117, 806, 857, 888]]<|/det|>
+Our reply: The metabolic network analysis, which is based on the RNAseq data, revealed that sphingolipid metabolism reactions show differential activity between MYb115 and MYb115 \(\Delta\) sgaA. We focused on this result and validated the transcriptomic data using lipidomic profiling. The other results of the RNAseq study are not the focus of this project. 3. Host sphingolipid metabolism interferes with Bt247 susceptibility  
+
+<|ref|>text<|/ref|><|det|>[[115, 888, 855, 903]]<|/det|>
+a) From the host lipidomics analysis shown in Fig. 4b, sgaAB in MYb115 seems to reduce
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[116, 82, 875, 247]]<|/det|>
+level of Cer, HexCer, DhCer and SM in the host. To test whether accumulation of these molecules reduces survival of animals upon Bt247 infection, the authors used different mutants in the sphingolipid metabolism pathway to test their susceptibility to Bt247. However, cgt- 1 and cerk- 1 mutant animals fed with OP50 show increased survival, although these animals are supposed to have Cer accumulation. In contrast, asm- 3, hyl- 1, hyl- 2, splt- 1 and splt- 3 mutants which probably have reduced Cer level are also resistant. Do these results argue against an important role of Cer in Bt247 susceptibility? In summary, from these Bt247 survival assay with different mutants in sphingolipid metabolism pathway, the author could not provide a clear answer which classes of host sphingolipid are responsible for the altered susceptibility to the pathogen.  
+
+<|ref|>text<|/ref|><|det|>[[116, 246, 880, 473]]<|/det|>
+Our reply: We thank the reviewer for this valid comment. We tested four ceramide metabolic gene mutants, all with an expected increased ceramide content, but two mutants (cerk- 1 and cgt- 1) show a variable survival phenotype and two mutants (asah- 1,2 and sms- 1) are more susceptible to infection. In response to the comments of reviewer #2 we now further assessed the role of specific sphingolipids in defense against Bt infection and supplemented Bt infected worms, using a set of new experiments with the commercially available sphingolipids ceramide, sphingomyelin, and sphingosine- 1- phosphate. We found that supplementation with C18 and C20 ceramide significantly improved survival rates, while C22 ceramide, C16 sphingomyelin, C18 sphingomyelin, d- sphingosine, and S1P did not affect survival. These new results and the phenotypic differences between the ceramide metabolic gene mutants imply that the inhibition of ceramide metabolism (and the associated increase in ceramide content) is not the only factor determining susceptibility to infection. We present and discuss these additional new results in the revised manuscript (line 390- 396. Figure S13 and line 483- 492).  
+
+<|ref|>text<|/ref|><|det|>[[118, 472, 872, 505]]<|/det|>
+b) What is account for Bt247 toxicity? Why should altered sphingolipid metabolism influence toxicity of Bt247?  
+
+<|ref|>text<|/ref|><|det|>[[116, 505, 877, 700]]<|/det|>
+Our reply: How altered host sphingolipid metabolism affect defense against Bt and thus Bt toxicity remains mysterious. Since sphingolipids are required for the integrity of cellular membranes, but can also act as bioactive signalling molecules involved in regulation of a myriad of cell activities, there also are a myriad of potential links between sphingolipid metabolism and defense against Bt247. In response to the reviewer's comments above, but also below, we further explored the two most apparent potential links between sphingolipids and C. elegans defense against Bt infection: The involvement of mitochondrial surveillance (see comment by the reviewer below) and the bre genes (see comment 1 a). We did not find any evidence of a link between Bt infection and mitochondrial surveillance or the bre genes. We present and discuss these results in the revised manuscript (line 426- 438. Figure S25 and line 504- 518). Based on our comprehensive data set, we discuss possible alternative routes of how sphingolipid metabolism interacts with Bt247 (line 502- 522).  
+
+<|ref|>text<|/ref|><|det|>[[118, 700, 872, 781]]<|/det|>
+A study from Ruvkun's lab (Liu et al., 2014, Nature) has shown that some nature habitat bacteria cause mitochondria dysfunction and animals respond with mitochondrial surveillance machinery which is ceramide dependent. At the same time some other bacteria could inhibit mitochondrial surveillance to render a more effective virulence. Does Bt247 modulate the mitochondria surveillance?  
+
+<|ref|>text<|/ref|><|det|>[[118, 781, 860, 878]]<|/det|>
+Our reply: Following the reviewer's suggestion, we tested if Bt247 infection induces the mitochondrial stress- induced hsp- 6p::gfp reporter that was also used in the study by Liu et al. Neither Bt247 infection, nor MYb115 induced expression of hsp- 6p::gfp, indicating that mitochondrial surveillance is not involved in MYb115- mediated protection against Bt247 infection. We added these results to the revised version of our manuscript (line 426- 438. Figure S25A).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 100, 765, 117]]<|/det|>
+4. Sphingolipid in MYb115 affect host sphingolipid biogenesis and metabolism.  
+
+<|ref|>text<|/ref|><|det|>[[117, 117, 875, 181]]<|/det|>
+There is no attempt to address how this could happen. Is uptake of the bacterial sphingolipid into worm intestine necessary? Can feeding worms with the identified sphinganines 1- 3 or sphingolipid 4- 6 impact Bt247 resistance? If yes, does this supplementation change sphingolipid composition of worms?  
+
+<|ref|>text<|/ref|><|det|>[[115, 181, 880, 472]]<|/det|>
+Our reply: We agree with the reviewer that the question how exactly MYb115- derived sphingolipids affects host sphingolipid metabolism is highly intriguing, yet its further analysis goes beyond the scope of the current study. Our study did reveal for the first time that P. fluorescens- derived sphingolipids affect C. elegans sphingolipid metabolism and that modulation of host sphingolipid metabolism increases tolerance towards Bt infection - among others. We now provide a comprehensive set of new data to substantiate the findings made. Supplementation experiments, as suggested by the reviewer, would indeed be helpful to differentiate between the effects of the identified sphinganines 1- 3 and PG- sphingolipids 4- 6. As also discussed above (comment 1 f), we unfortunately do not have the MYb115- derived sphingolipids purified, which we would need to do these supplementation experiments. However, in collaboration with Manuel Liebeke (Kiel University, now co- author) using MALDI mass spectrometry spot assays (https://doi.org/10.1038/s41596-023-00864-1), we visualized sphinganines 1- 3 and PG- sphingolipid 4- 6 in bacterial cultures, whose protective effect we then tested in survival analyses. These additional experiments demonstrate that protection significantly correlates with the abundance of sphinganines 1- 3 and PG- sphingolipid 4, indicating that host protection is dependent on these sphingolipids. In the revised manuscript, we now added and explain these new data (line 179- 192 and Figure 1H).  
+
+<|ref|>text<|/ref|><|det|>[[118, 473, 225, 488]]<|/det|>
+Minor issues  
+
+<|ref|>text<|/ref|><|det|>[[118, 489, 878, 553]]<|/det|>
+1. Zeile 377: Notably, in C. elegans, glucosylceramide deficiency was linked to an increase in autophagy which plays an important role in cellular defence after attack by certain Bt performing toxins (PFTs). Does Myb affect autophagy? If yes, could manipulating autophagy affect survival of animals to Bt247?  
+
+<|ref|>text<|/ref|><|det|>[[118, 554, 863, 636]]<|/det|>
+Our reply: We appreciate the reviewer's suggestion regarding the possible involvement of autophagy. This is indeed an interesting idea and could provide valuable insights for future work. However, due to time and resource constraints, we prioritized addressing the most important points in our revisions. Thus, we decided against an additional investigation of autophagy.  
+
+<|ref|>text<|/ref|><|det|>[[118, 636, 875, 717]]<|/det|>
+Fig. S6: Should A1 be compared with B1, A2 with B2 and so on? If yes, why is the mutant A10 in red line and B10 in yellow? The color used for asah- 1 and asah- 2 are too similar to be differentiated. Maybe just add the name of the mutant directly to the figures. Why are there several figures for some mutants, e.g. asah- 1 while for some others only one e.g. cgt- 1 on OP50? It does not seem to be biological replicates as \(n = 4\) for each figure.  
+
+<|ref|>text<|/ref|><|det|>[[117, 717, 876, 911]]<|/det|>
+Our reply: We appreciate the reviewer's feedback regarding the data representation. In response, we have revised the figure to improve clarity. Specifically, we have updated the color scheme: for all assays where worms were exposed to OP50, the lines are now depicted in grey (panel A), while survival assays with worms exposed to MYb115 are shown in blue (panel B). This adjustment ensures consistency with the color scheme used throughout the manuscript. Additionally, as suggested, we have directly labeled each facet with the name of the corresponding mutant strain, further enhancing the figure's clarity. Moreover, we have now added more survival assays, yielding a larger number of independent experiments for each C. elegans mutant strain, and thus much more robust insights into the role of the respective genes. For some mutants i.e. (cgt- 1 and cerk- 1) the additional experiments revealed that the survival phenotype is variable. We thus included the technical and biological replicates in the heat- map of Figure 5C, so that the reader can see
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 84, 870, 149]]<|/det|>
+at a glance how many experimental runs were done and identify variation across technical and biological replicates. The previous and the additional new data now provide an in- depth as well as statistically sound overview of the involvement of different branches of the entire sphingolipid metabolism.  
+
+<|ref|>text<|/ref|><|det|>[[117, 149, 866, 198]]<|/det|>
+3. Fig. 3B: Does the width of the boxes present the number of bacteria? Yes. Information is shown in the supplemental table. Please indicate this in the main figure to make understanding easier.  
+
+<|ref|>text<|/ref|><|det|>[[117, 198, 866, 231]]<|/det|>
+Our reply: We thank the reviewer for pointing out this ambiguity. We added the information in the legend of Figure 3.  
+
+<|ref|>text<|/ref|><|det|>[[120, 262, 430, 278]]<|/det|>
+Reviewer #3 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[117, 294, 875, 392]]<|/det|>
+In silico analyses revealed three biosynthetic gene clusters in P. fluorescens MYb115. Two were characterized, and the authors demonstrate that a PKS cluster produces sphingolipids, which alters sphingolipid metabolism in the host. An active PKS cluster in MYb115 is required to provide protection against Bacillus thuringiensis (Bt), which is (in part) mediated by host sphingolipid metabolism. Thus, a nice interplay between a natural member of the C. elegans microbiota, C. elegans and microbial pathogens has been uncovered.  
+
+<|ref|>text<|/ref|><|det|>[[118, 393, 266, 407]]<|/det|>
+Major Comments:  
+
+<|ref|>text<|/ref|><|det|>[[118, 408, 875, 456]]<|/det|>
+1. Three biosynthetic gene clusters were identified, yet only two were functionally characterized. For completeness, all three BGCs should be included in the analyses for their role in providing protection against Bt247;  
+
+<|ref|>text<|/ref|><|det|>[[117, 457, 875, 555]]<|/det|>
+Our reply: We thank the reviewer for their comments. Although interesting, analysis of the third BSG is not essential to prove involvement of the sgaAB BGC in sphingolipid metabolism and protection against pathogens. We decided to prioritize on other experiments which now yielded a comprehensive data base for obtaining critical additional insights into the role of sgaAB BGC (see also above and below replies), and thus omitted an analysis of the arylpolene pathway.  
+
+<|ref|>text<|/ref|><|det|>[[118, 555, 504, 571]]<|/det|>
+2. Are the protective effects specific for Bt247?  
+
+<|ref|>text<|/ref|><|det|>[[117, 571, 876, 667]]<|/det|>
+Our reply: We have previously already shown that MYb115 also protects against another Bt strain, Bt679 (Kissoyan and Peters et al. https://doi.org/10.3389/fcimb.2022.775728). We now mention this in the revised manuscript (line 272-274). The comments by reviewer #2 and #3 prompted us to test if protection against Bt679 is also lost on the MYb115 ΔsgaA mutant and indeed it is. We now included this result in Figure S17. We did not test other Gram-positive bacteria.  
+
+<|ref|>text<|/ref|><|det|>[[117, 668, 876, 732]]<|/det|>
+3. one cannot conclude that NRPS does not contribute to the protection against Bt247 with an active PKS. The experiments should be repeated in a Δsga background to address whether ΔsgaΔnrpA reduces survival compared to Δsga and/or ΔnrpA. The third BGC should also be included.  
+
+<|ref|>text<|/ref|><|det|>[[117, 733, 875, 911]]<|/det|>
+Our reply: We thank the reviewer for pointing out this ambiguity. Due to time and resource constraints, we prioritized addressing the complementation of the MYb115 ΔsgaA and ΔsgaB mutants requested by reviewer #2 and #3. We now refrain from drawing conclusions regarding the NRPS cluster and rephrased the sentence in line 117- 120 "While the PKS gene cluster affects MYb115- mediated protection, we did not observe significant differences in worm survival with or without arabinose supplementation on the MYb115 PBADnrpA strain (Figure 1B, Table S1), indicating that the MYb115 NRPS gene cluster is not involved in MYb115- mediated protection against Bt infection." to "While the PKS gene cluster affects MYb115- mediated protection, we did not observe significant differences in worm survival with or without arabinose supplementation on the MYb115 PBADnrpA strain (Figure 1B, Table S1). We therefore focused on the MYb115 PKS gene cluster in our subsequent analyses."
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 83, 866, 133]]<|/det|>
+Since we observed a complete loss of sphingolipid production and a clear decrease in host protection for the MYb115 \(\Delta sgaA\) and MYb115 \(\Delta sgaB\) single mutants, we think that it is justified to focus on the PKS cluster within the scope of the current project.  
+
+<|ref|>text<|/ref|><|det|>[[117, 133, 810, 164]]<|/det|>
+4. while there is a phenotype observed, what is known about leaky expression of the arabinose promoter in this system?  
+
+<|ref|>text<|/ref|><|det|>[[117, 164, 875, 295]]<|/det|>
+Our reply: We assessed MYb115 P \(_{BADsga}\) sphingolipid production in an induced (+ arabinose) and non- induced (- arabinose) state by LC- MS and did not detect any sphingolipids in the non- induced state (Figure 1C). In addition, in the revised manuscript, we included MYb115 P \(_{BADsga}\) in an induced (+ arabinose) and non- induced (- arabinose) state in an experiment to visualize production of the different sphingolipids (collaboration with Manuel Liebeke, Kiel University now co- author) and could confirm that P \(_{BAD}\) is not leaky in this system (datasets: MPIMM_514_QE_P https://metaspace2020. org/dataset/2025- 02- 27_13h37m58s).  
+
+<|ref|>text<|/ref|><|det|>[[117, 295, 728, 311]]<|/det|>
+5. It is expected that gene deletion mutants are complemented for function  
+
+<|ref|>text<|/ref|><|det|>[[117, 311, 878, 473]]<|/det|>
+Our reply (also see our response to the comment 1c of reviewer #2 above): In response to the comments of reviewer #2 and #3 we have conducted a series of additional new experiments to functionally complement the MYb115 \(\Delta sgaA\) mutant by inserting the vanillic acid inducible PvanCC promoter in front of sgaA or the complete sgaAB BGC on the plasmid pSEVA631, which was then introduced into the MYb115 mutants. Only the complementation that included the complete sgaAB BGC restored sphingolipid production (detection of compounds 1 (m/z 414.4 [M+H]+), 2 (m/z 386.4 [M+H]+), and 3 (m/z 442.4 [M+H]+) by LC- MS (Figure S27) and only in the MYb115 \(\Delta sgaA\) mutant. We observed a clear increase in resistance to Bt247 of worms on this complemented MYb115 \(\Delta sgaA\) mutant and present the results in Figure 1F of the revised manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[117, 473, 880, 624]]<|/det|>
+Of note: We obtained two complemented MYb115 \(\Delta sgaA\) mutant strains. While these strains were meant for targeted activation of the BGC, we realized that the vanillic acid inducible promoter was leaky, resulting in sphingolipid production also in the absence of vanillic acid (datasets: MPIMM_514_QE_P https://metaspace2020. org/dataset/2025- 02- 27_13h37m58s). However, we used this to show that variations in sphingolipid production are reflected in variations in the protective effect, providing further evidence that host protection is dependent on bacterial sphingolipid production. The whole set of obtained results do clearly demonstrate the role of the complete MYb115 sgaAB BGC in sphingolipid production and host protection, which we now emphasize in the revised manuscript (line 179- 192 and Figure 1H).  
+
+<|ref|>text<|/ref|><|det|>[[118, 625, 266, 639]]<|/det|>
+Minor Comments:  
+
+<|ref|>text<|/ref|><|det|>[[118, 639, 588, 655]]<|/det|>
+6. L82: it should read 'intraspecies' and not 'interspecies';  
+
+<|ref|>text<|/ref|><|det|>[[118, 655, 673, 670]]<|/det|>
+Our reply: We have now replaced "interspecies" with "intraspecies"  
+
+<|ref|>text<|/ref|><|det|>[[118, 671, 456, 685]]<|/det|>
+7. L109: this reads odd; please rephrase;  
+
+<|ref|>text<|/ref|><|det|>[[117, 686, 857, 766]]<|/det|>
+Our reply: The sentence: "Supplementation of the C. elegans laboratory food Escherichia coli OP50 with arabinose did not affect survival, showing that arabinose itself does not influence C. elegans resistance to Bt (Table S1)." has now been reworded to: "Arabinose supplementation had no effect on resistance of C. elegans to Bt infection on its standard laboratory food Escherichia coli OP50."  
+
+<|ref|>text<|/ref|><|det|>[[117, 767, 840, 798]]<|/det|>
+8. Fig. 1: since these are separate infection studies, and not consecutive sampling, data should be shown as bars instead of continuous lines.  
+
+<|ref|>text<|/ref|><|det|>[[117, 799, 805, 846]]<|/det|>
+Our reply: The continuous lines in Figure 1 A, B and E represent a dose- response relationship. As it is common practice for dose- response curves, the data points are connected.  
+
+<|ref|>text<|/ref|><|det|>[[117, 847, 713, 863]]<|/det|>
+9. Fig 5A: please show actual data rather than the abstract interpretation  
+
+<|ref|>text<|/ref|><|det|>[[117, 864, 874, 911]]<|/det|>
+Our reply: We show the survival data in Figure 5C as heatmaps to summarize the results of 32 and 24 survival assays on E. coli OP50 and P. fluorescens MYb115, respectively, and to facilitate the comparison of results between different knockout mutants. We now also
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 83, 864, 149]]<|/det|>
+included technical and biological replicates, so that the reader can see at a glance how many experimental runs were done and to facilitate the identification of variation across technical and biological replicates. For the interested reader, we do provide each individual survival curve in the supplementary Figure S23.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[119, 85, 459, 100]]<|/det|>
+Reviewer's Remarks to the author in grey  
+
+<|ref|>sub_title<|/ref|><|det|>[[119, 116, 310, 131]]<|/det|>
+## Our responses in blue  
+
+<|ref|>text<|/ref|><|det|>[[117, 163, 850, 180]]<|/det|>
+We thank all reviewers for thoughtful re- evaluation of the manuscript and their comments.  
+
+<|ref|>text<|/ref|><|det|>[[119, 205, 430, 221]]<|/det|>
+Reviewer #1 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[119, 238, 787, 270]]<|/det|>
+This is an interesting, important study. I do not have any further concerns with the manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[119, 270, 560, 286]]<|/det|>
+Our reply: We thank the reviewer for their comments.  
+
+<|ref|>text<|/ref|><|det|>[[119, 302, 430, 318]]<|/det|>
+Reviewer #2 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[117, 334, 875, 450]]<|/det|>
+The authors have revised their manuscript with great diligence and have addressed the majority of the concerns raised, particularly those related to the identification of bacterial SLs in the first section. I support the public communication of this work. However, I am still not fully satisfied with the organization of the figures. The order in which the figures are presented does not always correspond to the sequence in which they are referenced in the Results section. For example, Figure 1C is first mentioned in the second paragraph, while Figures 1D- 1F are already discussed in the first paragraph.  
+
+<|ref|>text<|/ref|><|det|>[[119, 450, 715, 465]]<|/det|>
+Our reply: We changed the text according to the reviewer's suggestions.  
+
+<|ref|>text<|/ref|><|det|>[[117, 466, 870, 546]]<|/det|>
+Additionally, the main figures present only a limited portion of the results, whereas the supplementary material contains 22 figures, some of which include highly relevant data. Given that many readers may not consult the supplementary information in detail, I strongly recommend that the authors incorporate some of these key findings into the main figures to enhance the accessibility and impact of the manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[117, 546, 874, 627]]<|/det|>
+Our reply: We thank the reviewer for pointing this out. We now incorporated the data shown in figures S13A into figure 2D, figure S18 B is now (new) figure 4A, figure S19 is now incorporated in figure 4 (new figure 4B and C), and figure S20 is now (new) figure 5. In addition, we combined several supplementary figures, so that we now have a total of 21 (instead of 27) supplementary figures.  
+
+<|ref|>text<|/ref|><|det|>[[119, 628, 312, 641]]<|/det|>
+Additional minor remark  
+
+<|ref|>text<|/ref|><|det|>[[119, 643, 410, 658]]<|/det|>
+line 212: two dots after "Figure S4)  
+
+<|ref|>text<|/ref|><|det|>[[119, 660, 390, 674]]<|/det|>
+Our reply: We removed one dot.  
+
+<|ref|>text<|/ref|><|det|>[[117, 676, 868, 723]]<|/det|>
+line 279: This may indicate that MYb115- derived SLs do not 280 strongly affect C. elegans on the transcript level, but more strongly influence the host on the proteome or metabolome level.  
+
+<|ref|>text<|/ref|><|det|>[[117, 725, 752, 756]]<|/det|>
+Our reply: We changed the sentence according to the reviewer's suggestion. line 305: "we \(\Delta \Delta\) integrated" what does \(\Delta \Delta\) integrated mean?  
+
+<|ref|>text<|/ref|><|det|>[[117, 757, 813, 773]]<|/det|>
+Our reply: We thank the reviewer for pointing out this mistake. We removed the ' \(\Delta \Delta\)  
+
+<|ref|>text<|/ref|><|det|>[[119, 775, 480, 789]]<|/det|>
+Similarly, line 303: colonisation \(\Delta\) propanoate  
+
+<|ref|>text<|/ref|><|det|>[[119, 790, 476, 805]]<|/det|>
+Our reply: We removed the 'colonisation \(\Delta\) .  
+
+<|ref|>text<|/ref|><|det|>[[119, 822, 430, 837]]<|/det|>
+Reviewer #3 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[117, 855, 860, 887]]<|/det|>
+The authors have performed a comprehensive revision of their original submission, and I believe that all of my original concerns have been fully addressed. The (new) data strongly
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 83, 870, 115]]<|/det|>
+support the authors' conclusions. Their findings present a major advance. The authors must be lauded for their efforts!  
+
+<|ref|>text<|/ref|><|det|>[[118, 116, 667, 132]]<|/det|>
+Our reply: We thank the reviewer for their encouraging comments.
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__7160e8ada422dde98fca0e9f5ca2a6bbd511db08fc6ba86516f46299acc3b0ef/images_list.json

@@ -0,0 +1,362 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_0.jpg",
+    "caption": "Figure R1. Heat wave event of 2023. A) Seasonal evolution of maximum (red) and minimum (blue) temperature for Phoenix, Arizona for the year 2023 from 1 June to 20 September. The long-term daily mean is shown by the dashed line whereas the \\(5^{\\mathrm{th}}\\) and \\(95^{\\mathrm{th}}\\) percentiles are shown by the shading region. Excess above the \\(95^{\\mathrm{th}}\\) percentile is shown by red shading plus hatching for both maximum and minimum temperatures. B) Observed histogram of maximum (red) and minimum (blue) temperature for Phoenix, Arizona for the period 1 June to 31 August from 1955-2023. The extremely warm temperatures of 19 July 2023, during the peak amplitude of the heat wave, are shown for reference. c) Five-day averaged surface temperature anomaly centered on 19 July 2023.",
+    "footnote": [],
+    "bbox": [
+      [
+        270,
+        95,
+        728,
+        600
+      ]
+    ],
+    "page_idx": 12
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_1.jpg",
+    "caption": "Figure R2. Heat budget analysis. Energy budget averaged every 10 days over the southwest U.S. and northwest Mexico from 6 June to 24 August 2023. A) 200 hPa geopotential height anomaly [gpm], 850 hPa temperature and 2-meter temperature anomalies [°C]. b) Vertically integrated anomalous heating rates from \\(975 - 800 \\mathrm{hPa}\\) layer [°C day \\(^{-1}\\) ], and c) surface heat fluxes (net surface shortwave and longwave radiation, sensible and latent heat fluxes) [Wm \\(^{-2}\\) ]. Daily anomalies are derived from the long-term monthly mean for the 1979-2023 period.",
+    "footnote": [],
+    "bbox": [
+      [
+        293,
+        315,
+        710,
+        730
+      ]
+    ],
+    "page_idx": 13
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_2.jpg",
+    "caption": "Figure R3. Cloud radiative effects computed from the differences between clear-sky minus all-sky radiative fluxes at the top of the atmosphere (TOA) for a) net fluxes, b) shortwave reflected fluxes, and c) outgoing longwave fluxes. These values represent the five-day average centered on 19 July 2023, during the middle of the heat wave. Regions with dark blue represent little or no clouds.",
+    "footnote": [],
+    "bbox": [
+      [
+        305,
+        95,
+        690,
+        630
+      ]
+    ],
+    "page_idx": 16
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_3.jpg",
+    "caption": "Figure R4. June – September (JJAS) precipitation anomaly during 2023. The red stipples indicate the North American monsoon region defined as where the local summer (May – September) precipitation minus the winter (November–March) precipitation exceeds 2.5 mm per day and the local summer precipitation exceeds \\(55\\%\\) of total annual precipitation. Dataset from the JRA55 reanalysis.",
+    "footnote": [],
+    "bbox": [
+      [
+        280,
+        494,
+        714,
+        740
+      ]
+    ],
+    "page_idx": 17
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_4.jpg",
+    "caption": "Figure R5. A) July 2023 maximum 2-meter temperature anomalies. B) Correlation of maximum 2-meter temperatures and the North American Monsoon (NAM) index for the 1979 – 2022 period (the year 2023",
+    "footnote": [],
+    "bbox": [
+      [
+        128,
+        451,
+        868,
+        866
+      ]
+    ],
+    "page_idx": 18
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_5.jpg",
+    "caption": "Figure R6. Time series of the North American monsoon index for June-August. Years of significant NAM index anomalies are highlighted by color-filled bars. Statistical significance is based on the 99-percentile significance level based on a Student's-t test.",
+    "footnote": [],
+    "bbox": [
+      [
+        113,
+        357,
+        882,
+        537
+      ]
+    ],
+    "page_idx": 19
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_6.jpg",
+    "caption": "Figure R7. Daily evolution of the North American Monsoon (NAM) indices for a) northern domain (blue region in the inserted map) and b) southern domain (red region in the inserted map).",
+    "footnote": [],
+    "bbox": [
+      [
+        196,
+        285,
+        800,
+        638
+      ]
+    ],
+    "page_idx": 20
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_7.jpg",
+    "caption": "Figure R8. a and b) June and July precipitation rate climatology respectively. Panels c) and d) show the precipitation rate anomalies for June and July 2023 respectively. Black stipples in panels c and d) indicate statistical significance at a 95% confidence level based on a Student’s t-test. The precipitation rate is measured in mm per day.",
+    "footnote": [],
+    "bbox": [
+      [
+        128,
+        401,
+        832,
+        760
+      ]
+    ],
+    "page_idx": 21
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_8.jpg",
+    "caption": "Figure R9. SST sensitivity from an atmospheric model experiment. Composite difference of simulated a) 200 hPa temperature (color) and 200 hPa streamfunction (black contour, 10° s⁻¹) and b) 2-meter air temperature from the AGCM experiment with prescribed 2023 global SSTs (GLB23). c) and d) are the same as a) and b) but for the AGCM experiments with 2023 SST forcing relative to an extended climatology (1958 – 2022). The composite differences are with respect to the control experiment (CTL) for JJA.",
+    "footnote": [],
+    "bbox": [
+      [
+        111,
+        346,
+        884,
+        661
+      ]
+    ],
+    "page_idx": 28
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_9.jpg",
+    "caption": "Figure R11. June - September (JJAS) precipitation anomaly for a) 2020 and b) 2023. The red stipples indicate the North American monsoon region defined as where the local summer (May - September) precipitation minus the winter (November–March) precipitation exceeds 2.5 mm per day and the local summer precipitation exceeds \\(55\\%\\) of total annual precipitation. Dataset from the JRA55 reanalysis.",
+    "footnote": [],
+    "bbox": [
+      [
+        112,
+        115,
+        884,
+        350
+      ]
+    ],
+    "page_idx": 35
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_10.jpg",
+    "caption": "Figure R12. Same as Fig. R11 but using the ERA5 reanalysis.",
+    "footnote": [],
+    "bbox": [
+      [
+        112,
+        457,
+        884,
+        690
+      ]
+    ],
+    "page_idx": 35
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_11.jpg",
+    "caption": "Figure R13. SST sensitivity from an atmospheric model experiment for the Rossby wave source (RWS) components and circulation. Composite difference of simulated a) \\(200\\mathrm{hPa}\\) RWS, b) \\(200\\mathrm{hPa}\\) streamfunction (contour, \\(10^{6}s^{-1}\\) ) and anomalous vorticity advection (color), c) \\(200\\mathrm{hPa}\\) velocity potential (contour, \\(10^{6}s^{-1}\\) ) and mean vorticity advection (color), d) \\(200\\mathrm{hPa}\\) rotational wind (vector, \\(ms^{-1}\\) ) and anomalous vortex stretching (color), and d) \\(200\\mathrm{hPa}\\) divergent wind component (vector, \\(ms^{-1}\\) ) and mean vortex stretching. The units for the RWS terms are \\(10^{-11}s^{-2}\\) , see Methods for definition. Composites are from the AGCM experiment with prescribed 2023 global SSTs (GLB23). The composite differences are with respect to the control experiment (CTL) for June-July-August.",
+    "footnote": [],
+    "bbox": [
+      [
+        168,
+        90,
+        833,
+        437
+      ]
+    ],
+    "page_idx": 39
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_12.jpg",
+    "caption": "Figure R14. Same as R13 but for the ATL23 model experiment which compares the sensitivity to Atlantic SST forcing.",
+    "footnote": [],
+    "bbox": [
+      [
+        171,
+        90,
+        830,
+        435
+      ]
+    ],
+    "page_idx": 40
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_13.jpg",
+    "caption": "Figure R15. Same as R13 but for the PAC23 model experiment which compares the sensitivity to Pacific SST forcing.",
+    "footnote": [],
+    "bbox": [
+      [
+        168,
+        496,
+        833,
+        842
+      ]
+    ],
+    "page_idx": 40
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_14.jpg",
+    "caption": "Figure R16. Heat wave event of 2023. Seasonal evolution of maximum (red) and minimum (blue) temperature for the year 2023 from 1 June to 20 September for a) Las Vegas, Nevada, b) Albuquerque, New Mexico, c) El Paso, Texas, and d) San Antonio, Texas. The long-term daily mean is shown by the dashed line whereas the \\(5^{\\text{th}}\\) and \\(95^{\\text{th}}\\) percentiles are shown by the shading region. Excess above the \\(95^{\\text{th}}\\)",
+    "footnote": [],
+    "bbox": [
+      [
+        140,
+        373,
+        855,
+        808
+      ]
+    ],
+    "page_idx": 42
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_15.jpg",
+    "caption": "Figure R17. Heat budget analysis. Energy budget averaged every 10 days over the southwest U.S. and northwest Mexico from 6 June to 24 August 2023. Each row represents (from the top): 200 hPa geopotential height anomaly [gpm], 850 hPa temperature and 2-meter temperature anomalies [°C], vertically integrated anomalous heating rates from 975 - 800 hPa [°C day⁻¹], and surface heat fluxes (net surface shortwave and longwave radiation, sensible and latent heat fluxes) [Wm⁻²]. Daily anomalies are derived from the long-term monthly mean for the 1979-2022 period.",
+    "footnote": [],
+    "bbox": [
+      [
+        252,
+        100,
+        750,
+        555
+      ]
+    ],
+    "page_idx": 44
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_16.jpg",
+    "caption": "Figure R18. Decomposition of the total anomalous temperature advection averaged every 10 days over the southwest U.S. and northwest Mexico from 6 June to 24 August 2023. Each row represents (from the top): anomalous temperature advection, anomalous temperature advection by the mean wind, mean temperature advection by the anomalous wind, anomalous temperature advection by the anomalous wind, the role of circulation-only, and the role of temperature gradient only in the horizontal advection. All values were vertically integrated from 975 - 800 hPa \\([^{\\circ}\\mathrm{C}\\) day \\(^{-1}\\) ]. Daily anomalies, denoted by primes, are derived from the long-term monthly mean climatology for the 1979-2022 period, denoted by the overbar. The entries within brackets are defined by the root mean square amplitude from June-August, effectively holding their contribution to the advection constant.",
+    "footnote": [],
+    "bbox": [
+      [
+        161,
+        231,
+        835,
+        455
+      ]
+    ],
+    "page_idx": 45
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Figure R19. Similar to figure 3 of the manuscript. a) Potential vorticity and wind at the \\(340\\mathrm{K}\\) isentropic level during the maximum amplitude of the heat wave on 20 June 2023. Thick vectors depict anti-cyclonic fluid trapping, a proxy for heat dome and air flow stagnation. The thick black line indicates the location of the dynamical tropopause. The blue star represents the location of San Antonio, Texas. b) Latitude-vertical cross-section along \\(100^{\\circ}\\mathrm{W}\\) on 20 June 2023 of anomalous potential temperature (color) and potential temperature (magenta 10K intervals). Also shown are zonal wind anomalies (light black contours at \\(3\\mathrm{m / s}\\) intervals), the tropopause level as measured by the 2 PVU (thick black line), and anti-cyclonic fluid trapping (circle hatching at \\(10^{-5}\\mathrm{s}^{-1}\\) ).",
+    "footnote": [],
+    "bbox": [
+      [
+        163,
+        90,
+        828,
+        425
+      ]
+    ],
+    "page_idx": 46
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_17.jpg",
+    "caption": "Figure R20. Same as figure R12 but for 06 August 2023. The cross section is shown along \\(100^{\\circ}\\mathrm{W}\\) , near El Paso, Texas.",
+    "footnote": [],
+    "bbox": [
+      [
+        128,
+        90,
+        780,
+        417
+      ]
+    ],
+    "page_idx": 47
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_18.jpg",
+    "caption": "Figure R21. Decomposition of the anomalous June-August 2023 vertically integrated moisture transport (vector, \\(K g m^{-1} s^{-1}\\) ) and its divergence (color, \\(K g m^{-2} s^{-1}\\) ), where negative values indicate convergence. a) Advection of mean moisture by anomalous wind, b) advection of anomalous moisture by the mean wind, and c) advection of anomalous moisture by anomalous wind. The overbars denote climatology and primes denote deviation from climatology computed from the departure from the 1979-2022 climatology.",
+    "footnote": [],
+    "bbox": [
+      [
+        112,
+        412,
+        884,
+        615
+      ]
+    ],
+    "page_idx": 48
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_19.jpg",
+    "caption": "Figure R22. (left-column) Zonal-vertical cross-section along \\(28^{\\circ}\\mathrm{N}\\) of the 2023 moisture transport anomaly (color). (Right column) is the same but for the meridional cross-section along \\(110^{\\circ}\\mathrm{W}\\) . The transport is decomposed into its components such as (a and b) advection of mean moisture by anomalous wind, (c and d) advection of anomalous moisture by the mean wind, and (e and f) advection of anomalous moisture by anomalous wind. The overbars denote climatology and primes denote deviation from climatology computed from the departure from the 1979-2022 climatology.",
+    "footnote": [],
+    "bbox": [
+      [
+        191,
+        87,
+        805,
+        544
+      ]
+    ],
+    "page_idx": 49
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_20.jpg",
+    "caption": "Figure R23. a) Climatological vertically integrated moisture transport (vector, \\(K K g m^{-1} s^{-1}\\) ) and its divergence (color, \\(K K g m^{-2} s^{-1}\\) ), where negative values indicate convergence. b) Same as a) but for the 2023 anomalies computed from the departure from the 1979-2022 climatology. Vertical integration is from the surface to \\(600\\mathrm{hPa}\\) .",
+    "footnote": [],
+    "bbox": [
+      [
+        111,
+        87,
+        884,
+        383
+      ]
+    ],
+    "page_idx": 50
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_21.jpg",
+    "caption": "Figure R24. a) \\(200\\mathrm{hPa}\\) streamfunction (black contour, \\(10^{6}s^{-1}\\) ) and velocity potential (color, \\(10^{6}s^{-1}\\) ) anomalies from the AGCM experiment with prescribed 2023 Atlantic SSTs (GLB23). The anomalies are defined with respect to the control experiment (CTL) for June-July-August.",
+    "footnote": [],
+    "bbox": [
+      [
+        260,
+        88,
+        714,
+        400
+      ]
+    ],
+    "page_idx": 51
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_22.jpg",
+    "caption": "Figure R25. a) Anomalous \\(200\\mathrm{hPa}\\) temperature (color) and streamfunction (black contour, \\(10^{6}\\mathrm{s}^{-1}\\) ) for June - August of 2023. b) same as a) but for 2-meter maximum air temperature anomaly. c) Regression coefficient of Niño3 plus TNA SSTs and \\(200\\mathrm{hPa}\\) temperature (color) and \\(200\\mathrm{hPa}\\) streamfunction (black contour, \\(10^{6}\\mathrm{s}^{-1}\\) ). d) Same as c) but for 2-meter maximum air temperatures. The regression coefficients are computed for June-August for the 1960-2023 period through partial regression and the units are per standard deviation of the SST anomalies.",
+    "footnote": [],
+    "bbox": [
+      [
+        111,
+        230,
+        883,
+        562
+      ]
+    ],
+    "page_idx": 52
+  }
+]

peer_reviews/supplementary_0_Transparent Peer Review file__71ad1fa120aa40741d8ca0d96a841c714c9dbb8bf44d0505f50d03c40cf6eefc/images_list.json

@@ -0,0 +1,32 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Figure 1. Association between mitochondrial count and age.",
+    "footnote": [],
+    "bbox": [
+      [
+        287,
+        410,
+        703,
+        651
+      ]
+    ],
+    "page_idx": 10
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Figure 2. Boxplot showing the distribution of mitochondrial count between males and females.",
+    "footnote": [],
+    "bbox": [
+      [
+        285,
+        83,
+        710,
+        308
+      ]
+    ],
+    "page_idx": 11
+  }
+]

peer_reviews/supplementary_0_Transparent Peer Review file__71c5664823cdb51d0090bce56bc9fa1f7a11a279819fc98c43c3bd438c64318c/images_list.json

@@ -0,0 +1 @@
+[]

peer_reviews/supplementary_0_Transparent Peer Review file__71c5664823cdb51d0090bce56bc9fa1f7a11a279819fc98c43c3bd438c64318c/supplementary_0_Transparent Peer Review file__71c5664823cdb51d0090bce56bc9fa1f7a11a279819fc98c43c3bd438c64318c.mmd

@@ -0,0 +1,309 @@
+
+# Learning Plasma Dynamics and Robust Rampdown Trajectories with Predict-First Experiments at TCV  
+
+Corresponding Author: Mr Allen Wang  
+
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+Version 0:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author)  
+
+The work revolves around the use of reinforcement learning (RL) for the optimization of the ramp- down trajectory at TCV, an experimental device in Switzerland. The idea of RL is that the model can learn how to interact with the system by trial and error and requires a sufficiently accurate model of the system. In this case, also the system is approximated by a Neural State Space Model (NSSM), that is a Neural Network which can approximate the state space representation mixing a data driven loss and numerical integration of the (net+system) dynamics.  
+
+The work is interesting but due to the scarcity of experiments, the successfully designed rampdowns are only 5, out of a total of 9 optimized ones. For this reason, a further analysis and replication of the experiments is recommended for making robust conclusions.  
+
+Moreover, I have the following comments on the work:  
+
+NSSM and training environment for the RL system  
+
+The article is quite dense, and the authors explain that one model (the NSSM) serves as the training environment for the RL model, however another neural model to predict the temperature and density profiles is mentioned as part of the NSSM. From section 2.2:  
+
+"The NSSM was initially developed with a neural network predictor for the kinetic profiles on the full p grid, and initial training runs found that the profile predictor can accurately predict kinetic profiles given the set of OD scalars listed in Table 2. Figure 5 provides an example comparison of predictions of the Te and ne profiles against Thomson measurements for a full shot in the validation dataset, showing accurate prediction across all phases of the shot. This result corroborates previous findings at NSTX- U that neural networks can accurately predict kinetic profiles given a set of similar OD scalars34. Given this result suggests most of the relevant profile information is implicitly captured by OD scalars, the profile predictor was disabled prior to running experiments to accelerate training, hence reported predictions of kinetic profiles are not predict- first. A noteworthy feature of this profile predictor that should be explored in future work is its ability to function as a Thomson upsampler, as the input variables are all sampled at a higher time resolution, 1ms, than the TCV Thomson Scattering system, which takes measurements every 17ms."  
+
+If the predictor has not been adopted, I suggest removing the parts connected with the profile predictor (e.g section 4.1.3, Figures 5, 9) to avoid confusion, otherwise I would suggest indicating for how many discharges the predictor has been used to train the RL method and if there was any benefit from this approach.  
+
+How were the discharges to train the NSSM model selected? What are the variations in terms of the (observation, action) space?  
+
+Connected to the previous point, it seems that the vertical instability growth rate \((\gamma_{- }\) vgr) reconstruction by the NSSM is very poor, even with a very low prediction Horizon (Figure 4). Since this parameter is connected to vertical plasma stability, this would explain the VDE event in shot #81751. How large is the uncertainty assumed for the \(\gamma_{- }\) vgr?  
+
+An interesting point could be made by running the control discharges in the NSSM and showing that they were close to disruptive boundaries/expected to disrupt. This would show, despite the low number of experiments, that the NSSM can well represent the system.  
+
+Is it possible to interpret the strategies adopted by the RL model by showing which parameters are kept far from critical boundaries?
+
+<--- Page Split --->
+
+Other comments  
+
+Some other works related to the use of RL in fusion are missing, as applications for ITER [1] and EAST [2]. Please add some details on the model adopted (number of layers, neurons) and on how they have been selected/optimized. I tried running the code and installing all the dependencies using poetry, but I could not manage to make it work. Most of the python notebooks seem to reproduce the figure and the main results, I did not find the scripts for training the models.  
+
+[1] S. Dubbioso, G. De Tommasi, A. Mele, G. Tartaglione, M. Ariola, and A. Pironti, 'A Deep Reinforcement Learning approach for Vertical Stabilization of tokamak plasmas', Fusion Engineering and Design, vol. 194, p. 113725, Sep. 2023, doi: 10.1016/j.fusengdes.2023.113725. [2] G. De Tommasi, S. Dubbioso, Y. Huang, Z. P. Luo, A. Mele, and B. J. Xiao, 'A RL- based Vertical Stabilization System for the EAST tokamak', in 2022 American Control Conference (ACC), Jun. 2022, pp. 5328- 5333. doi: 10.23919/ACC53348.2022.9867499.  
+
+(Remarks on code availability) Partial review of the code due to the difficulties in installing all the proper libraries The readme file instructs on using poetry, but after finding the correct version of poetry I ran into an error: Directory TokaGym\submodule's\contrax for contrax does not seem to be a Python package  
+
+Reviewer #2  
+
+(Remarks to the Author) General Comments:  
+
+This study leverages deep learning (NSSM) to predict plasma ramp- down dynamics in the TCV tokamak and applies reinforcement learning (RL) to design an optimal trajectory for stable termination during this phase. Notably, ramp- down in fusion reactors is physically and technically prone to instability, often associated with plasma disruptions. Thus, ensuring stable control and termination in this phase is critical for future large- scale devices such as ITER. This study combines a predictive model and RL on TCV to design a stable plasma termination trajectory and demonstrates robust outcomes through incremental re- training with actual experiments. This represents a significant milestone in fusion research. I believe that, provided the following comments are adequately addressed, the manuscript could be suitable for publication in Nature Communications.  
+
+## Major Comments:  
+
+1. While the significance and performance of the study are commendable, the storyline in the figures and descriptions is not very clear. Since Nature Communications is a multidisciplinary journal, the content should be understandable even to readers who are not fusion experts. From a non-expert's perspective, when reading the early part of the paper, the role of the proposed technique is not immediately intuitive (e.g., what constitutes a desired termination and how it differs from undesired outcomes). In particular, Figure 2E is the key illustration that visually demonstrates this difference, but it is not explained at all in the main text. I recommend moving Figure 2E to Figure 1 (right side of the original Fig 1) and introducing it in the Introduction or early in the Results section to clarify the role of the proposed method.  
+
+2. (Figures 6-8) These figures show only the actual plasma trajectories, but not the RL-designed or the manually programmed (planned) trajectories. However, the RL-designed trajectory, the manually programmed trajectory in the PCS, and the actual plasma response may all differ. For example, the authors state in Figure 6 that the trajectory was optimized to reduce the LFS gap to prevent VDEs, but from the figure alone, it is unclear whether this was truly suggested by the RL agent, manually reflected in PCS programming, or just occurred randomly regardless of the programmed trajectory. This distinction is crucial for interpreting and evaluating the value of the research. I recommend that the RL-designed trajectory and the PCS-programmed trajectory be also added to Figures 6 and 7 (e.g., using light-colored or dashed lines). In Figure 8, although multiple RL trajectories under uncertainty are shown as black lines, the actual programmed trajectory used in the experiment is not shown. While the caption mentions that the average trajectory was used, it would be helpful to indicate this explicitly in the figure as well.  
+
+3. (Page 24) Unlike general RL settings, the RL agent in this study observes only time (Equation 12). This implies that once trained, the agent will always produce the same trajectory, regardless of the initial condition or change of plasma states. Thus, the first question is: when computing the trajectory to input into TCV as shown in Figure 8, how was random uncertainty introduced (like just normal random added to the given trajectory)? Second, if the initial conditions or plasma current differ, does the RL agent need to be re-trained from scratch each time?  
+
+## Minor Comments:  
+
+4. (Page 2) In the Introduction, the authors emphasize that this study focuses on feedforward trajectory design before the
+
+<--- Page Split --->
+
+experiment, in contrast to previous RL- based feedback control studies. However, some previous works have also used deep learning- based predictive models in combination with RL to design operation trajectories in a feedforward manner (e.g., [J Seo et al., Nucl. Fusion 61 106010 (2021) & J Seo et al., Nucl. Fusion 62 086049 (2022)]). Although these earlier works may be more limited and less robust to uncertainty compared to the current study, they are similar in concept and deserve to be mentioned as related study.  
+
+5. (Figure 2A and others) In Figure 2A, the markers in the legend are difficult to distinguish from those in the graph itself. I suggest modifying the legend (like using boxed legends or other formatting) for easier visual distinction, including in other similar figures.  
+
+6. (Figure 2D) The background colors in this graph are not intuitive. At first glance, one might assume that green indicates successful terminations and red indicates failures. However, shot 81635, which was successfully terminated, is marked red, while shot 81741, which had a legacy software issue, is shown in green. Additionally, the x-axis does not follow a monotonic shot order (e.g., 81635 appears after 81741), and there is no explanation for the ambiguous non-red/non-green background colors. It would also be helpful to indicate the threshold values for successful termination (W_tot and I_p criteria) directly on the graph. Similar color ambiguity is present in Figure 3 (e.g., mixed green and blue in W_tot panel), but without explanation.  
+
+7. (Page 6) The authors aim to design trajectories that avoid user-specific limits associated with disruptions and so create models that predict the evolution of relevant quantities. For the readers to understand better, it would be helpful to explicitly define the key quantities correlated with disruptions and the criteria or limits that typically lead to disruptions. Although Table 3 later in the paper provides some indirect answers through the structure of the reward function, it would be better to introduce and explain these disruption criteria earlier in the manuscript. Fusion experts may already be familiar with them, but for a multidisciplinary audience like that of Nature Communications, such clarifications are important.  
+
+8. (Tables 1 and 2) Tables 1 and 2 both define input and output variables for the deep learning models. However, in Table 1 (for NSSM), the inputs (actions) are at the bottom and outputs (observations) are at the top, while Table 2 has inputs at the top and outputs at the bottom. This inconsistency may confuse readers, especially since in typical RL settings, observations are inputs and actions are outputs. Unless there is a reason for this ordering, I recommend standardizing the layout for clarity.  
+
+9. (Page 6 and Figures 5 & 9) The profile predictor appears to reconstruct 1D profiles from already predicted 0D parameters, rather than predicting the dynamics of the profiles directly. This implies that the information in the 1D profiles is already embedded in the 0D parameters. Therefore, using the reconstructed 1D profiles in NSSM or RL trajectory design might be unnecessary and only increase input complexity. In fact, as far as I understand, the profile predictor does not play a major role in the key process of this study. I'm wondering the role of this profile predictor regarding this study. If it is not essential to the main pipeline, I suggest removing it from the current paper and including it in a follow-up study.  
+
+10. (Figure 8) The black lines represent RL trajectories under uncertainty in the actuations. However, in practice, uncertainty (or noise) affects not only the level of the actions or trajectory but also introduces fluctuations in the actuations—like the red or magenta lines in the figure. However, the RL trajectories for a minor, remain flat except for level differences. Doesn’t fluctuating action noise in real experiments affect the robustness of the designed RL trajectories?  
+
+(Remarks on code availability)  
+
+Version 1:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author)  
+
+1) In the newly added text at page 4 the authors claim:  
+
+"At present, a comprehensive understanding of disruptive limits remains an open problem, motivating many works on machine-learning based prediction of disruptions 37-40."  
+
+Despite being valid, this statement is supported by many self- citing references. Further/different works could be referenced since ML- based prediction of disruptions is a well developed line of research in fusion.  
+
+2) Despite the statement of the authors, I consider the role of the profile predictor is still marginal in the description of the work, and recommend showing the plots discussing the 1D profile reconstruction either in the supplementary data or in a new work. On the other hand, for the purposes of the work, I find more interesting the discussion on Figures 16 and 20, now in the supplementary data. These figures allow to discuss the validity of the NSSM for the current application. For instance, in the two successful pulses of Figure 16, the yvgr is close to the same limit value as in the one of Figure 20. How is the distinction between these two sets of pulses made?  
+
+(Remarks on code availability)
+
+<--- Page Split --->
+
+Reviewer #2  
+
+(Remarks to the Author) Thanks for the authors' sincere replies to my previous comments. The authors appropriately answered my previous questions and revised the paper accordingly. Now, I recommend the paper for publication as it is.  
+
+(Remarks on code availability)  
+
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source. The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.  
+
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+
+## Responses to Reviewers  
+
+## Reviewer 1  
+
+Thank you for taking the time to provide helpful feedback. We concur that it would be beneficial to increase the statistics of the result; however, it is not straightforward to simply request further shots from an experimental fusion device, especially for this high performance scenario which is highly unusual for TCV.  
+
+We have, however, performed additional analyses to increase confidence in the results. This includes running a disruptive shots action trajectory in the RL environment, as you recommended. This also involved a further physics- based analysis confirming that the new RL design trajectories have more vertical stability in the presence of control errors.  
+
+We have also made changes to the introductory figures in the paper, in large part in response to feedback from your co- reviewer. We hope the new format makes the results more accessible to a broad audience.  
+
+Thank you for pointing out issues with the codebase. It appears we did not archive the key submodule; this issue has been addressed.  
+
+Further details are provided in the line item responses below.  
+
+Regards, Allen M. Wang On Behalf of Coauthors  
+
+## Line-Item Response  
+
+<table><tr><td>Comment</td><td>Response</td></tr><tr><td>The work is interesting but due to the scarcity of experiments, the successfully designed rampdowns are only 5, out of a total of 9 optimized ones. For this reason, a further analysis and replication of the experiments is recommended for making robust conclusions.</td><td>We would like to point out that:<br>1) 6 rampdowns were successful<br>2) Two shots were utilized explicitly to debug an unrelated issue caused by a legacy 1/lp^2 term in the PCS, leading to poor radial control during rampdowns. The fact that this issue existed and was finally fixed after the debug shots is shown in Figure 14 in the Supplementary Materials. We</td></tr></table>
+
+<--- Page Split --->
+
+
+<table><tr><td></td><td>have chosen to report these shots<br>nevertheless for transparency, and to<br>raise awareness on the often subtle,<br>but important, issues involved in<br>fusion experiments which are often<br>omitted from publications.<br>3) We performed statistical analysis both<br>with and without these two shots<br>included in the test set. Even with<br>these two shots included in the test<br>set, the results are encouraging.</td></tr><tr><td>The article is quite dense, and the authors explain that one model (the NSSM) serves as the training environment for the RL<br>model, however another neural model to predict the temperature and density profiles is mentioned as part of the NSSM. From<br>section 2.2:<br>"The NSSM was initially developed with a<br>neural network predictor for the kinetic<br>profiles on the full p grid, and initial training runs found that the profile predictor can<br>accurately predict kinetic profiles given the set of 0D scalars listed in Table 2.<br>Figure 5 provides an example comparison of predictions of the Te and ne profiles<br>against Thomson measurements for a full shot in the validation dataset, showing<br>accurate prediction across all phases of the shot. This result corroborates previous<br>findings at NSTX-U that neural networks<br>can accurately predict kinetic profiles given a set of similar 0D scalars34. Given this<br>result suggests most of the relevant profile information is implicitly captured by 0D<br>scalars, the profile predictor was disabled<br>prior to running experiments to accelerate training, hence reported predictions of<br>kinetic profiles are not predict-first. A<br>noteworthy feature of this profile predictor that should be explored in future work is its</td><td>We understand why removing the content on the profile predictor would help streamline the paper. However, we want to highlight the fact that the results of our work motivate further<br>research investment in combining established 0D transport simulation methods with neural networks. We believe this is a worthwhile<br>message to preserve, given the large number of ongoing efforts towards developing<br>principles-based transport simulators for<br>control. This message is highlighted in the<br>Discussion:<br>"The ability for a neural network to predict<br>kinetic profiles using 0D scalars,<br>demonstrated both in this work and in prior work43, suggests a<br>data-driven approach may be sufficient for<br>certain control tasks without principles-based transport simulation, which can be<br>extremely computationally expensive and<br>require strong assumptions on edge<br>temperature and density."<br>Changes Made: added a sentence earlier in<br>the paper (Section 2.2) for clarity of message:<br>"This result also suggests a structured<br>data-driven approach to modeling tokamak<br>transport merits further research, in parallel<br>with several ongoing principles-based efforts."</td></tr></table>
+
+<--- Page Split --->
+
+
+<table><tr><td>ability to function as a Thomson<br>up-sampler, as the input variables are all<br>sampled at a higher time resolution, 1ms,<br>than the TCV Thomson Scattering system,<br>which takes measurements every 17ms."<br>If the predictor has not been adopted, I<br>suggest removing the parts connected with<br>the profile predictor (e.g section 4.1.3,<br>Figures 5, 9) to avoid confusion, otherwise I<br>would suggest indicating for how many<br>discharges the predictor has been used to<br>train the RL method and if there was any<br>benefit from this approach.</td><td></td></tr><tr><td>How were the discharges to train the NSSM model selected? What are the variations in terms of the (observation, action) space?</td><td>We provide details in the subsection "Training Data Distribution" in the methods. To provide further clarity, we used the DEFUSE system to gather recent shots that simultaneously<br>have a partial rampdown and sufficient<br>diagnostic availability. A plot of the data<br>distribution in beta, Ip and greenwald fraction space is also shown in that subsection.<br>Changes Made: modified "Training Data<br>Distribution" to clarify that we gathered the<br>most recent shots with sufficient diagnostic<br>availability.</td></tr><tr><td>Connected to the previous point, it seems<br>that the vertical instability growth rate<br>(y_vgr) reconstruction by the NSSM is very<br>poor, even with a very low prediction<br>Horizon (Figure 4). Since this parameter is<br>connected to vertical plasma stability, this<br>would explain the VDE event in shot<br>#81751. How large is the uncertainty<br>assumed for the y_vgr?</td><td>We discuss the prediction accuracy of y_vgr<br>in Page 6 paragraph two:<br>"The percent errors for yvgr can be relatively large, but, as shown in Figure 16 this is largely attributable to the small value of yvgr of limited plasmas as the absolute error is relatively low."<br>The wide distribution you are seeing is due to<br>the fact that gamma_vgr on TCV is highly<br>sensitive to control errors, which is discussed in page 11.<br>The uncertainty distributions added to the<br>y_vgr and the variables that are input into the model predicting it were quantified from<br>model prediction error metrics, as discussed in page 23:<br>"In some cases, uncertainty distributions could easily be</td></tr></table>
+
+<--- Page Split --->
+
+
+<table><tr><td></td><td>quantified from past experimental data (such as<br>tracking error in the plasma current), or from model<br>prediction accuracy (such as vvgr),"</td></tr><tr><td>An interesting point could be made by<br>running the control discharges in the NSSM and showing that they were close to<br>disruptive boundaries/expected to disrupt.<br>This would show, despite the low number of experiments, that the NSSM can well<br>represent the system.</td><td>We performed this analysis, which shows that the training environment expects a large<br>vertical instability growth rate using the action<br>trajectories from 81101, one of the baseline<br>shots.<br>Given the length and scope of the main paper as it is, and the size of this plot, we believe<br>this result belongs in the supplementary<br>materials.<br>Changes Made: added a figure to the<br>supplementary materials showing NSSM<br>predictions given the action trajectory from a disruptive shot.</td></tr><tr><td>Is it possible to interpret the strategies<br>adopted by the RL model by showing which parameters are kept far from critical<br>boundaries?</td><td>The right hand side of Figure 8 is meant to<br>illustrate this.</td></tr><tr><td>Some other works related to the use of RL<br>in fusion are missing, as applications for<br>ITER [1] and EAST [2].</td><td>Thank you for pointing out these works.<br>Changes Made: included references to these two papers in the introduction.</td></tr><tr><td>Please add some details on the model<br>adopted (number of layers, neurons) and on how they have been selected/optimized.</td><td>Thank you for pointing out this omission.<br>Changes Made: Introduced "Training<br>Methods" to the NSSM section. Added<br>sentences added to "RL Methods" specifying model architecture.</td></tr><tr><td>I tried running the code and installing all the dependencies using poetry, but I could not<br>manage to make it work. Most of the<br>python notebooks seem to reproduce the<br>figure and the main results, I did not find the scripts for training the models.</td><td>Apologies, it appears that we didn't archive<br>the submodule containing the dynamics<br>model. We have taken care to make sure it is<br>available in the updated .tar.gz file and poetry install works if the user is using<br>python3.10/11.<br>While we have taken care to make sure<br>notebooks under "paper_plot_generation" are reproducible, execution of the full repo is not<br>possible due to dependencies on large<br>datasets that are under Eurofusion data<br>access restrictions, and thus cannot be made</td></tr></table>
+
+<--- Page Split --->
+
+<table><tr><td></td><td>public for the foreseeable future.<br/>The NSSM training main script is under:<br/>"submodules/contrax/contrax/examples/plasma/tcv/train_model.py"<br/>RL training scripts are under:<br/>"scripts/train_ppo.py"<br/>"scripts/train_es.py"<br/>Changes Made: updated .tar.gz containing codebase + data files.</td></tr></table>
+
+Reviewer 2
+
+Thank you for taking your time to provide helpful feedback in improving the quality of the manuscript, and informing us of important prior works on RL for trajectory optimization in fusion we were not previously aware of. We have adopted your recommendations, which has helped revamp the early presentation of the paper to be more accessible to the reader.
+
+In response to both you and your co-reviewer, we have also performed a further physics-based analysis which corroborates the improved robustness of RL-designed trajectories to vertical instability by introducing control errors into the shot preparation equilibrium solver and solving for the resulting gamma_vgr distributions.
+
+Further details are provided in the line item responses below.
+
+Regards,
+
+Allen M. Wang
+
+On Behalf of Coauthors
+
+Line-Item Response
+
+<table><tr><td>Comment</td><td>Response</td></tr><tr><td>1. While the significance and performance<br>of the study are commendable, the storyline in the figures and descriptions is not very<br>clear. Since Nature Communications is a<br>multidisciplinary journal, the content should be understandable even to readers who are not fusion experts. From a non-expert's</td><td>Thank you for this recommendation. We<br>concur that moving Figure 2E to Figure 1 on<br>the right side is helpful, and we have adopted this change. We have also updated the<br>original Figure 1 to a shot that disrupts at high Greenwald Fraction for further clarity.<br>Changes Made: Figure 2E is now part of<br>Figure 1, and Figure 1 updated to a shot that</td></tr></table>
+
+<--- Page Split --->
+
+
+<table><tr><td>perspective, when reading the early part of the paper, the role of the proposed<br>technique is not immediately intuitive (e.g.,<br>what constitutes a desired termination and how it differs from undesired outcomes). In particular, Figure 2E is the key illustration<br>that visually demonstrates this difference,<br>but it is not explained at all in the main text. I recommend moving Figure 2E to Figure 1<br>(right side of the original Fig 1) and<br>introducing it in the Introduction or early in<br>the Results section to clarify the role of the<br>proposed method.</td><td>more clearly illustrated the issue at hand.</td></tr><tr><td>2. (Figures 6-8) These figures show only<br>the actual plasma trajectories, but not the<br>RL-designed or the manually programmed<br>(planned) trajectories. However, the<br>RL-designed trajectory, the manually<br>programmed trajectory in the PCS, and the<br>actual plasma response may all differ. For<br>example, the authors state in Figure 6 that<br>the trajectory was optimized to reduce the<br>LFS gap to prevent VDEs, but from the<br>figure alone, it is unclear whether this was<br>truly suggested by the RL agent, manually<br>reflected in PCS programming, or just<br>occurred randomly regardless of the<br>programmed trajectory. This distinction is<br>crucial for interpreting and evaluating the<br>value of the research. I recommend that the<br>RL-designed trajectory and the<br>PCS-programmed trajectory be also added<br>to Figures 6 and 7 (e.g., using light-colored<br>or dashed lines).<br>In Figure 8, although multiple RL<br>trajectories under uncertainty are shown as<br>black lines, the actual programmed<br>trajectory used in the experiment is not<br>shown. While the caption mentions that the</td><td>Thank you for your feedback. We will respond by figure.<br>Changes Made:<br>Figure 6: thank you for your recommendation. We have added the RL-designed<br>PCS-programmed trajectory to this figure. To further confirm that the RL-designed<br>trajectory after adding control error is indeed more robust to the vertical instability, we have performed a further physics-based<br>robustness analysis by introducing minor<br>radius control errors to the RL designed<br>trajectories, solving for equilibria using the<br>free-boundary equilibrium solver used for<br>TCV shot preparation, and then calculating<br>gamma_vgr distributions for the resulting<br>equilibria. The results (shown in the new<br>Figure 7) show that indeed 82875 is more<br>robust than 81751.<br>Figure 7 (now Figure 8): it is more difficult to show the RL designed trajectories in<br>conjunction with the actual trajectories for this figure, given that the trajectories changed<br>from shot to shot. The purpose of the figure is to show that we progressed over the course<br>of the experiment. However, to provide the<br>information you suggested, we have included the equivalent of Figure 8 for the final two<br>140kA shots as a part of the supplementary<br>information.</td></tr></table>
+
+<--- Page Split --->
+
+
+<table><tr><td>average trajectory was used, it would be<br>helpful to indicate this explicitly in the figure as well.</td><td>Figure 8 (now Figure 9): this is a good<br>suggestion, we have included the mean of<br>the action trajectories as a solid thick black<br>line as well, and included a comment on this in the caption.</td></tr><tr><td>3. (Page 24) Unlike general RL settings, the RL agent in this study observes only time (Equation 12). This implies that once<br>trained, the agent will always produce the same trajectory, regardless of the initial<br>condition or change of plasma states.<br>Thus, the first question is: when computing the trajectory to input into TCV as shown in Figure 8, how was random uncertainty<br>introduced (like just normal random added to the given trajectory)? Second, if the initial conditions or plasma current differ, does<br>the RL agent need to be re-trained from<br>scratch each time?</td><td>This is a good point. The uncertainty<br>distribution for initial conditions specified is shown in Section 4.4: Uncertainty Model.<br>The idea is to have a single action trajectory that is likely to succeed across a range of<br>initial conditions.<br>If the initial conditions change by more than anticipated, then indeed retraining needs to happen (as was the case for 140kA vs.<br>170kA).<br>Changes Made: we have introduced a<br>sentence in 4.5 to make this more clear:<br>"Given that time is the only observable, but<br>there exists different physical conditions in<br>the parallel training environments which are unobservable to the policy, the reward<br>maximization process yields a trajectory that is designed to succeed across the different conditions specified in Subsection<br>vref{subsec:uncertainty_model}."</td></tr><tr><td>4. (Page 2) In the Introduction, the authors emphasize that this study focuses on<br>feedforward trajectory design before the<br>experiment, in contrast to previous<br>RL-based feedback control studies.<br>However, some previous works have also<br>used deep learning-based predictive<br>models in combination with RL to design<br>operation trajectories in a feedforward<br>manner (e.g., [J Seo et al., Nucl. Fusion 61<br>106010 (2021) & J Seo et al., Nucl. Fusion<br>62 086049 (2022)]). Although these earlier<br>works may be more limited and less robust to uncertainty compared to the current</td><td>Thank you for informing us of this work.<br>We have included references.<br>Changes Made: we have included both<br>citations and introduced the papers in the<br>introduction:<br>"A similar approach has previously been<br>demonstrated at KSTAR for designing<br>feed-forward trajectories that reach target<br>states<br>\\cite{seo2021feedforward,seo2022developme nt}."</td></tr></table>
+
+<--- Page Split --->
+
+
+<table><tr><td>study, they are similar in concept and<br>deserve to be mentioned as related study.</td><td></td></tr><tr><td>5. (Figure 2A and others) In Figure 2A, the markers in the legend are difficult to<br>distinguish from those in the graph itself. I suggest modifying the legend (like using<br>boxed legends or other formatting) for<br>easier visual distinction, including in other<br>similar figures.</td><td>Thank you for pointing out this deficiency.<br>Changes Made: added a box to the legend.</td></tr><tr><td>6. (Figure 2D) The background colors in this graph are not intuitive. At first glance, one<br>might assume that green indicates<br>successful terminations and red indicates<br>failures. However, shot 81635, which was<br>successfully terminated, is marked red,<br>while shot 81741, which had a legacy<br>software issue, is shown in green.<br>Additionally, the x-axis does not follow a<br>monotonic shot order (e.g., 81635 appears<br>after 81741), and there is no explanation for the ambiguous non-red/non-green<br>background colors. It would also be helpful to indicate the threshold values for<br>successful termination (W_tot and I_p<br>criteria) directly on the graph. Similar color ambiguity is present in Figure 3 (e.g., mixed green and blue in W_tot panel), but without explanation.</td><td>Thank you for pointing out the non-monotonic ordering. This was somewhat embarrassing<br>given that it was 81741 and 81745 that were<br>the debug shots. We have addressed this<br>issue.<br>Your feedback has helped us recognize there is a better way to communicate our results.<br>We have now replaced the histogram with a<br>scatter plot in (Wtot and Ip) space showing<br>the shots and a table of statistics embedded<br>in the plot. We realized it likely makes more<br>sense to have this plot in the "Overview" as<br>opposed to having the shot-by-shot<br>breakdown in the overview. The shot-by-shot<br>breakdown is now its own separate figure.<br>We have added the goal Ip to the scatter plot. As we note in a new footnote, we initially<br>chose a goal stored energy, 0.5kJ, that<br>proved unrealistic to diagnose as equilibrium<br>reconstruction no longer converges reliably at such low plasma current and stored energy.<br>Thus, we would like to primarily focus on<br>reaching the goal Ip as it can be diagnosed<br>relatively reliably without reconstruction.<br>Changes Made: 81635 + 81741 are now in<br>their proper order. Figure 3 is now a scatter<br>plot in Wtot and Ip space and is part of the<br>overview. The shot-by-shot breakdown is now a separate figure. Added footnote on how we chose an overtly optimistically low energy that can be reached.</td></tr></table>
+
+<--- Page Split --->
+
+
+<table><tr><td>7. (Page 6) The authors aim to design<br>trajectories that avoid user-specific limits<br>associated with disruptions and so create<br>models that predict the evolution of<br>relevant quantities. For the readers to<br>understand better, it would be helpful to<br>explicitly define the key quantities<br>correlated with disruptions and the criteria<br>or limits that typically lead to disruptions.<br>Although Table 3 later in the paper provides<br>some indirect answers through the<br>structure of the reward function, it would be<br>better to introduce and explain these<br>disruption criteria earlier in the manuscript.<br>Fusion experts may already be familiar with<br>them, but for a multidisciplinary audience<br>like that of Nature Communications, such<br>clarifications are important.</td><td>Thank you for the feedback. We agree an<br>earlier discussion would be beneficial.<br>Changes Made: We have amended the end<br>of paragraph 1 in 2.1 to introduce the issue of<br>disruptive limits and the constraints we chose<br>to consider in this work.</td></tr><tr><td>8. (Tables 1 and 2) Tables 1 and 2 both<br>define input and output variables for the<br>deep learning models. However, in Table 1<br>(for NSSM), the inputs (actions) are at the<br>bottom and outputs (observations) are at<br>the top, while Table 2 has inputs at the top<br>and outputs at the bottom. This<br>inconsistency may confuse readers,<br>especially since in typical RL settings,<br>observations are inputs and actions are<br>outputs. Unless there is a reason for this<br>ordering, I recommend standardizing the<br>layout for clarity.</td><td>Thank you for the suggestion.<br>Changes Made: We have edited Table 1 to<br>include actions on top and observations on<br>the bottom.</td></tr><tr><td>9. (Page 6 and Figures 5 & 9) The profile<br>predictor appears to reconstruct 1D profiles<br>from already predicted 0D parameters,<br>rather than predicting the dynamics of the<br>profiles directly. This implies that the<br>information in the 1D profiles is already</td><td>You are correct. One of the findings of this<br>study is that details of 1D kinetic profiles did<br>not appear significant to achieve this result.<br>Even though the 1D profile predictor is not<br>necessary to obtain the core result, we<br>believe its inclusion in the paper contributes<br>to ongoing dialogue about the priority of<br>developing high-fidelity transport simulators</td></tr></table>
+
+<--- Page Split --->
+
+
+<table><tr><td>embedded in the 0D parameters. Therefore, using the reconstructed 1D profiles in NSSM or RL trajectory design might be unnecessary and only increase input complexity. In fact, as far as I understand, the profile predictor does not play a major role in the key process of this study. I'm wondering the role of this profile predictor regarding this study. If it is not essential to the main pipeline, I suggest removing it from the current paper and including it in a follow-up study.</td><td>for plasma control tasks, given the large volume of effort currently expended in that direction [1, 2, 3]. As we discuss in the Discussion section:<br>"The ability for a neural network to predict kinetic profiles using 0D scalars, demonstrated both in this work and in prior work [36], suggests a data-driven approach may be sufficient for certain control tasks without principles-based transport simulation, which can be extremely computationally expensive and require strong assumptions on edge temperature and density."<br>However, your comment helped us realize it will be worthwhile to make this message clear earlier in the paper as well.<br>Changes Made: added a sentence earlier in the paper (Section 2.2) for clarity of message: "This result also suggests a structured data-driven approach to modeling tokamak transport merits further research, in parallel with several ongoing principles-based efforts."<br>[1] Citrin, Jonathan, et al. "Torax: A fast and differentiable tokamak transport simulator in jax." arXiv preprint arXiv:2406.06718 (2024). [2] Muraca, Marco, et al. "Reduced transport models for a tokamak flight simulator." Plasma Physics and Controlled Fusion 65.3 (2023): 035007.<br>[3] Meneghini, O., et al. "FUSE (Fusion Synthesis Engine): A Next Generation Framework for Integrated Design of Fusion Pilot Plants." arXiv preprint arXiv:2409.05894 (2024).</td></tr><tr><td>10. (Figure 8) The black lines represent RL trajectories under uncertainty in the actuations. However, in practice, uncertainty (or noise) affects not only the level of the actions or trajectory but also introduces fluctuations in the actuations—like the red or magenta lines in the figure. However, the RL trajectories for a_minor, remain flat except for level</td><td>You are correct. This is a limitation of our uncertainty model.<br>Changes Made:<br>Included the following sentence at the end of the "Uncertainty Model" subsection:<br>"In addition, the uncertainty model employed does not account for time-varying fluctuations in uncertain variables; future work should employ time-varying stochastic processes.</td></tr></table>
+
+<--- Page Split --->
+
+
+<table><tr><td>differences. Doesn&#x27;t fluctuating action noise in real experiments affect the robustness of the designed RL trajectories?</td><td>Both of these limitations further highlight the need to advance experimental uncertainty quantification and robust control in the context of fusion plasma control.”</td></tr></table>
+
+# Reviews
+
+# Reviewer #1 (Remarks to the Author):
+
+The work revolves around the use of reinforcement learning (RL) for the optimization of the ramp-down trajectory at TCV, an experimental device in Switzerland. The idea of RL is that the model can learn how to interact with the system by trial and error and requires a sufficiently accurate model of the system. In this case, also the system is approximated by a Neural State Space Model (NSSM), that is a Neural Network which can approximate the state space representation mixing a data driven loss and numerical integration of the (net+system) dynamics.
+
+The work is interesting but due to the scarcity of experiments, the successfully designed rampdowns are only 5, out of a total of 9 optimized ones. For this reason, a further analysis and replication of the experiments is recommended for making robust conclusions.
+
+Moreover, I have the following comments on the work:
+
+NSSM and training environment for the RL system
+
+The article is quite dense, and the authors explain that one model (the NSSM) serves as the training environment for the RL model, however another neural model to predict the temperature and density profiles is mentioned as part of the NSSM. From section 2.2:
+
+“The NSSM was initially developed with a neural network predictor for the kinetic profiles on the full p grid, and initial training runs found that the profile predictor can accurately predict kinetic profiles given the set of OD scalars listed in Table 2.
+
+Figure 5 provides an example comparison of predictions of the Te and ne profiles against Thomson measurements for a full shot in the validation dataset, showing accurate prediction across all phases of the shot. This result corroborates previous findings at NSTX-U that neural networks can accurately predict kinetic profiles given a set of similar OD scalars34. Given this result suggests most of the relevant profile information is implicitly captured by OD scalars, the profile predictor was disabled prior to running experiments to accelerate training, hence reported predictions of kinetic profiles are not predict-first. A noteworthy feature of this profile predictor that should be explored in future work is its ability to function as a Thomson up-sampler, as the input variables are all sampled at a
+
+<--- Page Split --->
+
+higher time resolution, 1ms, than the TCV Thomson Scattering system, which takes measurements every 17ms."  
+
+If the predictor has not been adopted, I suggest removing the parts connected with the profile predictor (e.g section 4.1.3, Figures 5, 9) to avoid confusion, otherwise I would suggest indicating for how many discharges the predictor has been used to train the RL method and if there was any benefit from this approach.  
+
+How were the discharges to train the NSSM model selected? What are the variations in terms of the (observation, action) space?  
+
+Connected to the previous point, it seems that the vertical instability growth rate \((\gamma_{- }vgr)\) reconstruction by the NSSM is very poor, even with a very low prediction Horizon (Figure 4). Since this parameter is connected to vertical plasma stability, this would explain the VDE event in shot #81751. How large is the uncertainty assumed for the \(\gamma_{- }vgr\) ?  
+
+An interesting point could be made by running the control discharges in the NSSM and showing that they were close to disruptive boundaries/expected to disrupt. This would show, despite the low number of experiments, that the NSSM can well represent the system.  
+
+Is it possible to interpret the strategies adopted by the RL model by showing which parameters are kept far from critical boundaries?  
+
+## Other comments  
+
+Some other works related to the use of RL in fusion are missing, as applications for ITER [1] and EAST [2]. Please add some details on the model adopted (number of layers, neurons) and on how they have been selected/optimized. I tried running the code and installing all the dependencies using poetry, but I could not manage to make it work. Most of the python notebooks seem to reproduce the figure and the main results, I did not find the scripts for training the models.  
+
+[1] S. Dubbioso, G. De Tommasi, A. Mele, G. Tartaglione, M. Ariola, and A. Pironti, 'A Deep Reinforcement Learning approach for Vertical Stabilization of tokamak plasmas', Fusion Engineering and Design, vol. 194, p. 113725, Sep. 2023, doi: 10.1016/j.fusengdes.2023.113725. [2] G. De Tommasi, S. Dubbioso, Y. Huang, Z. P. Luo, A. Mele, and B. J. Xiao, 'A RL- based Vertical Stabilization System for the EAST tokamak', in 2022 American Control Conference (ACC), Jun. 2022, pp. 5328- 5333. doi: 10.23919/ACC53348.2022.9867499.
+
+<--- Page Split --->
+
+Reviewer #1 (Remarks on code availability):  
+
+Partial review of the code due to the difficulties in installing all the proper libraries The readme file instructs on using poetry, but after finding the correct version of poetry I ran into an error:  
+
+Directory TokaGym\submodules\contrarx for contrarx does not seem to be a Python package  
+
+## Reviewer #2 (Remarks to the Author):  
+
+General Comments:  
+
+This study leverages deep learning (NSSM) to predict plasma ramp- down dynamics in the TCV tokamak and applies reinforcement learning (RL) to design an optimal trajectory for stable termination during this phase. Notably, ramp- down in fusion reactors is physically and technically prone to instability, often associated with plasma disruptions. Thus, ensuring stable control and termination in this phase is critical for future large- scale devices such as ITER. This study combines a predictive model and RL on TCV to design a stable plasma termination trajectory and demonstrates robust outcomes through incremental re- training with actual experiments. This represents a significant milestone in fusion research. I believe that, provided the following comments are adequately addressed, the manuscript could be suitable for publication in Nature Communications.  
+
+## Major Comments:  
+
+1. While the significance and performance of the study are commendable, the storyline in the figures and descriptions is not very clear. Since Nature Communications is a multidisciplinary journal, the content should be understandable even to readers who are not fusion experts. From a non-expert's perspective, when reading the early part of the paper, the role of the proposed technique is not immediately intuitive (e.g., what constitutes a desired termination and how it differs from undesired outcomes). In particular, Figure 2E is the key illustration that visually demonstrates this difference, but it is not explained at all in the main text. I recommend moving Figure 2E to Figure 1 (right side of the original Fig 1) and introducing it in the Introduction or early in the Results section to clarify the role of the proposed method.
+
+<--- Page Split --->
+
+2. (Figures 6-8) These figures show only the actual plasma trajectories, but not the RL-designed or the manually programmed (planned) trajectories. However, the RL-designed trajectory, the manually programmed trajectory in the PCS, and the actual plasma response may all differ. For example, the authors state in Figure 6 that the trajectory was optimized to reduce the LFS gap to prevent VDEs, but from the figure alone, it is unclear whether this was truly suggested by the RL agent, manually reflected in PCS programming, or just occurred randomly regardless of the programmed trajectory. This distinction is crucial for interpreting and evaluating the value of the research. I recommend that the RL-designed trajectory and the PCS-programmed trajectory be also added to Figures 6 and 7 (e.g., using light-colored or dashed lines).  
+
+In Figure 8, although multiple RL trajectories under uncertainty are shown as black lines, the actual programmed trajectory used in the experiment is not shown. While the caption mentions that the average trajectory was used, it would be helpful to indicate this explicitly in the figure as well.  
+
+3. (Page 24) Unlike general RL settings, the RL agent in this study observes only time (Equation 12). This implies that once trained, the agent will always produce the same trajectory, regardless of the initial condition or change of plasma states. Thus, the first question is: when computing the trajectory to input into TCV as shown in Figure 8, how was random uncertainty introduced (like just normal random added to the given trajectory)? Second, if the initial conditions or plasma current differ, does the RL agent need to be re-trained from scratch each time?  
+
+## Minor Comments:  
+
+4. (Page 2) In the Introduction, the authors emphasize that this study focuses on feedforward trajectory design before the experiment, in contrast to previous RL-based feedback control studies. However, some previous works have also used deep learning-based predictive models in combination with RL to design operation trajectories in a feedforward manner (e.g., [J Seo et al., Nucl. Fusion 61 106010 (2021) & J Seo et al., Nucl. Fusion 62 086049 (2022)]). Although these earlier works may be more limited and less robust to uncertainty compared to the current study, they are similar in concept and deserve to be mentioned as related study.  
+
+5. (Figure 2A and others) In Figure 2A, the markers in the legend are difficult to distinguish from those in the graph itself. I suggest modifying the legend (like using boxed legends or other formatting) for easier visual distinction, including in other similar figures.  
+
+6. (Figure 2D) The background colors in this graph are not intuitive. At first glance, one might assume that green indicates successful terminations and red indicates failures.
+
+<--- Page Split --->
+
+However, shot 81635, which was successfully terminated, is marked red, while shot 81741, which had a legacy software issue, is shown in green. Additionally, the x- axis does not follow a monotonic shot order (e.g., 81635 appears after 81741), and there is no explanation for the ambiguous non- red/non- green background colors. It would also be helpful to indicate the threshold values for successful termination (W_tot and I_p criteria) directly on the graph. Similar color ambiguity is present in Figure 3 (e.g., mixed green and blue in W_tot panel), but without explanation.  
+
+7. (Page 6) The authors aim to design trajectories that avoid user-specific limits associated with disruptions and so create models that predict the evolution of relevant quantities. For the readers to understand better, it would be helpful to explicitly define the key quantities correlated with disruptions and the criteria or limits that typically lead to disruptions. Although Table 3 later in the paper provides some indirect answers through the structure of the reward function, it would be better to introduce and explain these disruption criteria earlier in the manuscript. Fusion experts may already be familiar with them, but for a multidisciplinary audience like that of Nature Communications, such clarifications are important.  
+
+8. (Tables 1 and 2) Tables 1 and 2 both define input and output variables for the deep learning models. However, in Table 1 (for NSSM), the inputs (actions) are at the bottom and outputs (observations) are at the top, while Table 2 has inputs at the top and outputs at the bottom. This inconsistency may confuse readers, especially since in typical RL settings, observations are inputs and actions are outputs. Unless there is a reason for this ordering, I recommend standardizing the layout for clarity.  
+
+9. (Page 6 and Figures 5 & 9) The profile predictor appears to reconstruct 1D profiles from already predicted 0D parameters, rather than predicting the dynamics of the profiles directly. This implies that the information in the 1D profiles is already embedded in the 0D parameters. Therefore, using the reconstructed 1D profiles in NSSM or RL trajectory design might be unnecessary and only increase input complexity. In fact, as far as I understand, the profile predictor does not play a major role in the key process of this study. I'm wondering the role of this profile predictor regarding this study. If it is not essential to the main pipeline, I suggest removing it from the current paper and including it in a follow-up study.  
+
+10. (Figure 8) The black lines represent RL trajectories under uncertainty in the actuations. However, in practice, uncertainty (or noise) affects not only the level of the actions or trajectory but also introduces fluctuations in the actuations—like the red or magenta lines in the figure. However, the RL trajectories for a_minor, remain flat except for level differences. Doesn't fluctuating action noise in real experiments affect the robustness of the designed RL trajectories?
+
+<--- Page Split --->
+
+# 1. 1. 1. 1. 1. 1. 1. 2. 2. 2. 2. 2. 2. 2.
+
+<--- Page Split --->
+
+Reviewer Comment: 1) In the newly added text at page 4 the authors claim: "At present, a comprehensive understanding of disruptive limits remains an open problem, motivating many works on machine- learning based prediction of disruptions 37- 40."  
+
+Despite being valid, this statement is supported by many self- citing references. Further/different works could be referenced since ML- based prediction of disruptions is a well developed line of research in fusion.  
+
+Response: Thank you for your suggestion. We would first like to clarify that only two out of the four cited works could be fairly characterized as self- citation. 38 is an overview paper for the international project ITER, involving a large portion of the world's experts, which inevitably leads to involvement of authors of this paper as well. 40 does not involve any authors of this paper whatsoever. To further address the balance, we have added a reference to Vega et al., 2022 Nature Physics.  
+
+Reviewer Comment: 2) Despite the statement of the authors, I consider the role of the profile predictor is still marginal in the description of the work, and recommend showing the plots discussing the 1D profile reconstruction either in the supplementary data or in a new work. On the other hand, for the purposes of the work, I find more interesting the discussion on Figures 16 and 20, now in the supplementary data. These figures allow to discuss the validity of the NSSM for the current application. For instance, in the two successful pulses of Figure 16, the yvgr is close to the same limit value as in the one of Figure 20. How is the distinction between these two sets of pulses made?  
+
+Response: Thank you for your perspective. Regarding the profile prediction content, we would like to re- iterate our opinion that the result is informative for the community given the multiple current research efforts to do this with a principles- based approach [1, 2, 3].  
+
+Regarding Figures 16 and 20, we would like to highlight that they are new figures introduced as a part of the revision. We would like to point out that in Figure 20, the distribution travels much further into the soft constraint region than in Figure 16, especially in the milliseconds immediately preceding the disruption.  
+
+[1] Teplukhina, A. A., et al. "Simulation of profile evolution from ramp- up to ramp- down and optimization of tokamak plasma termination with the RAPTOR code." Plasma Physics and Controlled Fusion 59.12 (2017): 124004. [2] Citrin, Jonathan, et al. "TORAX: A fast and differentiable tokamak transport simulator in JAX." arXiv preprint arXiv:2406.06718 (2024). [3] Meneghini, O., et al. "FUSE (Fusion Synthesis Engine): A next generation framework for integrated design of fusion pilot plants." arXiv preprint arXiv:2409.05894 (2024).
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__71c5664823cdb51d0090bce56bc9fa1f7a11a279819fc98c43c3bd438c64318c/supplementary_0_Transparent Peer Review file__71c5664823cdb51d0090bce56bc9fa1f7a11a279819fc98c43c3bd438c64318c_det.mmd

@@ -0,0 +1,419 @@
+<|ref|>title<|/ref|><|det|>[[73, 160, 784, 210]]<|/det|>
+# Learning Plasma Dynamics and Robust Rampdown Trajectories with Predict-First Experiments at TCV  
+
+<|ref|>text<|/ref|><|det|>[[73, 223, 402, 241]]<|/det|>
+Corresponding Author: Mr Allen Wang  
+
+<|ref|>text<|/ref|><|det|>[[70, 275, 864, 290]]<|/det|>
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+<|ref|>text<|/ref|><|det|>[[73, 327, 150, 341]]<|/det|>
+Version 0:  
+
+<|ref|>text<|/ref|><|det|>[[73, 354, 220, 368]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[73, 380, 160, 394]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[73, 405, 238, 418]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 418, 920, 485]]<|/det|>
+The work revolves around the use of reinforcement learning (RL) for the optimization of the ramp- down trajectory at TCV, an experimental device in Switzerland. The idea of RL is that the model can learn how to interact with the system by trial and error and requires a sufficiently accurate model of the system. In this case, also the system is approximated by a Neural State Space Model (NSSM), that is a Neural Network which can approximate the state space representation mixing a data driven loss and numerical integration of the (net+system) dynamics.  
+
+<|ref|>text<|/ref|><|det|>[[72, 484, 920, 524]]<|/det|>
+The work is interesting but due to the scarcity of experiments, the successfully designed rampdowns are only 5, out of a total of 9 optimized ones. For this reason, a further analysis and replication of the experiments is recommended for making robust conclusions.  
+
+<|ref|>text<|/ref|><|det|>[[73, 523, 444, 536]]<|/det|>
+Moreover, I have the following comments on the work:  
+
+<|ref|>text<|/ref|><|det|>[[73, 536, 426, 549]]<|/det|>
+NSSM and training environment for the RL system  
+
+<|ref|>text<|/ref|><|det|>[[72, 549, 916, 589]]<|/det|>
+The article is quite dense, and the authors explain that one model (the NSSM) serves as the training environment for the RL model, however another neural model to predict the temperature and density profiles is mentioned as part of the NSSM. From section 2.2:  
+
+<|ref|>text<|/ref|><|det|>[[72, 588, 923, 720]]<|/det|>
+"The NSSM was initially developed with a neural network predictor for the kinetic profiles on the full p grid, and initial training runs found that the profile predictor can accurately predict kinetic profiles given the set of OD scalars listed in Table 2. Figure 5 provides an example comparison of predictions of the Te and ne profiles against Thomson measurements for a full shot in the validation dataset, showing accurate prediction across all phases of the shot. This result corroborates previous findings at NSTX- U that neural networks can accurately predict kinetic profiles given a set of similar OD scalars34. Given this result suggests most of the relevant profile information is implicitly captured by OD scalars, the profile predictor was disabled prior to running experiments to accelerate training, hence reported predictions of kinetic profiles are not predict- first. A noteworthy feature of this profile predictor that should be explored in future work is its ability to function as a Thomson upsampler, as the input variables are all sampled at a higher time resolution, 1ms, than the TCV Thomson Scattering system, which takes measurements every 17ms."  
+
+<|ref|>text<|/ref|><|det|>[[72, 720, 916, 760]]<|/det|>
+If the predictor has not been adopted, I suggest removing the parts connected with the profile predictor (e.g section 4.1.3, Figures 5, 9) to avoid confusion, otherwise I would suggest indicating for how many discharges the predictor has been used to train the RL method and if there was any benefit from this approach.  
+
+<|ref|>text<|/ref|><|det|>[[70, 770, 900, 797]]<|/det|>
+How were the discharges to train the NSSM model selected? What are the variations in terms of the (observation, action) space?  
+
+<|ref|>text<|/ref|><|det|>[[72, 808, 911, 850]]<|/det|>
+Connected to the previous point, it seems that the vertical instability growth rate \((\gamma_{- }\) vgr) reconstruction by the NSSM is very poor, even with a very low prediction Horizon (Figure 4). Since this parameter is connected to vertical plasma stability, this would explain the VDE event in shot #81751. How large is the uncertainty assumed for the \(\gamma_{- }\) vgr?  
+
+<|ref|>text<|/ref|><|det|>[[72, 861, 920, 901]]<|/det|>
+An interesting point could be made by running the control discharges in the NSSM and showing that they were close to disruptive boundaries/expected to disrupt. This would show, despite the low number of experiments, that the NSSM can well represent the system.  
+
+<|ref|>text<|/ref|><|det|>[[70, 911, 880, 939]]<|/det|>
+Is it possible to interpret the strategies adopted by the RL model by showing which parameters are kept far from critical boundaries?
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[73, 60, 187, 73]]<|/det|>
+Other comments  
+
+<|ref|>text<|/ref|><|det|>[[72, 85, 920, 140]]<|/det|>
+Some other works related to the use of RL in fusion are missing, as applications for ITER [1] and EAST [2]. Please add some details on the model adopted (number of layers, neurons) and on how they have been selected/optimized. I tried running the code and installing all the dependencies using poetry, but I could not manage to make it work. Most of the python notebooks seem to reproduce the figure and the main results, I did not find the scripts for training the models.  
+
+<|ref|>text<|/ref|><|det|>[[72, 164, 920, 243]]<|/det|>
+[1] S. Dubbioso, G. De Tommasi, A. Mele, G. Tartaglione, M. Ariola, and A. Pironti, 'A Deep Reinforcement Learning approach for Vertical Stabilization of tokamak plasmas', Fusion Engineering and Design, vol. 194, p. 113725, Sep. 2023, doi: 10.1016/j.fusengdes.2023.113725. [2] G. De Tommasi, S. Dubbioso, Y. Huang, Z. P. Luo, A. Mele, and B. J. Xiao, 'A RL- based Vertical Stabilization System for the EAST tokamak', in 2022 American Control Conference (ACC), Jun. 2022, pp. 5328- 5333. doi: 10.23919/ACC53348.2022.9867499.  
+
+<|ref|>text<|/ref|><|det|>[[73, 281, 785, 334]]<|/det|>
+(Remarks on code availability) Partial review of the code due to the difficulties in installing all the proper libraries The readme file instructs on using poetry, but after finding the correct version of poetry I ran into an error: Directory TokaGym\submodule's\contrax for contrax does not seem to be a Python package  
+
+<|ref|>text<|/ref|><|det|>[[73, 346, 163, 360]]<|/det|>
+Reviewer #2  
+
+<|ref|>text<|/ref|><|det|>[[73, 372, 237, 399]]<|/det|>
+(Remarks to the Author) General Comments:  
+
+<|ref|>text<|/ref|><|det|>[[72, 410, 916, 518]]<|/det|>
+This study leverages deep learning (NSSM) to predict plasma ramp- down dynamics in the TCV tokamak and applies reinforcement learning (RL) to design an optimal trajectory for stable termination during this phase. Notably, ramp- down in fusion reactors is physically and technically prone to instability, often associated with plasma disruptions. Thus, ensuring stable control and termination in this phase is critical for future large- scale devices such as ITER. This study combines a predictive model and RL on TCV to design a stable plasma termination trajectory and demonstrates robust outcomes through incremental re- training with actual experiments. This represents a significant milestone in fusion research. I believe that, provided the following comments are adequately addressed, the manuscript could be suitable for publication in Nature Communications.  
+
+<|ref|>sub_title<|/ref|><|det|>[[73, 541, 195, 555]]<|/det|>
+## Major Comments:  
+
+<|ref|>text<|/ref|><|det|>[[72, 566, 925, 660]]<|/det|>
+1. While the significance and performance of the study are commendable, the storyline in the figures and descriptions is not very clear. Since Nature Communications is a multidisciplinary journal, the content should be understandable even to readers who are not fusion experts. From a non-expert's perspective, when reading the early part of the paper, the role of the proposed technique is not immediately intuitive (e.g., what constitutes a desired termination and how it differs from undesired outcomes). In particular, Figure 2E is the key illustration that visually demonstrates this difference, but it is not explained at all in the main text. I recommend moving Figure 2E to Figure 1 (right side of the original Fig 1) and introducing it in the Introduction or early in the Results section to clarify the role of the proposed method.  
+
+<|ref|>text<|/ref|><|det|>[[72, 671, 920, 802]]<|/det|>
+2. (Figures 6-8) These figures show only the actual plasma trajectories, but not the RL-designed or the manually programmed (planned) trajectories. However, the RL-designed trajectory, the manually programmed trajectory in the PCS, and the actual plasma response may all differ. For example, the authors state in Figure 6 that the trajectory was optimized to reduce the LFS gap to prevent VDEs, but from the figure alone, it is unclear whether this was truly suggested by the RL agent, manually reflected in PCS programming, or just occurred randomly regardless of the programmed trajectory. This distinction is crucial for interpreting and evaluating the value of the research. I recommend that the RL-designed trajectory and the PCS-programmed trajectory be also added to Figures 6 and 7 (e.g., using light-colored or dashed lines). In Figure 8, although multiple RL trajectories under uncertainty are shown as black lines, the actual programmed trajectory used in the experiment is not shown. While the caption mentions that the average trajectory was used, it would be helpful to indicate this explicitly in the figure as well.  
+
+<|ref|>text<|/ref|><|det|>[[72, 813, 915, 880]]<|/det|>
+3. (Page 24) Unlike general RL settings, the RL agent in this study observes only time (Equation 12). This implies that once trained, the agent will always produce the same trajectory, regardless of the initial condition or change of plasma states. Thus, the first question is: when computing the trajectory to input into TCV as shown in Figure 8, how was random uncertainty introduced (like just normal random added to the given trajectory)? Second, if the initial conditions or plasma current differ, does the RL agent need to be re-trained from scratch each time?  
+
+<|ref|>sub_title<|/ref|><|det|>[[73, 906, 195, 919]]<|/det|>
+## Minor Comments:  
+
+<|ref|>text<|/ref|><|det|>[[70, 931, 891, 946]]<|/det|>
+4. (Page 2) In the Introduction, the authors emphasize that this study focuses on feedforward trajectory design before the
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 46, 922, 112]]<|/det|>
+experiment, in contrast to previous RL- based feedback control studies. However, some previous works have also used deep learning- based predictive models in combination with RL to design operation trajectories in a feedforward manner (e.g., [J Seo et al., Nucl. Fusion 61 106010 (2021) & J Seo et al., Nucl. Fusion 62 086049 (2022)]). Although these earlier works may be more limited and less robust to uncertainty compared to the current study, they are similar in concept and deserve to be mentioned as related study.  
+
+<|ref|>text<|/ref|><|det|>[[72, 124, 905, 166]]<|/det|>
+5. (Figure 2A and others) In Figure 2A, the markers in the legend are difficult to distinguish from those in the graph itself. I suggest modifying the legend (like using boxed legends or other formatting) for easier visual distinction, including in other similar figures.  
+
+<|ref|>text<|/ref|><|det|>[[72, 177, 922, 257]]<|/det|>
+6. (Figure 2D) The background colors in this graph are not intuitive. At first glance, one might assume that green indicates successful terminations and red indicates failures. However, shot 81635, which was successfully terminated, is marked red, while shot 81741, which had a legacy software issue, is shown in green. Additionally, the x-axis does not follow a monotonic shot order (e.g., 81635 appears after 81741), and there is no explanation for the ambiguous non-red/non-green background colors. It would also be helpful to indicate the threshold values for successful termination (W_tot and I_p criteria) directly on the graph. Similar color ambiguity is present in Figure 3 (e.g., mixed green and blue in W_tot panel), but without explanation.  
+
+<|ref|>text<|/ref|><|det|>[[72, 268, 920, 347]]<|/det|>
+7. (Page 6) The authors aim to design trajectories that avoid user-specific limits associated with disruptions and so create models that predict the evolution of relevant quantities. For the readers to understand better, it would be helpful to explicitly define the key quantities correlated with disruptions and the criteria or limits that typically lead to disruptions. Although Table 3 later in the paper provides some indirect answers through the structure of the reward function, it would be better to introduce and explain these disruption criteria earlier in the manuscript. Fusion experts may already be familiar with them, but for a multidisciplinary audience like that of Nature Communications, such clarifications are important.  
+
+<|ref|>text<|/ref|><|det|>[[72, 358, 917, 425]]<|/det|>
+8. (Tables 1 and 2) Tables 1 and 2 both define input and output variables for the deep learning models. However, in Table 1 (for NSSM), the inputs (actions) are at the bottom and outputs (observations) are at the top, while Table 2 has inputs at the top and outputs at the bottom. This inconsistency may confuse readers, especially since in typical RL settings, observations are inputs and actions are outputs. Unless there is a reason for this ordering, I recommend standardizing the layout for clarity.  
+
+<|ref|>text<|/ref|><|det|>[[72, 435, 919, 516]]<|/det|>
+9. (Page 6 and Figures 5 & 9) The profile predictor appears to reconstruct 1D profiles from already predicted 0D parameters, rather than predicting the dynamics of the profiles directly. This implies that the information in the 1D profiles is already embedded in the 0D parameters. Therefore, using the reconstructed 1D profiles in NSSM or RL trajectory design might be unnecessary and only increase input complexity. In fact, as far as I understand, the profile predictor does not play a major role in the key process of this study. I'm wondering the role of this profile predictor regarding this study. If it is not essential to the main pipeline, I suggest removing it from the current paper and including it in a follow-up study.  
+
+<|ref|>text<|/ref|><|det|>[[72, 527, 919, 581]]<|/det|>
+10. (Figure 8) The black lines represent RL trajectories under uncertainty in the actuations. However, in practice, uncertainty (or noise) affects not only the level of the actions or trajectory but also introduces fluctuations in the actuations—like the red or magenta lines in the figure. However, the RL trajectories for a minor, remain flat except for level differences. Doesn’t fluctuating action noise in real experiments affect the robustness of the designed RL trajectories?  
+
+<|ref|>text<|/ref|><|det|>[[73, 593, 282, 607]]<|/det|>
+(Remarks on code availability)  
+
+<|ref|>text<|/ref|><|det|>[[73, 632, 144, 645]]<|/det|>
+Version 1:  
+
+<|ref|>text<|/ref|><|det|>[[73, 657, 220, 671]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[73, 683, 161, 696]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[73, 709, 237, 722]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[73, 722, 444, 735]]<|/det|>
+1) In the newly added text at page 4 the authors claim:  
+
+<|ref|>text<|/ref|><|det|>[[72, 736, 870, 763]]<|/det|>
+"At present, a comprehensive understanding of disruptive limits remains an open problem, motivating many works on machine-learning based prediction of disruptions 37-40."  
+
+<|ref|>text<|/ref|><|det|>[[72, 763, 905, 789]]<|/det|>
+Despite being valid, this statement is supported by many self- citing references. Further/different works could be referenced since ML- based prediction of disruptions is a well developed line of research in fusion.  
+
+<|ref|>text<|/ref|><|det|>[[72, 801, 912, 880]]<|/det|>
+2) Despite the statement of the authors, I consider the role of the profile predictor is still marginal in the description of the work, and recommend showing the plots discussing the 1D profile reconstruction either in the supplementary data or in a new work. On the other hand, for the purposes of the work, I find more interesting the discussion on Figures 16 and 20, now in the supplementary data. These figures allow to discuss the validity of the NSSM for the current application. For instance, in the two successful pulses of Figure 16, the yvgr is close to the same limit value as in the one of Figure 20. How is the distinction between these two sets of pulses made?  
+
+<|ref|>text<|/ref|><|det|>[[73, 906, 282, 920]]<|/det|>
+(Remarks on code availability)
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 47, 163, 60]]<|/det|>
+Reviewer #2  
+
+<|ref|>text<|/ref|><|det|>[[72, 74, 875, 115]]<|/det|>
+(Remarks to the Author) Thanks for the authors' sincere replies to my previous comments. The authors appropriately answered my previous questions and revised the paper accordingly. Now, I recommend the paper for publication as it is.  
+
+<|ref|>text<|/ref|><|det|>[[72, 127, 283, 141]]<|/det|>
+(Remarks on code availability)  
+
+<|ref|>text<|/ref|><|det|>[[72, 505, 916, 560]]<|/det|>
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+<|ref|>text<|/ref|><|det|>[[72, 559, 915, 625]]<|/det|>
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source. The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.  
+
+<|ref|>text<|/ref|><|det|>[[72, 624, 618, 638]]<|/det|>
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[115, 118, 480, 144]]<|/det|>
+## Responses to Reviewers  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 172, 246, 193]]<|/det|>
+## Reviewer 1  
+
+<|ref|>text<|/ref|><|det|>[[115, 204, 878, 277]]<|/det|>
+Thank you for taking the time to provide helpful feedback. We concur that it would be beneficial to increase the statistics of the result; however, it is not straightforward to simply request further shots from an experimental fusion device, especially for this high performance scenario which is highly unusual for TCV.  
+
+<|ref|>text<|/ref|><|det|>[[115, 296, 866, 369]]<|/det|>
+We have, however, performed additional analyses to increase confidence in the results. This includes running a disruptive shots action trajectory in the RL environment, as you recommended. This also involved a further physics- based analysis confirming that the new RL design trajectories have more vertical stability in the presence of control errors.  
+
+<|ref|>text<|/ref|><|det|>[[115, 388, 878, 441]]<|/det|>
+We have also made changes to the introductory figures in the paper, in large part in response to feedback from your co- reviewer. We hope the new format makes the results more accessible to a broad audience.  
+
+<|ref|>text<|/ref|><|det|>[[115, 461, 835, 497]]<|/det|>
+Thank you for pointing out issues with the codebase. It appears we did not archive the key submodule; this issue has been addressed.  
+
+<|ref|>text<|/ref|><|det|>[[115, 516, 600, 533]]<|/det|>
+Further details are provided in the line item responses below.  
+
+<|ref|>text<|/ref|><|det|>[[115, 553, 306, 606]]<|/det|>
+Regards, Allen M. Wang On Behalf of Coauthors  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 648, 322, 667]]<|/det|>
+## Line-Item Response  
+
+<|ref|>table<|/ref|><|det|>[[115, 691, 883, 899]]<|/det|>
+
+<table><tr><td>Comment</td><td>Response</td></tr><tr><td>The work is interesting but due to the scarcity of experiments, the successfully designed rampdowns are only 5, out of a total of 9 optimized ones. For this reason, a further analysis and replication of the experiments is recommended for making robust conclusions.</td><td>We would like to point out that:<br>1) 6 rampdowns were successful<br>2) Two shots were utilized explicitly to debug an unrelated issue caused by a legacy 1/lp^2 term in the PCS, leading to poor radial control during rampdowns. The fact that this issue existed and was finally fixed after the debug shots is shown in Figure 14 in the Supplementary Materials. We</td></tr></table>
+
+<--- Page Split --->
+<|ref|>table<|/ref|><|det|>[[113, 87, 883, 897]]<|/det|>
+
+<table><tr><td></td><td>have chosen to report these shots<br>nevertheless for transparency, and to<br>raise awareness on the often subtle,<br>but important, issues involved in<br>fusion experiments which are often<br>omitted from publications.<br>3) We performed statistical analysis both<br>with and without these two shots<br>included in the test set. Even with<br>these two shots included in the test<br>set, the results are encouraging.</td></tr><tr><td>The article is quite dense, and the authors explain that one model (the NSSM) serves as the training environment for the RL<br>model, however another neural model to predict the temperature and density profiles is mentioned as part of the NSSM. From<br>section 2.2:<br>"The NSSM was initially developed with a<br>neural network predictor for the kinetic<br>profiles on the full p grid, and initial training runs found that the profile predictor can<br>accurately predict kinetic profiles given the set of 0D scalars listed in Table 2.<br>Figure 5 provides an example comparison of predictions of the Te and ne profiles<br>against Thomson measurements for a full shot in the validation dataset, showing<br>accurate prediction across all phases of the shot. This result corroborates previous<br>findings at NSTX-U that neural networks<br>can accurately predict kinetic profiles given a set of similar 0D scalars34. Given this<br>result suggests most of the relevant profile information is implicitly captured by 0D<br>scalars, the profile predictor was disabled<br>prior to running experiments to accelerate training, hence reported predictions of<br>kinetic profiles are not predict-first. A<br>noteworthy feature of this profile predictor that should be explored in future work is its</td><td>We understand why removing the content on the profile predictor would help streamline the paper. However, we want to highlight the fact that the results of our work motivate further<br>research investment in combining established 0D transport simulation methods with neural networks. We believe this is a worthwhile<br>message to preserve, given the large number of ongoing efforts towards developing<br>principles-based transport simulators for<br>control. This message is highlighted in the<br>Discussion:<br>"The ability for a neural network to predict<br>kinetic profiles using 0D scalars,<br>demonstrated both in this work and in prior work43, suggests a<br>data-driven approach may be sufficient for<br>certain control tasks without principles-based transport simulation, which can be<br>extremely computationally expensive and<br>require strong assumptions on edge<br>temperature and density."<br>Changes Made: added a sentence earlier in<br>the paper (Section 2.2) for clarity of message:<br>"This result also suggests a structured<br>data-driven approach to modeling tokamak<br>transport merits further research, in parallel<br>with several ongoing principles-based efforts."</td></tr></table>
+
+<--- Page Split --->
+<|ref|>table<|/ref|><|det|>[[113, 90, 884, 911]]<|/det|>
+
+<table><tr><td>ability to function as a Thomson<br>up-sampler, as the input variables are all<br>sampled at a higher time resolution, 1ms,<br>than the TCV Thomson Scattering system,<br>which takes measurements every 17ms."<br>If the predictor has not been adopted, I<br>suggest removing the parts connected with<br>the profile predictor (e.g section 4.1.3,<br>Figures 5, 9) to avoid confusion, otherwise I<br>would suggest indicating for how many<br>discharges the predictor has been used to<br>train the RL method and if there was any<br>benefit from this approach.</td><td></td></tr><tr><td>How were the discharges to train the NSSM model selected? What are the variations in terms of the (observation, action) space?</td><td>We provide details in the subsection "Training Data Distribution" in the methods. To provide further clarity, we used the DEFUSE system to gather recent shots that simultaneously<br>have a partial rampdown and sufficient<br>diagnostic availability. A plot of the data<br>distribution in beta, Ip and greenwald fraction space is also shown in that subsection.<br>Changes Made: modified "Training Data<br>Distribution" to clarify that we gathered the<br>most recent shots with sufficient diagnostic<br>availability.</td></tr><tr><td>Connected to the previous point, it seems<br>that the vertical instability growth rate<br>(y_vgr) reconstruction by the NSSM is very<br>poor, even with a very low prediction<br>Horizon (Figure 4). Since this parameter is<br>connected to vertical plasma stability, this<br>would explain the VDE event in shot<br>#81751. How large is the uncertainty<br>assumed for the y_vgr?</td><td>We discuss the prediction accuracy of y_vgr<br>in Page 6 paragraph two:<br>"The percent errors for yvgr can be relatively large, but, as shown in Figure 16 this is largely attributable to the small value of yvgr of limited plasmas as the absolute error is relatively low."<br>The wide distribution you are seeing is due to<br>the fact that gamma_vgr on TCV is highly<br>sensitive to control errors, which is discussed in page 11.<br>The uncertainty distributions added to the<br>y_vgr and the variables that are input into the model predicting it were quantified from<br>model prediction error metrics, as discussed in page 23:<br>"In some cases, uncertainty distributions could easily be</td></tr></table>
+
+<--- Page Split --->
+<|ref|>table<|/ref|><|det|>[[113, 87, 884, 901]]<|/det|>
+
+<table><tr><td></td><td>quantified from past experimental data (such as<br>tracking error in the plasma current), or from model<br>prediction accuracy (such as vvgr),"</td></tr><tr><td>An interesting point could be made by<br>running the control discharges in the NSSM and showing that they were close to<br>disruptive boundaries/expected to disrupt.<br>This would show, despite the low number of experiments, that the NSSM can well<br>represent the system.</td><td>We performed this analysis, which shows that the training environment expects a large<br>vertical instability growth rate using the action<br>trajectories from 81101, one of the baseline<br>shots.<br>Given the length and scope of the main paper as it is, and the size of this plot, we believe<br>this result belongs in the supplementary<br>materials.<br>Changes Made: added a figure to the<br>supplementary materials showing NSSM<br>predictions given the action trajectory from a disruptive shot.</td></tr><tr><td>Is it possible to interpret the strategies<br>adopted by the RL model by showing which parameters are kept far from critical<br>boundaries?</td><td>The right hand side of Figure 8 is meant to<br>illustrate this.</td></tr><tr><td>Some other works related to the use of RL<br>in fusion are missing, as applications for<br>ITER [1] and EAST [2].</td><td>Thank you for pointing out these works.<br>Changes Made: included references to these two papers in the introduction.</td></tr><tr><td>Please add some details on the model<br>adopted (number of layers, neurons) and on how they have been selected/optimized.</td><td>Thank you for pointing out this omission.<br>Changes Made: Introduced "Training<br>Methods" to the NSSM section. Added<br>sentences added to "RL Methods" specifying model architecture.</td></tr><tr><td>I tried running the code and installing all the dependencies using poetry, but I could not<br>manage to make it work. Most of the<br>python notebooks seem to reproduce the<br>figure and the main results, I did not find the scripts for training the models.</td><td>Apologies, it appears that we didn't archive<br>the submodule containing the dynamics<br>model. We have taken care to make sure it is<br>available in the updated .tar.gz file and poetry install works if the user is using<br>python3.10/11.<br>While we have taken care to make sure<br>notebooks under "paper_plot_generation" are reproducible, execution of the full repo is not<br>possible due to dependencies on large<br>datasets that are under Eurofusion data<br>access restrictions, and thus cannot be made</td></tr></table>
+
+<--- Page Split --->
+<|ref|>table<|/ref|><|det|>[[113, 88, 884, 298]]<|/det|>
+<table><tr><td></td><td>public for the foreseeable future.<br/>The NSSM training main script is under:<br/>"submodules/contrax/contrax/examples/plasma/tcv/train_model.py"<br/>RL training scripts are under:<br/>"scripts/train_ppo.py"<br/>"scripts/train_es.py"<br/>Changes Made: updated .tar.gz containing codebase + data files.</td></tr></table>
+
+<|ref|>title<|/ref|><|det|>[[115, 340, 250, 360]]<|/det|>
+Reviewer 2
+
+<|ref|>text<|/ref|><|det|>[[115, 373, 874, 443]]<|/det|>
+Thank you for taking your time to provide helpful feedback in improving the quality of the manuscript, and informing us of important prior works on RL for trajectory optimization in fusion we were not previously aware of. We have adopted your recommendations, which has helped revamp the early presentation of the paper to be more accessible to the reader.
+
+<|ref|>text<|/ref|><|det|>[[115, 465, 872, 536]]<|/det|>
+In response to both you and your co-reviewer, we have also performed a further physics-based analysis which corroborates the improved robustness of RL-designed trajectories to vertical instability by introducing control errors into the shot preparation equilibrium solver and solving for the resulting gamma_vgr distributions.
+
+<|ref|>text<|/ref|><|det|>[[115, 558, 602, 572]]<|/det|>
+Further details are provided in the line item responses below.
+
+<|ref|>text<|/ref|><|det|>[[115, 595, 193, 609]]<|/det|>
+Regards,
+
+<|ref|>text<|/ref|><|det|>[[115, 613, 237, 627]]<|/det|>
+Allen M. Wang
+
+<|ref|>text<|/ref|><|det|>[[115, 631, 308, 644]]<|/det|>
+On Behalf of Coauthors
+
+<|ref|>title<|/ref|><|det|>[[115, 670, 325, 686]]<|/det|>
+Line-Item Response
+
+<|ref|>table<|/ref|><|det|>[[115, 712, 884, 904]]<|/det|>
+
+<table><tr><td>Comment</td><td>Response</td></tr><tr><td>1. While the significance and performance<br>of the study are commendable, the storyline in the figures and descriptions is not very<br>clear. Since Nature Communications is a<br>multidisciplinary journal, the content should be understandable even to readers who are not fusion experts. From a non-expert's</td><td>Thank you for this recommendation. We<br>concur that moving Figure 2E to Figure 1 on<br>the right side is helpful, and we have adopted this change. We have also updated the<br>original Figure 1 to a shot that disrupts at high Greenwald Fraction for further clarity.<br>Changes Made: Figure 2E is now part of<br>Figure 1, and Figure 1 updated to a shot that</td></tr></table>
+
+<--- Page Split --->
+<|ref|>table<|/ref|><|det|>[[112, 88, 884, 911]]<|/det|>
+
+<table><tr><td>perspective, when reading the early part of the paper, the role of the proposed<br>technique is not immediately intuitive (e.g.,<br>what constitutes a desired termination and how it differs from undesired outcomes). In particular, Figure 2E is the key illustration<br>that visually demonstrates this difference,<br>but it is not explained at all in the main text. I recommend moving Figure 2E to Figure 1<br>(right side of the original Fig 1) and<br>introducing it in the Introduction or early in<br>the Results section to clarify the role of the<br>proposed method.</td><td>more clearly illustrated the issue at hand.</td></tr><tr><td>2. (Figures 6-8) These figures show only<br>the actual plasma trajectories, but not the<br>RL-designed or the manually programmed<br>(planned) trajectories. However, the<br>RL-designed trajectory, the manually<br>programmed trajectory in the PCS, and the<br>actual plasma response may all differ. For<br>example, the authors state in Figure 6 that<br>the trajectory was optimized to reduce the<br>LFS gap to prevent VDEs, but from the<br>figure alone, it is unclear whether this was<br>truly suggested by the RL agent, manually<br>reflected in PCS programming, or just<br>occurred randomly regardless of the<br>programmed trajectory. This distinction is<br>crucial for interpreting and evaluating the<br>value of the research. I recommend that the<br>RL-designed trajectory and the<br>PCS-programmed trajectory be also added<br>to Figures 6 and 7 (e.g., using light-colored<br>or dashed lines).<br>In Figure 8, although multiple RL<br>trajectories under uncertainty are shown as<br>black lines, the actual programmed<br>trajectory used in the experiment is not<br>shown. While the caption mentions that the</td><td>Thank you for your feedback. We will respond by figure.<br>Changes Made:<br>Figure 6: thank you for your recommendation. We have added the RL-designed<br>PCS-programmed trajectory to this figure. To further confirm that the RL-designed<br>trajectory after adding control error is indeed more robust to the vertical instability, we have performed a further physics-based<br>robustness analysis by introducing minor<br>radius control errors to the RL designed<br>trajectories, solving for equilibria using the<br>free-boundary equilibrium solver used for<br>TCV shot preparation, and then calculating<br>gamma_vgr distributions for the resulting<br>equilibria. The results (shown in the new<br>Figure 7) show that indeed 82875 is more<br>robust than 81751.<br>Figure 7 (now Figure 8): it is more difficult to show the RL designed trajectories in<br>conjunction with the actual trajectories for this figure, given that the trajectories changed<br>from shot to shot. The purpose of the figure is to show that we progressed over the course<br>of the experiment. However, to provide the<br>information you suggested, we have included the equivalent of Figure 8 for the final two<br>140kA shots as a part of the supplementary<br>information.</td></tr></table>
+
+<--- Page Split --->
+<|ref|>table<|/ref|><|det|>[[113, 88, 884, 878]]<|/det|>
+
+<table><tr><td>average trajectory was used, it would be<br>helpful to indicate this explicitly in the figure as well.</td><td>Figure 8 (now Figure 9): this is a good<br>suggestion, we have included the mean of<br>the action trajectories as a solid thick black<br>line as well, and included a comment on this in the caption.</td></tr><tr><td>3. (Page 24) Unlike general RL settings, the RL agent in this study observes only time (Equation 12). This implies that once<br>trained, the agent will always produce the same trajectory, regardless of the initial<br>condition or change of plasma states.<br>Thus, the first question is: when computing the trajectory to input into TCV as shown in Figure 8, how was random uncertainty<br>introduced (like just normal random added to the given trajectory)? Second, if the initial conditions or plasma current differ, does<br>the RL agent need to be re-trained from<br>scratch each time?</td><td>This is a good point. The uncertainty<br>distribution for initial conditions specified is shown in Section 4.4: Uncertainty Model.<br>The idea is to have a single action trajectory that is likely to succeed across a range of<br>initial conditions.<br>If the initial conditions change by more than anticipated, then indeed retraining needs to happen (as was the case for 140kA vs.<br>170kA).<br>Changes Made: we have introduced a<br>sentence in 4.5 to make this more clear:<br>"Given that time is the only observable, but<br>there exists different physical conditions in<br>the parallel training environments which are unobservable to the policy, the reward<br>maximization process yields a trajectory that is designed to succeed across the different conditions specified in Subsection<br>vref{subsec:uncertainty_model}."</td></tr><tr><td>4. (Page 2) In the Introduction, the authors emphasize that this study focuses on<br>feedforward trajectory design before the<br>experiment, in contrast to previous<br>RL-based feedback control studies.<br>However, some previous works have also<br>used deep learning-based predictive<br>models in combination with RL to design<br>operation trajectories in a feedforward<br>manner (e.g., [J Seo et al., Nucl. Fusion 61<br>106010 (2021) & J Seo et al., Nucl. Fusion<br>62 086049 (2022)]). Although these earlier<br>works may be more limited and less robust to uncertainty compared to the current</td><td>Thank you for informing us of this work.<br>We have included references.<br>Changes Made: we have included both<br>citations and introduced the papers in the<br>introduction:<br>"A similar approach has previously been<br>demonstrated at KSTAR for designing<br>feed-forward trajectories that reach target<br>states<br>\\cite{seo2021feedforward,seo2022developme nt}."</td></tr></table>
+
+<--- Page Split --->
+<|ref|>table<|/ref|><|det|>[[113, 87, 884, 875]]<|/det|>
+
+<table><tr><td>study, they are similar in concept and<br>deserve to be mentioned as related study.</td><td></td></tr><tr><td>5. (Figure 2A and others) In Figure 2A, the markers in the legend are difficult to<br>distinguish from those in the graph itself. I suggest modifying the legend (like using<br>boxed legends or other formatting) for<br>easier visual distinction, including in other<br>similar figures.</td><td>Thank you for pointing out this deficiency.<br>Changes Made: added a box to the legend.</td></tr><tr><td>6. (Figure 2D) The background colors in this graph are not intuitive. At first glance, one<br>might assume that green indicates<br>successful terminations and red indicates<br>failures. However, shot 81635, which was<br>successfully terminated, is marked red,<br>while shot 81741, which had a legacy<br>software issue, is shown in green.<br>Additionally, the x-axis does not follow a<br>monotonic shot order (e.g., 81635 appears<br>after 81741), and there is no explanation for the ambiguous non-red/non-green<br>background colors. It would also be helpful to indicate the threshold values for<br>successful termination (W_tot and I_p<br>criteria) directly on the graph. Similar color ambiguity is present in Figure 3 (e.g., mixed green and blue in W_tot panel), but without explanation.</td><td>Thank you for pointing out the non-monotonic ordering. This was somewhat embarrassing<br>given that it was 81741 and 81745 that were<br>the debug shots. We have addressed this<br>issue.<br>Your feedback has helped us recognize there is a better way to communicate our results.<br>We have now replaced the histogram with a<br>scatter plot in (Wtot and Ip) space showing<br>the shots and a table of statistics embedded<br>in the plot. We realized it likely makes more<br>sense to have this plot in the "Overview" as<br>opposed to having the shot-by-shot<br>breakdown in the overview. The shot-by-shot<br>breakdown is now its own separate figure.<br>We have added the goal Ip to the scatter plot. As we note in a new footnote, we initially<br>chose a goal stored energy, 0.5kJ, that<br>proved unrealistic to diagnose as equilibrium<br>reconstruction no longer converges reliably at such low plasma current and stored energy.<br>Thus, we would like to primarily focus on<br>reaching the goal Ip as it can be diagnosed<br>relatively reliably without reconstruction.<br>Changes Made: 81635 + 81741 are now in<br>their proper order. Figure 3 is now a scatter<br>plot in Wtot and Ip space and is part of the<br>overview. The shot-by-shot breakdown is now a separate figure. Added footnote on how we chose an overtly optimistically low energy that can be reached.</td></tr></table>
+
+<--- Page Split --->
+<|ref|>table<|/ref|><|det|>[[113, 88, 884, 901]]<|/det|>
+
+<table><tr><td>7. (Page 6) The authors aim to design<br>trajectories that avoid user-specific limits<br>associated with disruptions and so create<br>models that predict the evolution of<br>relevant quantities. For the readers to<br>understand better, it would be helpful to<br>explicitly define the key quantities<br>correlated with disruptions and the criteria<br>or limits that typically lead to disruptions.<br>Although Table 3 later in the paper provides<br>some indirect answers through the<br>structure of the reward function, it would be<br>better to introduce and explain these<br>disruption criteria earlier in the manuscript.<br>Fusion experts may already be familiar with<br>them, but for a multidisciplinary audience<br>like that of Nature Communications, such<br>clarifications are important.</td><td>Thank you for the feedback. We agree an<br>earlier discussion would be beneficial.<br>Changes Made: We have amended the end<br>of paragraph 1 in 2.1 to introduce the issue of<br>disruptive limits and the constraints we chose<br>to consider in this work.</td></tr><tr><td>8. (Tables 1 and 2) Tables 1 and 2 both<br>define input and output variables for the<br>deep learning models. However, in Table 1<br>(for NSSM), the inputs (actions) are at the<br>bottom and outputs (observations) are at<br>the top, while Table 2 has inputs at the top<br>and outputs at the bottom. This<br>inconsistency may confuse readers,<br>especially since in typical RL settings,<br>observations are inputs and actions are<br>outputs. Unless there is a reason for this<br>ordering, I recommend standardizing the<br>layout for clarity.</td><td>Thank you for the suggestion.<br>Changes Made: We have edited Table 1 to<br>include actions on top and observations on<br>the bottom.</td></tr><tr><td>9. (Page 6 and Figures 5 & 9) The profile<br>predictor appears to reconstruct 1D profiles<br>from already predicted 0D parameters,<br>rather than predicting the dynamics of the<br>profiles directly. This implies that the<br>information in the 1D profiles is already</td><td>You are correct. One of the findings of this<br>study is that details of 1D kinetic profiles did<br>not appear significant to achieve this result.<br>Even though the 1D profile predictor is not<br>necessary to obtain the core result, we<br>believe its inclusion in the paper contributes<br>to ongoing dialogue about the priority of<br>developing high-fidelity transport simulators</td></tr></table>
+
+<--- Page Split --->
+<|ref|>table<|/ref|><|det|>[[115, 88, 881, 910]]<|/det|>
+
+<table><tr><td>embedded in the 0D parameters. Therefore, using the reconstructed 1D profiles in NSSM or RL trajectory design might be unnecessary and only increase input complexity. In fact, as far as I understand, the profile predictor does not play a major role in the key process of this study. I'm wondering the role of this profile predictor regarding this study. If it is not essential to the main pipeline, I suggest removing it from the current paper and including it in a follow-up study.</td><td>for plasma control tasks, given the large volume of effort currently expended in that direction [1, 2, 3]. As we discuss in the Discussion section:<br>"The ability for a neural network to predict kinetic profiles using 0D scalars, demonstrated both in this work and in prior work [36], suggests a data-driven approach may be sufficient for certain control tasks without principles-based transport simulation, which can be extremely computationally expensive and require strong assumptions on edge temperature and density."<br>However, your comment helped us realize it will be worthwhile to make this message clear earlier in the paper as well.<br>Changes Made: added a sentence earlier in the paper (Section 2.2) for clarity of message: "This result also suggests a structured data-driven approach to modeling tokamak transport merits further research, in parallel with several ongoing principles-based efforts."<br>[1] Citrin, Jonathan, et al. "Torax: A fast and differentiable tokamak transport simulator in jax." arXiv preprint arXiv:2406.06718 (2024). [2] Muraca, Marco, et al. "Reduced transport models for a tokamak flight simulator." Plasma Physics and Controlled Fusion 65.3 (2023): 035007.<br>[3] Meneghini, O., et al. "FUSE (Fusion Synthesis Engine): A Next Generation Framework for Integrated Design of Fusion Pilot Plants." arXiv preprint arXiv:2409.05894 (2024).</td></tr><tr><td>10. (Figure 8) The black lines represent RL trajectories under uncertainty in the actuations. However, in practice, uncertainty (or noise) affects not only the level of the actions or trajectory but also introduces fluctuations in the actuations—like the red or magenta lines in the figure. However, the RL trajectories for a_minor, remain flat except for level</td><td>You are correct. This is a limitation of our uncertainty model.<br>Changes Made:<br>Included the following sentence at the end of the "Uncertainty Model" subsection:<br>"In addition, the uncertainty model employed does not account for time-varying fluctuations in uncertain variables; future work should employ time-varying stochastic processes.</td></tr></table>
+
+<--- Page Split --->
+<|ref|>table<|/ref|><|det|>[[115, 88, 883, 170]]<|/det|>
+
+<table><tr><td>differences. Doesn&#x27;t fluctuating action noise in real experiments affect the robustness of the designed RL trajectories?</td><td>Both of these limitations further highlight the need to advance experimental uncertainty quantification and robust control in the context of fusion plasma control.”</td></tr></table>
+
+<|ref|>title<|/ref|><|det|>[[115, 234, 242, 255]]<|/det|>
+# Reviews
+
+<|ref|>title<|/ref|><|det|>[[115, 293, 436, 306]]<|/det|>
+# Reviewer #1 (Remarks to the Author):
+
+<|ref|>text<|/ref|><|det|>[[114, 331, 875, 466]]<|/det|>
+The work revolves around the use of reinforcement learning (RL) for the optimization of the ramp-down trajectory at TCV, an experimental device in Switzerland. The idea of RL is that the model can learn how to interact with the system by trial and error and requires a sufficiently accurate model of the system. In this case, also the system is approximated by a Neural State Space Model (NSSM), that is a Neural Network which can approximate the state space representation mixing a data driven loss and numerical integration of the (net+system) dynamics.
+
+<|ref|>text<|/ref|><|det|>[[114, 472, 875, 526]]<|/det|>
+The work is interesting but due to the scarcity of experiments, the successfully designed rampdowns are only 5, out of a total of 9 optimized ones. For this reason, a further analysis and replication of the experiments is recommended for making robust conclusions.
+
+<|ref|>text<|/ref|><|det|>[[114, 532, 568, 546]]<|/det|>
+Moreover, I have the following comments on the work:
+
+<|ref|>text<|/ref|><|det|>[[114, 552, 535, 566]]<|/det|>
+NSSM and training environment for the RL system
+
+<|ref|>text<|/ref|><|det|>[[114, 572, 877, 625]]<|/det|>
+The article is quite dense, and the authors explain that one model (the NSSM) serves as the training environment for the RL model, however another neural model to predict the temperature and density profiles is mentioned as part of the NSSM. From section 2.2:
+
+<|ref|>text<|/ref|><|det|>[[114, 631, 877, 685]]<|/det|>
+“The NSSM was initially developed with a neural network predictor for the kinetic profiles on the full p grid, and initial training runs found that the profile predictor can accurately predict kinetic profiles given the set of OD scalars listed in Table 2.
+
+<|ref|>text<|/ref|><|det|>[[114, 691, 877, 866]]<|/det|>
+Figure 5 provides an example comparison of predictions of the Te and ne profiles against Thomson measurements for a full shot in the validation dataset, showing accurate prediction across all phases of the shot. This result corroborates previous findings at NSTX-U that neural networks can accurately predict kinetic profiles given a set of similar OD scalars34. Given this result suggests most of the relevant profile information is implicitly captured by OD scalars, the profile predictor was disabled prior to running experiments to accelerate training, hence reported predictions of kinetic profiles are not predict-first. A noteworthy feature of this profile predictor that should be explored in future work is its ability to function as a Thomson up-sampler, as the input variables are all sampled at a
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[113, 88, 812, 127]]<|/det|>
+higher time resolution, 1ms, than the TCV Thomson Scattering system, which takes measurements every 17ms."  
+
+<|ref|>text<|/ref|><|det|>[[114, 128, 844, 208]]<|/det|>
+If the predictor has not been adopted, I suggest removing the parts connected with the profile predictor (e.g section 4.1.3, Figures 5, 9) to avoid confusion, otherwise I would suggest indicating for how many discharges the predictor has been used to train the RL method and if there was any benefit from this approach.  
+
+<|ref|>text<|/ref|><|det|>[[114, 226, 840, 266]]<|/det|>
+How were the discharges to train the NSSM model selected? What are the variations in terms of the (observation, action) space?  
+
+<|ref|>text<|/ref|><|det|>[[114, 285, 874, 365]]<|/det|>
+Connected to the previous point, it seems that the vertical instability growth rate \((\gamma_{- }vgr)\) reconstruction by the NSSM is very poor, even with a very low prediction Horizon (Figure 4). Since this parameter is connected to vertical plasma stability, this would explain the VDE event in shot #81751. How large is the uncertainty assumed for the \(\gamma_{- }vgr\) ?  
+
+<|ref|>text<|/ref|><|det|>[[114, 383, 880, 444]]<|/det|>
+An interesting point could be made by running the control discharges in the NSSM and showing that they were close to disruptive boundaries/expected to disrupt. This would show, despite the low number of experiments, that the NSSM can well represent the system.  
+
+<|ref|>text<|/ref|><|det|>[[114, 461, 805, 501]]<|/det|>
+Is it possible to interpret the strategies adopted by the RL model by showing which parameters are kept far from critical boundaries?  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 521, 257, 539]]<|/det|>
+## Other comments  
+
+<|ref|>text<|/ref|><|det|>[[111, 558, 875, 699]]<|/det|>
+Some other works related to the use of RL in fusion are missing, as applications for ITER [1] and EAST [2]. Please add some details on the model adopted (number of layers, neurons) and on how they have been selected/optimized. I tried running the code and installing all the dependencies using poetry, but I could not manage to make it work. Most of the python notebooks seem to reproduce the figure and the main results, I did not find the scripts for training the models.  
+
+<|ref|>text<|/ref|><|det|>[[113, 735, 868, 875]]<|/det|>
+[1] S. Dubbioso, G. De Tommasi, A. Mele, G. Tartaglione, M. Ariola, and A. Pironti, 'A Deep Reinforcement Learning approach for Vertical Stabilization of tokamak plasmas', Fusion Engineering and Design, vol. 194, p. 113725, Sep. 2023, doi: 10.1016/j.fusengdes.2023.113725. [2] G. De Tommasi, S. Dubbioso, Y. Huang, Z. P. Luo, A. Mele, and B. J. Xiao, 'A RL- based Vertical Stabilization System for the EAST tokamak', in 2022 American Control Conference (ACC), Jun. 2022, pp. 5328- 5333. doi: 10.23919/ACC53348.2022.9867499.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 126, 480, 144]]<|/det|>
+Reviewer #1 (Remarks on code availability):  
+
+<|ref|>text<|/ref|><|det|>[[115, 164, 881, 222]]<|/det|>
+Partial review of the code due to the difficulties in installing all the proper libraries The readme file instructs on using poetry, but after finding the correct version of poetry I ran into an error:  
+
+<|ref|>text<|/ref|><|det|>[[115, 225, 880, 243]]<|/det|>
+Directory TokaGym\submodules\contrarx for contrarx does not seem to be a Python package  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 321, 434, 340]]<|/det|>
+## Reviewer #2 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[115, 359, 282, 377]]<|/det|>
+General Comments:  
+
+<|ref|>text<|/ref|><|det|>[[114, 397, 883, 597]]<|/det|>
+This study leverages deep learning (NSSM) to predict plasma ramp- down dynamics in the TCV tokamak and applies reinforcement learning (RL) to design an optimal trajectory for stable termination during this phase. Notably, ramp- down in fusion reactors is physically and technically prone to instability, often associated with plasma disruptions. Thus, ensuring stable control and termination in this phase is critical for future large- scale devices such as ITER. This study combines a predictive model and RL on TCV to design a stable plasma termination trajectory and demonstrates robust outcomes through incremental re- training with actual experiments. This represents a significant milestone in fusion research. I believe that, provided the following comments are adequately addressed, the manuscript could be suitable for publication in Nature Communications.  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 635, 266, 653]]<|/det|>
+## Major Comments:  
+
+<|ref|>text<|/ref|><|det|>[[114, 672, 878, 871]]<|/det|>
+1. While the significance and performance of the study are commendable, the storyline in the figures and descriptions is not very clear. Since Nature Communications is a multidisciplinary journal, the content should be understandable even to readers who are not fusion experts. From a non-expert's perspective, when reading the early part of the paper, the role of the proposed technique is not immediately intuitive (e.g., what constitutes a desired termination and how it differs from undesired outcomes). In particular, Figure 2E is the key illustration that visually demonstrates this difference, but it is not explained at all in the main text. I recommend moving Figure 2E to Figure 1 (right side of the original Fig 1) and introducing it in the Introduction or early in the Results section to clarify the role of the proposed method.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[113, 88, 883, 288]]<|/det|>
+2. (Figures 6-8) These figures show only the actual plasma trajectories, but not the RL-designed or the manually programmed (planned) trajectories. However, the RL-designed trajectory, the manually programmed trajectory in the PCS, and the actual plasma response may all differ. For example, the authors state in Figure 6 that the trajectory was optimized to reduce the LFS gap to prevent VDEs, but from the figure alone, it is unclear whether this was truly suggested by the RL agent, manually reflected in PCS programming, or just occurred randomly regardless of the programmed trajectory. This distinction is crucial for interpreting and evaluating the value of the research. I recommend that the RL-designed trajectory and the PCS-programmed trajectory be also added to Figures 6 and 7 (e.g., using light-colored or dashed lines).  
+
+<|ref|>text<|/ref|><|det|>[[114, 289, 877, 368]]<|/det|>
+In Figure 8, although multiple RL trajectories under uncertainty are shown as black lines, the actual programmed trajectory used in the experiment is not shown. While the caption mentions that the average trajectory was used, it would be helpful to indicate this explicitly in the figure as well.  
+
+<|ref|>text<|/ref|><|det|>[[114, 386, 876, 527]]<|/det|>
+3. (Page 24) Unlike general RL settings, the RL agent in this study observes only time (Equation 12). This implies that once trained, the agent will always produce the same trajectory, regardless of the initial condition or change of plasma states. Thus, the first question is: when computing the trajectory to input into TCV as shown in Figure 8, how was random uncertainty introduced (like just normal random added to the given trajectory)? Second, if the initial conditions or plasma current differ, does the RL agent need to be re-trained from scratch each time?  
+
+<|ref|>sub_title<|/ref|><|det|>[[115, 565, 266, 582]]<|/det|>
+## Minor Comments:  
+
+<|ref|>text<|/ref|><|det|>[[114, 602, 880, 763]]<|/det|>
+4. (Page 2) In the Introduction, the authors emphasize that this study focuses on feedforward trajectory design before the experiment, in contrast to previous RL-based feedback control studies. However, some previous works have also used deep learning-based predictive models in combination with RL to design operation trajectories in a feedforward manner (e.g., [J Seo et al., Nucl. Fusion 61 106010 (2021) & J Seo et al., Nucl. Fusion 62 086049 (2022)]). Although these earlier works may be more limited and less robust to uncertainty compared to the current study, they are similar in concept and deserve to be mentioned as related study.  
+
+<|ref|>text<|/ref|><|det|>[[115, 781, 865, 840]]<|/det|>
+5. (Figure 2A and others) In Figure 2A, the markers in the legend are difficult to distinguish from those in the graph itself. I suggest modifying the legend (like using boxed legends or other formatting) for easier visual distinction, including in other similar figures.  
+
+<|ref|>text<|/ref|><|det|>[[115, 859, 833, 899]]<|/det|>
+6. (Figure 2D) The background colors in this graph are not intuitive. At first glance, one might assume that green indicates successful terminations and red indicates failures.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[114, 88, 870, 228]]<|/det|>
+However, shot 81635, which was successfully terminated, is marked red, while shot 81741, which had a legacy software issue, is shown in green. Additionally, the x- axis does not follow a monotonic shot order (e.g., 81635 appears after 81741), and there is no explanation for the ambiguous non- red/non- green background colors. It would also be helpful to indicate the threshold values for successful termination (W_tot and I_p criteria) directly on the graph. Similar color ambiguity is present in Figure 3 (e.g., mixed green and blue in W_tot panel), but without explanation.  
+
+<|ref|>text<|/ref|><|det|>[[113, 247, 878, 427]]<|/det|>
+7. (Page 6) The authors aim to design trajectories that avoid user-specific limits associated with disruptions and so create models that predict the evolution of relevant quantities. For the readers to understand better, it would be helpful to explicitly define the key quantities correlated with disruptions and the criteria or limits that typically lead to disruptions. Although Table 3 later in the paper provides some indirect answers through the structure of the reward function, it would be better to introduce and explain these disruption criteria earlier in the manuscript. Fusion experts may already be familiar with them, but for a multidisciplinary audience like that of Nature Communications, such clarifications are important.  
+
+<|ref|>text<|/ref|><|det|>[[114, 446, 878, 565]]<|/det|>
+8. (Tables 1 and 2) Tables 1 and 2 both define input and output variables for the deep learning models. However, in Table 1 (for NSSM), the inputs (actions) are at the bottom and outputs (observations) are at the top, while Table 2 has inputs at the top and outputs at the bottom. This inconsistency may confuse readers, especially since in typical RL settings, observations are inputs and actions are outputs. Unless there is a reason for this ordering, I recommend standardizing the layout for clarity.  
+
+<|ref|>text<|/ref|><|det|>[[114, 584, 883, 744]]<|/det|>
+9. (Page 6 and Figures 5 & 9) The profile predictor appears to reconstruct 1D profiles from already predicted 0D parameters, rather than predicting the dynamics of the profiles directly. This implies that the information in the 1D profiles is already embedded in the 0D parameters. Therefore, using the reconstructed 1D profiles in NSSM or RL trajectory design might be unnecessary and only increase input complexity. In fact, as far as I understand, the profile predictor does not play a major role in the key process of this study. I'm wondering the role of this profile predictor regarding this study. If it is not essential to the main pipeline, I suggest removing it from the current paper and including it in a follow-up study.  
+
+<|ref|>text<|/ref|><|det|>[[114, 762, 880, 881]]<|/det|>
+10. (Figure 8) The black lines represent RL trajectories under uncertainty in the actuations. However, in practice, uncertainty (or noise) affects not only the level of the actions or trajectory but also introduces fluctuations in the actuations—like the red or magenta lines in the figure. However, the RL trajectories for a_minor, remain flat except for level differences. Doesn't fluctuating action noise in real experiments affect the robustness of the designed RL trajectories?
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[0, 0, 997, 997]]<|/det|>
+# 1. 1. 1. 1. 1. 1. 1. 2. 2. 2. 2. 2. 2. 2.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 90, 857, 160]]<|/det|>
+Reviewer Comment: 1) In the newly added text at page 4 the authors claim: "At present, a comprehensive understanding of disruptive limits remains an open problem, motivating many works on machine- learning based prediction of disruptions 37- 40."  
+
+<|ref|>text<|/ref|><|det|>[[115, 160, 870, 212]]<|/det|>
+Despite being valid, this statement is supported by many self- citing references. Further/different works could be referenced since ML- based prediction of disruptions is a well developed line of research in fusion.  
+
+<|ref|>text<|/ref|><|det|>[[115, 229, 881, 334]]<|/det|>
+Response: Thank you for your suggestion. We would first like to clarify that only two out of the four cited works could be fairly characterized as self- citation. 38 is an overview paper for the international project ITER, involving a large portion of the world's experts, which inevitably leads to involvement of authors of this paper as well. 40 does not involve any authors of this paper whatsoever. To further address the balance, we have added a reference to Vega et al., 2022 Nature Physics.  
+
+<|ref|>text<|/ref|><|det|>[[115, 350, 880, 492]]<|/det|>
+Reviewer Comment: 2) Despite the statement of the authors, I consider the role of the profile predictor is still marginal in the description of the work, and recommend showing the plots discussing the 1D profile reconstruction either in the supplementary data or in a new work. On the other hand, for the purposes of the work, I find more interesting the discussion on Figures 16 and 20, now in the supplementary data. These figures allow to discuss the validity of the NSSM for the current application. For instance, in the two successful pulses of Figure 16, the yvgr is close to the same limit value as in the one of Figure 20. How is the distinction between these two sets of pulses made?  
+
+<|ref|>text<|/ref|><|det|>[[115, 508, 881, 561]]<|/det|>
+Response: Thank you for your perspective. Regarding the profile prediction content, we would like to re- iterate our opinion that the result is informative for the community given the multiple current research efforts to do this with a principles- based approach [1, 2, 3].  
+
+<|ref|>text<|/ref|><|det|>[[115, 577, 840, 648]]<|/det|>
+Regarding Figures 16 and 20, we would like to highlight that they are new figures introduced as a part of the revision. We would like to point out that in Figure 20, the distribution travels much further into the soft constraint region than in Figure 16, especially in the milliseconds immediately preceding the disruption.  
+
+<|ref|>text<|/ref|><|det|>[[115, 682, 875, 802]]<|/det|>
+[1] Teplukhina, A. A., et al. "Simulation of profile evolution from ramp- up to ramp- down and optimization of tokamak plasma termination with the RAPTOR code." Plasma Physics and Controlled Fusion 59.12 (2017): 124004. [2] Citrin, Jonathan, et al. "TORAX: A fast and differentiable tokamak transport simulator in JAX." arXiv preprint arXiv:2406.06718 (2024). [3] Meneghini, O., et al. "FUSE (Fusion Synthesis Engine): A next generation framework for integrated design of fusion pilot plants." arXiv preprint arXiv:2409.05894 (2024).
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__7206689b0b629e8d0089cf12c07ed61d70499cbe7a6de498450a7cc3e71a5daa/images_list.json

@@ -0,0 +1,17 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Fig. 1 The distribution of differentially methylated regions (DMRs) for Tree 109 and Tree 171.",
+    "footnote": [],
+    "bbox": [
+      [
+        216,
+        230,
+        780,
+        619
+      ]
+    ],
+    "page_idx": 9
+  }
+]

peer_reviews/supplementary_0_Transparent Peer Review file__7206689b0b629e8d0089cf12c07ed61d70499cbe7a6de498450a7cc3e71a5daa/supplementary_0_Transparent Peer Review file__7206689b0b629e8d0089cf12c07ed61d70499cbe7a6de498450a7cc3e71a5daa.mmd

@@ -0,0 +1,334 @@
+
+# nature portfolio  
+
+Peer Review File  
+
+## Accelerated growth increases the somatic epimutation rate in trees  
+
+Corresponding Author: Professor Frank Johannes  
+
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+Version 0:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author) In this manuscript the authors investigate the relationship between growth rate and somatic epimutation rates in trees, using European beech (Fagus sylvatica) as a model. The researchers leveraged a 150- year- old thinning experiment to compare trees subjected to different thinning intensities, which resulted in varying growth rates. They found that thinning- induced growth acceleration was accompanied by increased cell division rates in the main stems, as confirmed by cell count assays. Trees with higher growth rates exhibited significantly higher somatic epimutation rates in both main stems and lateral branches. Although the results are interesting, the analysis of epimutation rates is based on a small number of trees (two representative trees for detailed analysis), which may limit the generalizability of the findings.  
+
+Here are several concerns and suggestions:  
+
+1. Line 99-111, Please provide a figure to show the difference in stem growth rates, diameter breast height, and crown size between the treatments at a population level.  
+
+2. How representative are the two selected trees of their respective treatments? Do they have a close genetic relationship or genotypes?  
+
+3. Line 136-136, are they significantly different?  
+
+4. Fig2. What's the full name of DMP? How do you calculate the CG/CHG divergence? How about CHH divergence?  
+
+5. Fig2B, the variation from R1 vs. R2 is too big, and can't tell which dots are from which treatment. Color coding them? Since the variation between neighboring replicate samples is big, not sure one sample from the polar opposite side is enough.  
+
+6. Fig. How do you construct the topology tree?  
+
+7. What distribution of these gained or lost epimutations (or DMR) across the genome? Are they randomly mutated? What kind of genes were under epimutations? Do these mutations affect gene expression levels? Are these epimutations neutral or deleterious?  
+
+8. Could differences in stress responses (due to thinning) independently affect methylation? Other factors, like local environment, soil, or the genotypic background of the tree itself may also affect methylation.  
+
+9. Mean coverage was \(\sim 20X\) was this sufficient for low-frequency epimutation detection?  
+
+10. Does the accelerated growth affect the genomic mutation rate?  
+
+11. Missing title in Table S2
+
+<--- Page Split --->
+
+(Remarks to the Author)  
+
+The study by Zhou et al. explored the impact of tree growth rates on epimutation frequencies. To this end, the authors took advantage of an experimental plot where beech trees have been grown for decades under distinct thinning strategies. Trees in heavily thinned regions grow faster compared to those in less thinned areas. The authors exploited this difference to study epimutations (changes in DNA methylation) associated with accelerated growth.  
+
+First, the authors demonstrated that faster growth is associated with increased cell division (and also with increased cell expansion). They then measured global methylation divergence between faster growing and "control" trees grown in less thinned areas by comparing vascular tissues from opposite sides of the trunk. The analysis revealed that faster- growing trees exhibit higher levels of methylation divergence, specifically in the CG context, but not in the CHG context.  
+
+Subsequently, the authors performed an intra- organismal experiment in which methylation divergence was calculated as a function of divergence over time. For this, they compared methylomes of distinct leaves whose positions in the tree could be tracked and also determined the ages of the branches on which these leaves were located. The results indicated that leaves from the faster- growing tree showed increased methylation divergence, supporting their previous finding when comparing vascular tissues.  
+
+This is a very interesting manuscript using a unique genetic material that allowed to report a clear positive correlation between plant growth rate and epimutation rate. While there is no direct evidence proving that increased cell division is responsible for the elevated epimutation rate, these findings align well with the previously proposed "mitotic- rate hypothesis".  
+
+Comments/Suggestions:  
+
+- An interesting point is that cells in the faster-growing tree are also larger. Could this phenotype contribute to the increase in epimutation rate? Perhaps increased metabolic during extended cell growth activity also alters the deposition/removal of DNA methylation. This question is difficult to address with the current experimental model but perhaps could be added to the discussion if the authors think it could be possible.  
+
+- Fig. 2B, C: Please provide a more detailed description in the Methods section of how global mC divergence was calculated. Also, indicate in the figure legend where the "n" values come from, and why certain dots are shown in red.  
+
+- Fig. 2B, C: Please indicate the statistical significance of the differences between measurements in the plots. 
+- Fig. 2C: Were the CG and CHG divergence values calculated across all C positions in the genome? Please describe this more clearly. Additionally, what is the rationale for showing CHG but not CHH methylation for non-CG contexts? Would including CHH methylation offer further insights or reinforce the CHG findings?  
+
+- Related to Fig. 2: The authors state that "methylation divergence between samples from opposite sides of the stems was 2.64-fold higher in trees." Please clarify how this number was calculated.  
+
+- The authors present global CG divergence between trees and leaves. Are these loss and gains enriched in specific genomic regions or associated with genes with particular features (e.g., gene length, intron number,...), or are the epimutations randomly distributed across the coding genome?  
+
+- Fig. 3C: Please specify which datasets were used to generate this plot.  
+
+- The authors mention: "Removing the low-quality sequencing samples, we had 18 cambium samples and 17 leaf samples." Please clarify which specific samples were excluded from the analyses.  
+
+- Supplementary Table 1: Please include information indicating which sample IDs correspond to the CB and CC samples. 
+- On page 5, please correct the typo: "METHYLSTRANSFERASE" should be "METHYLTRANSFERASE."  
+
+Version 1:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author)  
+
+I appreciate that the authors addressed most of my questions, but I still have some concerns, see below 1. Fig. 2b.c Please provide P- values and statistical method. What do the red dots indicate?  
+
+2. For my question 7 Can you make a figure to show the distribution of these DMRs?  
+
+The authors addressed my questions by citing some papers. Is there any study work on the same species as you did here? If not, it's unclear whether you can draw any conclusions.  
+
+3. For my question 9 In the citations, 20X coverage was used in Arabidopsis, not sure whether it's suitable for European beech.  
+
+I didn't find the paper by Shahyary et al. 2020 in the Reference
+
+<--- Page Split --->
+
+## Reviewer #2  
+
+(Remarks to the Author)  
+
+I would like to thank the authors for their comments and the corrections they provided. I am satisfied with their responses and the revised version. I have two comments regarding the revised manuscript that could be considered:  
+
+- Line 421: Please consider revising the sentence: "To more accurately quantify different thinning intensities induced methylation divergence, ..."  
+
+- Thank you for reporting the genomic distribution of CG divergence in both trees in Fig. 1 of the rebuttal. I wonder why this data was not included in the manuscript, considering both reviewers asked the same question. It is up to the authors to decide whether to include it, but I think it is an informative piece of data. If yes, please add to the legend that these are CG-context DMRs.  
+
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source. The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.  
+
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+
+## REVIEWER COMMENTS  
+
+## Reviewer #1 (Remarks to the Author):  
+
+In this manuscript the authors investigate the relationship between growth rate and somatic epimutation rates in trees, using European beech (Fagus sylvatica) as a model. The researchers leveraged a 150- year- old thinning experiment to compare trees subjected to different thinning intensities, which resulted in varying growth rates. They found that thinning- induced growth acceleration was accompanied by increased cell division rates in the main stems, as confirmed by cell count assays. Trees with higher growth rates exhibited significantly higher somatic epimutation rates in both main stems and lateral branches. Although the results are interesting, the analysis of epimutation rates is based on a small number of trees (two representative trees for detailed analysis), which may limit the generalizability of the findings.  
+
+Here are several concerns and suggestions:  
+
+1. Line 99-111, Please provide a figure to show the difference in stem growth rates, diameter breast height, and crown size between the treatments at a population level. We thank the reviewer for this suggestion. This analysis is now shown in the new Supplementary Fig. 1.  
+
+2. How representative are the two selected trees of their respective treatments? Do they have a close genetic relationship or genotypes?  
+
+The two selected trees, Tree 109 and Tree 171, are representative individuals of their respective treatments. As shown in Fig. 1c and in Supplementary Fig. 1, their growth trajectories, tree dimensions and other characteristics fall within the intermediate range of trees in their respective plots, indicating they reflect typical patterns for their thinning treatment. Furthermore, all trees within the FAB 15 experiment originated from natural regeneration following shelterwood cuttings in 1822. This silvicultural practice promotes regeneration from a common seed source under similar ecological conditions. While the exact genotypes of Tree 109 and Tree 171 have not yet been determined, there is no indication of artificial selection or planting, and all trees in the plots are presumed to stem from the same local gene pool. Therefore, although we cannot confirm a close genetic relationship between the two selected trees, it is reasonable to assume they are representative members of a naturally regenerated, locally adapted beech population subjected to the respective long-term thinning treatments.  
+
+3. Line 136-136, are they significantly different?  
+
+We observed a significantly greater CG methylation divergence in the heavily thinned plot compared to the moderately thinned plot, as assessed by a t- test (p- value \(< 0.05\) ). We have included the statistical significance of the differences in the plots of Fig. 2b and c. Although not all comparisons reached statistical significance, the relatively large Cohen's \(D\) effect sizes support an increasing trend from RR to moderately thinned treatment to heavily thinned treatment, which supports our central hypothesis. Please refer to updated Fig. 2 for details.  
+
+4. Fig2. What's the full name of DMP? How do you calculate the CG/CHG divergence? How about CHH divergence?  
+
+DMP stands for Differentially Methylated Position. We are now defining this term in the Method section. To calculate CG and CHG divergence, we took the following steps: For each tree disk, we calculated the methylation difference between two replicates located on the same side (R1 vs. R2), as well as the methylation difference between each replicate and the sample from the opposite side (R1 vs. S, R2 vs. S). To obtain a
+
+<--- Page Split --->
+
+measure of genome- wide methylation difference between samples, we followed our previously published approach (see e.g. van der Graaf et al. 2015, Hazarika et al. 2022, Shahryary et al. 2020). For each cytosine site \(i\) , we calculated the methylation difference between the two samples (e.g., R1 and S): \(D_{i} = |M(R1,i) - M(S,i)|\) , and then took the average over all covered cytosines in the genome:  
+
+\[D = \frac{1}{n}\sum_{i = 1}^{n}D_{i}\]  
+
+We thank the reviewer for bringing up this question. We are now detailing this approach in the Methods section.  
+
+5. Fig2B, the variation from R1 vs. R2 is too big, and can't tell which dots are from which treatment. Color coding them? Since the variation between neighboring replicate samples is big, not sure one sample from the polar opposite side is enough.  
+
+We appreciate the reviewer's suggestion. Indeed, the variation between neighboring replicate samples (e.g., R1 vs. R2) appears relatively large. However, when looking at the mean divergence differences, there is a clear trend showing that R1 vs. R2 (RR) are, on average, much less divergent. Although the small sample sizes may limit statistical significance, the relatively large Cohen's \(D\) effect sizes support an increasing trend from RR to moderately thinned treatment to heavily thinned treatment. Please refer to updated Fig. 2 for details.  
+
+6. Fig. How do you construct the topology tree?  
+
+To construct the tree's topological structure, we first divided the fallen tree into segments and systematically marked each branch with identifiers. These markings enabled precise tracking and mapping of the branches, allowing us to accurately reconstruct the tree's topology. This method ensured that we could reliably determine the spatial relationships between branches, ensuring an accurate representation of the tree's structure.  
+
+7. What distribution of these gained or lost epimutations (or DMR) across the genome? Are they randomly mutated? What kind of genes were under epimutations? Do these mutations affect gene expression levels? Are these epimutations neutral or deleterious? Overall, Tree 171 exhibits a higher number of CG-context DMRs compared to Tree 109, yet both share a similar distribution pattern across genomic regions (Fig. 1). The majority of CG-DMRs are located in non-coding regions (intergenic) and transposable elements (TEs), which is a typical pattern in plants (Schmitz et al. 2013; Niederhuth et al. 2016; Zhang et al. 2018). A substantial number of DMRs are found in CDS regions. Intronic regions also contain many DMRs. In contrast, DMRs in exonic regions are relatively sparse. Regarding the potential impact of the identified epimutations on gene expression levels, although we did not have gene expression data (such as RNA-seq) in this study, prior work has shown that gene body methylation (gbM) changes, which constitute the majority of somatic epimutations, typically have little to no effect on gene expression levels (Zhang et al. 2018; Bewick et al. 2019; Hofmeister et al. 2020). This provides strong support that the epimutations detected here are largely functionally neutral, and it ensures that the accumulation of epimutations is associated with cell divisions rather than driven by natural (i.e. cell lineage) selection.  
+
+8. Could differences in stress responses (due to thinning) independently affect methylation? Other factors, like local environment, soil, or the genotypic background of the tree itself may also affect methylation.  
+
+We thank the reviewer for raising this important point. Our experimental design aimed to minimize the influence of these variables. All three 0.4 ha plots lie in the immediate vicinity to each other on loamic Luvisols, displaying very similar soil characteristics. All
+
+<--- Page Split --->
+
+sampled trees were selected randomly within the different treatments, avoiding a systematic bias. While localized microclimatic variation cannot be entirely ruled out, particularly given that higher thinning intensity reduces standing stock and may slightly alter environmental conditions, such differences are considered to be a direct consequence of the treatment itself rather than an independent confounding factor. Finally, all trees originated from natural regeneration following a shelterwood cut, so they derive from the same locally adapted gene pool; there was no artificial planting or selection of distinct genotypes in any plot. Although we have not yet performed molecular genotyping on Trees 109 and 171, their shared origin and randomized selection make it reasonable to assume no systematic genotypic differences among treatments. Thus, based on our randomized selection and the uniform soil conditions, we minimize other factors that might affect methylation. Moreover, the "thinning" itself only involved clearing dead wood or dying trees from the plots, and is not expected to impose stress on the surrounding trees. Nonetheless, we appreciate the reviewer's thoughts here: The correlative link between increased rates of growth, cell division and epimutations should be further investigated in future studies to pin point the precise causal relationship. However, the careful observation that such a correlative link exists is an important first step that our study has taken.  
+
+9. Mean coverage was \(\sim 20X\) was this sufficient for low-frequency epimutation detection?  
+
+From our experience with similar experimental setups (Becker et al. 2011; Graaf et al. 2015; Schmitz et al. 2022), a \(20\times\) coverage is sufficient for the detection of most epimutations. Additionally, our method for estimating rates is relatively robust to ascertainment error as we explicitly model the error in the estimation procedure (see Shahryary et al. 2020).  
+
+10. Does the accelerated growth affect the genomic mutation rate?  
+
+We did not assess genomic mutation rates in this study, as our focus was specifically on epimutations. Evaluating whether accelerated growth also affects the genomic mutation rate would require whole-genome resequencing data, which was beyond the scope of our current work. However, having shown that epimutation rates are closely linked to cell division rates, we hypothesize that accelerated growth may influence genomic mutation rates similarly. Recent studies have demonstrated that both somatic mutation rates and epimutation rates scale with maturation age in perennial plants such as trees (Hanlon et al. 2019; Johannes 2025), suggesting a shared underlying biological mechanism likely associated with the division rates of shoot apical meristem (SAM) cells.  
+
+We thank the reviewer for bringing up this point. We have now included a brief discussion of this hypothesis in the discussion section.  
+
+11. Missing title in Table S2Thanks for pointing out the missing title in Table S2. We have now added a title to the table, and the revised version is included in the updated files.  
+
+## Reviewer #2 (Remarks to the Author):  
+
+The study by Zhou et al. explored the impact of tree growth rates on epimutation frequencies. To this end, the authors took advantage of an experimental plot where beech trees have been grown for decades under distinct thinning strategies. Trees in heavily thinned regions grow faster compared to those in less thinned areas. The authors exploited this difference to study epimutations (changes in DNA methylation) associated with accelerated growth.
+
+<--- Page Split --->
+
+First, the authors demonstrated that faster growth is associated with increased cell division (and also with increased cell expansion). They then measured global methylation divergence between faster growing and "control" trees grown in less thinned areas by comparing vascular tissues from opposite sides of the trunk. The analysis revealed that faster- growing trees exhibit higher levels of methylation divergence, specifically in the CG context, but not in the CHG context.  
+
+Subsequently, the authors performed an intra- organismal experiment in which methylation divergence was calculated as a function of divergence over time. For this, they compared methylomes of distinct leaves whose positions in the tree could be tracked and also determined the ages of the branches on which these leaves were located. The results indicated that leaves from the faster- growing tree showed increased methylation divergence, supporting their previous finding when comparing vascular tissues.  
+
+This is a very interesting manuscript using a unique genetic material that allowed to report a clear positive correlation between plant growth rate and epimutation rate. While there is no direct evidence proving that increased cell division is responsible for the elevated epimutation rate, these findings align well with the previously proposed "mitotic- rate hypothesis".  
+
+We thank this reviewer for the positive assessment of our work.  
+
+Comments/Suggestions:  
+
+- An interesting point is that cells in the faster-growing tree are also larger. Could this phenotype contribute to the increase in epimutation rate? Perhaps increased metabolic during extended cell growth activity also alters the deposition/removal of DNA methylation. This question is difficult to address with the current experimental model but perhaps could be added to the discussion if the authors think it could be possible. The potential link between cell size, metabolic activity, and epimutation rate is indeed an interesting hypothesis. This discussion aligns well with the recent observation that both somatic mutation and epimutation rates show a powerlaw scaling with maturation age in trees (Johannes 2025), suggesting that there is an allometric link between life history traits and (epi)genomic maintenance fidelity. It is therefore of theoretical interest to explore how the scaling connects with the allometric laws that were previously found to link cellular metabolic rate, body size, and growth (West et al. 2002). Our current study cannot dissect these links, but our observations are consistent with this hypothesis and highlight a potential direction for future work.  
+
+We thank the reviewer for the insightful suggestion. We have now included a brief discussion of this hypothesis in the discussion section.  
+
+- Fig. 2B, C: Please provide a more detailed description in the Methods section of how global mC divergence was calculated. Also, indicate in the figure legend where the "n" values come from, and why certain dots are shown in red.  
+
+Thanks for the suggestion. We have revised the manuscript accordingly. A more detailed description has been added to the Methods section. Additionally, we have updated the figure legend to indicate the source of the "n" values and to explain why certain dots are shown in red.  
+
+- Fig. 2B, C: Please indicate the statistical significance of the differences between measurements in the plots.  
+
+Thanks, we have included the statistical significance of the differences in the plots of
+
+<--- Page Split --->
+
+Fig. 2b, c.  
+
+- Fig. 2C: Were the CG and CHG divergence values calculated across all C positions in the genome? Please describe this more clearly. Additionally, what is the rationale for showing CHG but not CHH methylation for non-CG contexts? Would including CHH methylation offer further insights or reinforce the CHG findings?  
+
+We thank the reviewer for this helpful comment. The CG and CHG methylation divergence values in Fig. 2c were calculated genome-wide, based on all covered (posteriorMax \(\geq 0.99\) ) cytosines in the respective sequence contexts. We have clarified this in the Methods section of the revised manuscript. Consistent with previous work both in somatic systems (Hofmeister et al. 2020; Ibañez et al. 2023) and transgenerational studies (Quadrana and Colot 2016; Zheng et al. 2017; Li et al. 2020), we found no time-dependent accumulation in CHH epimutations. This is likely due to the fact that the methylation in CHH context is mainly targeted by de novo methylation pathways that prevent the accumulation of stable gains and losses over generations (Johannes and Schmitz, 2019).  
+
+- Related to Fig. 2: The authors state that "methylation divergence between samples from opposite sides of the stems was 2.64-fold higher in trees." Please clarify how this number was calculated.  
+
+Thanks for the question. We have added more detailed explanations in the Methods section. As shown in Fig. 2b, the first box represents methylation divergence between replicate samples on the same side of the stem (R1 vs. R2). The second and third boxes represent divergence between samples from opposite sides of the stem under moderately and heavily thinned treatments, respectively.  
+
+To more accurately quantify different thinning intensities induced methylation divergence, we normalized the opposite- side divergence by subtracting the baseline variation observed between same- side replicates. In other words, the methylation difference between two sides of the stem is defined as the "opposite sides methylation difference" minus the "same- side replicates methylation difference". We then calculated the fold change between treatments based on these adjusted values. Using the mean values shown in Fig. 2b: \(0.005221033\) (replicates), \(0.005485005\) (moderately), and \(0.005919238\) (heavily), the fold change was calculated as: \((0.005919238 - 0.005221033) / (0.005485005 - 0.005221033) = 2.64\) .  
+
+Based on this, we stated that "methylation divergence between samples from opposite sides of the stems was 2.64- fold higher in trees from the heavily thinned plot compared to trees from the moderately thinned plot."  
+
+- The authors present global CG divergence between trees and leaves. Are these loss and gains enriched in specific genomic regions or associated with genes with particular features (e.g., gene length, intron number,...), or are the epimutations randomly distributed across the coding genome?  
+
+Overall, Tree 171 exhibits a higher number of CG- context DMRs compared to Tree 109. Yet, both share a similar distribution pattern across genomic regions (Fig. 1). The majority of CG- DMRs are located in non- coding regions (intergenic) and transposable elements (TEs), which is a typical pattern in plants (Schmitz et al. 2013; Niederhuth et al. 2016; Zhang et al. 2018). A substantial number of DMRs are found in CDS regions. Intronic regions also contain many DMRs. In contrast, DMRs in exonic regions are relatively sparse.  
+
+- Fig. 3C: Please specify which datasets were used to generate this plot. Thanks for the comment. This information has now been clarified in the revised figure legend.  
+
+- The authors mention: "Removing the low-quality sequencing samples, we had 18
+
+<--- Page Split --->
+
+cambium samples and 17 leaf samples." Please clarify which specific samples were excluded from the analyses.  
+
+This was a misstatement; it should refer to samples that failed library construction. We have corrected this.  
+
+- Supplementary Table 1: Please include information indicating which sample IDs correspond to the CB and CC samples.  
+
+We thank the reviewer for pointing this out. We have now added information to Supplementary Data 1.  
+
+- On page 5, please correct the typo: "METHYLSTRANSFERASE" should be "METHYLTRANSFERASE."  
+
+Thanks for the correction. We have corrected "METHYLSTRANSFERASE" to "METHYLTRANSFERASE" in the revised manuscript.  
+
+## Reference  
+
+1. Schmitz, R., Schultz, M., Urich, M. et al. Patterns of population epigenomic diversity. Nature 495, 193-198 (2013).  
+
+2. Niederhuth, C. E., Bewick, A. J., Ji, L. et al. Widespread natural variation of DNA methylation within angiosperms. Genome Biol. 17, 194 (2016).  
+
+3. Zhang, H., Lang, Z. & Zhu, J. K. Dynamics and function of DNA methylation in plants. Nat. Rev. Mol. Cell Biol. 19, 489-506 (2018).  
+
+4. Bewick, A. J., Zhang, Y., Wendte, J. M., Zhang, X. & Schmitz, R. J. Evolutionary and experimental loss of gene body methylation and its consequence to gene expression. G3 Genes|Genomes|Genetics 9, 2441-2445 (2019).  
+
+5. Hofmeister, B. T., Denkena, J., Colomé-Tatché, M. et al. A genome assembly and the somatic genetic and epigenetic mutation rate in a wild long-lived perennial Populus trichocarpa. Genome Biol. 21, 259 (2020).  
+
+6. Becker, C., Hagmann, J., Müller, J. et al. Spontaneous epigenetic variation in the Arabidopsis thaliana methylome. Nature 480, 245-249 (2011).  
+
+7. van der Graaf, A., Wardenaar, R., Neumann, D. A. et al. Rate, spectrum, and evolutionary dynamics of spontaneous epimutations. Proc. Natl Acad. Sci. USA 112, 6676-6681 (2015).  
+
+8. Schmitz, R. J., Marand, A. P., Zhang, X. et al. Quality control and evaluation of plant epigenomics data. Plant Cell 34, 503-513 (2022).  
+
+9. Hanlon, V. C. T., Otto, S. P. & Aitken, S. N. Somatic mutations substantially increase the per-generation mutation rate in the conifer Picea sitchensis. Evol. Lett. 3, 348-358 (2019).  
+
+10. Johannes, F. Allometric scaling of somatic mutation and epimutation rates in trees. Evolution 79, 1-5 (2025).  
+
+11. West, G. B., Woodruff, W. H. & Brown, J. H. Allometric scaling of metabolic rate from molecules and mitochondria to cells and mammals. Proc. Natl Acad. Sci. USA 99 (Suppl 1), 2473-2478 (2002).  
+
+12. Noé Ibañez, V., van Antro, M., Peña-Ponton, C., Milanovic-Ivanovic, S., Wagemaker, C. A. M., Gawehns, F. & Verhoeven, K. J. F. Environmental and genealogical effects on DNA methylation in a widespread apomictic dandelion lineage. J. Evol. Biol. 36, 663-674 (2023).  
+
+13. Quadrana, L. & Colot, V. Plant transgenerational epigenetics. Annu. Rev. Genet. 50, 467-491 (2016).  
+
+14. Zheng, X., Chen, L., Xia, H. et al. Transgenerational epimutations induced by multi-generation drought imposition mediate rice plant's adaptation to drought condition. Sci. Rep. 7, 39843 (2017).
+
+<--- Page Split --->
+
+15. Li, J., Yang, D. L., Huang, H. et al. Epigenetic memory marks determine epiallele stability at loci targeted by de novo DNA methylation. Nat. Plants 6, 661–674 (2020). 
+
+16. Johannes, F. & Schmitz, R. J. Spontaneous epimutations in plants. New Phytol. 221, 1253–1259 (2019). 
+
+![](images/Figure_1.jpg)
+
+ 
+
+<center>Fig. 1 The distribution of differentially methylated regions (DMRs) for Tree 109 and Tree 171.</center>
+
+<--- Page Split --->
+
+## REVIEWER COMMENTS - updated  
+
+## Reviewer #1 (Remarks to the Author):  
+
+I appreciate that the authors addressed most of my questions, but I still have some concerns, see belowWe would like to thank the reviewer for insightful suggestions, which have greatly improved the quality and completeness of our manuscript.  
+
+1. Fig. 2b.c Please provide P-values and statistical method. What do the red dots indicate? We thank the reviewer for this very helpful suggestion. We have updated Figure 2b, c to include the exact P-value for the significant result, and provided the statistical method in the legend, and full statistical results, including non-significant differences, are provided in Supplementary Data 5. The red dots deviate from the overall data distribution and may be potential outliers; however, they do not affect the overall results or conclusions, so they are retained in the figure.  
+
+## 2. For my question 7  
+
+Can you make a figure to show the distribution of these DMRs?  
+
+The authors addressed my questions by citing some papers. Is there any study work on the same species as you did here? If not, it's unclear whether you can draw any conclusions.  
+
+We thank the reviewer for the follow- up question. To our knowledge, there is currently no published study on CG- DMR distribution in European beech. Nevertheless, our analysis provides a description of somatic CG- DMRs in this species, offering valuable insight into the genomic distribution of epimutations. While direct comparisons with previous work in the same species are not possible, the observed patterns are consistent with general trends reported in other plants (as cited in our initial response), supporting the robustness and interpretability of our data.  
+
+## 3. For my question 9  
+
+In the citations, 20X coverage was used in Arabidopsis, not sure whether it's suitable for European beech.  
+
+Thanks for your comment. The cited study utilized \(20\times\) sequencing depth in Arabidopsis, and \(20\times\) coverage has also been successfully applied in other plant species, for instance, in Zea mays and Pinus tabuliformis (Xu et al. 2020; Li et al. 2023). Increasing sequencing depth will improve sensitivity, but a \(20\times\) coverage is sufficient for the detection of most epimutations.  
+
+I didn't find the paper by Shahrvary et al. 2020 in the Reference  
+
+Shahrvary, Y., Symeonidi, A., Hazarika, R. R. et al. AlphaBeta: computational inference of epimutation rates and spectra from high- throughput DNA methylation data in plants. Genome Biol. 21, 260 (2020).  
+
+## Reference  
+
+1. Xu, G., Lyu, J., Li, Q. et al. Evolutionary and functional genomics of DNA methylation in maize domestication and improvement. Nat Commun 11, 5539 (2020).  
+2. Li, J., Han, F., Yuan, T. et al. The methylation landscape of giga-genome and the epigenetic timer of age in Chinese pine. Nat Commun 14, 1947 (2023).  
+3. Shahrvary, Y., Symeonidi, A., Hazarika, R. R. et al. AlphaBeta: computational inference of epimutation rates and spectra from high-throughput DNA methylation data in plants. Genome Biol. 21, 260 (2020).
+
+<--- Page Split --->
+
+## Reviewer #2 (Remarks to the Author):  
+
+I would like to thank the authors for their comments and the corrections they provided. I am satisfied with their responses and the revised version. I have two comments regarding the revised manuscript that could be considered:  
+
+We thank the reviewer for valuable suggestions, which have greatly improved the clarity and quality of our manuscript.  
+
+- Line 421: Please consider revising the sentence: "To more accurately quantify different thinning intensities induced methylation divergence, ..."  
+
+Thanks for your suggestion. We have revised the sentence for clarity and readability. The updated version now reads:  
+
+To more accurately quantify methylation divergence induced by different thinning intensities.  
+
+- Thank you for reporting the genomic distribution of CG divergence in both trees in Fig. 1 of the rebuttal. I wonder why this data was not included in the manuscript, considering both reviewers asked the same question. It is up to the authors to decide whether to include it, but I think it is an informative piece of data. If yes, please add to the legend that these are CG-context DMRs.  
+
+Thanks for your suggestion. We agree that the CG-context DMRs are informative data, and have incorporated the distribution of CG-DMRs into Supplementary Figure 3. We think this will give readers a more intuitive understanding of our data.
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__7206689b0b629e8d0089cf12c07ed61d70499cbe7a6de498450a7cc3e71a5daa/supplementary_0_Transparent Peer Review file__7206689b0b629e8d0089cf12c07ed61d70499cbe7a6de498450a7cc3e71a5daa_det.mmd

@@ -0,0 +1,472 @@
+<|ref|>title<|/ref|><|det|>[[72, 53, 296, 80]]<|/det|>
+# nature portfolio  
+
+<|ref|>text<|/ref|><|det|>[[74, 97, 296, 119]]<|/det|>
+Peer Review File  
+
+<|ref|>sub_title<|/ref|><|det|>[[72, 161, 891, 209]]<|/det|>
+## Accelerated growth increases the somatic epimutation rate in trees  
+
+<|ref|>text<|/ref|><|det|>[[72, 224, 503, 240]]<|/det|>
+Corresponding Author: Professor Frank Johannes  
+
+<|ref|>text<|/ref|><|det|>[[72, 274, 864, 289]]<|/det|>
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+<|ref|>text<|/ref|><|det|>[[72, 326, 141, 340]]<|/det|>
+Version 0:  
+
+<|ref|>text<|/ref|><|det|>[[72, 353, 220, 367]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[72, 379, 160, 393]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[72, 404, 922, 511]]<|/det|>
+(Remarks to the Author) In this manuscript the authors investigate the relationship between growth rate and somatic epimutation rates in trees, using European beech (Fagus sylvatica) as a model. The researchers leveraged a 150- year- old thinning experiment to compare trees subjected to different thinning intensities, which resulted in varying growth rates. They found that thinning- induced growth acceleration was accompanied by increased cell division rates in the main stems, as confirmed by cell count assays. Trees with higher growth rates exhibited significantly higher somatic epimutation rates in both main stems and lateral branches. Although the results are interesting, the analysis of epimutation rates is based on a small number of trees (two representative trees for detailed analysis), which may limit the generalizability of the findings.  
+
+<|ref|>text<|/ref|><|det|>[[73, 522, 377, 536]]<|/det|>
+Here are several concerns and suggestions:  
+
+<|ref|>text<|/ref|><|det|>[[70, 548, 907, 576]]<|/det|>
+1. Line 99-111, Please provide a figure to show the difference in stem growth rates, diameter breast height, and crown size between the treatments at a population level.  
+
+<|ref|>text<|/ref|><|det|>[[72, 588, 612, 616]]<|/det|>
+2. How representative are the two selected trees of their respective treatments? Do they have a close genetic relationship or genotypes?  
+
+<|ref|>text<|/ref|><|det|>[[72, 627, 397, 641]]<|/det|>
+3. Line 136-136, are they significantly different?  
+
+<|ref|>text<|/ref|><|det|>[[70, 652, 870, 667]]<|/det|>
+4. Fig2. What's the full name of DMP? How do you calculate the CG/CHG divergence? How about CHH divergence?  
+
+<|ref|>text<|/ref|><|det|>[[72, 678, 884, 719]]<|/det|>
+5. Fig2B, the variation from R1 vs. R2 is too big, and can't tell which dots are from which treatment. Color coding them? Since the variation between neighboring replicate samples is big, not sure one sample from the polar opposite side is enough.  
+
+<|ref|>text<|/ref|><|det|>[[72, 730, 395, 744]]<|/det|>
+6. Fig. How do you construct the topology tree?  
+
+<|ref|>text<|/ref|><|det|>[[72, 756, 910, 796]]<|/det|>
+7. What distribution of these gained or lost epimutations (or DMR) across the genome? Are they randomly mutated? What kind of genes were under epimutations? Do these mutations affect gene expression levels? Are these epimutations neutral or deleterious?  
+
+<|ref|>text<|/ref|><|det|>[[70, 808, 860, 835]]<|/det|>
+8. Could differences in stress responses (due to thinning) independently affect methylation? Other factors, like local environment, soil, or the genotypic background of the tree itself may also affect methylation.  
+
+<|ref|>text<|/ref|><|det|>[[70, 847, 689, 861]]<|/det|>
+9. Mean coverage was \(\sim 20X\) was this sufficient for low-frequency epimutation detection?  
+
+<|ref|>text<|/ref|><|det|>[[72, 873, 528, 887]]<|/det|>
+10. Does the accelerated growth affect the genomic mutation rate?  
+
+<|ref|>text<|/ref|><|det|>[[73, 899, 264, 912]]<|/det|>
+11. Missing title in Table S2
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[73, 72, 238, 85]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[73, 85, 921, 138]]<|/det|>
+The study by Zhou et al. explored the impact of tree growth rates on epimutation frequencies. To this end, the authors took advantage of an experimental plot where beech trees have been grown for decades under distinct thinning strategies. Trees in heavily thinned regions grow faster compared to those in less thinned areas. The authors exploited this difference to study epimutations (changes in DNA methylation) associated with accelerated growth.  
+
+<|ref|>text<|/ref|><|det|>[[73, 150, 900, 204]]<|/det|>
+First, the authors demonstrated that faster growth is associated with increased cell division (and also with increased cell expansion). They then measured global methylation divergence between faster growing and "control" trees grown in less thinned areas by comparing vascular tissues from opposite sides of the trunk. The analysis revealed that faster- growing trees exhibit higher levels of methylation divergence, specifically in the CG context, but not in the CHG context.  
+
+<|ref|>text<|/ref|><|det|>[[73, 215, 920, 281]]<|/det|>
+Subsequently, the authors performed an intra- organismal experiment in which methylation divergence was calculated as a function of divergence over time. For this, they compared methylomes of distinct leaves whose positions in the tree could be tracked and also determined the ages of the branches on which these leaves were located. The results indicated that leaves from the faster- growing tree showed increased methylation divergence, supporting their previous finding when comparing vascular tissues.  
+
+<|ref|>text<|/ref|><|det|>[[73, 293, 888, 346]]<|/det|>
+This is a very interesting manuscript using a unique genetic material that allowed to report a clear positive correlation between plant growth rate and epimutation rate. While there is no direct evidence proving that increased cell division is responsible for the elevated epimutation rate, these findings align well with the previously proposed "mitotic- rate hypothesis".  
+
+<|ref|>text<|/ref|><|det|>[[73, 359, 243, 372]]<|/det|>
+Comments/Suggestions:  
+
+<|ref|>text<|/ref|><|det|>[[72, 384, 920, 435]]<|/det|>
+- An interesting point is that cells in the faster-growing tree are also larger. Could this phenotype contribute to the increase in epimutation rate? Perhaps increased metabolic during extended cell growth activity also alters the deposition/removal of DNA methylation. This question is difficult to address with the current experimental model but perhaps could be added to the discussion if the authors think it could be possible.  
+
+<|ref|>text<|/ref|><|det|>[[72, 435, 911, 475]]<|/det|>
+- Fig. 2B, C: Please provide a more detailed description in the Methods section of how global mC divergence was calculated. Also, indicate in the figure legend where the "n" values come from, and why certain dots are shown in red.  
+
+<|ref|>text<|/ref|><|det|>[[72, 475, 911, 515]]<|/det|>
+- Fig. 2B, C: Please indicate the statistical significance of the differences between measurements in the plots. 
+- Fig. 2C: Were the CG and CHG divergence values calculated across all C positions in the genome? Please describe this more clearly. Additionally, what is the rationale for showing CHG but not CHH methylation for non-CG contexts? Would including CHH methylation offer further insights or reinforce the CHG findings?  
+
+<|ref|>text<|/ref|><|det|>[[72, 515, 904, 541]]<|/det|>
+- Related to Fig. 2: The authors state that "methylation divergence between samples from opposite sides of the stems was 2.64-fold higher in trees." Please clarify how this number was calculated.  
+
+<|ref|>text<|/ref|><|det|>[[72, 541, 870, 580]]<|/det|>
+- The authors present global CG divergence between trees and leaves. Are these loss and gains enriched in specific genomic regions or associated with genes with particular features (e.g., gene length, intron number,...), or are the epimutations randomly distributed across the coding genome?  
+
+<|ref|>text<|/ref|><|det|>[[72, 580, 567, 593]]<|/det|>
+- Fig. 3C: Please specify which datasets were used to generate this plot.  
+
+<|ref|>text<|/ref|><|det|>[[72, 593, 916, 619]]<|/det|>
+- The authors mention: "Removing the low-quality sequencing samples, we had 18 cambium samples and 17 leaf samples." Please clarify which specific samples were excluded from the analyses.  
+
+<|ref|>text<|/ref|><|det|>[[72, 619, 904, 647]]<|/det|>
+- Supplementary Table 1: Please include information indicating which sample IDs correspond to the CB and CC samples. 
+- On page 5, please correct the typo: "METHYLSTRANSFERASE" should be "METHYLTRANSFERASE."  
+
+<|ref|>text<|/ref|><|det|>[[73, 686, 144, 699]]<|/det|>
+Version 1:  
+
+<|ref|>text<|/ref|><|det|>[[73, 711, 219, 724]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[73, 736, 160, 749]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[73, 763, 238, 776]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 777, 777, 803]]<|/det|>
+I appreciate that the authors addressed most of my questions, but I still have some concerns, see below 1. Fig. 2b.c Please provide P- values and statistical method. What do the red dots indicate?  
+
+<|ref|>text<|/ref|><|det|>[[72, 816, 506, 842]]<|/det|>
+2. For my question 7 Can you make a figure to show the distribution of these DMRs?  
+
+<|ref|>text<|/ref|><|det|>[[72, 854, 923, 881]]<|/det|>
+The authors addressed my questions by citing some papers. Is there any study work on the same species as you did here? If not, it's unclear whether you can draw any conclusions.  
+
+<|ref|>text<|/ref|><|det|>[[72, 893, 797, 920]]<|/det|>
+3. For my question 9 In the citations, 20X coverage was used in Arabidopsis, not sure whether it's suitable for European beech.  
+
+<|ref|>text<|/ref|><|det|>[[72, 931, 506, 945]]<|/det|>
+I didn't find the paper by Shahyary et al. 2020 in the Reference
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[73, 73, 163, 86]]<|/det|>
+## Reviewer #2  
+
+<|ref|>text<|/ref|><|det|>[[73, 99, 238, 112]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[70, 112, 921, 140]]<|/det|>
+I would like to thank the authors for their comments and the corrections they provided. I am satisfied with their responses and the revised version. I have two comments regarding the revised manuscript that could be considered:  
+
+<|ref|>text<|/ref|><|det|>[[70, 152, 870, 180]]<|/det|>
+- Line 421: Please consider revising the sentence: "To more accurately quantify different thinning intensities induced methylation divergence, ..."  
+
+<|ref|>text<|/ref|><|det|>[[72, 190, 907, 245]]<|/det|>
+- Thank you for reporting the genomic distribution of CG divergence in both trees in Fig. 1 of the rebuttal. I wonder why this data was not included in the manuscript, considering both reviewers asked the same question. It is up to the authors to decide whether to include it, but I think it is an informative piece of data. If yes, please add to the legend that these are CG-context DMRs.  
+
+<|ref|>text<|/ref|><|det|>[[72, 636, 912, 688]]<|/det|>
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+<|ref|>text<|/ref|><|det|>[[72, 688, 916, 755]]<|/det|>
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source. The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.  
+
+<|ref|>text<|/ref|><|det|>[[73, 755, 618, 768]]<|/det|>
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[150, 83, 361, 99]]<|/det|>
+## REVIEWER COMMENTS  
+
+<|ref|>sub_title<|/ref|><|det|>[[150, 113, 477, 129]]<|/det|>
+## Reviewer #1 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[149, 143, 850, 295]]<|/det|>
+In this manuscript the authors investigate the relationship between growth rate and somatic epimutation rates in trees, using European beech (Fagus sylvatica) as a model. The researchers leveraged a 150- year- old thinning experiment to compare trees subjected to different thinning intensities, which resulted in varying growth rates. They found that thinning- induced growth acceleration was accompanied by increased cell division rates in the main stems, as confirmed by cell count assays. Trees with higher growth rates exhibited significantly higher somatic epimutation rates in both main stems and lateral branches. Although the results are interesting, the analysis of epimutation rates is based on a small number of trees (two representative trees for detailed analysis), which may limit the generalizability of the findings.  
+
+<|ref|>text<|/ref|><|det|>[[150, 309, 511, 324]]<|/det|>
+Here are several concerns and suggestions:  
+
+<|ref|>text<|/ref|><|det|>[[149, 339, 850, 401]]<|/det|>
+1. Line 99-111, Please provide a figure to show the difference in stem growth rates, diameter breast height, and crown size between the treatments at a population level. We thank the reviewer for this suggestion. This analysis is now shown in the new Supplementary Fig. 1.  
+
+<|ref|>text<|/ref|><|det|>[[149, 414, 800, 444]]<|/det|>
+2. How representative are the two selected trees of their respective treatments? Do they have a close genetic relationship or genotypes?  
+
+<|ref|>text<|/ref|><|det|>[[148, 444, 850, 655]]<|/det|>
+The two selected trees, Tree 109 and Tree 171, are representative individuals of their respective treatments. As shown in Fig. 1c and in Supplementary Fig. 1, their growth trajectories, tree dimensions and other characteristics fall within the intermediate range of trees in their respective plots, indicating they reflect typical patterns for their thinning treatment. Furthermore, all trees within the FAB 15 experiment originated from natural regeneration following shelterwood cuttings in 1822. This silvicultural practice promotes regeneration from a common seed source under similar ecological conditions. While the exact genotypes of Tree 109 and Tree 171 have not yet been determined, there is no indication of artificial selection or planting, and all trees in the plots are presumed to stem from the same local gene pool. Therefore, although we cannot confirm a close genetic relationship between the two selected trees, it is reasonable to assume they are representative members of a naturally regenerated, locally adapted beech population subjected to the respective long-term thinning treatments.  
+
+<|ref|>text<|/ref|><|det|>[[149, 670, 538, 685]]<|/det|>
+3. Line 136-136, are they significantly different?  
+
+<|ref|>text<|/ref|><|det|>[[149, 685, 850, 791]]<|/det|>
+We observed a significantly greater CG methylation divergence in the heavily thinned plot compared to the moderately thinned plot, as assessed by a t- test (p- value \(< 0.05\) ). We have included the statistical significance of the differences in the plots of Fig. 2b and c. Although not all comparisons reached statistical significance, the relatively large Cohen's \(D\) effect sizes support an increasing trend from RR to moderately thinned treatment to heavily thinned treatment, which supports our central hypothesis. Please refer to updated Fig. 2 for details.  
+
+<|ref|>text<|/ref|><|det|>[[148, 805, 848, 835]]<|/det|>
+4. Fig2. What's the full name of DMP? How do you calculate the CG/CHG divergence? How about CHH divergence?  
+
+<|ref|>text<|/ref|><|det|>[[149, 836, 849, 910]]<|/det|>
+DMP stands for Differentially Methylated Position. We are now defining this term in the Method section. To calculate CG and CHG divergence, we took the following steps: For each tree disk, we calculated the methylation difference between two replicates located on the same side (R1 vs. R2), as well as the methylation difference between each replicate and the sample from the opposite side (R1 vs. S, R2 vs. S). To obtain a
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[148, 84, 853, 160]]<|/det|>
+measure of genome- wide methylation difference between samples, we followed our previously published approach (see e.g. van der Graaf et al. 2015, Hazarika et al. 2022, Shahryary et al. 2020). For each cytosine site \(i\) , we calculated the methylation difference between the two samples (e.g., R1 and S): \(D_{i} = |M(R1,i) - M(S,i)|\) , and then took the average over all covered cytosines in the genome:  
+
+<|ref|>equation<|/ref|><|det|>[[147, 159, 258, 207]]<|/det|>
+\[D = \frac{1}{n}\sum_{i = 1}^{n}D_{i}\]  
+
+<|ref|>text<|/ref|><|det|>[[148, 218, 848, 248]]<|/det|>
+We thank the reviewer for bringing up this question. We are now detailing this approach in the Methods section.  
+
+<|ref|>text<|/ref|><|det|>[[148, 262, 850, 306]]<|/det|>
+5. Fig2B, the variation from R1 vs. R2 is too big, and can't tell which dots are from which treatment. Color coding them? Since the variation between neighboring replicate samples is big, not sure one sample from the polar opposite side is enough.  
+
+<|ref|>text<|/ref|><|det|>[[148, 307, 850, 412]]<|/det|>
+We appreciate the reviewer's suggestion. Indeed, the variation between neighboring replicate samples (e.g., R1 vs. R2) appears relatively large. However, when looking at the mean divergence differences, there is a clear trend showing that R1 vs. R2 (RR) are, on average, much less divergent. Although the small sample sizes may limit statistical significance, the relatively large Cohen's \(D\) effect sizes support an increasing trend from RR to moderately thinned treatment to heavily thinned treatment. Please refer to updated Fig. 2 for details.  
+
+<|ref|>text<|/ref|><|det|>[[149, 428, 537, 443]]<|/det|>
+6. Fig. How do you construct the topology tree?  
+
+<|ref|>text<|/ref|><|det|>[[148, 443, 850, 533]]<|/det|>
+To construct the tree's topological structure, we first divided the fallen tree into segments and systematically marked each branch with identifiers. These markings enabled precise tracking and mapping of the branches, allowing us to accurately reconstruct the tree's topology. This method ensured that we could reliably determine the spatial relationships between branches, ensuring an accurate representation of the tree's structure.  
+
+<|ref|>text<|/ref|><|det|>[[148, 548, 855, 803]]<|/det|>
+7. What distribution of these gained or lost epimutations (or DMR) across the genome? Are they randomly mutated? What kind of genes were under epimutations? Do these mutations affect gene expression levels? Are these epimutations neutral or deleterious? Overall, Tree 171 exhibits a higher number of CG-context DMRs compared to Tree 109, yet both share a similar distribution pattern across genomic regions (Fig. 1). The majority of CG-DMRs are located in non-coding regions (intergenic) and transposable elements (TEs), which is a typical pattern in plants (Schmitz et al. 2013; Niederhuth et al. 2016; Zhang et al. 2018). A substantial number of DMRs are found in CDS regions. Intronic regions also contain many DMRs. In contrast, DMRs in exonic regions are relatively sparse. Regarding the potential impact of the identified epimutations on gene expression levels, although we did not have gene expression data (such as RNA-seq) in this study, prior work has shown that gene body methylation (gbM) changes, which constitute the majority of somatic epimutations, typically have little to no effect on gene expression levels (Zhang et al. 2018; Bewick et al. 2019; Hofmeister et al. 2020). This provides strong support that the epimutations detected here are largely functionally neutral, and it ensures that the accumulation of epimutations is associated with cell divisions rather than driven by natural (i.e. cell lineage) selection.  
+
+<|ref|>text<|/ref|><|det|>[[149, 818, 850, 862]]<|/det|>
+8. Could differences in stress responses (due to thinning) independently affect methylation? Other factors, like local environment, soil, or the genotypic background of the tree itself may also affect methylation.  
+
+<|ref|>text<|/ref|><|det|>[[149, 863, 850, 907]]<|/det|>
+We thank the reviewer for raising this important point. Our experimental design aimed to minimize the influence of these variables. All three 0.4 ha plots lie in the immediate vicinity to each other on loamic Luvisols, displaying very similar soil characteristics. All
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[148, 84, 850, 355]]<|/det|>
+sampled trees were selected randomly within the different treatments, avoiding a systematic bias. While localized microclimatic variation cannot be entirely ruled out, particularly given that higher thinning intensity reduces standing stock and may slightly alter environmental conditions, such differences are considered to be a direct consequence of the treatment itself rather than an independent confounding factor. Finally, all trees originated from natural regeneration following a shelterwood cut, so they derive from the same locally adapted gene pool; there was no artificial planting or selection of distinct genotypes in any plot. Although we have not yet performed molecular genotyping on Trees 109 and 171, their shared origin and randomized selection make it reasonable to assume no systematic genotypic differences among treatments. Thus, based on our randomized selection and the uniform soil conditions, we minimize other factors that might affect methylation. Moreover, the "thinning" itself only involved clearing dead wood or dying trees from the plots, and is not expected to impose stress on the surrounding trees. Nonetheless, we appreciate the reviewer's thoughts here: The correlative link between increased rates of growth, cell division and epimutations should be further investigated in future studies to pin point the precise causal relationship. However, the careful observation that such a correlative link exists is an important first step that our study has taken.  
+
+<|ref|>text<|/ref|><|det|>[[147, 369, 848, 400]]<|/det|>
+9. Mean coverage was \(\sim 20X\) was this sufficient for low-frequency epimutation detection?  
+
+<|ref|>text<|/ref|><|det|>[[149, 400, 849, 476]]<|/det|>
+From our experience with similar experimental setups (Becker et al. 2011; Graaf et al. 2015; Schmitz et al. 2022), a \(20\times\) coverage is sufficient for the detection of most epimutations. Additionally, our method for estimating rates is relatively robust to ascertainment error as we explicitly model the error in the estimation procedure (see Shahryary et al. 2020).  
+
+<|ref|>text<|/ref|><|det|>[[150, 490, 696, 505]]<|/det|>
+10. Does the accelerated growth affect the genomic mutation rate?  
+
+<|ref|>text<|/ref|><|det|>[[149, 505, 850, 655]]<|/det|>
+We did not assess genomic mutation rates in this study, as our focus was specifically on epimutations. Evaluating whether accelerated growth also affects the genomic mutation rate would require whole-genome resequencing data, which was beyond the scope of our current work. However, having shown that epimutation rates are closely linked to cell division rates, we hypothesize that accelerated growth may influence genomic mutation rates similarly. Recent studies have demonstrated that both somatic mutation rates and epimutation rates scale with maturation age in perennial plants such as trees (Hanlon et al. 2019; Johannes 2025), suggesting a shared underlying biological mechanism likely associated with the division rates of shoot apical meristem (SAM) cells.  
+
+<|ref|>text<|/ref|><|det|>[[149, 669, 848, 700]]<|/det|>
+We thank the reviewer for bringing up this point. We have now included a brief discussion of this hypothesis in the discussion section.  
+
+<|ref|>text<|/ref|><|det|>[[150, 714, 850, 761]]<|/det|>
+11. Missing title in Table S2Thanks for pointing out the missing title in Table S2. We have now added a title to the table, and the revised version is included in the updated files.  
+
+<|ref|>sub_title<|/ref|><|det|>[[150, 790, 477, 806]]<|/det|>
+## Reviewer #2 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[149, 820, 850, 911]]<|/det|>
+The study by Zhou et al. explored the impact of tree growth rates on epimutation frequencies. To this end, the authors took advantage of an experimental plot where beech trees have been grown for decades under distinct thinning strategies. Trees in heavily thinned regions grow faster compared to those in less thinned areas. The authors exploited this difference to study epimutations (changes in DNA methylation) associated with accelerated growth.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[149, 98, 849, 189]]<|/det|>
+First, the authors demonstrated that faster growth is associated with increased cell division (and also with increased cell expansion). They then measured global methylation divergence between faster growing and "control" trees grown in less thinned areas by comparing vascular tissues from opposite sides of the trunk. The analysis revealed that faster- growing trees exhibit higher levels of methylation divergence, specifically in the CG context, but not in the CHG context.  
+
+<|ref|>text<|/ref|><|det|>[[149, 203, 849, 310]]<|/det|>
+Subsequently, the authors performed an intra- organismal experiment in which methylation divergence was calculated as a function of divergence over time. For this, they compared methylomes of distinct leaves whose positions in the tree could be tracked and also determined the ages of the branches on which these leaves were located. The results indicated that leaves from the faster- growing tree showed increased methylation divergence, supporting their previous finding when comparing vascular tissues.  
+
+<|ref|>text<|/ref|><|det|>[[149, 324, 849, 400]]<|/det|>
+This is a very interesting manuscript using a unique genetic material that allowed to report a clear positive correlation between plant growth rate and epimutation rate. While there is no direct evidence proving that increased cell division is responsible for the elevated epimutation rate, these findings align well with the previously proposed "mitotic- rate hypothesis".  
+
+<|ref|>text<|/ref|><|det|>[[149, 414, 666, 429]]<|/det|>
+We thank this reviewer for the positive assessment of our work.  
+
+<|ref|>text<|/ref|><|det|>[[150, 444, 351, 460]]<|/det|>
+Comments/Suggestions:  
+
+<|ref|>text<|/ref|><|det|>[[148, 474, 850, 685]]<|/det|>
+- An interesting point is that cells in the faster-growing tree are also larger. Could this phenotype contribute to the increase in epimutation rate? Perhaps increased metabolic during extended cell growth activity also alters the deposition/removal of DNA methylation. This question is difficult to address with the current experimental model but perhaps could be added to the discussion if the authors think it could be possible. The potential link between cell size, metabolic activity, and epimutation rate is indeed an interesting hypothesis. This discussion aligns well with the recent observation that both somatic mutation and epimutation rates show a powerlaw scaling with maturation age in trees (Johannes 2025), suggesting that there is an allometric link between life history traits and (epi)genomic maintenance fidelity. It is therefore of theoretical interest to explore how the scaling connects with the allometric laws that were previously found to link cellular metabolic rate, body size, and growth (West et al. 2002). Our current study cannot dissect these links, but our observations are consistent with this hypothesis and highlight a potential direction for future work.  
+
+<|ref|>text<|/ref|><|det|>[[149, 699, 848, 730]]<|/det|>
+We thank the reviewer for the insightful suggestion. We have now included a brief discussion of this hypothesis in the discussion section.  
+
+<|ref|>text<|/ref|><|det|>[[149, 744, 849, 788]]<|/det|>
+- Fig. 2B, C: Please provide a more detailed description in the Methods section of how global mC divergence was calculated. Also, indicate in the figure legend where the "n" values come from, and why certain dots are shown in red.  
+
+<|ref|>text<|/ref|><|det|>[[149, 789, 849, 849]]<|/det|>
+Thanks for the suggestion. We have revised the manuscript accordingly. A more detailed description has been added to the Methods section. Additionally, we have updated the figure legend to indicate the source of the "n" values and to explain why certain dots are shown in red.  
+
+<|ref|>text<|/ref|><|det|>[[149, 864, 849, 896]]<|/det|>
+- Fig. 2B, C: Please indicate the statistical significance of the differences between measurements in the plots.  
+
+<|ref|>text<|/ref|><|det|>[[149, 896, 848, 911]]<|/det|>
+Thanks, we have included the statistical significance of the differences in the plots of
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[149, 85, 230, 99]]<|/det|>
+Fig. 2b, c.  
+
+<|ref|>text<|/ref|><|det|>[[149, 113, 849, 175]]<|/det|>
+- Fig. 2C: Were the CG and CHG divergence values calculated across all C positions in the genome? Please describe this more clearly. Additionally, what is the rationale for showing CHG but not CHH methylation for non-CG contexts? Would including CHH methylation offer further insights or reinforce the CHG findings?  
+
+<|ref|>text<|/ref|><|det|>[[149, 174, 849, 325]]<|/det|>
+We thank the reviewer for this helpful comment. The CG and CHG methylation divergence values in Fig. 2c were calculated genome-wide, based on all covered (posteriorMax \(\geq 0.99\) ) cytosines in the respective sequence contexts. We have clarified this in the Methods section of the revised manuscript. Consistent with previous work both in somatic systems (Hofmeister et al. 2020; Ibañez et al. 2023) and transgenerational studies (Quadrana and Colot 2016; Zheng et al. 2017; Li et al. 2020), we found no time-dependent accumulation in CHH epimutations. This is likely due to the fact that the methylation in CHH context is mainly targeted by de novo methylation pathways that prevent the accumulation of stable gains and losses over generations (Johannes and Schmitz, 2019).  
+
+<|ref|>text<|/ref|><|det|>[[149, 339, 849, 384]]<|/det|>
+- Related to Fig. 2: The authors state that "methylation divergence between samples from opposite sides of the stems was 2.64-fold higher in trees." Please clarify how this number was calculated.  
+
+<|ref|>text<|/ref|><|det|>[[149, 385, 849, 460]]<|/det|>
+Thanks for the question. We have added more detailed explanations in the Methods section. As shown in Fig. 2b, the first box represents methylation divergence between replicate samples on the same side of the stem (R1 vs. R2). The second and third boxes represent divergence between samples from opposite sides of the stem under moderately and heavily thinned treatments, respectively.  
+
+<|ref|>text<|/ref|><|det|>[[149, 460, 849, 595]]<|/det|>
+To more accurately quantify different thinning intensities induced methylation divergence, we normalized the opposite- side divergence by subtracting the baseline variation observed between same- side replicates. In other words, the methylation difference between two sides of the stem is defined as the "opposite sides methylation difference" minus the "same- side replicates methylation difference". We then calculated the fold change between treatments based on these adjusted values. Using the mean values shown in Fig. 2b: \(0.005221033\) (replicates), \(0.005485005\) (moderately), and \(0.005919238\) (heavily), the fold change was calculated as: \((0.005919238 - 0.005221033) / (0.005485005 - 0.005221033) = 2.64\) .  
+
+<|ref|>text<|/ref|><|det|>[[149, 595, 849, 640]]<|/det|>
+Based on this, we stated that "methylation divergence between samples from opposite sides of the stems was 2.64- fold higher in trees from the heavily thinned plot compared to trees from the moderately thinned plot."  
+
+<|ref|>text<|/ref|><|det|>[[149, 655, 849, 714]]<|/det|>
+- The authors present global CG divergence between trees and leaves. Are these loss and gains enriched in specific genomic regions or associated with genes with particular features (e.g., gene length, intron number,...), or are the epimutations randomly distributed across the coding genome?  
+
+<|ref|>text<|/ref|><|det|>[[149, 715, 849, 820]]<|/det|>
+Overall, Tree 171 exhibits a higher number of CG- context DMRs compared to Tree 109. Yet, both share a similar distribution pattern across genomic regions (Fig. 1). The majority of CG- DMRs are located in non- coding regions (intergenic) and transposable elements (TEs), which is a typical pattern in plants (Schmitz et al. 2013; Niederhuth et al. 2016; Zhang et al. 2018). A substantial number of DMRs are found in CDS regions. Intronic regions also contain many DMRs. In contrast, DMRs in exonic regions are relatively sparse.  
+
+<|ref|>text<|/ref|><|det|>[[149, 835, 849, 880]]<|/det|>
+- Fig. 3C: Please specify which datasets were used to generate this plot. Thanks for the comment. This information has now been clarified in the revised figure legend.  
+
+<|ref|>text<|/ref|><|det|>[[147, 895, 849, 910]]<|/det|>
+- The authors mention: "Removing the low-quality sequencing samples, we had 18
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 84, 850, 113]]<|/det|>
+cambium samples and 17 leaf samples." Please clarify which specific samples were excluded from the analyses.  
+
+<|ref|>text<|/ref|><|det|>[[147, 114, 850, 144]]<|/det|>
+This was a misstatement; it should refer to samples that failed library construction. We have corrected this.  
+
+<|ref|>text<|/ref|><|det|>[[147, 158, 850, 187]]<|/det|>
+- Supplementary Table 1: Please include information indicating which sample IDs correspond to the CB and CC samples.  
+
+<|ref|>text<|/ref|><|det|>[[147, 188, 850, 218]]<|/det|>
+We thank the reviewer for pointing this out. We have now added information to Supplementary Data 1.  
+
+<|ref|>text<|/ref|><|det|>[[147, 234, 850, 264]]<|/det|>
+- On page 5, please correct the typo: "METHYLSTRANSFERASE" should be "METHYLTRANSFERASE."  
+
+<|ref|>text<|/ref|><|det|>[[147, 264, 850, 294]]<|/det|>
+Thanks for the correction. We have corrected "METHYLSTRANSFERASE" to "METHYLTRANSFERASE" in the revised manuscript.  
+
+<|ref|>sub_title<|/ref|><|det|>[[148, 339, 236, 354]]<|/det|>
+## Reference  
+
+<|ref|>text<|/ref|><|det|>[[177, 368, 850, 401]]<|/det|>
+1. Schmitz, R., Schultz, M., Urich, M. et al. Patterns of population epigenomic diversity. Nature 495, 193-198 (2013).  
+
+<|ref|>text<|/ref|><|det|>[[177, 400, 850, 431]]<|/det|>
+2. Niederhuth, C. E., Bewick, A. J., Ji, L. et al. Widespread natural variation of DNA methylation within angiosperms. Genome Biol. 17, 194 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[177, 430, 850, 461]]<|/det|>
+3. Zhang, H., Lang, Z. & Zhu, J. K. Dynamics and function of DNA methylation in plants. Nat. Rev. Mol. Cell Biol. 19, 489-506 (2018).  
+
+<|ref|>text<|/ref|><|det|>[[177, 460, 850, 506]]<|/det|>
+4. Bewick, A. J., Zhang, Y., Wendte, J. M., Zhang, X. & Schmitz, R. J. Evolutionary and experimental loss of gene body methylation and its consequence to gene expression. G3 Genes|Genomes|Genetics 9, 2441-2445 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[177, 505, 850, 551]]<|/det|>
+5. Hofmeister, B. T., Denkena, J., Colomé-Tatché, M. et al. A genome assembly and the somatic genetic and epigenetic mutation rate in a wild long-lived perennial Populus trichocarpa. Genome Biol. 21, 259 (2020).  
+
+<|ref|>text<|/ref|><|det|>[[177, 551, 850, 580]]<|/det|>
+6. Becker, C., Hagmann, J., Müller, J. et al. Spontaneous epigenetic variation in the Arabidopsis thaliana methylome. Nature 480, 245-249 (2011).  
+
+<|ref|>text<|/ref|><|det|>[[177, 580, 850, 625]]<|/det|>
+7. van der Graaf, A., Wardenaar, R., Neumann, D. A. et al. Rate, spectrum, and evolutionary dynamics of spontaneous epimutations. Proc. Natl Acad. Sci. USA 112, 6676-6681 (2015).  
+
+<|ref|>text<|/ref|><|det|>[[177, 625, 850, 655]]<|/det|>
+8. Schmitz, R. J., Marand, A. P., Zhang, X. et al. Quality control and evaluation of plant epigenomics data. Plant Cell 34, 503-513 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[177, 655, 850, 698]]<|/det|>
+9. Hanlon, V. C. T., Otto, S. P. & Aitken, S. N. Somatic mutations substantially increase the per-generation mutation rate in the conifer Picea sitchensis. Evol. Lett. 3, 348-358 (2019).  
+
+<|ref|>text<|/ref|><|det|>[[177, 699, 850, 728]]<|/det|>
+10. Johannes, F. Allometric scaling of somatic mutation and epimutation rates in trees. Evolution 79, 1-5 (2025).  
+
+<|ref|>text<|/ref|><|det|>[[177, 729, 850, 774]]<|/det|>
+11. West, G. B., Woodruff, W. H. & Brown, J. H. Allometric scaling of metabolic rate from molecules and mitochondria to cells and mammals. Proc. Natl Acad. Sci. USA 99 (Suppl 1), 2473-2478 (2002).  
+
+<|ref|>text<|/ref|><|det|>[[177, 775, 850, 832]]<|/det|>
+12. Noé Ibañez, V., van Antro, M., Peña-Ponton, C., Milanovic-Ivanovic, S., Wagemaker, C. A. M., Gawehns, F. & Verhoeven, K. J. F. Environmental and genealogical effects on DNA methylation in a widespread apomictic dandelion lineage. J. Evol. Biol. 36, 663-674 (2023).  
+
+<|ref|>text<|/ref|><|det|>[[177, 833, 850, 862]]<|/det|>
+13. Quadrana, L. & Colot, V. Plant transgenerational epigenetics. Annu. Rev. Genet. 50, 467-491 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[177, 863, 850, 908]]<|/det|>
+14. Zheng, X., Chen, L., Xia, H. et al. Transgenerational epimutations induced by multi-generation drought imposition mediate rice plant's adaptation to drought condition. Sci. Rep. 7, 39843 (2017).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[179, 84, 849, 130]]<|/det|>
+15. Li, J., Yang, D. L., Huang, H. et al. Epigenetic memory marks determine epiallele stability at loci targeted by de novo DNA methylation. Nat. Plants 6, 661–674 (2020). 
+
+<|ref|>text<|/ref|><|det|>[[179, 130, 849, 161]]<|/det|>
+16. Johannes, F. & Schmitz, R. J. Spontaneous epimutations in plants. New Phytol. 221, 1253–1259 (2019). 
+
+<|ref|>image<|/ref|><|det|>[[216, 230, 780, 619]]<|/det|>
+ 
+
+<|ref|>image_caption<|/ref|><|det|>[[203, 654, 797, 688]]<|/det|>
+<center>Fig. 1 The distribution of differentially methylated regions (DMRs) for Tree 109 and Tree 171.</center>
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[150, 83, 441, 99]]<|/det|>
+## REVIEWER COMMENTS - updated  
+
+<|ref|>sub_title<|/ref|><|det|>[[150, 112, 465, 128]]<|/det|>
+## Reviewer #1 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[149, 141, 849, 201]]<|/det|>
+I appreciate that the authors addressed most of my questions, but I still have some concerns, see belowWe would like to thank the reviewer for insightful suggestions, which have greatly improved the quality and completeness of our manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[149, 212, 850, 315]]<|/det|>
+1. Fig. 2b.c Please provide P-values and statistical method. What do the red dots indicate? We thank the reviewer for this very helpful suggestion. We have updated Figure 2b, c to include the exact P-value for the significant result, and provided the statistical method in the legend, and full statistical results, including non-significant differences, are provided in Supplementary Data 5. The red dots deviate from the overall data distribution and may be potential outliers; however, they do not affect the overall results or conclusions, so they are retained in the figure.  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 328, 312, 342]]<|/det|>
+## 2. For my question 7  
+
+<|ref|>text<|/ref|><|det|>[[150, 342, 644, 357]]<|/det|>
+Can you make a figure to show the distribution of these DMRs?  
+
+<|ref|>text<|/ref|><|det|>[[149, 370, 850, 414]]<|/det|>
+The authors addressed my questions by citing some papers. Is there any study work on the same species as you did here? If not, it's unclear whether you can draw any conclusions.  
+
+<|ref|>text<|/ref|><|det|>[[149, 414, 849, 515]]<|/det|>
+We thank the reviewer for the follow- up question. To our knowledge, there is currently no published study on CG- DMR distribution in European beech. Nevertheless, our analysis provides a description of somatic CG- DMRs in this species, offering valuable insight into the genomic distribution of epimutations. While direct comparisons with previous work in the same species are not possible, the observed patterns are consistent with general trends reported in other plants (as cited in our initial response), supporting the robustness and interpretability of our data.  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 529, 312, 543]]<|/det|>
+## 3. For my question 9  
+
+<|ref|>text<|/ref|><|det|>[[149, 543, 849, 572]]<|/det|>
+In the citations, 20X coverage was used in Arabidopsis, not sure whether it's suitable for European beech.  
+
+<|ref|>text<|/ref|><|det|>[[149, 572, 849, 645]]<|/det|>
+Thanks for your comment. The cited study utilized \(20\times\) sequencing depth in Arabidopsis, and \(20\times\) coverage has also been successfully applied in other plant species, for instance, in Zea mays and Pinus tabuliformis (Xu et al. 2020; Li et al. 2023). Increasing sequencing depth will improve sensitivity, but a \(20\times\) coverage is sufficient for the detection of most epimutations.  
+
+<|ref|>text<|/ref|><|det|>[[149, 658, 649, 673]]<|/det|>
+I didn't find the paper by Shahrvary et al. 2020 in the Reference  
+
+<|ref|>text<|/ref|><|det|>[[149, 673, 849, 717]]<|/det|>
+Shahrvary, Y., Symeonidi, A., Hazarika, R. R. et al. AlphaBeta: computational inference of epimutation rates and spectra from high- throughput DNA methylation data in plants. Genome Biol. 21, 260 (2020).  
+
+<|ref|>sub_title<|/ref|><|det|>[[149, 731, 232, 745]]<|/det|>
+## Reference  
+
+<|ref|>text<|/ref|><|det|>[[178, 758, 849, 874]]<|/det|>
+1. Xu, G., Lyu, J., Li, Q. et al. Evolutionary and functional genomics of DNA methylation in maize domestication and improvement. Nat Commun 11, 5539 (2020).  
+2. Li, J., Han, F., Yuan, T. et al. The methylation landscape of giga-genome and the epigenetic timer of age in Chinese pine. Nat Commun 14, 1947 (2023).  
+3. Shahrvary, Y., Symeonidi, A., Hazarika, R. R. et al. AlphaBeta: computational inference of epimutation rates and spectra from high-throughput DNA methylation data in plants. Genome Biol. 21, 260 (2020).
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[149, 84, 464, 99]]<|/det|>
+## Reviewer #2 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[149, 112, 849, 155]]<|/det|>
+I would like to thank the authors for their comments and the corrections they provided. I am satisfied with their responses and the revised version. I have two comments regarding the revised manuscript that could be considered:  
+
+<|ref|>text<|/ref|><|det|>[[149, 156, 848, 185]]<|/det|>
+We thank the reviewer for valuable suggestions, which have greatly improved the clarity and quality of our manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[149, 199, 848, 228]]<|/det|>
+- Line 421: Please consider revising the sentence: "To more accurately quantify different thinning intensities induced methylation divergence, ..."  
+
+<|ref|>text<|/ref|><|det|>[[149, 228, 848, 256]]<|/det|>
+Thanks for your suggestion. We have revised the sentence for clarity and readability. The updated version now reads:  
+
+<|ref|>text<|/ref|><|det|>[[149, 257, 848, 286]]<|/det|>
+To more accurately quantify methylation divergence induced by different thinning intensities.  
+
+<|ref|>text<|/ref|><|det|>[[149, 299, 849, 370]]<|/det|>
+- Thank you for reporting the genomic distribution of CG divergence in both trees in Fig. 1 of the rebuttal. I wonder why this data was not included in the manuscript, considering both reviewers asked the same question. It is up to the authors to decide whether to include it, but I think it is an informative piece of data. If yes, please add to the legend that these are CG-context DMRs.  
+
+<|ref|>text<|/ref|><|det|>[[149, 371, 848, 414]]<|/det|>
+Thanks for your suggestion. We agree that the CG-context DMRs are informative data, and have incorporated the distribution of CG-DMRs into Supplementary Figure 3. We think this will give readers a more intuitive understanding of our data.
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__7206d39aac3140f44125036e400400f2bc6214730597409b3eca3742f3fd110f/images_list.json

@@ -0,0 +1 @@
+[]

peer_reviews/supplementary_0_Transparent Peer Review file__7206d39aac3140f44125036e400400f2bc6214730597409b3eca3742f3fd110f/supplementary_0_Transparent Peer Review file__7206d39aac3140f44125036e400400f2bc6214730597409b3eca3742f3fd110f.mmd

@@ -0,0 +1,562 @@
+
+# nature portfolio  
+
+# Peer Review File  
+
+## Citizens' smartphones unravel earthquake shaking in urban areas  
+
+Corresponding Author: Professor Francesco Finazzi  
+
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+Attachments originally included by the reviewers as part of their assessment can be found at the end of this file.  
+
+Version 0:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author)  
+
+This manuscript presents a methodology to use peak acceleration observations from smartphones to complement shaking intensity maps (called ShakeMaps by the seismological community). The approach is applied to the Campi Flegrei region around Naples, Italy. Four earthquakes from a vigorous swarm are used to constrain the site amplifications, and a site amplification map is derived and used to enhance ShakeMaps for the earthquakes. In future earthquakes, the same amplification could be applied to enhance ShakeMaps.  
+
+The use of dense smartphone data presents a real opportunity to add constrains on earthquake shaking. However, as the authors note, there are challenges calibrating what the smartphones are measuring compared to traditional free field instrumentation. Their solution is to assume a radial decay in ground motion for both the INGV stations and the phones, and then determine amplification factors relative to the best- fit radial models.  
+
+This is a good approach, assuming the ground motion is radial. This is an OK assumption for small events like those modeled M<4.5. But for larger earthquakes we do not expect radial ground motions due to the finiteness of the source and directivity effects. It would be helpful to add this issue to the discussion. In addition, for larger earthquakes the amplification factors likely do not scale linearly. To be fair, this issue is not accounted for in other ShakeMap creation approaches, but still it would be helpful to note this issue as a challenge.  
+
+Current ShakeMaps don't just use a ground motion model plus seismic station data, they also use felt reports. Did the authors consider how the presented models/amplification compares to models that also use felt reports? Do they correlate? If they do, this provides validation. If they don't why not? Again, perhaps worth discussion.  
+
+Finally a detail. Why not plot the peak amplitudes of the phone observations on the ShakeMaps (e.g. Fig 3 and in supplement)? You plot the station observations.  
+
+This is a very interesting study and shows the potential of phone- based observations. I recommend publication once the authors have considered the above comments and hopefully added to the discussion.  
+
+(Remarks on code availability)  
+
+Reviewer #2  
+
+(Remarks to the Author)  
+
+The manuscript is very well written, presenting an interesting research topic and a comprehensive analysis. I believe that this research, along with similar studies, opens a new avenue for collecting seismological data, given the computing power currently available and the new algorithms, especially in machine learning, that enable us to manage large volumes of data. However, there are several questions that the authors need to address, which I outline below:
+
+<--- Page Split --->
+
+(1) Lines 305 – 306: “In equation (10) describes the potential amplification of the ground motion due to the site effects.” How can it be confirmed that the amplification observed in this study is solely attributable to site effects, rather than being influenced by path effects (such as azimuth dependence, 3d velocity structure) or source effects (e.g., radiation patterns)?  
+
+(2) It is challenging to interpret the uncertainty presented in Supplementary Figures 6 and 7. I recommend utilizing a single map that displays the epistemic uncertainty for the entire grid, rather than dividing it into two separate plots for the upper and lower boundaries.  
+
+(3) Supplementary Figure 4c has caused some confusion. Epistemic uncertainty appears to be analogous to the standard error, which makes Figure 4d comprehensible, as it illustrates lower epistemic uncertainty in regions with a greater amount of data. However, if a location is represented by only one station or without data, the epistemic uncertainty should be approximated to \(\sigma /N = 1\) .  
+
+Therefore, it is unclear why the smaller value \((\sim 0.05)\) is indicated in the region with only one station and 0.4 for the region without data in Figure 4c, which is unreasonable.  
+
+(4) Lines 221–222: Do the authors use the same correlation length (denoted as 0 in the paper) for both datasets? If so, this may not be reliable, as the data obtained from the smartphone is expected to exhibit a smaller correlation length.  
+
+(5) What kind of metadata can we expect to be accessible if the smartphone datasets are eventually released to the public? There is considerable potential for smartphones to transform early warning systems and enhance post-seismic damage distribution maps. However, if the metadata is limited or not comprehensively available, its utility for researchers may be substantially diminished.  
+
+(6) It would be valuable to identify the specific components of the filtering procedure that have contributed to the highest percentage of discarded waveforms. In particular, what is the primary source of poor-quality waveforms in data recorded via smartphones?  
+
+(7) Smartphone data can augment or build on free-field predictions, and with tremendous potential. It may be beneficial for future research to consider the relationship between MMI and smartphones.  
+
+(8) Please include the phiS2S (between-site sigma) and phiSS (within-site sigma) values from both models in Table 1, in addition to the tau values.  
+
+(9) The paper should include the values pertaining to the variability of Inst derived from the two datasets (stations and smartphones).  
+
+(10) Are the results obtained from smartphones within the epistemic uncertainty range of ergodic GMM? Please verify this and generate a corresponding plot. I expect that the results should lie within the epistemic uncertainty range, otherwise, this would suggest that they do not really help to improve the GMM.  
+
+(Remarks on code availability)  
+
+The record is publicly accessible, but files are restricted to users with access.  
+
+Reviewer #3  
+
+(Remarks to the Author)  
+
+The paper presents the integration of smartphone accelerometer data into seismic monitoring as a potential method to increase the spatial density of ground- motion observations. The study proposes an interesting concept, but uncertainties and methodological limitations reduce its reliability. The model serves as a useful exploratory tool; however, its ability to complement traditional recordings and site- response analysis requires further investigation and validation. The approach has potential but necessitates additional verification steps before it can be considered fully reliable. I request a major revision.  
+
+The study relies on smartphone accelerometers, which introduce significant bias and uncertainty. Unlike dedicated seismic instruments, smartphones exhibit higher noise levels and variable recording conditions. Firstly, as mentioned in the article, smartphones are not anchored to the ground, and their placement within buildings is unknown, introducing variability in the recorded signals. Additionally, ambient noise and user activity may further affect the accuracy and reliability of the recorded measurements. Finally, it should be specified whether the type and quality of accelerometers vary significantly across different smartphone models. The methodology should explicitly quantify systematic errors introduced by these uncertainties. It should also specify what kind of processing is applied to smartphone data and the time required for such processing. Given that smartphone data are likely to be very noisy, a rigorous preprocessing step is necessary to make them comparable with seismic station accelerometer data.  
+
+The paper states that smartphone accelerations were not calibrated against seismic station measurements, relying instead on spatial statistical models. While the statistical approach may help extract patterns, it does not fully validate the accuracy of smartphone- based measurements. Before this type of measurement can be fully utilized, a detailed comparison with high- quality seismic records is necessary. In the discussion, the following statement is made: "The methodology we have developed provides a solution that accounts for the bias in smartphone measurements relative to seismological station
+
+<--- Page Split --->
+
+measurements and that relies on the consistent spatial cross- correlation between smartphone and station measurements rather than on the absolute accelerations recorded by the smartphones. The methodology thus achieves data fusion without requiring any calibration between smartphones and stations." I am not entirely convinced by this claim. The authors have shown that the soil amplification maps generated using smartphone and seismic station data correlate \((r = 0.84)\) . This suggests that smartphones can capture similar spatial patterns, but it does not prove that the measured accelerations are comparable in absolute value. There is no direct comparison between smartphone and seismic station measurements at the exact locations or in close proximity. Therefore, stating that the methodology "accounts for the bias" seems somewhat overstated—while smartphone uncertainty has been modeled, without direct calibration between the two data sources, the accuracy of this correction remains uncertain.  
+
+Line 91—95: It is unclear how the spatial coverage of stations remained consistent across the four earthquakes, given that the number of smartphones varied between 56 and 441. Furthermore, it is not clear why only four earthquakes were used for calibration. Additionally, the use of validation earthquakes listed in the supplementary table is not well explained.  
+
+Figure S3: The log- amplification structure of the station data appears much more linear compared to the smartphone data, where red and blue values are adjacent to each other. This should be commented on. Figure S4 (panels c and d): It appears that the standard deviation is much lower in the case of smartphones than in seismic stations. I believe this is strongly influenced by the larger number of smartphone data points rather than an actual consistency between the measurements.  
+
+Amplification and geological structures: Can these amplification values (Figure 2 and Figures S3, S4, and S5) be associated with specific geological structures related to the caldera? This should be explored, particularly in light of Figure S9. It is not clear what the authors intend to convey with this figure.  
+
+Figures S6 and S7: The uncertainties appear very small, considering the intrinsic uncertainties in using smartphone data. Would it be possible to compare them with the sigma of a Ground Motion Model (GMM)? Regarding the higher spatial resolution of the acceleration map in Figure 3, it is important to remember that any additional data included in a map will modify it. What is missing in this work is a discussion on the reliability of the smartphone acceleration data. By the nature of smartphone data, I think it is essential to include it in the map by indicating uncertainty.  
+
+Ground Motion Model (GMM) Issues: I find it difficult to grasp the purpose and significance of the GMM developed in this study, as this analysis seems incomplete lacks a thorough explanation of how it was constructed and its practical utility. Iervolino et al. (2024) developed a GMM specifically for the Campi Flegrei area using a dataset comparable to the one in this study with a functional equation of similar form. The sigma of their model should be compared with the amplification- derived sigma obtained in this study. It is essential to provide more details on the number of data points used, as Supplementary Table 1 does not include this information. If a GMM is to be proposed, it should be accompanied by a thorough discussion and comparison with existing models. Based on these concerns, the conclusion stating: "The amplification map issued from the fusion of seismological stations and smartphone records can also be used to recalibrate conventional GMMs and reduce the random variability of seismic motion prediction." should be reconsidered.  
+
+For future operational implementation, the model must be validated across multiple earthquakes. Additionally, its performance should be analyzed for various magnitudes, epicentral distances, and site conditions to ensure its robustness and reliability across different seismic scenarios.  
+
+The terminology in the text should be refined for greater precision and to avoid ambiguity. For instance, even in the abstract:  
+
+Instead of making broad statements like "the ubiquity of smartphones", the revised version should specify the data type being used and how it contributes to seismic monitoring.  
+
+The term "red zone" is officially used by the Italian Civil Protection Department to designate high- risk volcanic areas in the Campi Flegrei region. This term identifies areas subject to preventive evacuation measures for public safety in case of an eruption. While it can be retained, it should be clearly explained or cited. Otherwise, using a more specific term would enhance the scientific clarity and rigor of the text.  
+
+In scientific writing, stating that amplification varies from 0.25 to 2.83 is insufficient unless the reference measure is specified and the calculation method explained.  
+
+Additionally, the paper lacks proper references to other works dealing with the rapid quantification of ground motion using crowdsourced or smartphone- based techniques and similar challenges.  
+
+I am not entirely sure that the current manuscript structure is appropriate. A significant portion of the results and figures are in the supplementary material, making it difficult to fully grasp the study without it. Additionally, postponing the Methods section in this way does not facilitate comprehension.  
+
+Line 128: Figure S9 is cited before S6, S7 and S8. Line 299: the correct citation should be number 20. Line 306: replace 'read' with 'red'. The caption of S8 has the wrong ID.
+
+<--- Page Split --->
+
+(Remarks on code availability)  
+
+Version 1:  
+
+Reviewer comments:  
+
+Reviewer #2  
+
+(Remarks to the Author)  
+
+Other modifications are acceptable, but I would like to address one point:  
+
+I understand that if the location has sufficient data, the epistemic uncertainty will show a lower value. However, how can a value lower than 0.1 be achieved when there are only a few records (fewer than 5)? This concerns my key question (Question 3). I am particularly interested in the phiS2S value; if phiS2S is around 0.2, it may be possible to reach a value below 0.1. However, the general phiS2S is approximately 0.3- 0.35, making obtaining a value under 0.1 almost impossible. I find it surprising that the authors state it is challenging to obtain phiS2S, but I understand if it is not available.  
+
+I believe the authors may underestimate the limitations associated with locations that have few records. I suggest that the authors consider using equation (29) from Lavrentiadis et al. (2023) to address this issue or at least mention it in their paper. This would be beneficial since readers should be aware of potential improvements.  
+
+## Reference:  
+
+Lavrentiadis, G., Abrahamson, N.A. (2023). A non- ergodic spectral acceleration ground motion model for California developed with random vibration theory. Bull Earthquake Eng, 21, 5265- 5291. https://doi.org/10.1007/s10518- 023- 01689- 9  
+
+Additionally, another new relevant paper from the US that can be cited is: "Ground- Motion Modeling Using MyShake Smartphone Peak Acceleration Data," BSSA, 2025, DOI: 10.1785/0120240209.  
+
+(Remarks on code availability)  
+
+Reviewer #3  
+
+(Remarks to the Author)  
+
+Decision Recommendation: Minor Revision  
+
+General Recommendation:  
+
+The revised manuscript shows meaningful improvements in clarity, methodological transparency, and justification of the approach. The authors have responded constructively to most reviewer comments and addressed several concerns regarding the use of smartphone data, the treatment of bias, and the statistical modelling of amplification. The integration of smartphone data remains an innovative and potentially impactful idea, and the manuscript now provides a more robust basis for such integration. The overall structure and figures have also improved, and key assumptions (e.g., spatial cross- correlation, non- reliance on calibration) are now better explained. However, a few important clarifications and adjustments are still needed before the manuscript can be considered for publication.  
+
+Publication.  
+
+Specific Requests (Required for Acceptance)  
+
+Clarify the spatial limitations of the amplification map, especially in areas with sparse smartphone data. It should be explicitly noted that fine- scale interpretation in such zones may be unreliable due to local uncertainty.  
+
+State more clearly the operational dependency of the method on the presence of a sufficiently dense and active (charging) smartphone network. The impact of uneven spatial distribution should be discussed more directly.  
+
+Include a table listing all seismic stations used, with a clear indication of whether they belong to the RAN or INGV networks. This will help readers assess the provenance and consistency of the data.  
+
+Correct the typographical error ("shanking" \(\rightarrow\) "shaking" on line 320) and review consistency in terminology across sections.  
+
+## Conclusion  
+
+I recommend a minor revision, conditional on the authors addressing the above points. The manuscript is promising and can make a valuable contribution to the field if these final clarifications are incorporated.  
+
+(Remarks on code availability)
+
+<--- Page Split --->
+
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source.  
+
+The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.  
+
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+
+## AE COMMENTS  
+
+You will see that the reviewers are interested in your study but also raise points that require improvement. In particular, they raise the issue of the assumed radial decay of the ground motion as well as the amplification being solely attributed to site effects and request a more detailed analysis and discussion about the uncertainties and methodological limitations. We also would like to highlight that our format allows 5000 words in the main text (without Methods) and a flexible length for the Methods section as well as 10 display items (figures and/or tables). This should allow you to move essential information from the SI to the main text.  
+
+## REVIEWER COMMENTS  
+
+Reviewer #1 (Remarks to the Author):  
+
+This manuscript presents a methodology to use peak acceleration observations from smartphones to complement shaking intensity maps (called ShakeMaps by the seismological community). The approach is applied to the Campi Flegrei region around Naples, Italy. Four earthquakes from a vigorous swarm are used to constrain the site amplifications, and a site amplification map is derived and used to enhance ShakeMaps for the earthquakes. In future earthquakes, the same amplification could be applied to enhance ShakeMaps.  
+
+The use of dense smartphone data presents a real opportunity to add constrains on earthquake shaking. However, as the authors note, there are challenges calibrating what the smartphones are measuring compared to traditional free field instrumentation. Their solution is to assume a radial decay in ground motion for both the INGV stations and the phones, and then determine amplification factors relative to the best- fit radial models.  
+
+This is a good approach, assuming the ground motion is radial. This is an OK assumption for small events like those modeled \(M< 4.5\) . But for larger earthquakes we do not expect radial ground motions due to the finiteness of the source and directivity effects. It would be helpful to add this issue to the discussion. In addition, for larger earthquakes the amplification factors likely do not scale linearly. To be fair, this issue is not accounted for in other ShakeMap creation approaches, but still it would be helpful to note this issue as a challenge.  
+
+The reviewer is correct with the fact that site amplification emerges as a model residual from the fitted radial decay. The amplification map must then be "learned" using events that have radial decay (and thus small or moderate magnitude). However, for any new event in the same region, the ShakeMap estimation is based on the model in Equation (11) which takes into account: 1) the amplification map learned from past earthquakes (i.e., S2S), and 2) the SW terms which describe the ground- motions anomalies not explained neither by the isotropic decay nor by the site amplification. These terms capture potential event- specific non- isotopic patterns and therefore directivity effects and possible non- linear effects.
+
+<--- Page Split --->
+
+Of course, such directivity effects can not be predicted for future earthquakes and then isotropic GMM are classically used in Engineering seismology. We then derived in Equation (16) a GMM based on an isotropic decay and on our amplification map. ShakeMaps generated using Equation (16) are not isotropic because the amplification map is not. We compared our GMM to similar, classical and isotropic GMM used in this region (in particular the one derived recently by Irevolino et al., 2024) and, as discussed below (see the answer to Reviewer #3) our model is performing better than previously published models. Our GMM can also be used for classical seismic hazard/risk analysis (as shown by Irevolino et al., 2024).  
+
+Current ShakeMaps don't just use a ground motion model plus seismic station data, they also use felt reports. Did the authors consider how the presented models/amplification compares to models that also use felt reports? Do they correlate? If they do, this provides validation. If they don't why not? Again, perhaps worth discussion.  
+
+The addition of felt report intensities as a new data source would add considerable complexity to the statistical modelling. Indeed, intensities are provided on a discrete scale that cannot be modelled using continuous Gaussian processes. A simple option would be to convert intensities to PGA measurements using relationships available in the literature (e.g., Olivet et al., 2022) but the uncertainty of the resulting ShakeMap would be wrongly quantified. Furthermore, the range of observed intensities for the events considered in this article remains limited due to the relatively small magnitude of the earthquakes.  
+
+Finally a detail. Why not plot the peak amplitudes of the phone observations on the ShakeMaps (e.g. Fig 3 and in supplement)? You plot the station observations.  
+
+The smartphone peak accelerations are higher than the PGA due to the bias shown in Figure 2 (up to 3 times). If the smartphone measurements were plotted on the same colour scale, the PGA would be almost flat.  
+
+This is a very interesting study and shows the potential of phone- based observations. I recommend publication once the authors have considered the above comments and hopefully added to the discussion.  
+
+Thanks for the useful comments.  
+
+Reviewer #2 (Remarks to the Author):  
+
+The manuscript is very well written, presenting an interesting research topic and a comprehensive analysis. I believe that this research, along with similar studies, opens a new avenue for collecting seismological data, given the computing power currently available and the new algorithms, especially in machine learning, that enable us to manage large volumes of data. However, there are several questions that the authors need to address, which I outline below:  
+
+(1) Lines 305 – 306: “In equation (10) describes the potential amplification of the ground motion due to the site effects.” How can it be confirmed that the amplification observed in this study is solely attributable to site effects, rather than being influenced by path effects
+
+<--- Page Split --->
+
+(such as azimuth dependence, 3d velocity structure) or source effects (e.g., radiation patterns)?  
+
+Separating source, path and site effects is a classical challenge of such studies. The amplification map developed here captures 3D site effects (if they are not azimuth dependent) which is an added value compared to classical site classification based solely on VS30. Such a "3D" amplification map is "learned" using many events and then event- specific path effects or radiation patterns have less and less influence when the number of learning events is increasing. Note also (as explained above) that the 8W terms which describe the ground- motions anomalies not explained neither by the isotropic decay nor by the site amplification capture potential event- specific non- isotopic patterns and therefore directivity effects and possible event- specific path effects.  
+
+(2) It is challenging to interpret the uncertainty presented in Supplementary Figures 6 and 7. I recommend utilizing a single map that displays the epistemic uncertainty for the entire grid, rather than dividing it into two separate plots for the upper and lower boundaries.  
+
+Thank you for the helpful suggestion. The new Figure 6 of the article now shows the PGA ShakeMap and its uncertainty in terms of standard deviation of the log PGA.  
+
+(3) Supplementary Figure 4c has caused some confusion. Epistemic uncertainty appears to be analogous to the standard error, which makes Figure 4d comprehensible, as it illustrates lower epistemic uncertainty in regions with a greater amount of data. However, if a location is represented by only one station or without data, the epistemic uncertainty should be approximated to \(\sigma /\sqrt{(\mathrm{N} = 1)}\) . Therefore, it is unclear why the smaller value \((\sim 0.05)\) is indicated in the region with only one station and 0.4 for the region without data in Figure 4c, which is unreasonable.  
+
+The uncertainty spatial pattern shown in Supplementary Figures 2(c), 2(d) and 2(b) (ex Figures 4 and 5) depends on three elements: the spatial distribution of the measurements, the measurement uncertainty and the spatial correlation. All maps are based on data: station data for 2(c), smartphone data for 2(d) and both station and smartphone data for 2(b).  
+
+In 2(c), the lowest uncertainty is at the spatial locations of the stations. This is because the station data has a low measurement uncertainty. The uncertainty then increases to 0.40 far from the stations, because far from the network the ShakeMap variance converges to the data marginal variance.  
+
+In 2(d), the lowest uncertainty is higher than 0.05 because the smartphone data have a higher measurement uncertainty. However, the network is denser and covers the red zone better. Thanks to the spatial correlation, the uncertainty never increases to 0.40 (it actually reaches 0.40 far from the smartphone network outside the red zone).  
+
+(4) Lines 221-222: Do the authors use the same correlation length (denoted as \(\theta\) in the paper) for both datasets? If so, this may not be reliable, as the data obtained from the smartphone is expected to exhibit a smaller correlation length.  
+
+Thanks for pointing this out. It is true that smartphones have a smaller correlation length, but this only happens at the observed peak smartphone accelerations. In our spatial models, the smartphone acceleration is described by 3 terms: the isotropic decay, the latent Gaussian
+
+<--- Page Split --->
+
+process, and the measurement error (which includes the variability induced by all factors affecting the smartphone measurement). 0 only describes the spatial correlation of the second term, which we use to model site amplification. The site amplification is the same for stations and smartphones, so the common 0 is not a strong assumption.  
+
+Furthermore, the common 0 is imposed by the spatial statistical model to obtain valid bivariate Gaussian processes (i.e. with positive definite correlation and cross- correlation spatial functions).  
+
+(5)What kind of metadata can we expect to be accessible if the smartphone datasets are eventually released to the public? There is considerable potential for smartphones to transform early warning systems and enhance post-seismic damage distribution maps. However, if the metadata is limited or not comprehensively available, its utility for researchers may be substantially diminished.  
+
+We plan to launch the service and release the high-resolution ShakeMaps in the near future, starting with Campi Flegrei. However, as smartphone location data is considered personal data, we first need to address potential privacy issues before openly publishing individual smartphone locations. For example, even if smartphone location data is published anonymously (i.e. without a smartphone ID), it can be used to identify which houses are occupied. And after a certain number of earthquake detections from the same area, some patterns of individual habits may emerge.  
+
+(6)It would be valuable to identify the specific components of the filtering procedure that have contributed to the highest percentage of discarded waveforms. In particular, what is the primary source of poor-quality waveforms in data recorded via smartphones?  
+
+The Earthquake Network app does not record waveforms but it directly provides the peak smartphone acceleration. Also, no data is discarded. The number of smartphone accelerations is lower than the number of users because, at any given time during the day, only a fraction of the smartphones are monitoring (those which are charging).  
+
+(7) Smartphone data can augment or build on free-field predictions, and with tremendous potential. It may be beneficial for future research to consider the relationship between MMI and smartphones.  
+
+The main advantage of smartphones is their large number, high density and spatial resolution. Their major weakness is the high uncertainty associated with each individual measurement. A strategy based on direct correlation between an individual smartphone measurement and another seismological parameter like MMI does not take advantage of the spatial density and will suffer from individual uncertainty. This is therefore not our strategy.  
+
+(8)Please include the phiS2S (between-site sigma) and phiSS (within-site sigma) values from both models in Table 1, in addition to the tau values.  
+
+In the original version of the article, we used the symbol tau to represent the standard deviation of the total residuals (log observation minus log prediction). We used this symbol to follow the formalism of Iervolino et al. (2024). In general, the symbol sigma is however used
+
+<--- Page Split --->
+
+for the standard deviation of the total residuals and tau is reserved for the between events variability. The reviewer therefore understood that we had evaluated the different components of variability (between- event, within- event, station- to- station) and that tau represented the between- event variability.  
+
+The calculation of the different components of variability is generally carried out for databases of several thousand recordings. Such a calculation therefore remains difficult here (which also explains why Iervolino et al. (2024) did not carry it out). Moreover, we do not see what the purpose of such a calculation would be. Indeed, we perform this calculation to verify that our model explains the data better than local models and therefore only a comparison with sigma (standard deviation of the total residuals) is possible.  
+
+We have therefore simply changed the notation and replaced tau with sigma to return to a more conventional notation.  
+
+(9)The paper should include the values pertaining to the variability of Inst derived from the two datasets (stations and smartphones).  
+
+This information is provided in the Supplementary Table 1 with the rest of the model parameters.  
+
+(10)Are the results obtained from smartphones within the epistemic uncertainty range of ergodic GMM? Please verify this and generate a corresponding plot. I expect that the results should lie within the epistemic uncertainty range, otherwise, this would suggest that they do not really help to improve the GMM.  
+
+The epistemic uncertainty of models for predicting seismic movements in volcanic areas is difficult to evaluate because there are few models, and these models are usually adjusted for each volcanic edifice. To answer the reviewer's question, we systematically compared our GMM in Equation (16) with the most recent GMM proposed in this region by Iervolino et al. (2024). The evaluation of performance is based on the comparison of the standard deviation of the residuals, as defined in engineering seismology (Log(observed) - Log(predicted)). These residuals are computed for all data, including the data used to calibrate the model, to follow the conventions of engineering seismologists. This comparison confirms that our model is performing better (the resulting sigma is lower, see Table 1).  
+
+Reviewer #2 (Remarks on code availability):  
+
+The record is publicly accessible, but files are restricted to users with access.  
+
+Code is now open. The access was restricted during the review process.  
+
+Reviewer #3 (Remarks to the Author):  
+
+The paper presents the integration of smartphone accelerometer data into seismic monitoring as a potential method to increase the spatial density of ground- motion observations. The study proposes an interesting concept, but uncertainties and methodological limitations reduce its reliability. The model serves as a useful exploratory tool; however, its ability to complement traditional recordings and site- response analysis
+
+<--- Page Split --->
+
+requires further investigation and validation. The approach has potential but necessitates additional verification steps before it can be considered fully reliable. I request a major revision.  
+
+The study relies on smartphone accelerometers, which introduce significant bias and uncertainty. Unlike dedicated seismic instruments, smartphones exhibit higher noise levels and variable recording conditions. Firstly, as mentioned in the article, smartphones are not anchored to the ground, and their placement within buildings is unknown, introducing variability in the recorded signals. Additionally, ambient noise and user activity may further affect the accuracy and reliability of the recorded measurements. Finally, it should be specified whether the type and quality of accelerometers vary significantly across different smartphone models. The methodology should explicitly quantify systematic errors introduced by these uncertainties. It should also specify what kind of processing is applied to smartphone data and the time required for such processing. Given that smartphone data are likely to be very noisy, a rigorous preprocessing step is necessary to make them comparable with seismic station accelerometer data.  
+
+The paper states that smartphone accelerations were not calibrated against seismic station measurements, relying instead on spatial statistical models. While the statistical approach may help extract patterns, it does not fully validate the accuracy of smartphone- based measurements. Before this type of measurement can be fully utilized, a detailed comparison with high- quality seismic records is necessary.  
+
+Thank you for raising these points. Over the last decade or so, many authors have focused on assessing the reliability of smartphones in seismic monitoring. This has been done in controlled experiments using shake tables, or by comparing smartphone and scientific- grade accelerometers during an earthquake with instruments available at the same site. All the tests done show that smartphone and accelerometer readings are comparable, at least for earthquake detection and preliminary quantification of the earthquake parameters (see D'Alessandro and D'Anna, 2013; Cascone et al., 2021). However, these results can only be partially extended to a citizen science network and real conditions, where, as mentioned by the reviewer, nothing is known about the following factors:  
+
+- The sensor characteristics (model, manufacturer, etc.) and its behaviour- The object above which the smartphone is placed- The smartphone case (its shape may affect the smartphone vibration)- The orientation of the smartphone in space (this information is actually retrievable, but with large uncertainty)- The smartphone location within the building- The building type  
+
+A transfer function relating the PGA to the peak smartphone acceleration (PSmA) certainly exists, but learning it to quantify biases and uncertainties is impractical. Note also that we have to protect the privacy of smartphone owners and we will never be able to document the factors listed above. Our strategy is different.  
+
+The role of our statistical methodology is to average the above factors, to allow the cross- correlation between PGAs and PSmA to emerge. This is possible thanks to:
+
+<--- Page Split --->
+
+- The relatively large number of smartphones- The use of a statistical model based on a Gaussian spatial process- The data collection from multiple events (with different contributing smartphones for each event)  
+
+Our method does not remove the bias. With a new analysis (results in Figure 3) we show that the bias between PGA and PSmA is even amplitude dependent. However, this bias only exists when station and smartphone measurements are compared in absolute value. Instead, we look for the cross- correlation when the isotropic decay is removed (independently for PGA and PSmA). If this correlation exists (i.e., if it is found during maximum likelihood model estimation), then the smartphone information contributes to the quantification of the site amplification and to the generation of the PGA ShakeMaps. The first key result confirming our strategy is the fact that this cross- correlation exists which means that despite the biases that can affect any single measure the number of smartphones is large enough to capture it.  
+
+Correlation does not imply causation, but when we use our amplification map as a site effect in a regional ground motion model (Equation 16), we show that the standard deviation of the residual/unexplained PGA decreases (using 34 events, not just the 4 used to estimate the amplification map). In our opinion, this is the second key result confirming our strategy and, for us, a validation of our amplification map (at least for the red zone of Campi Flegrei and for earthquakes in the same magnitude range) and a proof that our methodology is sound.  
+
+Finally, the data fusion of data with different biases and uncertainties is well consolidated in the statistical modelling literature and it is routinely used in many environmental applications (e.g., Berrocal et al., 2012; Nguyen et al., 2012; Di Curzio et al., 2021).  
+
+In the discussion, the following statement is made: "The methodology we have developed provides a solution that accounts for the bias in smartphone measurements relative to seismological station measurements and that relies on the consistent spatial cross- correlation between smartphone and station measurements rather than on the absolute accelerations recorded by the smartphones. The methodology thus achieves data fusion without requiring any calibration between smartphones and stations." I am not entirely convinced by this claim. The authors have shown that the soil amplification maps generated using smartphone and seismic station data correlate ( \(r = 0.84\) ). This suggests that smartphones can capture similar spatial patterns, but it does not prove that the measured accelerations are comparable in absolute value.  
+
+The new bias analysis that we carried out shows that PGA and PSmA are not comparable in absolute value. This is also found in other studies (e.g., Marcou et al., 2024). As explained above, however, our methodology does not rely on absolute values. Instead, it is important that, once the large- scale decay has been removed, station and smartphone measurements are locally cross- correlated. Cross- correlation means that, when a station residual (with respect to the decay) is above the average, the smartphone residual is also above the average. If this cross- correlation exists, then the smartphone information MAY be useful to improve the amplification map. That "MAY be useful" becomes "IS useful" once we validated the amplification map by showing that the use of the amplification maps is providing better
+
+<--- Page Split --->
+
+predictions compared to the most recent ground- motion models based on classical soil classes.  
+
+There is no direct comparison between smartphone and seismic station measurements at the exact locations or in close proximity. Therefore, stating that the methodology "accounts for the bias" seems somewhat overstated—while smartphone uncertainty has been modeled, without direct calibration between the two data sources, the accuracy of this correction remains uncertain.  
+
+What we actually meant was that the methodology does not require the bias to be included as a model term. For example, we have actually been able to quantify the bias, but this information is not used by our spatial statistical models. We have clarified this in the text.  
+
+Line 91—95: It is unclear how the spatial coverage of stations remained consistent across the four earthquakes, given that the number of smartphones varied between 56 and 441.  
+
+It is important that the area covered by stations and smartphones does not change too much. This is because the average amplifications to which we are comparing should be the same for all the events considered. The number of smartphones is less important.  
+
+Furthermore, it is not clear why only four earthquakes were used for calibration.  
+
+Although EQN detected many earthquakes in the red zone of Campi Flegrei, we selected only those events that guarantee a similar (and good) coverage of the red zone. The possibility of obtaining a good coverage of the red zone depends on the magnitude and epicentre of the earthquake, as well as on the geometry of the smartphone network at the time of the event. Note that the good coverage requirement only affects the learning of the amplification map, not the ability to generate a ShakeMap for any future event once the amplification map has been learned.  
+
+Additionally, the use of validation earthquakes listed in the supplementary table is not well explained.  
+
+Thanks, we have split the original table into two tables to improve clarity.  
+
+Figure S3: The log- amplification structure of the station data appears much more linear compared to the smartphone data, where red and blue values are adjacent to each other. This should be commented on.  
+
+This happens because any two smartphones may be close in space but the above listed factors may be very different. We commented this in the main article where the figure (now Figure 4) has been moved.  
+
+Figure S4 (panels c and d): It appears that the standard deviation is much lower in the case of smartphones than in seismic stations. I believe this is strongly influenced by the larger number of smartphone data points rather than an actual consistency between the measurements.
+
+<--- Page Split --->
+
+Yes, for panels c and d the standard deviation depends only on the network because the two log- amplification maps have been obtained independently. We commented this in the figure caption.  
+
+Amplification and geological structures: Can these amplification values (Figure 2 and Figures S3, S4, and S5) be associated with specific geological structures related to the caldera? This should be explored, particularly in light of Figure S9. It is not clear what the authors intend to convey with this figure.  
+
+We have explored this correlation. None of the existing geological maps (e.g., Vitale and Isaia, 2014) was showing a clear correlation between the observed amplification and the surface geology. Such results are not totally surprising given the complexity of geological structures in such a volcanic area, and the fact that amplification results from unknown and three dimensional geological structures at depth.  
+
+Figures S6 and S7: The uncertainties appear very small, considering the intrinsic uncertainties in using smartphone data. Would it be possible to compare them with the sigma of a Ground Motion Model (GMM)?  
+
+The comparison with the GMM is unfair (for the GMM) because the ShakeMap of Figure 6 is "event- specific" and based on measurements (from stations and smartphones), while ShakeMaps obtained by a classic GMM is based only on magnitude, epicentral distance and possibly depth (meaning that the event- specific decay or ground- motion anomalies are not taken into account). Also, the ShakeMap uncertainty is spatially varying and is affected by the station and smartphone networks, while the sigma of a GMM is a single number.  
+
+Regarding the higher spatial resolution of the acceleration map in Figure 3, it is important to remember that any additional data included in a map will modify it. What is missing in this work is a discussion on the reliability of the smartphone acceleration data.  
+
+We only partially agree with the first statement, as each map we generate (either the amplification map or the PGA ShakeMaps) is accompanied by the corresponding uncertainty map. Adding new data may change the map numerically, but not in a statistically significant way. This is particularly true for smartphone measurements: due to their high uncertainty, the contribution of a single measurement to the final map is generally very small. On the contrary, adding station measurements is likely to change the final map significantly.  
+
+In the article, we emphasise that when smartphone measurements are added, the final amplification map is significantly different from 1 (or 0 on the logarithmic scale) over a larger area of the red zone. For us, this is even more important than the high resolution of the map itself, as we get relevant information for a larger area (and thus more people and more buildings).  
+
+Finally, the spatial statistical model equations are such that, if the smartphone measurements were not reliable (i.e. not spatially cross- correlated with the station measurements), the smartphone information would not "flow" onto the PGA and the PGA ShakeMap would be based on station measurements only.
+
+<--- Page Split --->
+
+By the nature of smartphone data, I think it is essential to include it in the map by indicating uncertainty.  
+
+The uncertainty has been included in Figure 6b (ex Figure 3).  
+
+Ground Motion Model (GMM) Issues: I find it difficult to grasp the purpose and significance of the GMM developed in this study, as this analysis seems incomplete and lacks a thorough explanation of how it was constructed and its practical utility. Iervolino et al. (2024) developed a GMM specifically for the Campi Flegrei area using a dataset comparable to the one in this study with a functional equation of similar form. The sigma of their model should be compared with the amplification- derived sigma obtained in this study. It is essential to provide more details on the number of data points used, as Supplementary Table 1 does not include this information. If a GMM is to be proposed, it should be accompanied by a thorough discussion and comparison with existing models. Based on these concerns, the conclusion stating: "The amplification map issued from the fusion of seismological stations and smartphone records can also be used to recalibrate conventional GMMs and reduce the random variability of seismic motion prediction." should be reconsidered.  
+
+We thank the reviewer for this feedback. Our approach can be used for different purposes:  
+
+- The model in Equation (11) takes into account: 1) an event-specific decay, 2) the amplification map resulting from the lessons learned from past earthquakes (δS2S), and 3) the δW terms which represent the event-specific anomaly not explained neither by the isotropic decay nor by the site amplification (e.g. directivity). To use this model, the determination of the event magnitude is not needed (the ShakeMaps are only calibrated on observed PGA measured both by the seismological stations and the smartphones). This model, event-specific and magnitude agnostic, is then fully adapted to the needs of ShakeMaps generation when a new earthquake occurs (as illustrated on Figure 6).  
+
+- The GMM described in Equation (16) predicts ground-shaking according to magnitude, distance and the amplification map learned from past earthquakes, and it is used to predict the ground-shaking for any future event. The decay of this GMM is not event-dependent, but ShakeMaps are not isotropic because the amplification map is not (as illustrated on Figure 7).  
+
+This GMM was derived, first, to validate the smartphone- based amplification map, and second, to test the benefit of including the amplification map as a term of the GMM compared to classical GMMs used in this region (in particular the one recently derived by Iervolino et al., 2024). Our GMM is also useful for classical seismic hazard/risk analysis (as shown by Iervolino et al., 2024) to predict the effects of future earthquakes.  
+
+- As suggested by the reviewer, we have computed the sigma (standard deviation of the residuals between observations and predictions) for the GMM in Equation (16) considering 34 events restricted to the red zone and compared it with Iervolino et al. (2024). The sigma obtained by our model (0.36) is significantly lower than the sigma obtained by the Iervolino et al. (2024) model (0.41). However, the Iervolino et al. (2024) model was calibrated using different earthquakes that were partly outside our
+
+<--- Page Split --->
+
+study area. We then recalibrated the Iervolino et al. (2024) model using the same 34 earthquakes and seismological data as those used to calibrate our GMM. The recalibrated Iervolino et al. (2024) GMM, again shows a significantly lower performance than our model (sigma = 0.40), which, in our opinion, demonstrates the added value of the high- resolution amplification map derived in this paper compared to a classical amplification model based on soil classes.  
+
+These explanations and the fact that we propose various types of models adapted to different needs have been added to the paper and explained in the captions of Figure 6 and 7.  
+
+For future operational implementation, the model must be validated across multiple earthquakes. Additionally, its performance should be analyzed for various magnitudes, epicentral distances, and site conditions to ensure its robustness and reliability across different seismic scenarios.  
+
+We have been following the suggestion of the reviewer (explanation above) and tested the performance (sigma) for all INGV stations of the area and all available large earthquakes (34 events) by comparing the obtained sigma with the one of Iervolino et al. (2024).  
+
+The terminology in the text should be refined for greater precision and to avoid ambiguity. For instance, even in the abstract:  
+
+Instead of making broad statements like "the ubiquity of smartphones", the revised version should specify the data type being used and how it contributes to seismic monitoring.  
+
+We rewritten the specific sentence and we added the "Station and smartphone measurements" subsection in the Methods section to precisely describe the station and smartphone measurements and how they are collected.  
+
+The term "red zone" is officially used by the Italian Civil Protection Department to designate high- risk volcanic areas in the Campi Flegrei region. This term identifies areas subject to preventive evacuation measures for public safety in case of an eruption. While it can be retained, it should be clearly explained or cited. Otherwise, using a more specific term would enhance the scientific clarity and rigor of the text.  
+
+We have clarified the definition of the red zone and the reason for its existence. We have kept the term because it is well known to scientists and the public.  
+
+In scientific writing, stating that amplification varies from 0.25 to 2.83 is insufficient unless the reference measure is specified and the calculation method explained.  
+
+We agree: amplification factors depend on the chosen reference (a classical issue and discussion in engineering seismology). A seismic amplification map depends on the reference chosen. In our approach, site amplification and deamplification is relative to the average site amplification in the region. This differs from conventional amplification factors used in engineering seismology, where the site amplification is relative to a rock reference, but is very similar to amplification maps recently derived at the regional scale for European
+
+<--- Page Split --->
+
+risk models 7. This amplification map can only be compared with those of other regions if it has been verified that the average amplification effects due to subsurface layers are similar in both regions. Furthermore, the amplification values in our map can be rescaled relative to a specific station or subset of stations—provided these are included in the calibration dataset—thereby enabling more tailored interpretations. These explanations have been added in the discussion.  
+
+Additionally, the paper lacks proper references to other works dealing with the rapid quantification of ground motion using crowdsourced or smartphone- based techniques and similar challenges.  
+
+We have added the references of Kong et al. (2016) and Voosen (2021) on initiatives similar to the Earthquake Network. However, only in Marcou et al. (2024) is there a first attempt to model ground motion using smartphone data.  
+
+I am not entirely sure that the current manuscript structure is appropriate. A significant portion of the results and figures are in the supplementary material, making it difficult to fully grasp the study without it.  
+
+We moved some of the figures from the supplementary material to the main article.  
+
+Additionally, postponing the Methods section in this way does not facilitate comprehension.  
+
+We follow Nature Communication guidelines.  
+
+Line 128: Figure S9 is cited before S6, S7 and S8. Line 299: the correct citation should be number 20. Line 306: replace 'read' with 'red'. The caption of S8 has the wrong ID.  
+
+Fixed, thanks.  
+
+## REFERENCES  
+
+Berrocal, V.J., Gelfand, A.E. and Holland, D.M. Space- time data fusion under error in computer model output: an application to modeling air quality. Biometrics, 68, 837- 848 (2012).  
+
+Cascone, V., Boaga, J. and Cassiani, G. Small local earthquake detection using low- cost MEMS accelerometers: Examples in northern and central Italy. The Seismic Record, 1, 20- 26 (2021).  
+
+D'Alessandro, A. and D'Anna, G. Suitability of low- cost three- axis MEMS accelerometers in strong- motion seismology: Tests on the LIS331DLH (iPhone) accelerometer. Bulletin of the Seismological Society of America, 103, 2906- 2913 (2013).
+
+<--- Page Split --->
+
+Di Curzio, D., Castrignanò, A., Fountas, S., Romić, M. and Rossel, R.A.V. Multi- source data fusion of big spatial- temporal data in soil, geo- engineering and environmental studies. Science of the Total Environment, 788, 147842 (2021).  
+
+Kong, Q., Allen, R.M., Schreier, L. and Kwon, Y.W. MyShake: A smartphone seismic network for earthquake early warning and beyond. Science Advances, 2, 1501055 (2016).  
+
+Iervolino, I., Cito, P., De Falco, M., Festa, G., Herrmann, M., Lomax, A., ... & Zollo, A. Seismic risk mitigation at Campi Flegrei in volcanic unrest. Nature Communications 15, 1- 14 (2024).  
+
+Marcou, S., Allen, R.M., Abrahamson, N.A. and Sung, C.H. Ground- Motion Modeling Using MyShake Smartphone Peak Acceleration Data. Bulletin of the Seismological Society of America, 115, 86- 105 (2025).  
+
+Nguyen, H., Cressie, N. and Braverman, A. Spatial statistical data fusion for remote sensing applications. Journal of the American Statistical Association, 107, 1004- 1018 (2012).  
+
+Oliveti, I., Faenza, L. and Michelini, A. New reversible relationships between ground motion parameters and macrosismic intensity for Italy and their application in ShakeMap. Geophysical Journal International, 23, 1117- 1137 (2022).  
+
+Vitale, S. and Isaia, R. Fractures and faults in volcanic rocks (Campi Flegrei, southern Italy): insight into volcano- tectonic processes. International Journal of Earth Sciences, 103, 801- 819 (2014).  
+
+Voosen, P. New Google effort uses cellphones to detect earthquakes. Science, 48, 101721 (2021).
+
+<--- Page Split --->
+
+Dear Professor Finazzi,  
+
+Thank you again for submitting your manuscript "Citizens' smartphones unravel earthquake shaking in urban areas" to Nature Communications. We have now received reports from 2 reviewers and, based on their comments, we have decided to invite a revision of your work. Your revision should address all the points raised by our reviewers (see their reports below). Please also ensure to implement the discussion points raised by ref#1 in the last review round in the manuscript text, additionally to your reply in the response letter.  
+
+We have explained at line 451 of the new version of the manuscript that the model of Equation (14) is able to capture event- specific directivity or non- linear effects. Additionally, at line 43 we have explained why we are not considering crowdsourced felt reports in our current work.  
+
+When resubmitting, you must provide a point- by- point response to the reviewers' comments. Please show all changes in the manuscript text file with track changes or colour highlighting.  
+
+Along with the new manuscript file we submitted a track changes version.  
+
+If you are unable to address specific reviewer requests or find any points invalid, please explain why in the point- by- point response.  
+
+## REVIEWER COMMENTS  
+
+Reviewer #2 (Remarks to the Author):  
+
+Other modifications are acceptable, but I would like to address one point:  
+
+I understand that if the location has sufficient data, the epistemic uncertainty will show a lower value. However, how can a value lower than 0.1 be achieved when there are only a few records (fewer than 5)? This concerns my key question (Question 3). I am particularly interested in the phiS2S value; if phiS2S is around 0.2, it may be possible to reach a value below 0.1. However, the general phiS2S is approximately 0.3- 0.35, making obtaining a value under 0.1 almost impossible. I find it surprising that the authors state it is challenging to obtain phiS2S, but I understand if it is not available.  
+
+What we show in Supplementary Figure 3(c- d) is not the site- to- site variability (phiS2S as described in Al Atik et al., 2010) but rather the uncertainty of the Gaussian processes \(\delta \mathrm{S2S}\) given by Equations (8). The Gaussian process actually captures some of the residual variability (the spatially correlated part of the residuals with respect to the isotropic decay) that would otherwise be included in the uncertainty modelled by the terms \(\delta \mathrm{lnst}\) in Equations (1) and (2). We discussed this better in section "Smartphone and station data modelling" at line 338 and following.  
+
+The type of uncertainty that the reviewer likely has in mind (similar to classical GMPE uncertainty evaluation) is for instance expressed by \(\sigma\) in Table 1. When \(\delta \mathrm{S2S}\) is used in the GMM in Equation (19), the standard deviation (total aleatory sigma as expressed in
+
+<--- Page Split --->
+
+engineering seismology, as described by Al Atik et al., 2010) on the predicted PGA decreases from 0.4107 to 0.3569 (on the log10 scale).  
+
+We obtained a measure of site- to- site variability (similar to phiS2S as described by Al Atik et al., 2010) by calculating the standard deviation of the estimated residual variability at seismological stations, before and after adjusting for \(\delta \mathrm{S2S}\) (Equations 12 and 13, respectively). This resulted in a reduction of \(71.4\%\) , with the standard deviation moving from 0.6389 (0.2775 on the log10 scale) to 0.1827 (0.0793 on the log10 scale). This is discussed at line 176 and 437.  
+
+Atik, L. A., Abrahamson, N., Bommer, J. J., Scherbaum, F., Cotton, F., & Kuehn, N. (2010). The variability of ground- motion prediction models and its components. Seismological Research Letters, 81(5), 794- 801.  
+
+I believe the authors may underestimate the limitations associated with locations that have few records.  
+
+In areas with few records or where there are many uncertain records (e.g. only smartphone measurements and no station measurements), we can still estimate \(\delta \mathrm{S2S}\) , but the uncertainty is such that we cannot statistically claim that \(\delta \mathrm{S2S}\) is significantly different from the average amplification of the region. This has been explained at line 168 and following.  
+
+I suggest that the authors consider using equation (29) from Lavrentiadis et al. (2023) to address this issue or at least mention it in their paper. This would be beneficial since readers should be aware of potential improvements.  
+
+## Reference:  
+
+Lavrentiadis, G., Abrahamson, N.A. (2023). A non- ergodic spectral acceleration ground motion model for California developed with random vibration theory. Bull Earthquake Eng, 21, 5265- 5291. https://doi.org/10.1007/s10518- 023- 01689- 9  
+
+We have recovered the article, but could not find equation (29). We understand that the reviewer would like more information about the residual analysis performed in this paper, and this has been taken into account in the updated version (see the response to the first point).  
+
+Additionally, another new relevant paper from the US that can be cited is: "Ground- Motion Modeling Using MyShake Smartphone Peak Acceleration Data," BSSA, 2025, DOI: 10.1785/0120240209.  
+
+This paper has been on our reference list since the first draft of the manuscript was written. The confusion may lie in the year, which is given as 2024 in the BSSA journal.  
+
+## Reviewer #3 (Remarks to the Author):  
+
+Decision Recommendation: Minor Revision  
+
+General Recommendation:
+
+<--- Page Split --->
+
+The revised manuscript shows meaningful improvements in clarity, methodological transparency, and justification of the approach. The authors have responded constructively to most reviewer comments and addressed several concerns regarding the use of smartphone data, the treatment of bias, and the statistical modelling of amplification.  
+
+The integration of smartphone data remains an innovative and potentially impactful idea, and the manuscript now provides a more robust basis for such integration. The overall structure and figures have also improved, and key assumptions (e.g., spatial cross- correlation, non- reliance on calibration) are now better explained.  
+
+However, a few important clarifications and adjustments are still needed before the manuscript can be considered for publication.  
+
+Specific Requests (Required for Acceptance)  
+
+Clarify the spatial limitations of the amplification map, especially in areas with sparse smartphone data. It should be explicitly noted that fine- scale interpretation in such zones may be unreliable due to local uncertainty.  
+
+This point is now discussed at line 284 and following.  
+
+State more clearly the operational dependency of the method on the presence of a sufficiently dense and active (charging) smartphone network. The impact of uneven spatial distribution should be discussed more directly.  
+
+As emphasised at lines 168 and 284, uncertainty in the model output is as important as the output itself. An uneven spatial distribution of stations and/or smartphones does not present a methodological or algorithmic problem. In areas with a low density of measurements, the uncertainty in the model output is higher and it is less reliable. Therefore, it is important to be able to assess the uncertainty so that decision makers can make informed decisions based on both the model output and its uncertainty. In the updated version, we have emphasised the importance of uncertainty quantification and explained how uncertainty is quantified in more detail.  
+
+Include a table listing all seismic stations used, with a clear indication of whether they belong to the RAN or INGV networks. This will help readers assess the provenance and consistency of the data.  
+
+We recovered the station codes, and added Supplementary Figure 1 and Supplementary Dataset 1 which contain information on the networks and on which station detected which event.  
+
+Correct the typographical error ("shanking" \(\rightarrow\) "shaking" on line 320) and review consistency in terminology across sections.  
+
+Thanks.  
+
+## Conclusion  
+
+I recommend a minor revision, conditional on the authors addressing the above points. The manuscript is promising and can make a valuable contribution to the field if these final clarifications are incorporated.
+
+<--- Page Split --->
+
+The manuscript is very well written, presenting an interesting research topic and a comprehensive analysis. I believe that this research, along with similar studies, opens a new avenue for collecting seismological data, given the computing power currently available and the new algorithms, especially in machine learning, that enable us to manage large volumes of data. However, there are several questions that the authors need to address, which I outline below:  
+
+(1) Lines 305-306: "In equation (10) describes the potential amplification of the ground motion due to the site effects." How can it be confirmed that the amplification observed in this study is solely attributable to site effects, rather than being influenced by path effects (such as azimuth dependence, 3d velocity structure) or source effects (e.g., radiation patterns)?  
+
+(2) It is challenging to interpret the uncertainty presented in Supplementary Figures 6 and 7. I recommend utilizing a single map that displays the epistemic uncertainty for the entire grid, rather than dividing it into two separate plots for the upper and lower boundaries.  
+
+(3) Supplementary Figure 4c has caused some confusion. Epistemic uncertainty appears to be analogous to the standard error, which makes Figure 4d comprehensible, as it illustrates lower epistemic uncertainty in regions with a greater amount of data. However, if a location is represented by only one station or without data, the epistemic uncertainty should be approximated to \(\frac{\sigma}{\sqrt{N = 1}}\) . Therefore, it is unclear why the smaller value \((\sim 0.05)\) is indicated in the region with only one station and 0.4 for the region without data in Figure 4c, which is unreasonable.  
+
+(4) Lines 221-222: Do the authors use the same correlation length (denoted as \(\theta\) in the paper) for both datasets? If so, this may not be reliable, as the data obtained from the smartphone is expected to exhibit a smaller correlation length.  
+
+(5) What kind of metadata can we expect to be accessible if the smartphone datasets are eventually released to the public? There is considerable potential for smartphones to transform early warning systems and enhance post-seismic damage distribution maps. However, if the metadata is limited or not comprehensively available, its utility for researchers may be substantially
+
+<--- Page Split --->
+
+diminished.  
+
+(6) It would be valuable to identify the specific components of the filtering procedure that have contributed to the highest percentage of discarded waveforms. In particular, what is the primary source of poor-quality waveforms in data recorded via smartphones?  
+
+(7) In my view, smartphone data can augment or build-on free-field predictions, and with tremendous potential. It may be beneficial for future research to consider the relationship between MMI and smartphones.  
+
+(8) Please include the phiS2S (between-site sigma) and phiSS (within-site sigma) values from both models in Table 1, in addition to the tau values.  
+
+(9) The paper should include the values pertaining to the variability of Inst derived from the two datasets (stations and smartphones).  
+
+(10) Are the results obtained from smartphones within the epistemic uncertainty range of ergodic GMM? Please verify this and generate a corresponding plot. I expect that the results should lie within the epistemic uncertainty range, otherwise, this would suggest that they do not really help to improve the GMM.
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__7206d39aac3140f44125036e400400f2bc6214730597409b3eca3742f3fd110f/supplementary_0_Transparent Peer Review file__7206d39aac3140f44125036e400400f2bc6214730597409b3eca3742f3fd110f_det.mmd

@@ -0,0 +1,797 @@
+<|ref|>title<|/ref|><|det|>[[72, 53, 295, 80]]<|/det|>
+# nature portfolio  
+
+<|ref|>title<|/ref|><|det|>[[74, 96, 296, 118]]<|/det|>
+# Peer Review File  
+
+<|ref|>sub_title<|/ref|><|det|>[[72, 161, 899, 209]]<|/det|>
+## Citizens' smartphones unravel earthquake shaking in urban areas  
+
+<|ref|>text<|/ref|><|det|>[[73, 225, 521, 241]]<|/det|>
+Corresponding Author: Professor Francesco Finazzi  
+
+<|ref|>text<|/ref|><|det|>[[70, 274, 866, 289]]<|/det|>
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+<|ref|>text<|/ref|><|det|>[[70, 301, 890, 315]]<|/det|>
+Attachments originally included by the reviewers as part of their assessment can be found at the end of this file.  
+
+<|ref|>text<|/ref|><|det|>[[73, 353, 144, 366]]<|/det|>
+Version 0:  
+
+<|ref|>text<|/ref|><|det|>[[73, 379, 219, 393]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[73, 404, 160, 418]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[73, 431, 238, 444]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[73, 444, 910, 508]]<|/det|>
+This manuscript presents a methodology to use peak acceleration observations from smartphones to complement shaking intensity maps (called ShakeMaps by the seismological community). The approach is applied to the Campi Flegrei region around Naples, Italy. Four earthquakes from a vigorous swarm are used to constrain the site amplifications, and a site amplification map is derived and used to enhance ShakeMaps for the earthquakes. In future earthquakes, the same amplification could be applied to enhance ShakeMaps.  
+
+<|ref|>text<|/ref|><|det|>[[73, 520, 916, 574]]<|/det|>
+The use of dense smartphone data presents a real opportunity to add constrains on earthquake shaking. However, as the authors note, there are challenges calibrating what the smartphones are measuring compared to traditional free field instrumentation. Their solution is to assume a radial decay in ground motion for both the INGV stations and the phones, and then determine amplification factors relative to the best- fit radial models.  
+
+<|ref|>text<|/ref|><|det|>[[73, 586, 918, 653]]<|/det|>
+This is a good approach, assuming the ground motion is radial. This is an OK assumption for small events like those modeled M<4.5. But for larger earthquakes we do not expect radial ground motions due to the finiteness of the source and directivity effects. It would be helpful to add this issue to the discussion. In addition, for larger earthquakes the amplification factors likely do not scale linearly. To be fair, this issue is not accounted for in other ShakeMap creation approaches, but still it would be helpful to note this issue as a challenge.  
+
+<|ref|>text<|/ref|><|det|>[[73, 665, 913, 705]]<|/det|>
+Current ShakeMaps don't just use a ground motion model plus seismic station data, they also use felt reports. Did the authors consider how the presented models/amplification compares to models that also use felt reports? Do they correlate? If they do, this provides validation. If they don't why not? Again, perhaps worth discussion.  
+
+<|ref|>text<|/ref|><|det|>[[73, 717, 844, 744]]<|/det|>
+Finally a detail. Why not plot the peak amplitudes of the phone observations on the ShakeMaps (e.g. Fig 3 and in supplement)? You plot the station observations.  
+
+<|ref|>text<|/ref|><|det|>[[70, 756, 891, 783]]<|/det|>
+This is a very interesting study and shows the potential of phone- based observations. I recommend publication once the authors have considered the above comments and hopefully added to the discussion.  
+
+<|ref|>text<|/ref|><|det|>[[73, 795, 282, 808]]<|/det|>
+(Remarks on code availability)  
+
+<|ref|>text<|/ref|><|det|>[[73, 834, 161, 848]]<|/det|>
+Reviewer #2  
+
+<|ref|>text<|/ref|><|det|>[[73, 861, 238, 874]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[73, 875, 916, 927]]<|/det|>
+The manuscript is very well written, presenting an interesting research topic and a comprehensive analysis. I believe that this research, along with similar studies, opens a new avenue for collecting seismological data, given the computing power currently available and the new algorithms, especially in machine learning, that enable us to manage large volumes of data. However, there are several questions that the authors need to address, which I outline below:
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 46, 913, 88]]<|/det|>
+(1) Lines 305 – 306: “In equation (10) describes the potential amplification of the ground motion due to the site effects.” How can it be confirmed that the amplification observed in this study is solely attributable to site effects, rather than being influenced by path effects (such as azimuth dependence, 3d velocity structure) or source effects (e.g., radiation patterns)?  
+
+<|ref|>text<|/ref|><|det|>[[72, 99, 917, 140]]<|/det|>
+(2) It is challenging to interpret the uncertainty presented in Supplementary Figures 6 and 7. I recommend utilizing a single map that displays the epistemic uncertainty for the entire grid, rather than dividing it into two separate plots for the upper and lower boundaries.  
+
+<|ref|>text<|/ref|><|det|>[[72, 152, 911, 206]]<|/det|>
+(3) Supplementary Figure 4c has caused some confusion. Epistemic uncertainty appears to be analogous to the standard error, which makes Figure 4d comprehensible, as it illustrates lower epistemic uncertainty in regions with a greater amount of data. However, if a location is represented by only one station or without data, the epistemic uncertainty should be approximated to \(\sigma /N = 1\) .  
+
+<|ref|>text<|/ref|><|det|>[[72, 204, 905, 232]]<|/det|>
+Therefore, it is unclear why the smaller value \((\sim 0.05)\) is indicated in the region with only one station and 0.4 for the region without data in Figure 4c, which is unreasonable.  
+
+<|ref|>text<|/ref|><|det|>[[72, 242, 905, 271]]<|/det|>
+(4) Lines 221–222: Do the authors use the same correlation length (denoted as 0 in the paper) for both datasets? If so, this may not be reliable, as the data obtained from the smartphone is expected to exhibit a smaller correlation length.  
+
+<|ref|>text<|/ref|><|det|>[[72, 281, 910, 335]]<|/det|>
+(5) What kind of metadata can we expect to be accessible if the smartphone datasets are eventually released to the public? There is considerable potential for smartphones to transform early warning systems and enhance post-seismic damage distribution maps. However, if the metadata is limited or not comprehensively available, its utility for researchers may be substantially diminished.  
+
+<|ref|>text<|/ref|><|det|>[[72, 346, 914, 387]]<|/det|>
+(6) It would be valuable to identify the specific components of the filtering procedure that have contributed to the highest percentage of discarded waveforms. In particular, what is the primary source of poor-quality waveforms in data recorded via smartphones?  
+
+<|ref|>text<|/ref|><|det|>[[72, 397, 905, 425]]<|/det|>
+(7) Smartphone data can augment or build on free-field predictions, and with tremendous potential. It may be beneficial for future research to consider the relationship between MMI and smartphones.  
+
+<|ref|>text<|/ref|><|det|>[[72, 436, 897, 464]]<|/det|>
+(8) Please include the phiS2S (between-site sigma) and phiSS (within-site sigma) values from both models in Table 1, in addition to the tau values.  
+
+<|ref|>text<|/ref|><|det|>[[72, 475, 870, 504]]<|/det|>
+(9) The paper should include the values pertaining to the variability of Inst derived from the two datasets (stations and smartphones).  
+
+<|ref|>text<|/ref|><|det|>[[72, 514, 915, 556]]<|/det|>
+(10) Are the results obtained from smartphones within the epistemic uncertainty range of ergodic GMM? Please verify this and generate a corresponding plot. I expect that the results should lie within the epistemic uncertainty range, otherwise, this would suggest that they do not really help to improve the GMM.  
+
+<|ref|>text<|/ref|><|det|>[[72, 580, 285, 593]]<|/det|>
+(Remarks on code availability)  
+
+<|ref|>text<|/ref|><|det|>[[72, 593, 599, 606]]<|/det|>
+The record is publicly accessible, but files are restricted to users with access.  
+
+<|ref|>text<|/ref|><|det|>[[72, 618, 162, 631]]<|/det|>
+Reviewer #3  
+
+<|ref|>text<|/ref|><|det|>[[72, 645, 237, 658]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 658, 924, 739]]<|/det|>
+The paper presents the integration of smartphone accelerometer data into seismic monitoring as a potential method to increase the spatial density of ground- motion observations. The study proposes an interesting concept, but uncertainties and methodological limitations reduce its reliability. The model serves as a useful exploratory tool; however, its ability to complement traditional recordings and site- response analysis requires further investigation and validation. The approach has potential but necessitates additional verification steps before it can be considered fully reliable. I request a major revision.  
+
+<|ref|>text<|/ref|><|det|>[[72, 750, 920, 867]]<|/det|>
+The study relies on smartphone accelerometers, which introduce significant bias and uncertainty. Unlike dedicated seismic instruments, smartphones exhibit higher noise levels and variable recording conditions. Firstly, as mentioned in the article, smartphones are not anchored to the ground, and their placement within buildings is unknown, introducing variability in the recorded signals. Additionally, ambient noise and user activity may further affect the accuracy and reliability of the recorded measurements. Finally, it should be specified whether the type and quality of accelerometers vary significantly across different smartphone models. The methodology should explicitly quantify systematic errors introduced by these uncertainties. It should also specify what kind of processing is applied to smartphone data and the time required for such processing. Given that smartphone data are likely to be very noisy, a rigorous preprocessing step is necessary to make them comparable with seismic station accelerometer data.  
+
+<|ref|>text<|/ref|><|det|>[[72, 879, 920, 947]]<|/det|>
+The paper states that smartphone accelerations were not calibrated against seismic station measurements, relying instead on spatial statistical models. While the statistical approach may help extract patterns, it does not fully validate the accuracy of smartphone- based measurements. Before this type of measurement can be fully utilized, a detailed comparison with high- quality seismic records is necessary. In the discussion, the following statement is made: "The methodology we have developed provides a solution that accounts for the bias in smartphone measurements relative to seismological station
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 47, 920, 166]]<|/det|>
+measurements and that relies on the consistent spatial cross- correlation between smartphone and station measurements rather than on the absolute accelerations recorded by the smartphones. The methodology thus achieves data fusion without requiring any calibration between smartphones and stations." I am not entirely convinced by this claim. The authors have shown that the soil amplification maps generated using smartphone and seismic station data correlate \((r = 0.84)\) . This suggests that smartphones can capture similar spatial patterns, but it does not prove that the measured accelerations are comparable in absolute value. There is no direct comparison between smartphone and seismic station measurements at the exact locations or in close proximity. Therefore, stating that the methodology "accounts for the bias" seems somewhat overstated—while smartphone uncertainty has been modeled, without direct calibration between the two data sources, the accuracy of this correction remains uncertain.  
+
+<|ref|>text<|/ref|><|det|>[[72, 177, 920, 218]]<|/det|>
+Line 91—95: It is unclear how the spatial coverage of stations remained consistent across the four earthquakes, given that the number of smartphones varied between 56 and 441. Furthermore, it is not clear why only four earthquakes were used for calibration. Additionally, the use of validation earthquakes listed in the supplementary table is not well explained.  
+
+<|ref|>text<|/ref|><|det|>[[72, 229, 920, 283]]<|/det|>
+Figure S3: The log- amplification structure of the station data appears much more linear compared to the smartphone data, where red and blue values are adjacent to each other. This should be commented on. Figure S4 (panels c and d): It appears that the standard deviation is much lower in the case of smartphones than in seismic stations. I believe this is strongly influenced by the larger number of smartphone data points rather than an actual consistency between the measurements.  
+
+<|ref|>text<|/ref|><|det|>[[72, 294, 923, 335]]<|/det|>
+Amplification and geological structures: Can these amplification values (Figure 2 and Figures S3, S4, and S5) be associated with specific geological structures related to the caldera? This should be explored, particularly in light of Figure S9. It is not clear what the authors intend to convey with this figure.  
+
+<|ref|>text<|/ref|><|det|>[[72, 346, 915, 412]]<|/det|>
+Figures S6 and S7: The uncertainties appear very small, considering the intrinsic uncertainties in using smartphone data. Would it be possible to compare them with the sigma of a Ground Motion Model (GMM)? Regarding the higher spatial resolution of the acceleration map in Figure 3, it is important to remember that any additional data included in a map will modify it. What is missing in this work is a discussion on the reliability of the smartphone acceleration data. By the nature of smartphone data, I think it is essential to include it in the map by indicating uncertainty.  
+
+<|ref|>text<|/ref|><|det|>[[72, 423, 923, 543]]<|/det|>
+Ground Motion Model (GMM) Issues: I find it difficult to grasp the purpose and significance of the GMM developed in this study, as this analysis seems incomplete lacks a thorough explanation of how it was constructed and its practical utility. Iervolino et al. (2024) developed a GMM specifically for the Campi Flegrei area using a dataset comparable to the one in this study with a functional equation of similar form. The sigma of their model should be compared with the amplification- derived sigma obtained in this study. It is essential to provide more details on the number of data points used, as Supplementary Table 1 does not include this information. If a GMM is to be proposed, it should be accompanied by a thorough discussion and comparison with existing models. Based on these concerns, the conclusion stating: "The amplification map issued from the fusion of seismological stations and smartphone records can also be used to recalibrate conventional GMMs and reduce the random variability of seismic motion prediction." should be reconsidered.  
+
+<|ref|>text<|/ref|><|det|>[[72, 554, 911, 595]]<|/det|>
+For future operational implementation, the model must be validated across multiple earthquakes. Additionally, its performance should be analyzed for various magnitudes, epicentral distances, and site conditions to ensure its robustness and reliability across different seismic scenarios.  
+
+<|ref|>text<|/ref|><|det|>[[72, 606, 915, 621]]<|/det|>
+The terminology in the text should be refined for greater precision and to avoid ambiguity. For instance, even in the abstract:  
+
+<|ref|>text<|/ref|><|det|>[[72, 632, 884, 660]]<|/det|>
+Instead of making broad statements like "the ubiquity of smartphones", the revised version should specify the data type being used and how it contributes to seismic monitoring.  
+
+<|ref|>text<|/ref|><|det|>[[72, 671, 905, 725]]<|/det|>
+The term "red zone" is officially used by the Italian Civil Protection Department to designate high- risk volcanic areas in the Campi Flegrei region. This term identifies areas subject to preventive evacuation measures for public safety in case of an eruption. While it can be retained, it should be clearly explained or cited. Otherwise, using a more specific term would enhance the scientific clarity and rigor of the text.  
+
+<|ref|>text<|/ref|><|det|>[[72, 736, 918, 764]]<|/det|>
+In scientific writing, stating that amplification varies from 0.25 to 2.83 is insufficient unless the reference measure is specified and the calculation method explained.  
+
+<|ref|>text<|/ref|><|det|>[[72, 775, 904, 804]]<|/det|>
+Additionally, the paper lacks proper references to other works dealing with the rapid quantification of ground motion using crowdsourced or smartphone- based techniques and similar challenges.  
+
+<|ref|>text<|/ref|><|det|>[[72, 815, 920, 856]]<|/det|>
+I am not entirely sure that the current manuscript structure is appropriate. A significant portion of the results and figures are in the supplementary material, making it difficult to fully grasp the study without it. Additionally, postponing the Methods section in this way does not facilitate comprehension.  
+
+<|ref|>text<|/ref|><|det|>[[72, 867, 420, 920]]<|/det|>
+Line 128: Figure S9 is cited before S6, S7 and S8. Line 299: the correct citation should be number 20. Line 306: replace 'read' with 'red'. The caption of S8 has the wrong ID.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[73, 47, 283, 61]]<|/det|>
+(Remarks on code availability)  
+
+<|ref|>text<|/ref|><|det|>[[73, 86, 144, 99]]<|/det|>
+Version 1:  
+
+<|ref|>text<|/ref|><|det|>[[73, 112, 220, 125]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[73, 138, 162, 151]]<|/det|>
+Reviewer #2  
+
+<|ref|>text<|/ref|><|det|>[[73, 164, 228, 177]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[73, 177, 570, 191]]<|/det|>
+Other modifications are acceptable, but I would like to address one point:  
+
+<|ref|>text<|/ref|><|det|>[[73, 203, 916, 269]]<|/det|>
+I understand that if the location has sufficient data, the epistemic uncertainty will show a lower value. However, how can a value lower than 0.1 be achieved when there are only a few records (fewer than 5)? This concerns my key question (Question 3). I am particularly interested in the phiS2S value; if phiS2S is around 0.2, it may be possible to reach a value below 0.1. However, the general phiS2S is approximately 0.3- 0.35, making obtaining a value under 0.1 almost impossible. I find it surprising that the authors state it is challenging to obtain phiS2S, but I understand if it is not available.  
+
+<|ref|>text<|/ref|><|det|>[[73, 281, 916, 321]]<|/det|>
+I believe the authors may underestimate the limitations associated with locations that have few records. I suggest that the authors consider using equation (29) from Lavrentiadis et al. (2023) to address this issue or at least mention it in their paper. This would be beneficial since readers should be aware of potential improvements.  
+
+<|ref|>sub_title<|/ref|><|det|>[[73, 334, 150, 346]]<|/det|>
+## Reference:  
+
+<|ref|>text<|/ref|><|det|>[[73, 347, 916, 374]]<|/det|>
+Lavrentiadis, G., Abrahamson, N.A. (2023). A non- ergodic spectral acceleration ground motion model for California developed with random vibration theory. Bull Earthquake Eng, 21, 5265- 5291. https://doi.org/10.1007/s10518- 023- 01689- 9  
+
+<|ref|>text<|/ref|><|det|>[[73, 385, 868, 412]]<|/det|>
+Additionally, another new relevant paper from the US that can be cited is: "Ground- Motion Modeling Using MyShake Smartphone Peak Acceleration Data," BSSA, 2025, DOI: 10.1785/0120240209.  
+
+<|ref|>text<|/ref|><|det|>[[73, 424, 283, 438]]<|/det|>
+(Remarks on code availability)  
+
+<|ref|>text<|/ref|><|det|>[[73, 463, 162, 476]]<|/det|>
+Reviewer #3  
+
+<|ref|>text<|/ref|><|det|>[[73, 490, 223, 502]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[73, 503, 373, 515]]<|/det|>
+Decision Recommendation: Minor Revision  
+
+<|ref|>text<|/ref|><|det|>[[73, 516, 264, 528]]<|/det|>
+General Recommendation:  
+
+<|ref|>text<|/ref|><|det|>[[72, 528, 916, 625]]<|/det|>
+The revised manuscript shows meaningful improvements in clarity, methodological transparency, and justification of the approach. The authors have responded constructively to most reviewer comments and addressed several concerns regarding the use of smartphone data, the treatment of bias, and the statistical modelling of amplification. The integration of smartphone data remains an innovative and potentially impactful idea, and the manuscript now provides a more robust basis for such integration. The overall structure and figures have also improved, and key assumptions (e.g., spatial cross- correlation, non- reliance on calibration) are now better explained. However, a few important clarifications and adjustments are still needed before the manuscript can be considered for publication.  
+
+<|ref|>text<|/ref|><|det|>[[73, 625, 166, 638]]<|/det|>
+Publication.  
+
+<|ref|>text<|/ref|><|det|>[[73, 648, 384, 661]]<|/det|>
+Specific Requests (Required for Acceptance)  
+
+<|ref|>text<|/ref|><|det|>[[72, 672, 923, 699]]<|/det|>
+Clarify the spatial limitations of the amplification map, especially in areas with sparse smartphone data. It should be explicitly noted that fine- scale interpretation in such zones may be unreliable due to local uncertainty.  
+
+<|ref|>text<|/ref|><|det|>[[72, 710, 910, 737]]<|/det|>
+State more clearly the operational dependency of the method on the presence of a sufficiently dense and active (charging) smartphone network. The impact of uneven spatial distribution should be discussed more directly.  
+
+<|ref|>text<|/ref|><|det|>[[72, 749, 916, 777]]<|/det|>
+Include a table listing all seismic stations used, with a clear indication of whether they belong to the RAN or INGV networks. This will help readers assess the provenance and consistency of the data.  
+
+<|ref|>text<|/ref|><|det|>[[72, 777, 920, 789]]<|/det|>
+Correct the typographical error ("shanking" \(\rightarrow\) "shaking" on line 320) and review consistency in terminology across sections.  
+
+<|ref|>sub_title<|/ref|><|det|>[[73, 802, 153, 814]]<|/det|>
+## Conclusion  
+
+<|ref|>text<|/ref|><|det|>[[72, 815, 920, 841]]<|/det|>
+I recommend a minor revision, conditional on the authors addressing the above points. The manuscript is promising and can make a valuable contribution to the field if these final clarifications are incorporated.  
+
+<|ref|>text<|/ref|><|det|>[[73, 866, 283, 879]]<|/det|>
+(Remarks on code availability)
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 45, 916, 99]]<|/det|>
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+<|ref|>text<|/ref|><|det|>[[72, 99, 796, 113]]<|/det|>
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source.  
+
+<|ref|>text<|/ref|><|det|>[[72, 112, 910, 165]]<|/det|>
+The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.  
+
+<|ref|>text<|/ref|><|det|>[[72, 165, 618, 179]]<|/det|>
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[118, 83, 258, 100]]<|/det|>
+## AE COMMENTS  
+
+<|ref|>text<|/ref|><|det|>[[118, 117, 879, 256]]<|/det|>
+You will see that the reviewers are interested in your study but also raise points that require improvement. In particular, they raise the issue of the assumed radial decay of the ground motion as well as the amplification being solely attributed to site effects and request a more detailed analysis and discussion about the uncertainties and methodological limitations. We also would like to highlight that our format allows 5000 words in the main text (without Methods) and a flexible length for the Methods section as well as 10 display items (figures and/or tables). This should allow you to move essential information from the SI to the main text.  
+
+<|ref|>sub_title<|/ref|><|det|>[[119, 273, 330, 290]]<|/det|>
+## REVIEWER COMMENTS  
+
+<|ref|>text<|/ref|><|det|>[[120, 308, 430, 325]]<|/det|>
+Reviewer #1 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[118, 342, 879, 464]]<|/det|>
+This manuscript presents a methodology to use peak acceleration observations from smartphones to complement shaking intensity maps (called ShakeMaps by the seismological community). The approach is applied to the Campi Flegrei region around Naples, Italy. Four earthquakes from a vigorous swarm are used to constrain the site amplifications, and a site amplification map is derived and used to enhance ShakeMaps for the earthquakes. In future earthquakes, the same amplification could be applied to enhance ShakeMaps.  
+
+<|ref|>text<|/ref|><|det|>[[118, 480, 879, 567]]<|/det|>
+The use of dense smartphone data presents a real opportunity to add constrains on earthquake shaking. However, as the authors note, there are challenges calibrating what the smartphones are measuring compared to traditional free field instrumentation. Their solution is to assume a radial decay in ground motion for both the INGV stations and the phones, and then determine amplification factors relative to the best- fit radial models.  
+
+<|ref|>text<|/ref|><|det|>[[118, 583, 879, 688]]<|/det|>
+This is a good approach, assuming the ground motion is radial. This is an OK assumption for small events like those modeled \(M< 4.5\) . But for larger earthquakes we do not expect radial ground motions due to the finiteness of the source and directivity effects. It would be helpful to add this issue to the discussion. In addition, for larger earthquakes the amplification factors likely do not scale linearly. To be fair, this issue is not accounted for in other ShakeMap creation approaches, but still it would be helpful to note this issue as a challenge.  
+
+<|ref|>text<|/ref|><|det|>[[117, 704, 879, 844]]<|/det|>
+The reviewer is correct with the fact that site amplification emerges as a model residual from the fitted radial decay. The amplification map must then be "learned" using events that have radial decay (and thus small or moderate magnitude). However, for any new event in the same region, the ShakeMap estimation is based on the model in Equation (11) which takes into account: 1) the amplification map learned from past earthquakes (i.e., S2S), and 2) the SW terms which describe the ground- motions anomalies not explained neither by the isotropic decay nor by the site amplification. These terms capture potential event- specific non- isotopic patterns and therefore directivity effects and possible non- linear effects.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 82, 879, 240]]<|/det|>
+Of course, such directivity effects can not be predicted for future earthquakes and then isotropic GMM are classically used in Engineering seismology. We then derived in Equation (16) a GMM based on an isotropic decay and on our amplification map. ShakeMaps generated using Equation (16) are not isotropic because the amplification map is not. We compared our GMM to similar, classical and isotropic GMM used in this region (in particular the one derived recently by Irevolino et al., 2024) and, as discussed below (see the answer to Reviewer #3) our model is performing better than previously published models. Our GMM can also be used for classical seismic hazard/risk analysis (as shown by Irevolino et al., 2024).  
+
+<|ref|>text<|/ref|><|det|>[[118, 251, 878, 321]]<|/det|>
+Current ShakeMaps don't just use a ground motion model plus seismic station data, they also use felt reports. Did the authors consider how the presented models/amplification compares to models that also use felt reports? Do they correlate? If they do, this provides validation. If they don't why not? Again, perhaps worth discussion.  
+
+<|ref|>text<|/ref|><|det|>[[118, 334, 879, 456]]<|/det|>
+The addition of felt report intensities as a new data source would add considerable complexity to the statistical modelling. Indeed, intensities are provided on a discrete scale that cannot be modelled using continuous Gaussian processes. A simple option would be to convert intensities to PGA measurements using relationships available in the literature (e.g., Olivet et al., 2022) but the uncertainty of the resulting ShakeMap would be wrongly quantified. Furthermore, the range of observed intensities for the events considered in this article remains limited due to the relatively small magnitude of the earthquakes.  
+
+<|ref|>text<|/ref|><|det|>[[118, 470, 877, 506]]<|/det|>
+Finally a detail. Why not plot the peak amplitudes of the phone observations on the ShakeMaps (e.g. Fig 3 and in supplement)? You plot the station observations.  
+
+<|ref|>text<|/ref|><|det|>[[118, 518, 878, 571]]<|/det|>
+The smartphone peak accelerations are higher than the PGA due to the bias shown in Figure 2 (up to 3 times). If the smartphone measurements were plotted on the same colour scale, the PGA would be almost flat.  
+
+<|ref|>text<|/ref|><|det|>[[118, 584, 878, 638]]<|/det|>
+This is a very interesting study and shows the potential of phone- based observations. I recommend publication once the authors have considered the above comments and hopefully added to the discussion.  
+
+<|ref|>text<|/ref|><|det|>[[120, 654, 384, 671]]<|/det|>
+Thanks for the useful comments.  
+
+<|ref|>text<|/ref|><|det|>[[120, 689, 430, 706]]<|/det|>
+Reviewer #2 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[118, 722, 879, 828]]<|/det|>
+The manuscript is very well written, presenting an interesting research topic and a comprehensive analysis. I believe that this research, along with similar studies, opens a new avenue for collecting seismological data, given the computing power currently available and the new algorithms, especially in machine learning, that enable us to manage large volumes of data. However, there are several questions that the authors need to address, which I outline below:  
+
+<|ref|>text<|/ref|><|det|>[[118, 843, 878, 897]]<|/det|>
+(1) Lines 305 – 306: “In equation (10) describes the potential amplification of the ground motion due to the site effects.” How can it be confirmed that the amplification observed in this study is solely attributable to site effects, rather than being influenced by path effects
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 876, 118]]<|/det|>
+(such as azimuth dependence, 3d velocity structure) or source effects (e.g., radiation patterns)?  
+
+<|ref|>text<|/ref|><|det|>[[117, 135, 879, 291]]<|/det|>
+Separating source, path and site effects is a classical challenge of such studies. The amplification map developed here captures 3D site effects (if they are not azimuth dependent) which is an added value compared to classical site classification based solely on VS30. Such a "3D" amplification map is "learned" using many events and then event- specific path effects or radiation patterns have less and less influence when the number of learning events is increasing. Note also (as explained above) that the 8W terms which describe the ground- motions anomalies not explained neither by the isotropic decay nor by the site amplification capture potential event- specific non- isotopic patterns and therefore directivity effects and possible event- specific path effects.  
+
+<|ref|>text<|/ref|><|det|>[[118, 308, 878, 360]]<|/det|>
+(2) It is challenging to interpret the uncertainty presented in Supplementary Figures 6 and 7. I recommend utilizing a single map that displays the epistemic uncertainty for the entire grid, rather than dividing it into two separate plots for the upper and lower boundaries.  
+
+<|ref|>text<|/ref|><|det|>[[118, 377, 876, 411]]<|/det|>
+Thank you for the helpful suggestion. The new Figure 6 of the article now shows the PGA ShakeMap and its uncertainty in terms of standard deviation of the log PGA.  
+
+<|ref|>text<|/ref|><|det|>[[117, 428, 879, 550]]<|/det|>
+(3) Supplementary Figure 4c has caused some confusion. Epistemic uncertainty appears to be analogous to the standard error, which makes Figure 4d comprehensible, as it illustrates lower epistemic uncertainty in regions with a greater amount of data. However, if a location is represented by only one station or without data, the epistemic uncertainty should be approximated to \(\sigma /\sqrt{(\mathrm{N} = 1)}\) . Therefore, it is unclear why the smaller value \((\sim 0.05)\) is indicated in the region with only one station and 0.4 for the region without data in Figure 4c, which is unreasonable.  
+
+<|ref|>text<|/ref|><|det|>[[117, 567, 879, 635]]<|/det|>
+The uncertainty spatial pattern shown in Supplementary Figures 2(c), 2(d) and 2(b) (ex Figures 4 and 5) depends on three elements: the spatial distribution of the measurements, the measurement uncertainty and the spatial correlation. All maps are based on data: station data for 2(c), smartphone data for 2(d) and both station and smartphone data for 2(b).  
+
+<|ref|>text<|/ref|><|det|>[[117, 636, 878, 704]]<|/det|>
+In 2(c), the lowest uncertainty is at the spatial locations of the stations. This is because the station data has a low measurement uncertainty. The uncertainty then increases to 0.40 far from the stations, because far from the network the ShakeMap variance converges to the data marginal variance.  
+
+<|ref|>text<|/ref|><|det|>[[117, 706, 878, 774]]<|/det|>
+In 2(d), the lowest uncertainty is higher than 0.05 because the smartphone data have a higher measurement uncertainty. However, the network is denser and covers the red zone better. Thanks to the spatial correlation, the uncertainty never increases to 0.40 (it actually reaches 0.40 far from the smartphone network outside the red zone).  
+
+<|ref|>text<|/ref|><|det|>[[117, 791, 878, 843]]<|/det|>
+(4) Lines 221-222: Do the authors use the same correlation length (denoted as \(\theta\) in the paper) for both datasets? If so, this may not be reliable, as the data obtained from the smartphone is expected to exhibit a smaller correlation length.  
+
+<|ref|>text<|/ref|><|det|>[[117, 859, 878, 910]]<|/det|>
+Thanks for pointing this out. It is true that smartphones have a smaller correlation length, but this only happens at the observed peak smartphone accelerations. In our spatial models, the smartphone acceleration is described by 3 terms: the isotropic decay, the latent Gaussian
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 83, 878, 152]]<|/det|>
+process, and the measurement error (which includes the variability induced by all factors affecting the smartphone measurement). 0 only describes the spatial correlation of the second term, which we use to model site amplification. The site amplification is the same for stations and smartphones, so the common 0 is not a strong assumption.  
+
+<|ref|>text<|/ref|><|det|>[[118, 167, 878, 218]]<|/det|>
+Furthermore, the common 0 is imposed by the spatial statistical model to obtain valid bivariate Gaussian processes (i.e. with positive definite correlation and cross- correlation spatial functions).  
+
+<|ref|>text<|/ref|><|det|>[[118, 232, 878, 319]]<|/det|>
+(5)What kind of metadata can we expect to be accessible if the smartphone datasets are eventually released to the public? There is considerable potential for smartphones to transform early warning systems and enhance post-seismic damage distribution maps. However, if the metadata is limited or not comprehensively available, its utility for researchers may be substantially diminished.  
+
+<|ref|>text<|/ref|><|det|>[[117, 333, 879, 454]]<|/det|>
+We plan to launch the service and release the high-resolution ShakeMaps in the near future, starting with Campi Flegrei. However, as smartphone location data is considered personal data, we first need to address potential privacy issues before openly publishing individual smartphone locations. For example, even if smartphone location data is published anonymously (i.e. without a smartphone ID), it can be used to identify which houses are occupied. And after a certain number of earthquake detections from the same area, some patterns of individual habits may emerge.  
+
+<|ref|>text<|/ref|><|det|>[[118, 468, 878, 520]]<|/det|>
+(6)It would be valuable to identify the specific components of the filtering procedure that have contributed to the highest percentage of discarded waveforms. In particular, what is the primary source of poor-quality waveforms in data recorded via smartphones?  
+
+<|ref|>text<|/ref|><|det|>[[118, 536, 878, 606]]<|/det|>
+The Earthquake Network app does not record waveforms but it directly provides the peak smartphone acceleration. Also, no data is discarded. The number of smartphone accelerations is lower than the number of users because, at any given time during the day, only a fraction of the smartphones are monitoring (those which are charging).  
+
+<|ref|>text<|/ref|><|det|>[[118, 622, 878, 674]]<|/det|>
+(7) Smartphone data can augment or build on free-field predictions, and with tremendous potential. It may be beneficial for future research to consider the relationship between MMI and smartphones.  
+
+<|ref|>text<|/ref|><|det|>[[118, 692, 878, 778]]<|/det|>
+The main advantage of smartphones is their large number, high density and spatial resolution. Their major weakness is the high uncertainty associated with each individual measurement. A strategy based on direct correlation between an individual smartphone measurement and another seismological parameter like MMI does not take advantage of the spatial density and will suffer from individual uncertainty. This is therefore not our strategy.  
+
+<|ref|>text<|/ref|><|det|>[[115, 794, 876, 829]]<|/det|>
+(8)Please include the phiS2S (between-site sigma) and phiSS (within-site sigma) values from both models in Table 1, in addition to the tau values.  
+
+<|ref|>text<|/ref|><|det|>[[118, 848, 878, 900]]<|/det|>
+In the original version of the article, we used the symbol tau to represent the standard deviation of the total residuals (log observation minus log prediction). We used this symbol to follow the formalism of Iervolino et al. (2024). In general, the symbol sigma is however used
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 83, 878, 152]]<|/det|>
+for the standard deviation of the total residuals and tau is reserved for the between events variability. The reviewer therefore understood that we had evaluated the different components of variability (between- event, within- event, station- to- station) and that tau represented the between- event variability.  
+
+<|ref|>text<|/ref|><|det|>[[117, 169, 879, 270]]<|/det|>
+The calculation of the different components of variability is generally carried out for databases of several thousand recordings. Such a calculation therefore remains difficult here (which also explains why Iervolino et al. (2024) did not carry it out). Moreover, we do not see what the purpose of such a calculation would be. Indeed, we perform this calculation to verify that our model explains the data better than local models and therefore only a comparison with sigma (standard deviation of the total residuals) is possible.  
+
+<|ref|>text<|/ref|><|det|>[[118, 272, 878, 306]]<|/det|>
+We have therefore simply changed the notation and replaced tau with sigma to return to a more conventional notation.  
+
+<|ref|>text<|/ref|><|det|>[[118, 324, 878, 359]]<|/det|>
+(9)The paper should include the values pertaining to the variability of Inst derived from the two datasets (stations and smartphones).  
+
+<|ref|>text<|/ref|><|det|>[[118, 376, 877, 411]]<|/det|>
+This information is provided in the Supplementary Table 1 with the rest of the model parameters.  
+
+<|ref|>text<|/ref|><|det|>[[118, 428, 879, 498]]<|/det|>
+(10)Are the results obtained from smartphones within the epistemic uncertainty range of ergodic GMM? Please verify this and generate a corresponding plot. I expect that the results should lie within the epistemic uncertainty range, otherwise, this would suggest that they do not really help to improve the GMM.  
+
+<|ref|>text<|/ref|><|det|>[[117, 514, 879, 671]]<|/det|>
+The epistemic uncertainty of models for predicting seismic movements in volcanic areas is difficult to evaluate because there are few models, and these models are usually adjusted for each volcanic edifice. To answer the reviewer's question, we systematically compared our GMM in Equation (16) with the most recent GMM proposed in this region by Iervolino et al. (2024). The evaluation of performance is based on the comparison of the standard deviation of the residuals, as defined in engineering seismology (Log(observed) - Log(predicted)). These residuals are computed for all data, including the data used to calibrate the model, to follow the conventions of engineering seismologists. This comparison confirms that our model is performing better (the resulting sigma is lower, see Table 1).  
+
+<|ref|>text<|/ref|><|det|>[[118, 688, 480, 704]]<|/det|>
+Reviewer #2 (Remarks on code availability):  
+
+<|ref|>text<|/ref|><|det|>[[118, 722, 748, 739]]<|/det|>
+The record is publicly accessible, but files are restricted to users with access.  
+
+<|ref|>text<|/ref|><|det|>[[118, 756, 710, 773]]<|/det|>
+Code is now open. The access was restricted during the review process.  
+
+<|ref|>text<|/ref|><|det|>[[118, 791, 430, 807]]<|/det|>
+Reviewer #3 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[118, 826, 879, 912]]<|/det|>
+The paper presents the integration of smartphone accelerometer data into seismic monitoring as a potential method to increase the spatial density of ground- motion observations. The study proposes an interesting concept, but uncertainties and methodological limitations reduce its reliability. The model serves as a useful exploratory tool; however, its ability to complement traditional recordings and site- response analysis
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 83, 878, 135]]<|/det|>
+requires further investigation and validation. The approach has potential but necessitates additional verification steps before it can be considered fully reliable. I request a major revision.  
+
+<|ref|>text<|/ref|><|det|>[[117, 153, 879, 359]]<|/det|>
+The study relies on smartphone accelerometers, which introduce significant bias and uncertainty. Unlike dedicated seismic instruments, smartphones exhibit higher noise levels and variable recording conditions. Firstly, as mentioned in the article, smartphones are not anchored to the ground, and their placement within buildings is unknown, introducing variability in the recorded signals. Additionally, ambient noise and user activity may further affect the accuracy and reliability of the recorded measurements. Finally, it should be specified whether the type and quality of accelerometers vary significantly across different smartphone models. The methodology should explicitly quantify systematic errors introduced by these uncertainties. It should also specify what kind of processing is applied to smartphone data and the time required for such processing. Given that smartphone data are likely to be very noisy, a rigorous preprocessing step is necessary to make them comparable with seismic station accelerometer data.  
+
+<|ref|>text<|/ref|><|det|>[[118, 360, 878, 445]]<|/det|>
+The paper states that smartphone accelerations were not calibrated against seismic station measurements, relying instead on spatial statistical models. While the statistical approach may help extract patterns, it does not fully validate the accuracy of smartphone- based measurements. Before this type of measurement can be fully utilized, a detailed comparison with high- quality seismic records is necessary.  
+
+<|ref|>text<|/ref|><|det|>[[117, 460, 879, 615]]<|/det|>
+Thank you for raising these points. Over the last decade or so, many authors have focused on assessing the reliability of smartphones in seismic monitoring. This has been done in controlled experiments using shake tables, or by comparing smartphone and scientific- grade accelerometers during an earthquake with instruments available at the same site. All the tests done show that smartphone and accelerometer readings are comparable, at least for earthquake detection and preliminary quantification of the earthquake parameters (see D'Alessandro and D'Anna, 2013; Cascone et al., 2021). However, these results can only be partially extended to a citizen science network and real conditions, where, as mentioned by the reviewer, nothing is known about the following factors:  
+
+<|ref|>text<|/ref|><|det|>[[147, 630, 878, 751]]<|/det|>
+- The sensor characteristics (model, manufacturer, etc.) and its behaviour- The object above which the smartphone is placed- The smartphone case (its shape may affect the smartphone vibration)- The orientation of the smartphone in space (this information is actually retrievable, but with large uncertainty)- The smartphone location within the building- The building type  
+
+<|ref|>text<|/ref|><|det|>[[118, 765, 878, 834]]<|/det|>
+A transfer function relating the PGA to the peak smartphone acceleration (PSmA) certainly exists, but learning it to quantify biases and uncertainties is impractical. Note also that we have to protect the privacy of smartphone owners and we will never be able to document the factors listed above. Our strategy is different.  
+
+<|ref|>text<|/ref|><|det|>[[118, 848, 878, 883]]<|/det|>
+The role of our statistical methodology is to average the above factors, to allow the cross- correlation between PGAs and PSmA to emerge. This is possible thanks to:
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[148, 84, 879, 153]]<|/det|>
+- The relatively large number of smartphones- The use of a statistical model based on a Gaussian spatial process- The data collection from multiple events (with different contributing smartphones for each event)  
+
+<|ref|>text<|/ref|><|det|>[[117, 167, 880, 340]]<|/det|>
+Our method does not remove the bias. With a new analysis (results in Figure 3) we show that the bias between PGA and PSmA is even amplitude dependent. However, this bias only exists when station and smartphone measurements are compared in absolute value. Instead, we look for the cross- correlation when the isotropic decay is removed (independently for PGA and PSmA). If this correlation exists (i.e., if it is found during maximum likelihood model estimation), then the smartphone information contributes to the quantification of the site amplification and to the generation of the PGA ShakeMaps. The first key result confirming our strategy is the fact that this cross- correlation exists which means that despite the biases that can affect any single measure the number of smartphones is large enough to capture it.  
+
+<|ref|>text<|/ref|><|det|>[[118, 354, 879, 457]]<|/det|>
+Correlation does not imply causation, but when we use our amplification map as a site effect in a regional ground motion model (Equation 16), we show that the standard deviation of the residual/unexplained PGA decreases (using 34 events, not just the 4 used to estimate the amplification map). In our opinion, this is the second key result confirming our strategy and, for us, a validation of our amplification map (at least for the red zone of Campi Flegrei and for earthquakes in the same magnitude range) and a proof that our methodology is sound.  
+
+<|ref|>text<|/ref|><|det|>[[118, 470, 879, 523]]<|/det|>
+Finally, the data fusion of data with different biases and uncertainties is well consolidated in the statistical modelling literature and it is routinely used in many environmental applications (e.g., Berrocal et al., 2012; Nguyen et al., 2012; Di Curzio et al., 2021).  
+
+<|ref|>text<|/ref|><|det|>[[118, 536, 880, 711]]<|/det|>
+In the discussion, the following statement is made: "The methodology we have developed provides a solution that accounts for the bias in smartphone measurements relative to seismological station measurements and that relies on the consistent spatial cross- correlation between smartphone and station measurements rather than on the absolute accelerations recorded by the smartphones. The methodology thus achieves data fusion without requiring any calibration between smartphones and stations." I am not entirely convinced by this claim. The authors have shown that the soil amplification maps generated using smartphone and seismic station data correlate ( \(r = 0.84\) ). This suggests that smartphones can capture similar spatial patterns, but it does not prove that the measured accelerations are comparable in absolute value.  
+
+<|ref|>text<|/ref|><|det|>[[118, 727, 880, 883]]<|/det|>
+The new bias analysis that we carried out shows that PGA and PSmA are not comparable in absolute value. This is also found in other studies (e.g., Marcou et al., 2024). As explained above, however, our methodology does not rely on absolute values. Instead, it is important that, once the large- scale decay has been removed, station and smartphone measurements are locally cross- correlated. Cross- correlation means that, when a station residual (with respect to the decay) is above the average, the smartphone residual is also above the average. If this cross- correlation exists, then the smartphone information MAY be useful to improve the amplification map. That "MAY be useful" becomes "IS useful" once we validated the amplification map by showing that the use of the amplification maps is providing better
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 84, 877, 118]]<|/det|>
+predictions compared to the most recent ground- motion models based on classical soil classes.  
+
+<|ref|>text<|/ref|><|det|>[[118, 135, 878, 221]]<|/det|>
+There is no direct comparison between smartphone and seismic station measurements at the exact locations or in close proximity. Therefore, stating that the methodology "accounts for the bias" seems somewhat overstated—while smartphone uncertainty has been modeled, without direct calibration between the two data sources, the accuracy of this correction remains uncertain.  
+
+<|ref|>text<|/ref|><|det|>[[118, 236, 878, 288]]<|/det|>
+What we actually meant was that the methodology does not require the bias to be included as a model term. For example, we have actually been able to quantify the bias, but this information is not used by our spatial statistical models. We have clarified this in the text.  
+
+<|ref|>text<|/ref|><|det|>[[118, 302, 878, 337]]<|/det|>
+Line 91—95: It is unclear how the spatial coverage of stations remained consistent across the four earthquakes, given that the number of smartphones varied between 56 and 441.  
+
+<|ref|>text<|/ref|><|det|>[[118, 351, 878, 403]]<|/det|>
+It is important that the area covered by stations and smartphones does not change too much. This is because the average amplifications to which we are comparing should be the same for all the events considered. The number of smartphones is less important.  
+
+<|ref|>text<|/ref|><|det|>[[118, 417, 765, 434]]<|/det|>
+Furthermore, it is not clear why only four earthquakes were used for calibration.  
+
+<|ref|>text<|/ref|><|det|>[[118, 448, 879, 569]]<|/det|>
+Although EQN detected many earthquakes in the red zone of Campi Flegrei, we selected only those events that guarantee a similar (and good) coverage of the red zone. The possibility of obtaining a good coverage of the red zone depends on the magnitude and epicentre of the earthquake, as well as on the geometry of the smartphone network at the time of the event. Note that the good coverage requirement only affects the learning of the amplification map, not the ability to generate a ShakeMap for any future event once the amplification map has been learned.  
+
+<|ref|>text<|/ref|><|det|>[[118, 583, 877, 618]]<|/det|>
+Additionally, the use of validation earthquakes listed in the supplementary table is not well explained.  
+
+<|ref|>text<|/ref|><|det|>[[118, 635, 708, 652]]<|/det|>
+Thanks, we have split the original table into two tables to improve clarity.  
+
+<|ref|>text<|/ref|><|det|>[[118, 669, 878, 721]]<|/det|>
+Figure S3: The log- amplification structure of the station data appears much more linear compared to the smartphone data, where red and blue values are adjacent to each other. This should be commented on.  
+
+<|ref|>text<|/ref|><|det|>[[118, 739, 878, 790]]<|/det|>
+This happens because any two smartphones may be close in space but the above listed factors may be very different. We commented this in the main article where the figure (now Figure 4) has been moved.  
+
+<|ref|>text<|/ref|><|det|>[[118, 808, 878, 877]]<|/det|>
+Figure S4 (panels c and d): It appears that the standard deviation is much lower in the case of smartphones than in seismic stations. I believe this is strongly influenced by the larger number of smartphone data points rather than an actual consistency between the measurements.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 878, 136]]<|/det|>
+Yes, for panels c and d the standard deviation depends only on the network because the two log- amplification maps have been obtained independently. We commented this in the figure caption.  
+
+<|ref|>text<|/ref|><|det|>[[118, 150, 878, 219]]<|/det|>
+Amplification and geological structures: Can these amplification values (Figure 2 and Figures S3, S4, and S5) be associated with specific geological structures related to the caldera? This should be explored, particularly in light of Figure S9. It is not clear what the authors intend to convey with this figure.  
+
+<|ref|>text<|/ref|><|det|>[[118, 236, 878, 322]]<|/det|>
+We have explored this correlation. None of the existing geological maps (e.g., Vitale and Isaia, 2014) was showing a clear correlation between the observed amplification and the surface geology. Such results are not totally surprising given the complexity of geological structures in such a volcanic area, and the fact that amplification results from unknown and three dimensional geological structures at depth.  
+
+<|ref|>text<|/ref|><|det|>[[118, 339, 878, 391]]<|/det|>
+Figures S6 and S7: The uncertainties appear very small, considering the intrinsic uncertainties in using smartphone data. Would it be possible to compare them with the sigma of a Ground Motion Model (GMM)?  
+
+<|ref|>text<|/ref|><|det|>[[118, 408, 879, 512]]<|/det|>
+The comparison with the GMM is unfair (for the GMM) because the ShakeMap of Figure 6 is "event- specific" and based on measurements (from stations and smartphones), while ShakeMaps obtained by a classic GMM is based only on magnitude, epicentral distance and possibly depth (meaning that the event- specific decay or ground- motion anomalies are not taken into account). Also, the ShakeMap uncertainty is spatially varying and is affected by the station and smartphone networks, while the sigma of a GMM is a single number.  
+
+<|ref|>text<|/ref|><|det|>[[118, 529, 878, 581]]<|/det|>
+Regarding the higher spatial resolution of the acceleration map in Figure 3, it is important to remember that any additional data included in a map will modify it. What is missing in this work is a discussion on the reliability of the smartphone acceleration data.  
+
+<|ref|>text<|/ref|><|det|>[[118, 596, 879, 699]]<|/det|>
+We only partially agree with the first statement, as each map we generate (either the amplification map or the PGA ShakeMaps) is accompanied by the corresponding uncertainty map. Adding new data may change the map numerically, but not in a statistically significant way. This is particularly true for smartphone measurements: due to their high uncertainty, the contribution of a single measurement to the final map is generally very small. On the contrary, adding station measurements is likely to change the final map significantly.  
+
+<|ref|>text<|/ref|><|det|>[[118, 714, 878, 800]]<|/det|>
+In the article, we emphasise that when smartphone measurements are added, the final amplification map is significantly different from 1 (or 0 on the logarithmic scale) over a larger area of the red zone. For us, this is even more important than the high resolution of the map itself, as we get relevant information for a larger area (and thus more people and more buildings).  
+
+<|ref|>text<|/ref|><|det|>[[118, 814, 878, 883]]<|/det|>
+Finally, the spatial statistical model equations are such that, if the smartphone measurements were not reliable (i.e. not spatially cross- correlated with the station measurements), the smartphone information would not "flow" onto the PGA and the PGA ShakeMap would be based on station measurements only.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 83, 877, 118]]<|/det|>
+By the nature of smartphone data, I think it is essential to include it in the map by indicating uncertainty.  
+
+<|ref|>text<|/ref|><|det|>[[118, 135, 618, 152]]<|/det|>
+The uncertainty has been included in Figure 6b (ex Figure 3).  
+
+<|ref|>text<|/ref|><|det|>[[117, 169, 880, 377]]<|/det|>
+Ground Motion Model (GMM) Issues: I find it difficult to grasp the purpose and significance of the GMM developed in this study, as this analysis seems incomplete and lacks a thorough explanation of how it was constructed and its practical utility. Iervolino et al. (2024) developed a GMM specifically for the Campi Flegrei area using a dataset comparable to the one in this study with a functional equation of similar form. The sigma of their model should be compared with the amplification- derived sigma obtained in this study. It is essential to provide more details on the number of data points used, as Supplementary Table 1 does not include this information. If a GMM is to be proposed, it should be accompanied by a thorough discussion and comparison with existing models. Based on these concerns, the conclusion stating: "The amplification map issued from the fusion of seismological stations and smartphone records can also be used to recalibrate conventional GMMs and reduce the random variability of seismic motion prediction." should be reconsidered.  
+
+<|ref|>text<|/ref|><|det|>[[118, 394, 857, 411]]<|/det|>
+We thank the reviewer for this feedback. Our approach can be used for different purposes:  
+
+<|ref|>text<|/ref|><|det|>[[147, 428, 880, 584]]<|/det|>
+- The model in Equation (11) takes into account: 1) an event-specific decay, 2) the amplification map resulting from the lessons learned from past earthquakes (δS2S), and 3) the δW terms which represent the event-specific anomaly not explained neither by the isotropic decay nor by the site amplification (e.g. directivity). To use this model, the determination of the event magnitude is not needed (the ShakeMaps are only calibrated on observed PGA measured both by the seismological stations and the smartphones). This model, event-specific and magnitude agnostic, is then fully adapted to the needs of ShakeMaps generation when a new earthquake occurs (as illustrated on Figure 6).  
+
+<|ref|>text<|/ref|><|det|>[[148, 601, 880, 686]]<|/det|>
+- The GMM described in Equation (16) predicts ground-shaking according to magnitude, distance and the amplification map learned from past earthquakes, and it is used to predict the ground-shaking for any future event. The decay of this GMM is not event-dependent, but ShakeMaps are not isotropic because the amplification map is not (as illustrated on Figure 7).  
+
+<|ref|>text<|/ref|><|det|>[[177, 688, 879, 790]]<|/det|>
+This GMM was derived, first, to validate the smartphone- based amplification map, and second, to test the benefit of including the amplification map as a term of the GMM compared to classical GMMs used in this region (in particular the one recently derived by Iervolino et al., 2024). Our GMM is also useful for classical seismic hazard/risk analysis (as shown by Iervolino et al., 2024) to predict the effects of future earthquakes.  
+
+<|ref|>text<|/ref|><|det|>[[149, 809, 880, 912]]<|/det|>
+- As suggested by the reviewer, we have computed the sigma (standard deviation of the residuals between observations and predictions) for the GMM in Equation (16) considering 34 events restricted to the red zone and compared it with Iervolino et al. (2024). The sigma obtained by our model (0.36) is significantly lower than the sigma obtained by the Iervolino et al. (2024) model (0.41). However, the Iervolino et al. (2024) model was calibrated using different earthquakes that were partly outside our
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[178, 83, 879, 187]]<|/det|>
+study area. We then recalibrated the Iervolino et al. (2024) model using the same 34 earthquakes and seismological data as those used to calibrate our GMM. The recalibrated Iervolino et al. (2024) GMM, again shows a significantly lower performance than our model (sigma = 0.40), which, in our opinion, demonstrates the added value of the high- resolution amplification map derived in this paper compared to a classical amplification model based on soil classes.  
+
+<|ref|>text<|/ref|><|det|>[[118, 204, 878, 256]]<|/det|>
+These explanations and the fact that we propose various types of models adapted to different needs have been added to the paper and explained in the captions of Figure 6 and 7.  
+
+<|ref|>text<|/ref|><|det|>[[118, 273, 878, 343]]<|/det|>
+For future operational implementation, the model must be validated across multiple earthquakes. Additionally, its performance should be analyzed for various magnitudes, epicentral distances, and site conditions to ensure its robustness and reliability across different seismic scenarios.  
+
+<|ref|>text<|/ref|><|det|>[[118, 359, 878, 412]]<|/det|>
+We have been following the suggestion of the reviewer (explanation above) and tested the performance (sigma) for all INGV stations of the area and all available large earthquakes (34 events) by comparing the obtained sigma with the one of Iervolino et al. (2024).  
+
+<|ref|>text<|/ref|><|det|>[[118, 428, 875, 464]]<|/det|>
+The terminology in the text should be refined for greater precision and to avoid ambiguity. For instance, even in the abstract:  
+
+<|ref|>text<|/ref|><|det|>[[118, 481, 875, 516]]<|/det|>
+Instead of making broad statements like "the ubiquity of smartphones", the revised version should specify the data type being used and how it contributes to seismic monitoring.  
+
+<|ref|>text<|/ref|><|det|>[[118, 532, 878, 584]]<|/det|>
+We rewritten the specific sentence and we added the "Station and smartphone measurements" subsection in the Methods section to precisely describe the station and smartphone measurements and how they are collected.  
+
+<|ref|>text<|/ref|><|det|>[[118, 600, 878, 688]]<|/det|>
+The term "red zone" is officially used by the Italian Civil Protection Department to designate high- risk volcanic areas in the Campi Flegrei region. This term identifies areas subject to preventive evacuation measures for public safety in case of an eruption. While it can be retained, it should be clearly explained or cited. Otherwise, using a more specific term would enhance the scientific clarity and rigor of the text.  
+
+<|ref|>text<|/ref|><|det|>[[118, 703, 878, 737]]<|/det|>
+We have clarified the definition of the red zone and the reason for its existence. We have kept the term because it is well known to scientists and the public.  
+
+<|ref|>text<|/ref|><|det|>[[118, 751, 878, 785]]<|/det|>
+In scientific writing, stating that amplification varies from 0.25 to 2.83 is insufficient unless the reference measure is specified and the calculation method explained.  
+
+<|ref|>text<|/ref|><|det|>[[118, 802, 879, 907]]<|/det|>
+We agree: amplification factors depend on the chosen reference (a classical issue and discussion in engineering seismology). A seismic amplification map depends on the reference chosen. In our approach, site amplification and deamplification is relative to the average site amplification in the region. This differs from conventional amplification factors used in engineering seismology, where the site amplification is relative to a rock reference, but is very similar to amplification maps recently derived at the regional scale for European
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 83, 879, 186]]<|/det|>
+risk models 7. This amplification map can only be compared with those of other regions if it has been verified that the average amplification effects due to subsurface layers are similar in both regions. Furthermore, the amplification values in our map can be rescaled relative to a specific station or subset of stations—provided these are included in the calibration dataset—thereby enabling more tailored interpretations. These explanations have been added in the discussion.  
+
+<|ref|>text<|/ref|><|det|>[[118, 204, 878, 256]]<|/det|>
+Additionally, the paper lacks proper references to other works dealing with the rapid quantification of ground motion using crowdsourced or smartphone- based techniques and similar challenges.  
+
+<|ref|>text<|/ref|><|det|>[[118, 270, 878, 322]]<|/det|>
+We have added the references of Kong et al. (2016) and Voosen (2021) on initiatives similar to the Earthquake Network. However, only in Marcou et al. (2024) is there a first attempt to model ground motion using smartphone data.  
+
+<|ref|>text<|/ref|><|det|>[[118, 336, 879, 388]]<|/det|>
+I am not entirely sure that the current manuscript structure is appropriate. A significant portion of the results and figures are in the supplementary material, making it difficult to fully grasp the study without it.  
+
+<|ref|>text<|/ref|><|det|>[[118, 404, 794, 421]]<|/det|>
+We moved some of the figures from the supplementary material to the main article.  
+
+<|ref|>text<|/ref|><|det|>[[118, 439, 860, 457]]<|/det|>
+Additionally, postponing the Methods section in this way does not facilitate comprehension.  
+
+<|ref|>text<|/ref|><|det|>[[118, 475, 486, 491]]<|/det|>
+We follow Nature Communication guidelines.  
+
+<|ref|>text<|/ref|><|det|>[[117, 508, 536, 578]]<|/det|>
+Line 128: Figure S9 is cited before S6, S7 and S8. Line 299: the correct citation should be number 20. Line 306: replace 'read' with 'red'. The caption of S8 has the wrong ID.  
+
+<|ref|>text<|/ref|><|det|>[[118, 595, 234, 611]]<|/det|>
+Fixed, thanks.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 648, 246, 664]]<|/det|>
+## REFERENCES  
+
+<|ref|>text<|/ref|><|det|>[[118, 679, 878, 731]]<|/det|>
+Berrocal, V.J., Gelfand, A.E. and Holland, D.M. Space- time data fusion under error in computer model output: an application to modeling air quality. Biometrics, 68, 837- 848 (2012).  
+
+<|ref|>text<|/ref|><|det|>[[118, 745, 878, 797]]<|/det|>
+Cascone, V., Boaga, J. and Cassiani, G. Small local earthquake detection using low- cost MEMS accelerometers: Examples in northern and central Italy. The Seismic Record, 1, 20- 26 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[118, 811, 878, 863]]<|/det|>
+D'Alessandro, A. and D'Anna, G. Suitability of low- cost three- axis MEMS accelerometers in strong- motion seismology: Tests on the LIS331DLH (iPhone) accelerometer. Bulletin of the Seismological Society of America, 103, 2906- 2913 (2013).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 82, 880, 137]]<|/det|>
+Di Curzio, D., Castrignanò, A., Fountas, S., Romić, M. and Rossel, R.A.V. Multi- source data fusion of big spatial- temporal data in soil, geo- engineering and environmental studies. Science of the Total Environment, 788, 147842 (2021).  
+
+<|ref|>text<|/ref|><|det|>[[115, 149, 880, 185]]<|/det|>
+Kong, Q., Allen, R.M., Schreier, L. and Kwon, Y.W. MyShake: A smartphone seismic network for earthquake early warning and beyond. Science Advances, 2, 1501055 (2016).  
+
+<|ref|>text<|/ref|><|det|>[[117, 200, 880, 253]]<|/det|>
+Iervolino, I., Cito, P., De Falco, M., Festa, G., Herrmann, M., Lomax, A., ... & Zollo, A. Seismic risk mitigation at Campi Flegrei in volcanic unrest. Nature Communications 15, 1- 14 (2024).  
+
+<|ref|>text<|/ref|><|det|>[[117, 269, 880, 323]]<|/det|>
+Marcou, S., Allen, R.M., Abrahamson, N.A. and Sung, C.H. Ground- Motion Modeling Using MyShake Smartphone Peak Acceleration Data. Bulletin of the Seismological Society of America, 115, 86- 105 (2025).  
+
+<|ref|>text<|/ref|><|det|>[[117, 336, 880, 373]]<|/det|>
+Nguyen, H., Cressie, N. and Braverman, A. Spatial statistical data fusion for remote sensing applications. Journal of the American Statistical Association, 107, 1004- 1018 (2012).  
+
+<|ref|>text<|/ref|><|det|>[[117, 385, 880, 439]]<|/det|>
+Oliveti, I., Faenza, L. and Michelini, A. New reversible relationships between ground motion parameters and macrosismic intensity for Italy and their application in ShakeMap. Geophysical Journal International, 23, 1117- 1137 (2022).  
+
+<|ref|>text<|/ref|><|det|>[[117, 451, 880, 504]]<|/det|>
+Vitale, S. and Isaia, R. Fractures and faults in volcanic rocks (Campi Flegrei, southern Italy): insight into volcano- tectonic processes. International Journal of Earth Sciences, 103, 801- 819 (2014).  
+
+<|ref|>text<|/ref|><|det|>[[115, 519, 876, 555]]<|/det|>
+Voosen, P. New Google effort uses cellphones to detect earthquakes. Science, 48, 101721 (2021).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 312, 99]]<|/det|>
+Dear Professor Finazzi,  
+
+<|ref|>text<|/ref|><|det|>[[118, 117, 880, 221]]<|/det|>
+Thank you again for submitting your manuscript "Citizens' smartphones unravel earthquake shaking in urban areas" to Nature Communications. We have now received reports from 2 reviewers and, based on their comments, we have decided to invite a revision of your work. Your revision should address all the points raised by our reviewers (see their reports below). Please also ensure to implement the discussion points raised by ref#1 in the last review round in the manuscript text, additionally to your reply in the response letter.  
+
+<|ref|>text<|/ref|><|det|>[[118, 239, 880, 290]]<|/det|>
+We have explained at line 451 of the new version of the manuscript that the model of Equation (14) is able to capture event- specific directivity or non- linear effects. Additionally, at line 43 we have explained why we are not considering crowdsourced felt reports in our current work.  
+
+<|ref|>text<|/ref|><|det|>[[118, 308, 880, 342]]<|/det|>
+When resubmitting, you must provide a point- by- point response to the reviewers' comments. Please show all changes in the manuscript text file with track changes or colour highlighting.  
+
+<|ref|>text<|/ref|><|det|>[[118, 359, 717, 376]]<|/det|>
+Along with the new manuscript file we submitted a track changes version.  
+
+<|ref|>text<|/ref|><|det|>[[118, 394, 880, 428]]<|/det|>
+If you are unable to address specific reviewer requests or find any points invalid, please explain why in the point- by- point response.  
+
+<|ref|>sub_title<|/ref|><|det|>[[120, 450, 370, 468]]<|/det|>
+## REVIEWER COMMENTS  
+
+<|ref|>text<|/ref|><|det|>[[120, 487, 449, 504]]<|/det|>
+Reviewer #2 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[120, 522, 715, 539]]<|/det|>
+Other modifications are acceptable, but I would like to address one point:  
+
+<|ref|>text<|/ref|><|det|>[[118, 556, 880, 676]]<|/det|>
+I understand that if the location has sufficient data, the epistemic uncertainty will show a lower value. However, how can a value lower than 0.1 be achieved when there are only a few records (fewer than 5)? This concerns my key question (Question 3). I am particularly interested in the phiS2S value; if phiS2S is around 0.2, it may be possible to reach a value below 0.1. However, the general phiS2S is approximately 0.3- 0.35, making obtaining a value under 0.1 almost impossible. I find it surprising that the authors state it is challenging to obtain phiS2S, but I understand if it is not available.  
+
+<|ref|>text<|/ref|><|det|>[[118, 694, 880, 816]]<|/det|>
+What we show in Supplementary Figure 3(c- d) is not the site- to- site variability (phiS2S as described in Al Atik et al., 2010) but rather the uncertainty of the Gaussian processes \(\delta \mathrm{S2S}\) given by Equations (8). The Gaussian process actually captures some of the residual variability (the spatially correlated part of the residuals with respect to the isotropic decay) that would otherwise be included in the uncertainty modelled by the terms \(\delta \mathrm{lnst}\) in Equations (1) and (2). We discussed this better in section "Smartphone and station data modelling" at line 338 and following.  
+
+<|ref|>text<|/ref|><|det|>[[118, 833, 880, 884]]<|/det|>
+The type of uncertainty that the reviewer likely has in mind (similar to classical GMPE uncertainty evaluation) is for instance expressed by \(\sigma\) in Table 1. When \(\delta \mathrm{S2S}\) is used in the GMM in Equation (19), the standard deviation (total aleatory sigma as expressed in
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 878, 117]]<|/det|>
+engineering seismology, as described by Al Atik et al., 2010) on the predicted PGA decreases from 0.4107 to 0.3569 (on the log10 scale).  
+
+<|ref|>text<|/ref|><|det|>[[118, 135, 880, 238]]<|/det|>
+We obtained a measure of site- to- site variability (similar to phiS2S as described by Al Atik et al., 2010) by calculating the standard deviation of the estimated residual variability at seismological stations, before and after adjusting for \(\delta \mathrm{S2S}\) (Equations 12 and 13, respectively). This resulted in a reduction of \(71.4\%\) , with the standard deviation moving from 0.6389 (0.2775 on the log10 scale) to 0.1827 (0.0793 on the log10 scale). This is discussed at line 176 and 437.  
+
+<|ref|>text<|/ref|><|det|>[[118, 256, 879, 307]]<|/det|>
+Atik, L. A., Abrahamson, N., Bommer, J. J., Scherbaum, F., Cotton, F., & Kuehn, N. (2010). The variability of ground- motion prediction models and its components. Seismological Research Letters, 81(5), 794- 801.  
+
+<|ref|>text<|/ref|><|det|>[[118, 325, 878, 357]]<|/det|>
+I believe the authors may underestimate the limitations associated with locations that have few records.  
+
+<|ref|>text<|/ref|><|det|>[[118, 373, 879, 442]]<|/det|>
+In areas with few records or where there are many uncertain records (e.g. only smartphone measurements and no station measurements), we can still estimate \(\delta \mathrm{S2S}\) , but the uncertainty is such that we cannot statistically claim that \(\delta \mathrm{S2S}\) is significantly different from the average amplification of the region. This has been explained at line 168 and following.  
+
+<|ref|>text<|/ref|><|det|>[[118, 458, 879, 508]]<|/det|>
+I suggest that the authors consider using equation (29) from Lavrentiadis et al. (2023) to address this issue or at least mention it in their paper. This would be beneficial since readers should be aware of potential improvements.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 526, 210, 541]]<|/det|>
+## Reference:  
+
+<|ref|>text<|/ref|><|det|>[[118, 544, 879, 594]]<|/det|>
+Lavrentiadis, G., Abrahamson, N.A. (2023). A non- ergodic spectral acceleration ground motion model for California developed with random vibration theory. Bull Earthquake Eng, 21, 5265- 5291. https://doi.org/10.1007/s10518- 023- 01689- 9  
+
+<|ref|>text<|/ref|><|det|>[[118, 610, 879, 660]]<|/det|>
+We have recovered the article, but could not find equation (29). We understand that the reviewer would like more information about the residual analysis performed in this paper, and this has been taken into account in the updated version (see the response to the first point).  
+
+<|ref|>text<|/ref|><|det|>[[118, 676, 879, 726]]<|/det|>
+Additionally, another new relevant paper from the US that can be cited is: "Ground- Motion Modeling Using MyShake Smartphone Peak Acceleration Data," BSSA, 2025, DOI: 10.1785/0120240209.  
+
+<|ref|>text<|/ref|><|det|>[[118, 742, 875, 774]]<|/det|>
+This paper has been on our reference list since the first draft of the manuscript was written. The confusion may lie in the year, which is given as 2024 in the BSSA journal.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 790, 449, 806]]<|/det|>
+## Reviewer #3 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[118, 825, 473, 840]]<|/det|>
+Decision Recommendation: Minor Revision  
+
+<|ref|>text<|/ref|><|det|>[[118, 860, 341, 875]]<|/det|>
+General Recommendation:
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 83, 880, 152]]<|/det|>
+The revised manuscript shows meaningful improvements in clarity, methodological transparency, and justification of the approach. The authors have responded constructively to most reviewer comments and addressed several concerns regarding the use of smartphone data, the treatment of bias, and the statistical modelling of amplification.  
+
+<|ref|>text<|/ref|><|det|>[[118, 153, 880, 220]]<|/det|>
+The integration of smartphone data remains an innovative and potentially impactful idea, and the manuscript now provides a more robust basis for such integration. The overall structure and figures have also improved, and key assumptions (e.g., spatial cross- correlation, non- reliance on calibration) are now better explained.  
+
+<|ref|>text<|/ref|><|det|>[[118, 221, 880, 255]]<|/det|>
+However, a few important clarifications and adjustments are still needed before the manuscript can be considered for publication.  
+
+<|ref|>text<|/ref|><|det|>[[118, 273, 490, 290]]<|/det|>
+Specific Requests (Required for Acceptance)  
+
+<|ref|>text<|/ref|><|det|>[[118, 308, 880, 358]]<|/det|>
+Clarify the spatial limitations of the amplification map, especially in areas with sparse smartphone data. It should be explicitly noted that fine- scale interpretation in such zones may be unreliable due to local uncertainty.  
+
+<|ref|>text<|/ref|><|det|>[[118, 374, 552, 390]]<|/det|>
+This point is now discussed at line 284 and following.  
+
+<|ref|>text<|/ref|><|det|>[[118, 405, 880, 456]]<|/det|>
+State more clearly the operational dependency of the method on the presence of a sufficiently dense and active (charging) smartphone network. The impact of uneven spatial distribution should be discussed more directly.  
+
+<|ref|>text<|/ref|><|det|>[[117, 471, 880, 608]]<|/det|>
+As emphasised at lines 168 and 284, uncertainty in the model output is as important as the output itself. An uneven spatial distribution of stations and/or smartphones does not present a methodological or algorithmic problem. In areas with a low density of measurements, the uncertainty in the model output is higher and it is less reliable. Therefore, it is important to be able to assess the uncertainty so that decision makers can make informed decisions based on both the model output and its uncertainty. In the updated version, we have emphasised the importance of uncertainty quantification and explained how uncertainty is quantified in more detail.  
+
+<|ref|>text<|/ref|><|det|>[[118, 624, 880, 675]]<|/det|>
+Include a table listing all seismic stations used, with a clear indication of whether they belong to the RAN or INGV networks. This will help readers assess the provenance and consistency of the data.  
+
+<|ref|>text<|/ref|><|det|>[[118, 690, 880, 740]]<|/det|>
+We recovered the station codes, and added Supplementary Figure 1 and Supplementary Dataset 1 which contain information on the networks and on which station detected which event.  
+
+<|ref|>text<|/ref|><|det|>[[118, 756, 880, 789]]<|/det|>
+Correct the typographical error ("shanking" \(\rightarrow\) "shaking" on line 320) and review consistency in terminology across sections.  
+
+<|ref|>text<|/ref|><|det|>[[118, 808, 185, 823]]<|/det|>
+Thanks.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 843, 212, 857]]<|/det|>
+## Conclusion  
+
+<|ref|>text<|/ref|><|det|>[[118, 860, 880, 910]]<|/det|>
+I recommend a minor revision, conditional on the authors addressing the above points. The manuscript is promising and can make a valuable contribution to the field if these final clarifications are incorporated.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[147, 87, 852, 229]]<|/det|>
+The manuscript is very well written, presenting an interesting research topic and a comprehensive analysis. I believe that this research, along with similar studies, opens a new avenue for collecting seismological data, given the computing power currently available and the new algorithms, especially in machine learning, that enable us to manage large volumes of data. However, there are several questions that the authors need to address, which I outline below:  
+
+<|ref|>text<|/ref|><|det|>[[147, 245, 852, 363]]<|/det|>
+(1) Lines 305-306: "In equation (10) describes the potential amplification of the ground motion due to the site effects." How can it be confirmed that the amplification observed in this study is solely attributable to site effects, rather than being influenced by path effects (such as azimuth dependence, 3d velocity structure) or source effects (e.g., radiation patterns)?  
+
+<|ref|>text<|/ref|><|det|>[[147, 369, 852, 461]]<|/det|>
+(2) It is challenging to interpret the uncertainty presented in Supplementary Figures 6 and 7. I recommend utilizing a single map that displays the epistemic uncertainty for the entire grid, rather than dividing it into two separate plots for the upper and lower boundaries.  
+
+<|ref|>text<|/ref|><|det|>[[147, 467, 852, 679]]<|/det|>
+(3) Supplementary Figure 4c has caused some confusion. Epistemic uncertainty appears to be analogous to the standard error, which makes Figure 4d comprehensible, as it illustrates lower epistemic uncertainty in regions with a greater amount of data. However, if a location is represented by only one station or without data, the epistemic uncertainty should be approximated to \(\frac{\sigma}{\sqrt{N = 1}}\) . Therefore, it is unclear why the smaller value \((\sim 0.05)\) is indicated in the region with only one station and 0.4 for the region without data in Figure 4c, which is unreasonable.  
+
+<|ref|>text<|/ref|><|det|>[[147, 686, 850, 777]]<|/det|>
+(4) Lines 221-222: Do the authors use the same correlation length (denoted as \(\theta\) in the paper) for both datasets? If so, this may not be reliable, as the data obtained from the smartphone is expected to exhibit a smaller correlation length.  
+
+<|ref|>text<|/ref|><|det|>[[147, 784, 852, 902]]<|/det|>
+(5) What kind of metadata can we expect to be accessible if the smartphone datasets are eventually released to the public? There is considerable potential for smartphones to transform early warning systems and enhance post-seismic damage distribution maps. However, if the metadata is limited or not comprehensively available, its utility for researchers may be substantially
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[178, 90, 283, 105]]<|/det|>
+diminished.  
+
+<|ref|>text<|/ref|><|det|>[[147, 113, 852, 201]]<|/det|>
+(6) It would be valuable to identify the specific components of the filtering procedure that have contributed to the highest percentage of discarded waveforms. In particular, what is the primary source of poor-quality waveforms in data recorded via smartphones?  
+
+<|ref|>text<|/ref|><|det|>[[148, 210, 851, 279]]<|/det|>
+(7) In my view, smartphone data can augment or build-on free-field predictions, and with tremendous potential. It may be beneficial for future research to consider the relationship between MMI and smartphones.  
+
+<|ref|>text<|/ref|><|det|>[[147, 285, 850, 328]]<|/det|>
+(8) Please include the phiS2S (between-site sigma) and phiSS (within-site sigma) values from both models in Table 1, in addition to the tau values.  
+
+<|ref|>text<|/ref|><|det|>[[147, 335, 850, 378]]<|/det|>
+(9) The paper should include the values pertaining to the variability of Inst derived from the two datasets (stations and smartphones).  
+
+<|ref|>text<|/ref|><|det|>[[148, 384, 851, 477]]<|/det|>
+(10) Are the results obtained from smartphones within the epistemic uncertainty range of ergodic GMM? Please verify this and generate a corresponding plot. I expect that the results should lie within the epistemic uncertainty range, otherwise, this would suggest that they do not really help to improve the GMM.
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__725a24fa23db0a0160e585b53860a42133f92ec1af04267fd9d24707fe1d1a6c/images_list.json

@@ -0,0 +1,47 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Response Figure 2. GTP hydrolysis assays for different mutants. GTP hydrolysis assays for WT\\* and mutants R291A (teal), R293A, and R291AR293A (violet) as a function of receptor concentrations from \\(0.25 \\mu \\mathrm{M}\\) to \\(1 \\mu \\mathrm{M}\\) . Of note, the references were normalized in two sets of measurements. Data with error bars are presented as mean±SEM of four independent experiments. Statistical analyses were performed using the ordinary oneway ANOVA followed the two-sides sidak's post-hoc test in PRISM 9.3.1, \\*\\*\\*p< 0.001, in comparison to the WT\\*.",
+    "footnote": [],
+    "bbox": [
+      [
+        248,
+        145,
+        730,
+        451
+      ]
+    ],
+    "page_idx": 11
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_3.jpg",
+    "caption": "Response Figure 3. Cavity volume comparison between S4 and S5 states using KVFinder. a the cavity in structure A (cyan) representing the S4 state has a volume of 293.33 ų. b the cavity in structure B (blue) representing the S5 state shows a larger volume of 534.38 ų. Calculations were performed using probes with inner and outer radii of 5 Å and 10 Å, respectively.",
+    "footnote": [],
+    "bbox": [
+      [
+        238,
+        579,
+        732,
+        789
+      ]
+    ],
+    "page_idx": 12
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Response Figure 4. Representative conformation of the 'Partially Released' state is depicted. Key interacting residues are shown in stick representation, with red dotted lines indicating their interactions.",
+    "footnote": [],
+    "bbox": [
+      [
+        323,
+        410,
+        663,
+        613
+      ]
+    ],
+    "page_idx": 16
+  }
+]

peer_reviews/supplementary_0_Transparent Peer Review file__725a24fa23db0a0160e585b53860a42133f92ec1af04267fd9d24707fe1d1a6c/supplementary_0_Transparent Peer Review file__725a24fa23db0a0160e585b53860a42133f92ec1af04267fd9d24707fe1d1a6c.mmd

@@ -0,0 +1,417 @@
+
+# nature portfolio  
+
+Peer Review File  
+
+# Structure and function of a near fully-activated intermediate GPCR-Goβγ complex  
+
+Corresponding Author: Dr Libin Ye  
+
+Parts of this Peer Review File have been redacted as indicated to remove third- party material.  
+
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+Version 0:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author)  
+
+The manuscript by Bi M. et al. characterizes the intermediate S4 state of adenosine \(\mathsf{A}_{2\mathsf{A}}\) receptor \((\mathsf{A}_{2\mathsf{A}}\mathsf{R})\) using the conformation biased mutant R291A, which they have previously shown to trap the receptor in the S4 state. The authors found that the R291A mutant was able to bind Goβγ with a decreased GTP hydrolysis and nucleotide exchange rate that is independent of ligand. Using both cryo- EM and Gaussian accelerated molecular dynamics simulation (GaMD), the authors determined the structure of the R291A \(\mathsf{A}_{2\mathsf{A}}\mathsf{R}\) in complex with mini- Goβγ and proposed a mechanism by which the S4 state carry out the nucleotide exchange and transition to the fully activated state. This paper provides a great follow- up to the authors' previous paper on trapping the intermediate states of a GPCR using the mutants R291A and R293A. This work is interesting because the authors showed in atomic detail the orientation and the interactions of the relevant residues involved in the intermediate state. The results of this study increase our understanding on the features of the intermediate state and filled in the gap on the structural details of the GPCR activation landscape.  
+
+I have some comments that I think would help improve the paper if addressed:  
+
+Comments:  
+
+1. The authors found that the R291A-Goβγ complex has a lower Michaelis-Menten constant Km for GTP hydrolysis than the fully activated state (WT-Goβγ-NECA). Can the authors elaborate on the implication of this?  
+
+2. On page 5, lines 141-142, the authors said that "BODIPY-FL-GTP showed more efficient binding to the S5-mediated G protein with a higher \(K_D\) value." What do the authors mean by this? In Figures 2d and 2e, based on the difference in \(K_D\) , it seems that R291A-Goβγ has a stronger affinity toward GTP or GDP than the S5-mediated one.  
+
+3. I want to make sure that I understand it correctly. In Figure 2g, the intermediate state has a slightly faster rate of GDP exchange with GTP than the fully activated state, but since the fluorescence does not go to zero, it suggests that the BODIPY-FL-GDP is still bound, unlike the S5 state? Maybe the authors can include an additional statement explaining this observation on the text.  
+
+4. If I understand it correctly, the authors are trying to make a case that the 2nd association step observed in the BODIPY-FL-GDP binding assay with R291A is due to the limited access of the nucleotide to the second site by demonstrating that the association step is also observed with S5 state when BODIPY-FL-GTPyS is used since the bulky BODIPY group is on the y-phosphate side. However, this wasn't conveyed very clearly in the main text. The authors may consider revising their statements in the manuscript to better communicate the argument.  
+
+5. When the nucleotide accessed the second site in the R291A-Goβγ complex, the authors posit that the complex remains in the partially open S4 state. Did the authors try collecting the \(^{19}\mathrm{F}\) NMR spectrum of the R291A-Goβγ complex with GDP or GTP to see if the receptor does not sample a different state?
+
+<--- Page Split --->
+
+6. Since the intermediate state of the \(\mathsf{A}_{2\mathsf{A}}\mathsf{R} - \mathsf{G}\alpha \beta \gamma\) complex can now be trapped and characterized, the application of this is that it is possible to design a therapeutic agent that can target this specific conformation of the receptor. Do the authors know of any disease or condition where this could be applied?  
+
+Minor Comments:  
+
+7. In the Methods section, GTPase hydrolysis assay, page 19, lines 436-442, change the description of the methods to passive voice.  
+
+8. Figure 2a, indicating which ligand is a full, partial, or inverse agonist on the figure would help the readers.  
+
+9. Figure 2c, the y-axis label is incorrect. I believe it should be velocity instead of GTP concentration.  
+
+10. Figure 2f, the legend label should be R291A+ Goβγ and WT\*+Goβγ+NECA for consistency with the legends in the other panels.  
+
+11. The ordering of the panels in Figures 3 and 4 could be rearranged as they are referenced in the text. In the current version of the manuscript, after Fig. 3b, Fig. 3f is then referenced, followed by Fig. 3d, 3e, 3g, then Fig. 3c. Same with Figure 4, Fig. 4f is referenced, followed by Fig. 4e, 4d, 4g-i.  
+
+12. Extended Figure 2a, missing legend for red and teal data points.  
+
+13. Kindly provide clearer images for Extended Data Figures 3c and 3d.  
+
+14. Extended Figure 3b, the bands for \(\mathsf{A}_{2\mathsf{A}}\mathsf{R}\) R291A and NB35 are not that visible on the SDS-PAGE gel for the most dominant peak (Fractions 18-20).  
+
+15. Extended Data Figure 4 caption: Schematic flowchart, instead of flow-chat.  
+
+16. The F376-M60 interaction was mentioned toward the end of the Results section, that it plays a significant role in the closed state of AHD, but this interaction was broken in both S4 and S5. Since this interaction is not a major key player in the transition between S4 and S5 states, Figures 4g, h, and I can probably be moved to the supplementary.  
+
+## Reviewer #2  
+
+(Remarks to the Author) Review of Bi et al.  
+
+The manuscript by Bi and coworkers reports the structure of mutant structure of A2A in complex with a mini-Gs. The aim of this work is to describe what the authors term an 'intermediate' transition state. This work was mostly guided by F19 NMR experiments where they observed an incomplete chemical shift to a fully active state structure. While the experiments are largely well conceived and performed, this reviewer has some fundamental issues with the interpretation of the data. I will say at the outset that I'm not a molecular dynamics expert and will not review that portion of the manuscript.  
+
+Firstly, I am absolutely not convinced that they have isolated a ligand free structure. In the maps the authors shared and one that I calculated (from their half maps), there is very clear density of something that hasn't been modelled in the active site of their structure. It is clearly not a ligand free structure, in fact Adenosine perfectly fits in the unmodelled density. It is not clear to me at all that the authors have isolated and solved a structure of an 'intermediate' S4 state. Their evidence that they have rests largely on their F19 NMR studies. While these NMR experiments are a first step in assessing the bulk chemical environments that the Fluorine samples in their different constructs it may not be the most robust way to determine that they have fully trapped an S4 state.  
+
+I'm also not sure the GTP exchange assays are fully formed. While there is clear phenotypic differences between their WT\* and the R291A mutant, it is not clear to this reviewer that this is due to some intermediate state. As no G-protein association assays (BRET or FRET based) or G-protein turnover assays were performed this reviewer is concerned that the phenotypic differences could be more due to differential G-protein association rather than trapping an S4 state.  
+
+Furthermore, during the processing of the SPA data in their final 3D classification prior to consensus map refinement it appears that they may have 'cherry-picked' a particular dynamic state which may or may not reflect an S4 state. As all their 3D classes looked to be of high quality (2nd last row in Sup Figure 4), their final consensus reconstruction may not be reflective of the totality of states captured during the SPA experiment. Also, there was much discussion around relatively minor differences between their structure and another active state A2A structure, but the authors performed no 3D variability to assess the extent of dynamic states captured. In this reviewers opinion all the particles in the 3D classification step should be included in a 3DVA analysis (either cryoSPARC or cryoDRGN) to judge if their analysis of the PDB coordinates is justified or if they have simply captured one of many stable subsets of an active state structure. Even a comparison to their other 3D classes would be interesting to see if any of them look more like the active state structure that was compared against (6GBG).
+
+<--- Page Split --->
+
+To summarize, I'm not convinced that the authors have captured a ligand free structure and that they have captured an intermediate state structure. Much more analysis of their SPA data is warranted and perhaps some pharmacology type experiments to tease out what is going on with G- protein association to more robustly interpret their GTP experiments would be necessary to confirm the claims this manuscript makes.  
+
+Some minor issues:  
+
+Some minor issues:1. Many side chains are unmodelled in the PDB, many of which have interpretable density (for example M177 and Y271 in receptor, to name just a few).2. The Figure Legend in Figure 1 is mislabelled (d and f)3. I don't know what the authors mean in Page 6. Line 183 "Closer inspection of the Tm6.... Using F19 ... revealed a clockwise rotation..." This needs some explanation as I can't see what the authors mean here.4. I would suggest to show the nucleotide exchange data for with the addition of agonists and inverse agonists for the mutant (maybe just in the sup material)5. There are some more modelling errors. The unmodelled loop between 148 and 166 in the receptor will most likely have the N-terminus pointing in the wrong direction. As its currently modelled the only way for this loop to be completed is by forming a protein knot through ECL3.6. This reviewer has major issues with the use of DeepEMhancer for modelling. The map that the authors originally provided is a prime example of Al hallucination. The Al model started to turn the detergent micelle into protein looking density and should NOT be trusted. The map I calculated from the half maps and the authors traditionally sharpened map is more than suitable for PDB modelling. EM enhanced maps also SHOULD NOT be the main map deposited in the PDB, as is taints the PDB and EMDB with data that is not suitable for interpretation.  
+
+Matthew Belousoff, Monash School of Pharmaceutical Sciences  
+
+## Reviewer #3  
+
+(Remarks to the Author)  
+
+This manuscript reports the structure and function of a ligand- free GPCR- G\(\alpha \beta \gamma\) intermediate complex. The authors determined the cryo- EM structure of an intermediate A2AR- mini- G\(\alpha \beta \gamma\) complex. They presented experimental and computational evidences that the intermediate complex (S4- G\(\alpha \beta \gamma\) ) initiates a rate- limited nucleotide exchange without progressing to the fully activated complex (S5- G\(\alpha \beta \gamma\) ), in which the \(\alpha\) - helical domain (AHD) of the G\(\alpha \beta \gamma\) is partially open engaged by a second nucleotide. Based on these, they proposed a mechanistic model for the rate- limited nucleotide exchange in the intermediate GPCR- G\(\alpha \beta \gamma\) complex.  
+
+The experimental method presented in this paper is suitable and the MD simulations are well- designed. The approach can be applied to characterize transient intermediate states of other GPCRs. However, there are a few short comings in this study. I recommend that this work is publishable with revision as follows:  
+
+1. The authors determined the cryo-EM structure of an intermediate A2AR-mini-G\(\alpha \beta \gamma\) complex. Has there any other intermediate state of other GPCRs reported in other papers? What's the similarity and difference between this intermediate structure and the others? It would be better to discuss this intermediate A2AR-mini-G\(\alpha \beta \gamma\) structure in the context of the literature.  
+
+2. In the determined cryo-EM structure of the A2AR-mini-G\(\alpha \beta \gamma\) complex, the AHD of G\(\alpha \beta \gamma\) was missing. The mini-G\(\alpha \beta \gamma\) is different from the full-length G\(\alpha \beta \gamma\) protein, will this influence the binding of G\(\alpha \beta \gamma\) ?  
+
+3. It is seen from Fig.1a that the most obvious difference between the intermediate complex (S4-G\(\alpha \beta \gamma\) ) and other complex (S1-G\(\alpha \beta \gamma\) , S2-G\(\alpha \beta \gamma\) , S3-G\(\alpha \beta \gamma\) and S5-G\(\alpha \beta \gamma\) ) is whether the AHD of G\(\alpha \beta \gamma\) is partially open and whether there is GDP bound in G\(\alpha \beta \gamma\) . It would be nice if the A2AR bound by the full-length G\(\alpha \beta \gamma\) protein (with GDP bound in G\(\alpha \beta \gamma\) ) is characterized by cryo-EM.  
+
+4. The authors proposed two nucleotide-bound sites (Fig. 5, site 1 and site 2) in the transition process from the intermediate to the fully activated GPCR-G\(\alpha \beta \gamma\) complex. They performed additional GaMD simulations to examine the GDP release process in the cS4-G\(\alpha \beta \gamma\) system. GDP bound in site 2 of G\(\alpha \beta \gamma\) in the starting structure of GaMD simulations of GDP release. In the simulation of the GDP release process, can the GDP be observed to make a short stay at site 1?  
+
+5. Free energy calculations indicated a "GDP Released" state in the S5-G\(\alpha \beta \gamma\) system and only a "Partially Released" state in the S4-G\(\alpha \beta \gamma\) system (Extended Data Fig. 8d and 8e). In this "Partially Released" state, is the GDP still binds in G\(\alpha \beta \gamma\) ? Is there a possibility that it binds in the proposed "site 1"?  
+
+6. Page 8 Lines 246: "Extended Data Figs. 8a-c". It should be Extended Data Figs. 8d-e.  
+
+Page 8 Lines 244- 249: "Extended Data Figs. 8d-e". It should be Extended Data Figs. 8a-c.  
+
+Version 1:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author)  
+
+The authors have addressed all my concerns and have updated the manuscript accordingly.  
+
+Some minor corrections:
+
+<--- Page Split --->
+
+1. Extended Data Figure 2, page 33: The figure caption seems to be mislabeled. I think the red data points represent the line width of S1 resonance and the teal data point represent the line width of the S4 resonance.  
+
+2. Page 5 lines 134-135: Omit the word domain after AHD  
+
+3. Page 5 line 134-135: formed by Ras-like domain and AHD resulting from a partial opening of AHD.  
+
+## Reviewer #2  
+
+(Remarks to the Author)  
+
+Firstly, I wish to commend the authors on their significant efforts to reanalyse the cryoEM data. I am of the opinion that this has vastly improved the authors manuscript. I do not fundamentally disagree with the overall experimental design and overall analysis but I have a comment that the authors may wish to consider.  
+
+Our research group has performed 3DVA on over 300 GPCR ternary complexes and we observe incredibly similar dynamic sampling of both G- protein engagement angles and G- protein conformations. However, does this mean their interpretation is incorrect? I do not think so, it is more likely that they have found an appropriate experimental system to more fully explore and interpret the dynamics and 3DVA data that are commonly observed across all active state GPCR structures. Put another way, essentially all reported structures of active state GPCRs are an average of the S4- S5 state, but the authors in this manuscript have perhaps skewed the overall population towards the S4 state and provided a great hypothesis for interpretation of the observed dynamic landscape in active state GPCR structures.  
+
+I would recommend this paper for publication as it is my opinion that the authors have appropriately addressed the reviewers concerns.  
+
+Matthew Belousoff, Monash Institute of Pharmaceutical Sciences  
+
+Reviewer #3  
+
+(Remarks to the Author)  
+
+The authors have addressed all my previous concerns with detailed explanations and appropriate revisions. The revisions have improved the overall quality of the manuscript. I have no further issues with the current version of the manuscript, and I believe it can be published.  
+
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source.  
+
+The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain
+
+<--- Page Split --->
+
+permission directly from the copyright holder. To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+
+Reviewer #1 (Remarks to the Author):  
+
+The manuscript by Bi M. et al. characterizes the intermediate S4 state of adenosine \(\mathsf{A}_{2\mathsf{A}}\) receptor \((\mathsf{A}_{2\mathsf{A}}\mathsf{R})\) using the conformation biased mutant R291A, which they have previously shown to trap the receptor in the S4 state. The authors found that the R291A mutant was able to bind \(\mathsf{G}\mathsf{a}\beta \gamma\) with a decreased GTP hydrolysis and nucleotide exchange rate that is independent of ligand. Using both cryo- EM and Gaussian accelerated molecular dynamics simulation (GaMD), the authors determined the structure of the R291A \(\mathsf{A}_{2\mathsf{A}}\mathsf{R}\) in complex with mini- \(\mathsf{G}\mathsf{a}\beta \gamma\) and proposed a mechanism by which the S4 state carry out the nucleotide exchange and transition to the fully activated state. This paper provides a great follow- up to the authors' previous paper on trapping the intermediate states of a GPCR using the mutants R291A and R293A. This work is interesting because the authors showed in atomic detail the orientation and the interactions of the relevant residues involved in the intermediate state. The results of this study increase our understanding on the features of the intermediate state and filled in the gap on the structural details of the GPCR activation landscape.  
+
+Thanks for the appreciation on our new findings.  
+
+I have some comments that I think would help improve the paper if addressed:  
+
+Comments:  
+
+1. The authors found that the R291A- \(\mathsf{G}\mathsf{a}\beta \gamma\) complex has a lower Michaelis-Menten constant Km for GTP hydrolysis than the fully activated state (WT- \(\mathsf{G}\mathsf{a}\beta \gamma\) -NECA). Can the authors elaborate on the implication of this?  
+
+The lower Km implies that the GTP- GDP conversion rate at the non- canonical nucleotide binding site 2 in the R291A- \(\mathsf{G}\mathsf{a}\beta \gamma\) complex was much smaller than at the canonical site (site 1) in the WT- \(\mathsf{G}\mathsf{a}\beta \gamma\) -NECA, indicating a limited capacity of the intermediate complex in regulating GTP hydrolysis while the fully activated complex holds the full capacity. Of note, we redefine the canonical binding site as the site 1 while the noncanonical binding site as the site 2. This also implies that the GTP doesn't bind to the canonical nucleotide binding site in the R291A- \(\mathsf{G}\mathsf{a}\beta \gamma\) complex.  
+
+2. On page 5, lines 141-142, the authors said that "BODIPY-FL-GTP showed more efficient binding to the S5-mediated G protein with a higher \(K_D\) value." What do the authors mean by this? In Figures 2d and 2e, based on the difference in \(K_D\) , it seems that R291A- \(\mathsf{G}\mathsf{a}\beta \gamma\) has a stronger affinity toward GTP or GDP than the S5-mediated one.  
+
+We appreciate the reviewer's insight. Our interpretation is that the GTP binding site in the S5- mediated G protein is freely accessible because of the full open of AHD domain, resulting in a robust BODIPY- FL- GTP association- dissociation process. In contrast,
+
+<--- Page Split --->
+
+due to the space limitation caused by a partial open pose of AHD domain in the S4- mediated G protein, the binding of BODIPY- FL- GTP to the canonical nucleotide binding site is not accessible. The binding process is halted at the stage that Ras- like domain and AHD domain hold the nucleotide en route to the fully open state. We think this nucleotide binding site 2 is close to the canonical nucleotide binding site 1, which allows it to exhibit a limited GTP hydrolysis capacity, along with a slow dissociation rate as presented in the figures. This also results in the R291A- Gαβγ exhibiting a stronger affinity toward GTP or GDP than the S5- mediated one, as pointed out by the reviewer.  
+
+3. I want to make sure that I understand it correctly. In Figure 2g, the intermediate state has a slightly faster rate of GDP exchange with GTP than the fully activated state, but since the fluorescence does not go to zero, it suggests that the BODIPY- FL- GDP is still bound, unlike the S5 state? Maybe the authors can include an additional statement explaining this observation on the text.  
+
+Yes. The data in Fig. 2 indicated that the release number of BODIPY- FL- GDP from the S4 state was much less than the S5 state, but it reached the equilibrium much faster, resulting in the \(K_{\text{off}}\) actually 0.09/min vs 0.05/min (Page 5, lines 153- 156). This was also consistent with our earlier findings in Figs. 2d and 2e showing that a substantial proportion of BODIPY- FL- GDP/GTP remained bound in the S4 state mediated G proteins.  
+
+4. If I understand it correctly, the authors are trying to make a case that the 2nd association step observed in the BODIPY- FL- GDP binding assay with R291A is due to the limited access of the nucleotide to the second site by demonstrating that the association step is also observed with S5 state when BODIPY- FL- GTPγS is used since the bulky BODIPY group is on the γ-phosphate side. However, this wasn't conveyed very clearly in the main text. The authors may consider revising their statements in the manuscript to better communicate the argument.  
+
+Yes, our conclusion is well articulated by the reviewer. We used the BODIPY- FL- GTP- γ- S as a case to confirm that the non- canonical binding site 2 is formed by the Ras- like domain and a partial open AHD domain that limits nucleotide to access canonical site 1. (Page 5, lines 148- 152).  
+
+5. When the nucleotide accessed the second site in the R291A- Gαβγ complex, the authors posit that the complex remains in the partially open S4 state. Did the authors try collecting the \(^{19}\mathrm{F}\) NMR spectrum of the R291A- Gαβγ complex with GDP or GTP to see if the receptor does not sample a different state  
+
+Yes, the \(^{19}\mathrm{F}\) - NMR spectra were acquired with GDP included. We did see a small portion ( \(\sim 10\%\) ) of S3 state in the sample of R291A- Gαβγ while the receptor was unable to shift to the S5 state (Fig. 1b and Extended Data Fig.1).
+
+<--- Page Split --->
+
+6. Since the intermediate state of the \(\mathsf{A}_{2\mathsf{A}}\mathsf{R} - \mathsf{G}\mathsf{a}\beta \mathsf{y}\) complex can now be trapped and characterized, the application of this is that it is possible to design a therapeutic agent that can target this specific conformation of the receptor. Do the authors know of any disease or condition where this could be applied?  
+
+Thanks for the insights regarding the correlation between conformation and disease. So far, we didn't explore the details and the distinction of signaling pathways that S4 and S5 could respectively regulate, considering the complexity of signaling involved with various G proteins, GRKs, and \(\beta\) - arrestins. However, we acknowledge the potential for therapeutic applications and this type of work has been in our plan. In May 2024, a clinic variance R291C (NCBI, SNP: rs745714462). was reported, despite the functional data of this variance is still missing in the report. Considering the similarity of R291C and R291A from the molecular interaction level, we will continue and expand our study to use the knock-in mouse model to study the correlation between the conformation and disease in the future.  
+
+## Minor Comments:  
+
+7. In the Methods section, GTPase hydrolysis assay, page 19, lines 436-442, change the description of the methods to passive voice.  
+
+We have revised the description in the GTPase hydrolysis assay section to passive voice to improve clarity and consistency. (Pages 21- 22, lines 479- 487)  
+
+8. Figure 2a, indicating which ligand is a full, partial, or inverse agonist on the figure would help the readers.  
+
+We have updated Fig. 2a to clearly indicate which ligands are full, partial, or inverse agonists, making it easier for readers to interpret the data. (Page 14, Fig. 2a)  
+
+9. Figure 2c, the y-axis label is incorrect. I believe it should be velocity instead of GTP concentration.  
+
+We have corrected the y-axis label in Fig. 2c to "velocity" for accuracy. (Page 14, Fig. 2c).  
+
+10. Figure 2f, the legend label should be R291A+ Gaβγ and WT\*+Gaβγ+NECA for consistency with the legends in the other panels.  
+
+We have revised the legend in Fig. 2f for consistency with the other panels. (Page 14, Fig. 2f)  
+
+11. The ordering of the panels in Figures 3 and 4 could be rearranged as they are referenced in the text. In the current version of the manuscript, after Fig. 3b, Fig. 3f is
+
+<--- Page Split --->
+
+then referenced, followed by Fig. 3d, 3e, 3g, then Fig. 3c. Same with Figure 4, Fig. 4f is referenced, followed by Fig. 4e, 4d, 4g- i.  
+
+Thank you for pointing this out. A new figure was added before the previous version of Figure 3, resulting in a renumbering of figures. We have carefully reviewed the figure references and revised the order of the panels in Figs. 4 and 5 to match the sequence in which they are discussed in the text (Pages 8- 10). This adjustment ensures better flow and alignment between the manuscript and the figures. We appreciate your feedback in helping to improve the clarity and organization of the manuscript.  
+
+12. Extended Figure 2a, missing legend for red and teal data points.  
+
+We appreciate you bringing this to our attention. The missing legend for the red and teal data points in Extended Data Fig. 2a has been added. (Page 33, Extended Data Fig. 2).  
+
+13. Kindly provide clearer images for Extended Data Figures 3c and 3d.  
+
+We have reprocessed and replaced Extended Data Figures 4c and 4d with higher- resolution images to improve clarity and ensure the details are more visible. (Page 35, Extended Data Fig.4).  
+
+14. Extended Figure 3b, the bands for \(\mathsf{A}_{2\mathsf{A}}\mathsf{R}\) R291A and NB35 are not that visible on the SDS-PAGE gel for the most dominant peak (Fractions 18-20).  
+
+The reduced visibility of the NB35 band was because we didn't heat the samples but merely incubated the sample with a loading buffer for better visibility of the receptor. This process could be the reason for the intensity decrease of soluble proteins like NB35. The same pattern was observed in others' reports as well, please refer to Response Fig. \(1^{1,2}\) .  
+
+Response Figure 1. SDS- PAGE analysis of GPCR complex. a GPR52 receptor complex, including G \(\beta \gamma\) , mini- G \(\alpha_{s}\) , and NB35 (Lin et al., 2020, Nature). b the purified components of the GLP- 1R complex include G \(\alpha_{s}\) , G \(\beta \gamma\) and NB35 (Cary et al., 2022, Nature Chemical Biology).
+
+<--- Page Split --->
+
+15. Extended Data Figure 4 caption: Schematic flowchart, instead of flow-chat.  
+
+Thank you for catching this typo. The caption for Extended Data Fig. 6 in the revised version has been corrected to "Schematic flowchart." (Page 37, Extended Data Fig. 6).  
+
+16. The F376-M60 interaction was mentioned toward the end of the Results section, that it plays a significant role in the closed state of AHD, but this interaction was broken in both S4 and S5. Since this interaction is not a major key player in the transition between S4 and S5 states, Figures 4g, h, and I can probably be moved to the supplementary.  
+
+Thank you for your suggestion. We have moved these panels to the supplementary and labeled them as Extended Data Figs. 12a, b, and c, as recommended.  
+
+Reviewer #2 (Remarks to the Author):  
+
+Review of Bi et al.  
+
+The manuscript by Bi and coworkers reports the structure of mutant structure of A2A in complex with a mini- Gs. The aim of this work is to describe what the authors term an 'intermediate' transition state. This work was mostly guided by F19 NMR experiments where they observed an incomplete chemical shift to a fully active state structure. While the experiments are largely well conceived and performed, this reviewer has some fundamental issues with the interpretation of the data. I will say at the outset that I'm not a molecular dynamics expert and will not review that portion of the manuscript.  
+
+1. Firstly, I am absolutely not convinced that they have isolated a ligand free structure. In the maps the authors shared and one that I calculated (from their half maps), there is very clear density of something that hasn't been modelled in the active site of their structure. It is clearly not a ligand free structure, in fact Adenosine perfectly fits in the unmodelled density.  
+
+We thank the reviewer for pointing this out. Our previous interpretation that the structure is ligand- free was based on experiments in which no exogenous ligands were added, and we did not consider the endogenous adenosine binding to the structure. We appreciate the reviewer's insight. Following his suggestion, we conducted additional analysis using a focused mask and confirmed that there is a clear extra density within the ligand- binding pocket, which can be modeled as adenosine. Subsequent mass spectrometry (Extended Data Fig.5 in the main text) confirmed the presence of adenosine in our sample. These findings suggest that the R291A mutant
+
+<--- Page Split --->
+
+stabilizes the receptor in a ligand- bound intermediate S4 state, distinct from the fully activated, ligand- bound S5 state. We have revised the text and figures accordingly and included the mass spectrometry result in the manuscript.  
+
+2. It is not clear to me at all that the authors have isolated and solved a structure of an 'intermediate' S4 state. Their evidence that they have rests largely on their F19 NMR studies. While these NMR experiments are a first step in assessing the bulk chemical environments that the Fluorine samples in their different constructs it may not be the most robust way to determine that they have fully trapped an S4 state.  
+
+With respect, we disagree with the reviewer's comment. \(^{19}\mathrm{F}\) NMR is a widely accepted and robust method for mapping conformational ensembles or landscapes. As demonstrated in our previous publication \(^{3}\) , the R291A mutation significantly enriched the population of the S4 state. While it is theoretically possible to capture intermediate states in SPA through classification alone, this specific mutation greatly enriches the S4 population, which in turn enables both detailed structural analysis and functional measurements of GTP hydrolysis at this intermediate state. In response to the reviewer's concern, we performed thorough classification, which revealed dynamic structural fluctuations centered around a predominant conformation, distinct from the well-defined S5 state. We interpret this predominant conformation as the S4 state, consistent with our \(^{19}\mathrm{F}\) NMR results.  
+
+As an orthogonal approach, we validated the existence of this intermediate S4 state through MD simulations, using the 6GDG structure with the R291A mutation and a full G- protein. These simulations, conducted independently of the cryo- EM data, confirmed a high structural similarity between the simulated S4 state and our cryo- EM structure. This provides further confidence that we have indeed captured the intermediate S4 state. We appreciate the reviewer's suggestion and acknowledge that the term "trapped" may have implied a more homogenous population than intended. We have revised the manuscript accordingly to clarify this.  
+
+3. I'm also not sure the GTP exchange assays are fully formed. While there is clear phenotypic differences between their WT\* and the R291A mutant, it is not clear to this reviewer that this is due to some intermediate state. As no G-protein association assays (BRET or FRET based) or G-protein turnover assays were performed this reviewer is concerned that the phenotypic differences could be more due to differential G-protein association rather than trapping an S4 state.  
+
+We have conducted GTP turnover assays in our previous publication (Xudong Wang, et al. Nature Communications, 2023, Supplement Fig. 7a) \(^{3}\) . We also attach the figure here (Response Fig. 2) as well as the current research (Figs. 2a- c) in addition to nucleotide exchange experiments (Figs. 2d- g). The experiments clearly showed that R291A mutant exhibited limited capacities of GTP turnover and nucleotide exchanges.
+
+<--- Page Split --->
+![](images/Figure_2.jpg)
+
+<center>Response Figure 2. GTP hydrolysis assays for different mutants. GTP hydrolysis assays for WT\* and mutants R291A (teal), R293A, and R291AR293A (violet) as a function of receptor concentrations from \(0.25 \mu \mathrm{M}\) to \(1 \mu \mathrm{M}\) . Of note, the references were normalized in two sets of measurements. Data with error bars are presented as mean±SEM of four independent experiments. Statistical analyses were performed using the ordinary oneway ANOVA followed the two-sides sidak's post-hoc test in PRISM 9.3.1, \*\*\*p< 0.001, in comparison to the WT\*. </center>  
+
+4. Furthermore, during the processing of the SPA data in their final 3D classification prior to consensus map refinement it appears that they may have 'cherry-picked' a particular dynamic state which may or may not reflect an S4 state. As all their 3D classes looked to be of high quality (2nd last row in Sup Figure 4), their final consensus reconstruction may not be reflective of the totality of states captured during the SPA experiment. Also, there was much discussion around relatively minor differences between their structure and another active state A2A structure, but the authors performed no 3D variability to assess the extent of dynamic states captured. In this reviewers opinion all the particles in the 3D classification step should be included in a 3DVA analysis (either cryoSPARC or cryoDRGN) to judge if their analysis of the PDB coordinates is justified or if they have simply captured one of many stable subsets of an active state structure. Even a comparison to their other 3D classes would be interesting to see if any of them look more like the active state structure that was compared against (6GBG). To summarize, I'm not convinced that the authors have captured a ligand free structure and that they have captured an intermediate state structure. Much more analysis of their SPA data is warranted and perhaps some
+
+<--- Page Split --->
+
+pharmacology type experiments to tease out what is going on with G- protein association to more robustly interpret their GTP experiments would be necessary to confirm the claims this manuscript makes.  
+
+We thank the reviewer for these comments, which led us to conduct a more detailed re- analysis of our data. In response, we revisited the entire particle set and all noted classes. Through extensive classification and 3D variability analysis (3DVA), we reconfirmed that the S4 conformation identified in our initial submission remains the predominant conformation. This conformation shows clear distinctions from the well- defined S5 state (6GDG). C- α displacement analysis across all classes indicated that the conformational snapshots cluster around this predominant S4 conformation, with most fluctuations occurring in the Gy subunit. This finding is consistent with our \(^{19}\mathrm{F}\) NMR data, which revealed an enrichment centered around a defined peak, further supporting the identification of S4 as a distinct intermediate state on the trajectory to full activation. Additional receptor pocket calculations are provided in Response Fig.3.  
+
+We focused on two snapshots that show the largest deviations from the predominant S4 conformation, which we refer to as S4d1 and S4d2. The remaining snapshots exhibit only minor variations in the G protein, while in S4d1, the outermost region of the G protein swings by up to 7.5 Å compared to the S4 state, likely due to intermediate G- protein engagement. However, the receptor itself displays only subtle variations within the overall S4 conformation. We have updated the text and figures to include these findings and have illustrated the conformational similarities among the classes using per- residue C- α displacement calculations, as shown in Extended Data Fig. 9 (Page 40).  
+
+![](images/Figure_3.jpg)
+
+<center>Response Figure 3. Cavity volume comparison between S4 and S5 states using KVFinder. a the cavity in structure A (cyan) representing the S4 state has a volume of 293.33 ų. b the cavity in structure B (blue) representing the S5 state shows a larger volume of 534.38 ų. Calculations were performed using probes with inner and outer radii of 5 Å and 10 Å, respectively. </center>
+
+<--- Page Split --->
+
+Some minor issues:  
+
+Some minor issues:1. Many side chains are unmodelled in the PDB, many of which have interpretable density (for example M177 and Y271 in receptor, to name just a few). Thank you for pointing this out. We have now modeled the side chains for M177 and Y271, and we have manually inspected all relevant residues to ensure the best fit. This includes adding residues such as V164, A165, among others.  
+
+2. The Figure Legend in Figure 1 is mislabelled (d and f) Thank you for your careful inspection. We have corrected the figure legend for Figure 1.  
+
+3. I don't know what the authors mean in Page 6. Line 183 "Closer inspection of the Tm6.... Using F19 ... revealed a clockwise rotation..." This needs some explanation as I can't see what the authors mean here.  
+
+We have revised the wording on page 8, lines 222- 226 for clarity. We used the \(^{19}\mathrm{F}\) - tag labeling site as a reference to track conformational changes and to illustrate our \(^{19}\mathrm{F}\) - qNMR data. The \(^{19}\mathrm{F}\) probe reveals a clockwise rotation as the receptor transitions towards activation, with the NMR signal for the S4 state appearing at a lower field compared to the S5 state, as shown in Fig. 4c.  
+
+4. I would suggest to show the nucleotide exchange data for with the addition of agonists and inverse agonists for the mutant (maybe just in the sup material)  
+
+Thank you for the suggestion. We have indeed performed nucleotide exchange experiments with the addition of both agonists and inverse agonists for the mutant. The data has been included in the revised manuscript and can be found in the supplementary material (Extended Data Fig. 3).  
+
+5. There are some more modelling errors. The unmodelled loop between 148 and 166 in the receptor will most likely have the N-terminus pointing in the wrong direction. As its currently modelled the only way for this loop to be completed is by forming a protein knot through ECL3.  
+
+Thank you for your detailed inspection. We were able to model V164, A165 into the density map, which provided a clear direction for the loop between residues 147- 164. Additionally, I corrected the fitting errors for F70 and C71. The updated model now clearly shows that the loop is not forming a knot through ECL3.  
+
+6. This reviewer has major issues with the use of DeepEMhancer for modelling. The map that the authors originally provided is a prime example of Al hallucination. The Al model started to turn the detergent micelle into protein looking density and should NOT be trusted. The map I calculated from the half maps and the authors traditionally sharpened map is more than suitable for PDB modelling. EM enhanced maps also
+
+<--- Page Split --->
+
+SHOULD NOT be the main map deposited in the PDB, as is taints the PDB and EMDB with data that is not suitable for interpretation.  
+
+We used DeepEMhancer map solely for illustration and initial modeling purposes. However, we have confirmed that all model refinements in Phenix were conducted using the raw map without the use of DeepEMhancer. For each EMDB and PDB entry, we have deposit both the sharpened map (sharpened by - 10Å in cisTEM) and the raw unsharpened map. In the revised manuscript, we have removed all references to the DeepEMhancer map.  
+
+Matthew Belousoff, Monash School of Pharmaceutical Sciences  
+
+Reviewer #3 (Remarks to the Author):  
+
+This manuscript reports the structure and function of a ligand- free GPCR- Gasβy intermediate complex. The authors determined the cryo- EM structure of an intermediate A2AR- mini- Gasβy complex. They presented experimental and computational evidences that the intermediate complex (S4- Gasβy) initiates a rate- limited nucleotide exchange without progressing to the fully activated complex (S5- Gasβy), in which the α- helical domain (AHD) of the Gas is partially open engaged by a second nucleotide. Based on these, they proposed a mechanistic model for the rate- limited nucleotide exchange in the intermediate GPCR- Gasβy complex.  
+
+Thanks for the commendation on our manuscript.  
+
+The experimental method presented in this paper is suitable and the MD simulations are well- designed. The approach can be applied to characterize transient intermediate states of other GPCRs. However, there are a few short comings in this study. I recommend that this work is publishable with revision as follows:  
+
+1. The authors determined the cryo-EM structure of an intermediate A2AR-mini- Gasβy complex. Has there any other intermediate state of other GPCRs reported in other papers? What's the similarity and difference between this intermediate structure and the others? It would be better to discuss this intermediate A2AR-mini-Gasβy structure in the context of the literature.  
+
+Thank you for the thoughtful suggestion. While no intermediate complex of A2AR has been reported so far, a partially active structure has been reported with the binding of LUF5833 and LUF5834<sup>4,5</sup>. In response to the reviewer's suggestion, we have added a discussion of other GPCR intermediate states in the literature to contextualize our findings, please refer to page 11, lines 318- 328.  
+
+Additionally, we compare our structure determined in this manuscript with all resolved
+
+<--- Page Split --->
+
+receptor- Gs complex for family A GPCR with the assistance of MD simulation (Fig.4c and 4d). Our results clearly indicate the uniqueness of our structure representing an intermediate position from the G protein perspective on both Ras- like domain (resolved structure) and AHD domain (simulated structure).  
+
+2. In the determined cryo-EM structure of the A2AR-mini-Gasβγ complex, the AHD of Gαs was missing. The mini-Gasβγ is different from the full-length Gs protein, will this influence the binding of Gs and A2AR?  
+
+We agree that using the mini- Gasβγ instead of the full- length Gs protein may influence binding. However, obtaining an intermediate GPCR- G protein complex with the full- length Gs protein remains challenging. We believe that protein- engineering of the G protein, similar to what we did for the receptor, would help stabilize an intermediate G protein conformation, and this is part of our future plans.  
+
+3. It is seen from Fig.1a that the most obvious difference between the intermediate complex (S4-Gasβγ) and other complex (S1-Gasβγ, S2-Gasβγ, S3-Gasβγ and S5-Gasβγ) is whether the AHD of Gs is partially open and whether there is GDP bound in Gs. It would be nice if the A2AR bound by the full-length Gs protein (with GDP bound in Gs) is characterized by cryo-EM.  
+
+Thank you for the suggestion. As mentioned above, we agree that characterizing the full- length Gs protein in complex with A2AR would provide more insight. This is part of our ongoing research, and we plan to explore this by engineering a stable full- length G protein in future studies.  
+
+4. The authors proposed two nucleotide-bound sites (Fig. 5, site 1 and site 2) in the transition process from the intermediate to the fully activated GPCR-Gasβγ complex. They performed additional GaMD simulations to examine the GDP release process in the cS4-Gasβγ system. GDP bound in site 2 of Gasβγ in the starting structure of GaMD simulations of GDP release. In the simulation of the GDP release process, can the GDP be observed to make a short stay at site 1?  
+
+Thank you for your detailed analysis. As shown in Fig. 5c, GDP rapidly transitions from the bound to the unbound state, with no intermediate state observed during this process. Similarly, as indicated in Extended Data Fig. 11e, no other low- energy states were detected. Additionally, the initial state of the G protein starts in a fully open conformation, and during GDP release, it remains largely open rather than adopting a partially open state. This could be the caveat of the MD here to capture the nucleotide binding to the site 2 (we redefine the original site 2 as the site 1 and the original site 1 as site 2) that was observed in our bench experiments. We will modify our model in the future or waiting for the structure of an intermediate A2AR- full length G protein resolved to revisit this phenomenon.
+
+<--- Page Split --->
+
+5. Free energy calculations indicated a "GDP Released" state in the S5-Gasβγ system and only a "Partially Released" state in the S4-Gasβγ system (Extended Data Fig. 8d and 8e). In this "Partially Released" state, is the GDP still binds in Gasβγ? Is there a possibility that it binds in the proposed "site 1"?  
+
+The representative structure of the 'Partially Released' state from the current research is shown below (Response Fig. 4). In this state, the phosphate group remains bound within the binding site, interacting with residues G52, S54, T55, and K293. Currently, we lack the data to confirm whether this corresponds to the proposed 'site 2.' Further studies will be conducted to investigate this hypothesis as mentioned above till we have the intermediate A2A-R-full length G protein complex and using it as a starting model for simulation.  
+
+![](images/Figure_4.jpg)
+
+<center>Response Figure 4. Representative conformation of the 'Partially Released' state is depicted. Key interacting residues are shown in stick representation, with red dotted lines indicating their interactions. </center>  
+
+6. Page 8 Lines 246: "Extended Data Figs. 8a-c". It should be Extended Data Figs. 8d-e. Page 8 Lines 244-249: "Extended Data Figs. 8d-e". It should be Extended Data Figs. 8a-c.  
+
+Thank you for pointing out these discrepancies. We have corrected the references to the figures (Page 10, line 286 and line 289).  
+
+## References:
+
+<--- Page Split --->
+
+receptor. Nat. Chem. Biol 18, 256- 263 (2022). https://doi.org/10.1038/s41589- 021- 00945- w2 Lin, X. et al. Structural basis of ligand recognition and self- activation of orphan GPR52. Nature 579, 152- 157 (2020). https://doi.org/10.1038/s41586- 020- 2019- 03Wang, X., Neale, C., Kim, S.- K., Goddard, W. A. & Ye, L. Intermediate- state- trapped mutants pinpoint G protein- coupled receptor conformational allostery. Nat. Commun. 14, 1325 (2023). https://doi.org/10.1038/s41467- 023- 36971- 64Claff, T. et al. Structural Insights into Partial Activation of the Prototypic G Protein- Coupled Adenosine A2A Receptor. ACS Pharmacol. Transl. Sci.. 7, 1415- 1425 (2024). https://doi.org/10.1021/acsptsci.4c00051Amelia, T. et al. Crystal Structure and Subsequent Ligand Design of a Nonriboside Partial Agonist Bound to the Adenosine A2A Receptor. J. Med. Chem. 64, 3827- 3842 (2021). https://doi.org/10.1021/acs.jmedchem.0c01856
+
+<--- Page Split --->
+
+We would like to thank all reviewers for their positive responses to our previous revision. In this final revision, we address all the remaining comments.  
+
+Reviewer #1 (Remarks to the Author):  
+
+The authors have addressed all my concerns and have updated the manuscript accordingly.  
+
+Some minor corrections:  
+
+1. Extended Data Figure 2, page 33: The figure caption seems to be mislabeled. I think the red data points represent the line width of S1 resonance and the teal data point represent the line width of the S4 resonance.  
+
+Thanks for pointing out this and we have corrected all these accordingly.  
+
+2. Page 5 lines 134-135: Omit the word domain after AHD  
+
+We have omitted the word "domain" after AHD and corrected all these throughout the manuscript.  
+
+3. Page 5 line 134-135: formed by Ras-like domain and AHD resulting from a partial opening of AHD.  
+
+We have changed the sentence as the reviewer suggested.  
+
+Reviewer #2 (Remarks to the Author):  
+
+Firstly, I wish to commend the authors on their significant efforts to reanalyse the cryoEM data. I am of the opinion that this has vastly improved the authors manuscript. I do not fundamentally disagree with the overall experimental design and overall analysis but I have a comment that the authors may wish to consider.  
+
+Our research group has performed 3DVA on over 300 GPCR ternary complexes and we observe incredibly similar dynamic sampling of both G- protein engagement angles and G- protein conformations. However, does this mean their interpretation is incorrect? I do not think so, it is more likely that they have found an appropriate experimental system to more fully explore and interpret the dynamics and 3DVA data that are commonly observed across all active state GPCR structures. Put another way, essentially all reported structures of active state GPCRs are an average of the S4- S5 state, but the
+
+<--- Page Split --->
+
+authors in this manuscript have perhaps skewed the overall population towards the S4 state and provided a great hypothesis for interpretation of the observed dynamic landscape in active state GPCR structures.  
+
+We appreciate this comment. Indeed, it is likely that our approach of combining 19F NMR and point mutation to bias the conformation towards S4 enabled us to characterize this intermediate state more thoroughly. We revised the discussion to reflect this point.  
+
+I would recommend this paper for publication as it is my opinion that the authors have appropriately addressed the reviewers concerns.  
+
+We thank Dr. Belousoff for his support in publishing our manuscript and appreciate his comments, which helped us improve it.  
+
+Matthew Belousoff, Monash Institute of Pharmaceutical Sciences  
+
+Reviewer #3 (Remarks to the Author):  
+
+The authors have addressed all my previous concerns with detailed explanations and appropriate revisions. The revisions have improved the overall quality of the manuscript. I have no further issues with the current version of the manuscript, and I believe it can be published.  
+
+We thank the Reviewer for his/her support.
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__725a24fa23db0a0160e585b53860a42133f92ec1af04267fd9d24707fe1d1a6c/supplementary_0_Transparent Peer Review file__725a24fa23db0a0160e585b53860a42133f92ec1af04267fd9d24707fe1d1a6c_det.mmd

@@ -0,0 +1,581 @@
+<|ref|>title<|/ref|><|det|>[[72, 53, 295, 80]]<|/det|>
+# nature portfolio  
+
+<|ref|>text<|/ref|><|det|>[[74, 96, 297, 118]]<|/det|>
+Peer Review File  
+
+<|ref|>title<|/ref|><|det|>[[73, 161, 900, 211]]<|/det|>
+# Structure and function of a near fully-activated intermediate GPCR-Goβγ complex  
+
+<|ref|>text<|/ref|><|det|>[[74, 224, 370, 240]]<|/det|>
+Corresponding Author: Dr Libin Ye  
+
+<|ref|>text<|/ref|><|det|>[[73, 274, 712, 289]]<|/det|>
+Parts of this Peer Review File have been redacted as indicated to remove third- party material.  
+
+<|ref|>text<|/ref|><|det|>[[70, 301, 866, 315]]<|/det|>
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+<|ref|>text<|/ref|><|det|>[[73, 353, 144, 366]]<|/det|>
+Version 0:  
+
+<|ref|>text<|/ref|><|det|>[[73, 379, 219, 393]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[73, 404, 160, 418]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[73, 431, 238, 444]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 444, 923, 580]]<|/det|>
+The manuscript by Bi M. et al. characterizes the intermediate S4 state of adenosine \(\mathsf{A}_{2\mathsf{A}}\) receptor \((\mathsf{A}_{2\mathsf{A}}\mathsf{R})\) using the conformation biased mutant R291A, which they have previously shown to trap the receptor in the S4 state. The authors found that the R291A mutant was able to bind Goβγ with a decreased GTP hydrolysis and nucleotide exchange rate that is independent of ligand. Using both cryo- EM and Gaussian accelerated molecular dynamics simulation (GaMD), the authors determined the structure of the R291A \(\mathsf{A}_{2\mathsf{A}}\mathsf{R}\) in complex with mini- Goβγ and proposed a mechanism by which the S4 state carry out the nucleotide exchange and transition to the fully activated state. This paper provides a great follow- up to the authors' previous paper on trapping the intermediate states of a GPCR using the mutants R291A and R293A. This work is interesting because the authors showed in atomic detail the orientation and the interactions of the relevant residues involved in the intermediate state. The results of this study increase our understanding on the features of the intermediate state and filled in the gap on the structural details of the GPCR activation landscape.  
+
+<|ref|>text<|/ref|><|det|>[[72, 592, 605, 606]]<|/det|>
+I have some comments that I think would help improve the paper if addressed:  
+
+<|ref|>text<|/ref|><|det|>[[73, 618, 153, 631]]<|/det|>
+Comments:  
+
+<|ref|>text<|/ref|><|det|>[[70, 643, 914, 670]]<|/det|>
+1. The authors found that the R291A-Goβγ complex has a lower Michaelis-Menten constant Km for GTP hydrolysis than the fully activated state (WT-Goβγ-NECA). Can the authors elaborate on the implication of this?  
+
+<|ref|>text<|/ref|><|det|>[[72, 681, 904, 724]]<|/det|>
+2. On page 5, lines 141-142, the authors said that "BODIPY-FL-GTP showed more efficient binding to the S5-mediated G protein with a higher \(K_D\) value." What do the authors mean by this? In Figures 2d and 2e, based on the difference in \(K_D\) , it seems that R291A-Goβγ has a stronger affinity toward GTP or GDP than the S5-mediated one.  
+
+<|ref|>text<|/ref|><|det|>[[72, 736, 912, 790]]<|/det|>
+3. I want to make sure that I understand it correctly. In Figure 2g, the intermediate state has a slightly faster rate of GDP exchange with GTP than the fully activated state, but since the fluorescence does not go to zero, it suggests that the BODIPY-FL-GDP is still bound, unlike the S5 state? Maybe the authors can include an additional statement explaining this observation on the text.  
+
+<|ref|>text<|/ref|><|det|>[[72, 802, 921, 868]]<|/det|>
+4. If I understand it correctly, the authors are trying to make a case that the 2nd association step observed in the BODIPY-FL-GDP binding assay with R291A is due to the limited access of the nucleotide to the second site by demonstrating that the association step is also observed with S5 state when BODIPY-FL-GTPyS is used since the bulky BODIPY group is on the y-phosphate side. However, this wasn't conveyed very clearly in the main text. The authors may consider revising their statements in the manuscript to better communicate the argument.  
+
+<|ref|>text<|/ref|><|det|>[[72, 879, 920, 922]]<|/det|>
+5. When the nucleotide accessed the second site in the R291A-Goβγ complex, the authors posit that the complex remains in the partially open S4 state. Did the authors try collecting the \(^{19}\mathrm{F}\) NMR spectrum of the R291A-Goβγ complex with GDP or GTP to see if the receptor does not sample a different state?
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 46, 920, 90]]<|/det|>
+6. Since the intermediate state of the \(\mathsf{A}_{2\mathsf{A}}\mathsf{R} - \mathsf{G}\alpha \beta \gamma\) complex can now be trapped and characterized, the application of this is that it is possible to design a therapeutic agent that can target this specific conformation of the receptor. Do the authors know of any disease or condition where this could be applied?  
+
+<|ref|>text<|/ref|><|det|>[[72, 101, 196, 114]]<|/det|>
+Minor Comments:  
+
+<|ref|>text<|/ref|><|det|>[[72, 114, 880, 142]]<|/det|>
+7. In the Methods section, GTPase hydrolysis assay, page 19, lines 436-442, change the description of the methods to passive voice.  
+
+<|ref|>text<|/ref|><|det|>[[72, 153, 808, 168]]<|/det|>
+8. Figure 2a, indicating which ligand is a full, partial, or inverse agonist on the figure would help the readers.  
+
+<|ref|>text<|/ref|><|det|>[[72, 179, 758, 194]]<|/det|>
+9. Figure 2c, the y-axis label is incorrect. I believe it should be velocity instead of GTP concentration.  
+
+<|ref|>text<|/ref|><|det|>[[72, 205, 918, 233]]<|/det|>
+10. Figure 2f, the legend label should be R291A+ Goβγ and WT\*+Goβγ+NECA for consistency with the legends in the other panels.  
+
+<|ref|>text<|/ref|><|det|>[[72, 244, 920, 286]]<|/det|>
+11. The ordering of the panels in Figures 3 and 4 could be rearranged as they are referenced in the text. In the current version of the manuscript, after Fig. 3b, Fig. 3f is then referenced, followed by Fig. 3d, 3e, 3g, then Fig. 3c. Same with Figure 4, Fig. 4f is referenced, followed by Fig. 4e, 4d, 4g-i.  
+
+<|ref|>text<|/ref|><|det|>[[72, 297, 537, 311]]<|/det|>
+12. Extended Figure 2a, missing legend for red and teal data points.  
+
+<|ref|>text<|/ref|><|det|>[[72, 323, 562, 337]]<|/det|>
+13. Kindly provide clearer images for Extended Data Figures 3c and 3d.  
+
+<|ref|>text<|/ref|><|det|>[[72, 348, 872, 377]]<|/det|>
+14. Extended Figure 3b, the bands for \(\mathsf{A}_{2\mathsf{A}}\mathsf{R}\) R291A and NB35 are not that visible on the SDS-PAGE gel for the most dominant peak (Fractions 18-20).  
+
+<|ref|>text<|/ref|><|det|>[[72, 389, 608, 403]]<|/det|>
+15. Extended Data Figure 4 caption: Schematic flowchart, instead of flow-chat.  
+
+<|ref|>text<|/ref|><|det|>[[72, 415, 920, 456]]<|/det|>
+16. The F376-M60 interaction was mentioned toward the end of the Results section, that it plays a significant role in the closed state of AHD, but this interaction was broken in both S4 and S5. Since this interaction is not a major key player in the transition between S4 and S5 states, Figures 4g, h, and I can probably be moved to the supplementary.  
+
+<|ref|>sub_title<|/ref|><|det|>[[72, 480, 162, 493]]<|/det|>
+## Reviewer #2  
+
+<|ref|>text<|/ref|><|det|>[[72, 505, 238, 531]]<|/det|>
+(Remarks to the Author) Review of Bi et al.  
+
+<|ref|>text<|/ref|><|det|>[[72, 544, 905, 612]]<|/det|>
+The manuscript by Bi and coworkers reports the structure of mutant structure of A2A in complex with a mini-Gs. The aim of this work is to describe what the authors term an 'intermediate' transition state. This work was mostly guided by F19 NMR experiments where they observed an incomplete chemical shift to a fully active state structure. While the experiments are largely well conceived and performed, this reviewer has some fundamental issues with the interpretation of the data. I will say at the outset that I'm not a molecular dynamics expert and will not review that portion of the manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[72, 623, 918, 715]]<|/det|>
+Firstly, I am absolutely not convinced that they have isolated a ligand free structure. In the maps the authors shared and one that I calculated (from their half maps), there is very clear density of something that hasn't been modelled in the active site of their structure. It is clearly not a ligand free structure, in fact Adenosine perfectly fits in the unmodelled density. It is not clear to me at all that the authors have isolated and solved a structure of an 'intermediate' S4 state. Their evidence that they have rests largely on their F19 NMR studies. While these NMR experiments are a first step in assessing the bulk chemical environments that the Fluorine samples in their different constructs it may not be the most robust way to determine that they have fully trapped an S4 state.  
+
+<|ref|>text<|/ref|><|det|>[[72, 726, 916, 780]]<|/det|>
+I'm also not sure the GTP exchange assays are fully formed. While there is clear phenotypic differences between their WT\* and the R291A mutant, it is not clear to this reviewer that this is due to some intermediate state. As no G-protein association assays (BRET or FRET based) or G-protein turnover assays were performed this reviewer is concerned that the phenotypic differences could be more due to differential G-protein association rather than trapping an S4 state.  
+
+<|ref|>text<|/ref|><|det|>[[72, 791, 920, 923]]<|/det|>
+Furthermore, during the processing of the SPA data in their final 3D classification prior to consensus map refinement it appears that they may have 'cherry-picked' a particular dynamic state which may or may not reflect an S4 state. As all their 3D classes looked to be of high quality (2nd last row in Sup Figure 4), their final consensus reconstruction may not be reflective of the totality of states captured during the SPA experiment. Also, there was much discussion around relatively minor differences between their structure and another active state A2A structure, but the authors performed no 3D variability to assess the extent of dynamic states captured. In this reviewers opinion all the particles in the 3D classification step should be included in a 3DVA analysis (either cryoSPARC or cryoDRGN) to judge if their analysis of the PDB coordinates is justified or if they have simply captured one of many stable subsets of an active state structure. Even a comparison to their other 3D classes would be interesting to see if any of them look more like the active state structure that was compared against (6GBG).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[73, 46, 920, 100]]<|/det|>
+To summarize, I'm not convinced that the authors have captured a ligand free structure and that they have captured an intermediate state structure. Much more analysis of their SPA data is warranted and perhaps some pharmacology type experiments to tease out what is going on with G- protein association to more robustly interpret their GTP experiments would be necessary to confirm the claims this manuscript makes.  
+
+<|ref|>text<|/ref|><|det|>[[73, 113, 210, 126]]<|/det|>
+Some minor issues:  
+
+<|ref|>text<|/ref|><|det|>[[70, 125, 925, 321]]<|/det|>
+Some minor issues:1. Many side chains are unmodelled in the PDB, many of which have interpretable density (for example M177 and Y271 in receptor, to name just a few).2. The Figure Legend in Figure 1 is mislabelled (d and f)3. I don't know what the authors mean in Page 6. Line 183 "Closer inspection of the Tm6.... Using F19 ... revealed a clockwise rotation..." This needs some explanation as I can't see what the authors mean here.4. I would suggest to show the nucleotide exchange data for with the addition of agonists and inverse agonists for the mutant (maybe just in the sup material)5. There are some more modelling errors. The unmodelled loop between 148 and 166 in the receptor will most likely have the N-terminus pointing in the wrong direction. As its currently modelled the only way for this loop to be completed is by forming a protein knot through ECL3.6. This reviewer has major issues with the use of DeepEMhancer for modelling. The map that the authors originally provided is a prime example of Al hallucination. The Al model started to turn the detergent micelle into protein looking density and should NOT be trusted. The map I calculated from the half maps and the authors traditionally sharpened map is more than suitable for PDB modelling. EM enhanced maps also SHOULD NOT be the main map deposited in the PDB, as is taints the PDB and EMDB with data that is not suitable for interpretation.  
+
+<|ref|>text<|/ref|><|det|>[[73, 333, 512, 347]]<|/det|>
+Matthew Belousoff, Monash School of Pharmaceutical Sciences  
+
+<|ref|>sub_title<|/ref|><|det|>[[73, 372, 162, 386]]<|/det|>
+## Reviewer #3  
+
+<|ref|>text<|/ref|><|det|>[[73, 398, 238, 411]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 411, 890, 490]]<|/det|>
+This manuscript reports the structure and function of a ligand- free GPCR- G\(\alpha \beta \gamma\) intermediate complex. The authors determined the cryo- EM structure of an intermediate A2AR- mini- G\(\alpha \beta \gamma\) complex. They presented experimental and computational evidences that the intermediate complex (S4- G\(\alpha \beta \gamma\) ) initiates a rate- limited nucleotide exchange without progressing to the fully activated complex (S5- G\(\alpha \beta \gamma\) ), in which the \(\alpha\) - helical domain (AHD) of the G\(\alpha \beta \gamma\) is partially open engaged by a second nucleotide. Based on these, they proposed a mechanistic model for the rate- limited nucleotide exchange in the intermediate GPCR- G\(\alpha \beta \gamma\) complex.  
+
+<|ref|>text<|/ref|><|det|>[[72, 489, 910, 529]]<|/det|>
+The experimental method presented in this paper is suitable and the MD simulations are well- designed. The approach can be applied to characterize transient intermediate states of other GPCRs. However, there are a few short comings in this study. I recommend that this work is publishable with revision as follows:  
+
+<|ref|>text<|/ref|><|det|>[[72, 528, 910, 580]]<|/det|>
+1. The authors determined the cryo-EM structure of an intermediate A2AR-mini-G\(\alpha \beta \gamma\) complex. Has there any other intermediate state of other GPCRs reported in other papers? What's the similarity and difference between this intermediate structure and the others? It would be better to discuss this intermediate A2AR-mini-G\(\alpha \beta \gamma\) structure in the context of the literature.  
+
+<|ref|>text<|/ref|><|det|>[[70, 579, 910, 608]]<|/det|>
+2. In the determined cryo-EM structure of the A2AR-mini-G\(\alpha \beta \gamma\) complex, the AHD of G\(\alpha \beta \gamma\) was missing. The mini-G\(\alpha \beta \gamma\) is different from the full-length G\(\alpha \beta \gamma\) protein, will this influence the binding of G\(\alpha \beta \gamma\) ?  
+
+<|ref|>text<|/ref|><|det|>[[72, 607, 910, 660]]<|/det|>
+3. It is seen from Fig.1a that the most obvious difference between the intermediate complex (S4-G\(\alpha \beta \gamma\) ) and other complex (S1-G\(\alpha \beta \gamma\) , S2-G\(\alpha \beta \gamma\) , S3-G\(\alpha \beta \gamma\) and S5-G\(\alpha \beta \gamma\) ) is whether the AHD of G\(\alpha \beta \gamma\) is partially open and whether there is GDP bound in G\(\alpha \beta \gamma\) . It would be nice if the A2AR bound by the full-length G\(\alpha \beta \gamma\) protein (with GDP bound in G\(\alpha \beta \gamma\) ) is characterized by cryo-EM.  
+
+<|ref|>text<|/ref|><|det|>[[72, 658, 915, 712]]<|/det|>
+4. The authors proposed two nucleotide-bound sites (Fig. 5, site 1 and site 2) in the transition process from the intermediate to the fully activated GPCR-G\(\alpha \beta \gamma\) complex. They performed additional GaMD simulations to examine the GDP release process in the cS4-G\(\alpha \beta \gamma\) system. GDP bound in site 2 of G\(\alpha \beta \gamma\) in the starting structure of GaMD simulations of GDP release. In the simulation of the GDP release process, can the GDP be observed to make a short stay at site 1?  
+
+<|ref|>text<|/ref|><|det|>[[72, 711, 920, 751]]<|/det|>
+5. Free energy calculations indicated a "GDP Released" state in the S5-G\(\alpha \beta \gamma\) system and only a "Partially Released" state in the S4-G\(\alpha \beta \gamma\) system (Extended Data Fig. 8d and 8e). In this "Partially Released" state, is the GDP still binds in G\(\alpha \beta \gamma\) ? Is there a possibility that it binds in the proposed "site 1"?  
+
+<|ref|>text<|/ref|><|det|>[[72, 750, 688, 764]]<|/det|>
+6. Page 8 Lines 246: "Extended Data Figs. 8a-c". It should be Extended Data Figs. 8d-e.  
+
+<|ref|>text<|/ref|><|det|>[[72, 763, 688, 777]]<|/det|>
+Page 8 Lines 244- 249: "Extended Data Figs. 8d-e". It should be Extended Data Figs. 8a-c.  
+
+<|ref|>text<|/ref|><|det|>[[73, 802, 144, 815]]<|/det|>
+Version 1:  
+
+<|ref|>text<|/ref|><|det|>[[73, 828, 219, 841]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[73, 854, 160, 867]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[73, 880, 238, 892]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 893, 701, 906]]<|/det|>
+The authors have addressed all my concerns and have updated the manuscript accordingly.  
+
+<|ref|>text<|/ref|><|det|>[[73, 919, 241, 932]]<|/det|>
+Some minor corrections:
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[70, 46, 920, 75]]<|/det|>
+1. Extended Data Figure 2, page 33: The figure caption seems to be mislabeled. I think the red data points represent the line width of S1 resonance and the teal data point represent the line width of the S4 resonance.  
+
+<|ref|>text<|/ref|><|det|>[[72, 86, 468, 101]]<|/det|>
+2. Page 5 lines 134-135: Omit the word domain after AHD  
+
+<|ref|>text<|/ref|><|det|>[[70, 111, 763, 127]]<|/det|>
+3. Page 5 line 134-135: formed by Ras-like domain and AHD resulting from a partial opening of AHD.  
+
+<|ref|>sub_title<|/ref|><|det|>[[72, 152, 162, 165]]<|/det|>
+## Reviewer #2  
+
+<|ref|>text<|/ref|><|det|>[[72, 177, 238, 191]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 191, 904, 231]]<|/det|>
+Firstly, I wish to commend the authors on their significant efforts to reanalyse the cryoEM data. I am of the opinion that this has vastly improved the authors manuscript. I do not fundamentally disagree with the overall experimental design and overall analysis but I have a comment that the authors may wish to consider.  
+
+<|ref|>text<|/ref|><|det|>[[72, 242, 921, 333]]<|/det|>
+Our research group has performed 3DVA on over 300 GPCR ternary complexes and we observe incredibly similar dynamic sampling of both G- protein engagement angles and G- protein conformations. However, does this mean their interpretation is incorrect? I do not think so, it is more likely that they have found an appropriate experimental system to more fully explore and interpret the dynamics and 3DVA data that are commonly observed across all active state GPCR structures. Put another way, essentially all reported structures of active state GPCRs are an average of the S4- S5 state, but the authors in this manuscript have perhaps skewed the overall population towards the S4 state and provided a great hypothesis for interpretation of the observed dynamic landscape in active state GPCR structures.  
+
+<|ref|>text<|/ref|><|det|>[[70, 345, 921, 373]]<|/det|>
+I would recommend this paper for publication as it is my opinion that the authors have appropriately addressed the reviewers concerns.  
+
+<|ref|>text<|/ref|><|det|>[[72, 384, 518, 399]]<|/det|>
+Matthew Belousoff, Monash Institute of Pharmaceutical Sciences  
+
+<|ref|>text<|/ref|><|det|>[[72, 411, 162, 424]]<|/det|>
+Reviewer #3  
+
+<|ref|>text<|/ref|><|det|>[[72, 437, 238, 450]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 450, 911, 490]]<|/det|>
+The authors have addressed all my previous concerns with detailed explanations and appropriate revisions. The revisions have improved the overall quality of the manuscript. I have no further issues with the current version of the manuscript, and I believe it can be published.  
+
+<|ref|>text<|/ref|><|det|>[[72, 830, 916, 883]]<|/det|>
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+<|ref|>text<|/ref|><|det|>[[72, 883, 796, 897]]<|/det|>
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source.  
+
+<|ref|>text<|/ref|><|det|>[[72, 897, 911, 936]]<|/det|>
+The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 48, 620, 75]]<|/det|>
+permission directly from the copyright holder. To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[119, 85, 457, 101]]<|/det|>
+Reviewer #1 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[118, 121, 880, 380]]<|/det|>
+The manuscript by Bi M. et al. characterizes the intermediate S4 state of adenosine \(\mathsf{A}_{2\mathsf{A}}\) receptor \((\mathsf{A}_{2\mathsf{A}}\mathsf{R})\) using the conformation biased mutant R291A, which they have previously shown to trap the receptor in the S4 state. The authors found that the R291A mutant was able to bind \(\mathsf{G}\mathsf{a}\beta \gamma\) with a decreased GTP hydrolysis and nucleotide exchange rate that is independent of ligand. Using both cryo- EM and Gaussian accelerated molecular dynamics simulation (GaMD), the authors determined the structure of the R291A \(\mathsf{A}_{2\mathsf{A}}\mathsf{R}\) in complex with mini- \(\mathsf{G}\mathsf{a}\beta \gamma\) and proposed a mechanism by which the S4 state carry out the nucleotide exchange and transition to the fully activated state. This paper provides a great follow- up to the authors' previous paper on trapping the intermediate states of a GPCR using the mutants R291A and R293A. This work is interesting because the authors showed in atomic detail the orientation and the interactions of the relevant residues involved in the intermediate state. The results of this study increase our understanding on the features of the intermediate state and filled in the gap on the structural details of the GPCR activation landscape.  
+
+<|ref|>text<|/ref|><|det|>[[120, 400, 549, 417]]<|/det|>
+Thanks for the appreciation on our new findings.  
+
+<|ref|>text<|/ref|><|det|>[[120, 436, 817, 454]]<|/det|>
+I have some comments that I think would help improve the paper if addressed:  
+
+<|ref|>text<|/ref|><|det|>[[119, 474, 222, 490]]<|/det|>
+Comments:  
+
+<|ref|>text<|/ref|><|det|>[[119, 510, 879, 565]]<|/det|>
+1. The authors found that the R291A- \(\mathsf{G}\mathsf{a}\beta \gamma\) complex has a lower Michaelis-Menten constant Km for GTP hydrolysis than the fully activated state (WT- \(\mathsf{G}\mathsf{a}\beta \gamma\) -NECA). Can the authors elaborate on the implication of this?  
+
+<|ref|>text<|/ref|><|det|>[[118, 585, 880, 714]]<|/det|>
+The lower Km implies that the GTP- GDP conversion rate at the non- canonical nucleotide binding site 2 in the R291A- \(\mathsf{G}\mathsf{a}\beta \gamma\) complex was much smaller than at the canonical site (site 1) in the WT- \(\mathsf{G}\mathsf{a}\beta \gamma\) -NECA, indicating a limited capacity of the intermediate complex in regulating GTP hydrolysis while the fully activated complex holds the full capacity. Of note, we redefine the canonical binding site as the site 1 while the noncanonical binding site as the site 2. This also implies that the GTP doesn't bind to the canonical nucleotide binding site in the R291A- \(\mathsf{G}\mathsf{a}\beta \gamma\) complex.  
+
+<|ref|>text<|/ref|><|det|>[[118, 732, 880, 824]]<|/det|>
+2. On page 5, lines 141-142, the authors said that "BODIPY-FL-GTP showed more efficient binding to the S5-mediated G protein with a higher \(K_D\) value." What do the authors mean by this? In Figures 2d and 2e, based on the difference in \(K_D\) , it seems that R291A- \(\mathsf{G}\mathsf{a}\beta \gamma\) has a stronger affinity toward GTP or GDP than the S5-mediated one.  
+
+<|ref|>text<|/ref|><|det|>[[118, 844, 881, 899]]<|/det|>
+We appreciate the reviewer's insight. Our interpretation is that the GTP binding site in the S5- mediated G protein is freely accessible because of the full open of AHD domain, resulting in a robust BODIPY- FL- GTP association- dissociation process. In contrast,
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 84, 880, 251]]<|/det|>
+due to the space limitation caused by a partial open pose of AHD domain in the S4- mediated G protein, the binding of BODIPY- FL- GTP to the canonical nucleotide binding site is not accessible. The binding process is halted at the stage that Ras- like domain and AHD domain hold the nucleotide en route to the fully open state. We think this nucleotide binding site 2 is close to the canonical nucleotide binding site 1, which allows it to exhibit a limited GTP hydrolysis capacity, along with a slow dissociation rate as presented in the figures. This also results in the R291A- Gαβγ exhibiting a stronger affinity toward GTP or GDP than the S5- mediated one, as pointed out by the reviewer.  
+
+<|ref|>text<|/ref|><|det|>[[117, 270, 880, 363]]<|/det|>
+3. I want to make sure that I understand it correctly. In Figure 2g, the intermediate state has a slightly faster rate of GDP exchange with GTP than the fully activated state, but since the fluorescence does not go to zero, it suggests that the BODIPY- FL- GDP is still bound, unlike the S5 state? Maybe the authors can include an additional statement explaining this observation on the text.  
+
+<|ref|>text<|/ref|><|det|>[[117, 380, 880, 491]]<|/det|>
+Yes. The data in Fig. 2 indicated that the release number of BODIPY- FL- GDP from the S4 state was much less than the S5 state, but it reached the equilibrium much faster, resulting in the \(K_{\text{off}}\) actually 0.09/min vs 0.05/min (Page 5, lines 153- 156). This was also consistent with our earlier findings in Figs. 2d and 2e showing that a substantial proportion of BODIPY- FL- GDP/GTP remained bound in the S4 state mediated G proteins.  
+
+<|ref|>text<|/ref|><|det|>[[117, 510, 880, 640]]<|/det|>
+4. If I understand it correctly, the authors are trying to make a case that the 2nd association step observed in the BODIPY- FL- GDP binding assay with R291A is due to the limited access of the nucleotide to the second site by demonstrating that the association step is also observed with S5 state when BODIPY- FL- GTPγS is used since the bulky BODIPY group is on the γ-phosphate side. However, this wasn't conveyed very clearly in the main text. The authors may consider revising their statements in the manuscript to better communicate the argument.  
+
+<|ref|>text<|/ref|><|det|>[[117, 658, 879, 732]]<|/det|>
+Yes, our conclusion is well articulated by the reviewer. We used the BODIPY- FL- GTP- γ- S as a case to confirm that the non- canonical binding site 2 is formed by the Ras- like domain and a partial open AHD domain that limits nucleotide to access canonical site 1. (Page 5, lines 148- 152).  
+
+<|ref|>text<|/ref|><|det|>[[117, 751, 879, 824]]<|/det|>
+5. When the nucleotide accessed the second site in the R291A- Gαβγ complex, the authors posit that the complex remains in the partially open S4 state. Did the authors try collecting the \(^{19}\mathrm{F}\) NMR spectrum of the R291A- Gαβγ complex with GDP or GTP to see if the receptor does not sample a different state  
+
+<|ref|>text<|/ref|><|det|>[[117, 843, 879, 898]]<|/det|>
+Yes, the \(^{19}\mathrm{F}\) - NMR spectra were acquired with GDP included. We did see a small portion ( \(\sim 10\%\) ) of S3 state in the sample of R291A- Gαβγ while the receptor was unable to shift to the S5 state (Fig. 1b and Extended Data Fig.1).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 101, 880, 177]]<|/det|>
+6. Since the intermediate state of the \(\mathsf{A}_{2\mathsf{A}}\mathsf{R} - \mathsf{G}\mathsf{a}\beta \mathsf{y}\) complex can now be trapped and characterized, the application of this is that it is possible to design a therapeutic agent that can target this specific conformation of the receptor. Do the authors know of any disease or condition where this could be applied?  
+
+<|ref|>text<|/ref|><|det|>[[117, 195, 880, 380]]<|/det|>
+Thanks for the insights regarding the correlation between conformation and disease. So far, we didn't explore the details and the distinction of signaling pathways that S4 and S5 could respectively regulate, considering the complexity of signaling involved with various G proteins, GRKs, and \(\beta\) - arrestins. However, we acknowledge the potential for therapeutic applications and this type of work has been in our plan. In May 2024, a clinic variance R291C (NCBI, SNP: rs745714462). was reported, despite the functional data of this variance is still missing in the report. Considering the similarity of R291C and R291A from the molecular interaction level, we will continue and expand our study to use the knock-in mouse model to study the correlation between the conformation and disease in the future.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 400, 280, 417]]<|/det|>
+## Minor Comments:  
+
+<|ref|>text<|/ref|><|det|>[[118, 418, 880, 454]]<|/det|>
+7. In the Methods section, GTPase hydrolysis assay, page 19, lines 436-442, change the description of the methods to passive voice.  
+
+<|ref|>text<|/ref|><|det|>[[118, 473, 880, 510]]<|/det|>
+We have revised the description in the GTPase hydrolysis assay section to passive voice to improve clarity and consistency. (Pages 21- 22, lines 479- 487)  
+
+<|ref|>text<|/ref|><|det|>[[118, 528, 880, 565]]<|/det|>
+8. Figure 2a, indicating which ligand is a full, partial, or inverse agonist on the figure would help the readers.  
+
+<|ref|>text<|/ref|><|det|>[[118, 584, 880, 621]]<|/det|>
+We have updated Fig. 2a to clearly indicate which ligands are full, partial, or inverse agonists, making it easier for readers to interpret the data. (Page 14, Fig. 2a)  
+
+<|ref|>text<|/ref|><|det|>[[118, 639, 880, 675]]<|/det|>
+9. Figure 2c, the y-axis label is incorrect. I believe it should be velocity instead of GTP concentration.  
+
+<|ref|>text<|/ref|><|det|>[[118, 695, 880, 732]]<|/det|>
+We have corrected the y-axis label in Fig. 2c to "velocity" for accuracy. (Page 14, Fig. 2c).  
+
+<|ref|>text<|/ref|><|det|>[[118, 750, 880, 787]]<|/det|>
+10. Figure 2f, the legend label should be R291A+ Gaβγ and WT\*+Gaβγ+NECA for consistency with the legends in the other panels.  
+
+<|ref|>text<|/ref|><|det|>[[118, 806, 880, 843]]<|/det|>
+We have revised the legend in Fig. 2f for consistency with the other panels. (Page 14, Fig. 2f)  
+
+<|ref|>text<|/ref|><|det|>[[118, 862, 880, 899]]<|/det|>
+11. The ordering of the panels in Figures 3 and 4 could be rearranged as they are referenced in the text. In the current version of the manuscript, after Fig. 3b, Fig. 3f is
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 880, 120]]<|/det|>
+then referenced, followed by Fig. 3d, 3e, 3g, then Fig. 3c. Same with Figure 4, Fig. 4f is referenced, followed by Fig. 4e, 4d, 4g- i.  
+
+<|ref|>text<|/ref|><|det|>[[118, 139, 880, 250]]<|/det|>
+Thank you for pointing this out. A new figure was added before the previous version of Figure 3, resulting in a renumbering of figures. We have carefully reviewed the figure references and revised the order of the panels in Figs. 4 and 5 to match the sequence in which they are discussed in the text (Pages 8- 10). This adjustment ensures better flow and alignment between the manuscript and the figures. We appreciate your feedback in helping to improve the clarity and organization of the manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[120, 269, 725, 287]]<|/det|>
+12. Extended Figure 2a, missing legend for red and teal data points.  
+
+<|ref|>text<|/ref|><|det|>[[118, 306, 880, 361]]<|/det|>
+We appreciate you bringing this to our attention. The missing legend for the red and teal data points in Extended Data Fig. 2a has been added. (Page 33, Extended Data Fig. 2).  
+
+<|ref|>text<|/ref|><|det|>[[120, 380, 757, 398]]<|/det|>
+13. Kindly provide clearer images for Extended Data Figures 3c and 3d.  
+
+<|ref|>text<|/ref|><|det|>[[118, 417, 879, 472]]<|/det|>
+We have reprocessed and replaced Extended Data Figures 4c and 4d with higher- resolution images to improve clarity and ensure the details are more visible. (Page 35, Extended Data Fig.4).  
+
+<|ref|>text<|/ref|><|det|>[[118, 492, 880, 528]]<|/det|>
+14. Extended Figure 3b, the bands for \(\mathsf{A}_{2\mathsf{A}}\mathsf{R}\) R291A and NB35 are not that visible on the SDS-PAGE gel for the most dominant peak (Fractions 18-20).  
+
+<|ref|>text<|/ref|><|det|>[[118, 547, 880, 639]]<|/det|>
+The reduced visibility of the NB35 band was because we didn't heat the samples but merely incubated the sample with a loading buffer for better visibility of the receptor. This process could be the reason for the intensity decrease of soluble proteins like NB35. The same pattern was observed in others' reports as well, please refer to Response Fig. \(1^{1,2}\) .  
+
+<|ref|>text<|/ref|><|det|>[[118, 828, 884, 880]]<|/det|>
+Response Figure 1. SDS- PAGE analysis of GPCR complex. a GPR52 receptor complex, including G \(\beta \gamma\) , mini- G \(\alpha_{s}\) , and NB35 (Lin et al., 2020, Nature). b the purified components of the GLP- 1R complex include G \(\alpha_{s}\) , G \(\beta \gamma\) and NB35 (Cary et al., 2022, Nature Chemical Biology).
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[120, 85, 819, 102]]<|/det|>
+15. Extended Data Figure 4 caption: Schematic flowchart, instead of flow-chat.  
+
+<|ref|>text<|/ref|><|det|>[[118, 122, 880, 177]]<|/det|>
+Thank you for catching this typo. The caption for Extended Data Fig. 6 in the revised version has been corrected to "Schematic flowchart." (Page 37, Extended Data Fig. 6).  
+
+<|ref|>text<|/ref|><|det|>[[118, 196, 880, 288]]<|/det|>
+16. The F376-M60 interaction was mentioned toward the end of the Results section, that it plays a significant role in the closed state of AHD, but this interaction was broken in both S4 and S5. Since this interaction is not a major key player in the transition between S4 and S5 states, Figures 4g, h, and I can probably be moved to the supplementary.  
+
+<|ref|>text<|/ref|><|det|>[[118, 308, 880, 343]]<|/det|>
+Thank you for your suggestion. We have moved these panels to the supplementary and labeled them as Extended Data Figs. 12a, b, and c, as recommended.  
+
+<|ref|>text<|/ref|><|det|>[[118, 400, 457, 416]]<|/det|>
+Reviewer #2 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[118, 437, 280, 453]]<|/det|>
+Review of Bi et al.  
+
+<|ref|>text<|/ref|><|det|>[[117, 473, 881, 621]]<|/det|>
+The manuscript by Bi and coworkers reports the structure of mutant structure of A2A in complex with a mini- Gs. The aim of this work is to describe what the authors term an 'intermediate' transition state. This work was mostly guided by F19 NMR experiments where they observed an incomplete chemical shift to a fully active state structure. While the experiments are largely well conceived and performed, this reviewer has some fundamental issues with the interpretation of the data. I will say at the outset that I'm not a molecular dynamics expert and will not review that portion of the manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[117, 640, 880, 731]]<|/det|>
+1. Firstly, I am absolutely not convinced that they have isolated a ligand free structure. In the maps the authors shared and one that I calculated (from their half maps), there is very clear density of something that hasn't been modelled in the active site of their structure. It is clearly not a ligand free structure, in fact Adenosine perfectly fits in the unmodelled density.  
+
+<|ref|>text<|/ref|><|det|>[[117, 751, 880, 899]]<|/det|>
+We thank the reviewer for pointing this out. Our previous interpretation that the structure is ligand- free was based on experiments in which no exogenous ligands were added, and we did not consider the endogenous adenosine binding to the structure. We appreciate the reviewer's insight. Following his suggestion, we conducted additional analysis using a focused mask and confirmed that there is a clear extra density within the ligand- binding pocket, which can be modeled as adenosine. Subsequent mass spectrometry (Extended Data Fig.5 in the main text) confirmed the presence of adenosine in our sample. These findings suggest that the R291A mutant
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 880, 140]]<|/det|>
+stabilizes the receptor in a ligand- bound intermediate S4 state, distinct from the fully activated, ligand- bound S5 state. We have revised the text and figures accordingly and included the mass spectrometry result in the manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[118, 176, 880, 270]]<|/det|>
+2. It is not clear to me at all that the authors have isolated and solved a structure of an 'intermediate' S4 state. Their evidence that they have rests largely on their F19 NMR studies. While these NMR experiments are a first step in assessing the bulk chemical environments that the Fluorine samples in their different constructs it may not be the most robust way to determine that they have fully trapped an S4 state.  
+
+<|ref|>text<|/ref|><|det|>[[117, 287, 880, 491]]<|/det|>
+With respect, we disagree with the reviewer's comment. \(^{19}\mathrm{F}\) NMR is a widely accepted and robust method for mapping conformational ensembles or landscapes. As demonstrated in our previous publication \(^{3}\) , the R291A mutation significantly enriched the population of the S4 state. While it is theoretically possible to capture intermediate states in SPA through classification alone, this specific mutation greatly enriches the S4 population, which in turn enables both detailed structural analysis and functional measurements of GTP hydrolysis at this intermediate state. In response to the reviewer's concern, we performed thorough classification, which revealed dynamic structural fluctuations centered around a predominant conformation, distinct from the well-defined S5 state. We interpret this predominant conformation as the S4 state, consistent with our \(^{19}\mathrm{F}\) NMR results.  
+
+<|ref|>text<|/ref|><|det|>[[117, 510, 880, 658]]<|/det|>
+As an orthogonal approach, we validated the existence of this intermediate S4 state through MD simulations, using the 6GDG structure with the R291A mutation and a full G- protein. These simulations, conducted independently of the cryo- EM data, confirmed a high structural similarity between the simulated S4 state and our cryo- EM structure. This provides further confidence that we have indeed captured the intermediate S4 state. We appreciate the reviewer's suggestion and acknowledge that the term "trapped" may have implied a more homogenous population than intended. We have revised the manuscript accordingly to clarify this.  
+
+<|ref|>text<|/ref|><|det|>[[117, 677, 880, 788]]<|/det|>
+3. I'm also not sure the GTP exchange assays are fully formed. While there is clear phenotypic differences between their WT\* and the R291A mutant, it is not clear to this reviewer that this is due to some intermediate state. As no G-protein association assays (BRET or FRET based) or G-protein turnover assays were performed this reviewer is concerned that the phenotypic differences could be more due to differential G-protein association rather than trapping an S4 state.  
+
+<|ref|>text<|/ref|><|det|>[[117, 807, 880, 899]]<|/det|>
+We have conducted GTP turnover assays in our previous publication (Xudong Wang, et al. Nature Communications, 2023, Supplement Fig. 7a) \(^{3}\) . We also attach the figure here (Response Fig. 2) as well as the current research (Figs. 2a- c) in addition to nucleotide exchange experiments (Figs. 2d- g). The experiments clearly showed that R291A mutant exhibited limited capacities of GTP turnover and nucleotide exchanges.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[248, 145, 730, 451]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[118, 464, 880, 573]]<|/det|>
+<center>Response Figure 2. GTP hydrolysis assays for different mutants. GTP hydrolysis assays for WT\* and mutants R291A (teal), R293A, and R291AR293A (violet) as a function of receptor concentrations from \(0.25 \mu \mathrm{M}\) to \(1 \mu \mathrm{M}\) . Of note, the references were normalized in two sets of measurements. Data with error bars are presented as mean±SEM of four independent experiments. Statistical analyses were performed using the ordinary oneway ANOVA followed the two-sides sidak's post-hoc test in PRISM 9.3.1, \*\*\*p< 0.001, in comparison to the WT\*. </center>  
+
+<|ref|>text<|/ref|><|det|>[[117, 602, 880, 899]]<|/det|>
+4. Furthermore, during the processing of the SPA data in their final 3D classification prior to consensus map refinement it appears that they may have 'cherry-picked' a particular dynamic state which may or may not reflect an S4 state. As all their 3D classes looked to be of high quality (2nd last row in Sup Figure 4), their final consensus reconstruction may not be reflective of the totality of states captured during the SPA experiment. Also, there was much discussion around relatively minor differences between their structure and another active state A2A structure, but the authors performed no 3D variability to assess the extent of dynamic states captured. In this reviewers opinion all the particles in the 3D classification step should be included in a 3DVA analysis (either cryoSPARC or cryoDRGN) to judge if their analysis of the PDB coordinates is justified or if they have simply captured one of many stable subsets of an active state structure. Even a comparison to their other 3D classes would be interesting to see if any of them look more like the active state structure that was compared against (6GBG). To summarize, I'm not convinced that the authors have captured a ligand free structure and that they have captured an intermediate state structure. Much more analysis of their SPA data is warranted and perhaps some
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 879, 140]]<|/det|>
+pharmacology type experiments to tease out what is going on with G- protein association to more robustly interpret their GTP experiments would be necessary to confirm the claims this manuscript makes.  
+
+<|ref|>text<|/ref|><|det|>[[117, 158, 881, 362]]<|/det|>
+We thank the reviewer for these comments, which led us to conduct a more detailed re- analysis of our data. In response, we revisited the entire particle set and all noted classes. Through extensive classification and 3D variability analysis (3DVA), we reconfirmed that the S4 conformation identified in our initial submission remains the predominant conformation. This conformation shows clear distinctions from the well- defined S5 state (6GDG). C- α displacement analysis across all classes indicated that the conformational snapshots cluster around this predominant S4 conformation, with most fluctuations occurring in the Gy subunit. This finding is consistent with our \(^{19}\mathrm{F}\) NMR data, which revealed an enrichment centered around a defined peak, further supporting the identification of S4 as a distinct intermediate state on the trajectory to full activation. Additional receptor pocket calculations are provided in Response Fig.3.  
+
+<|ref|>text<|/ref|><|det|>[[117, 380, 881, 548]]<|/det|>
+We focused on two snapshots that show the largest deviations from the predominant S4 conformation, which we refer to as S4d1 and S4d2. The remaining snapshots exhibit only minor variations in the G protein, while in S4d1, the outermost region of the G protein swings by up to 7.5 Å compared to the S4 state, likely due to intermediate G- protein engagement. However, the receptor itself displays only subtle variations within the overall S4 conformation. We have updated the text and figures to include these findings and have illustrated the conformational similarities among the classes using per- residue C- α displacement calculations, as shown in Extended Data Fig. 9 (Page 40).  
+
+<|ref|>image<|/ref|><|det|>[[238, 579, 732, 789]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[115, 812, 875, 885]]<|/det|>
+<center>Response Figure 3. Cavity volume comparison between S4 and S5 states using KVFinder. a the cavity in structure A (cyan) representing the S4 state has a volume of 293.33 ų. b the cavity in structure B (blue) representing the S5 state shows a larger volume of 534.38 ų. Calculations were performed using probes with inner and outer radii of 5 Å and 10 Å, respectively. </center>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 103, 297, 119]]<|/det|>
+Some minor issues:  
+
+<|ref|>text<|/ref|><|det|>[[117, 121, 880, 214]]<|/det|>
+Some minor issues:1. Many side chains are unmodelled in the PDB, many of which have interpretable density (for example M177 and Y271 in receptor, to name just a few). Thank you for pointing this out. We have now modeled the side chains for M177 and Y271, and we have manually inspected all relevant residues to ensure the best fit. This includes adding residues such as V164, A165, among others.  
+
+<|ref|>text<|/ref|><|det|>[[117, 232, 880, 287]]<|/det|>
+2. The Figure Legend in Figure 1 is mislabelled (d and f) Thank you for your careful inspection. We have corrected the figure legend for Figure 1.  
+
+<|ref|>text<|/ref|><|det|>[[117, 306, 880, 360]]<|/det|>
+3. I don't know what the authors mean in Page 6. Line 183 "Closer inspection of the Tm6.... Using F19 ... revealed a clockwise rotation..." This needs some explanation as I can't see what the authors mean here.  
+
+<|ref|>text<|/ref|><|det|>[[117, 361, 880, 453]]<|/det|>
+We have revised the wording on page 8, lines 222- 226 for clarity. We used the \(^{19}\mathrm{F}\) - tag labeling site as a reference to track conformational changes and to illustrate our \(^{19}\mathrm{F}\) - qNMR data. The \(^{19}\mathrm{F}\) probe reveals a clockwise rotation as the receptor transitions towards activation, with the NMR signal for the S4 state appearing at a lower field compared to the S5 state, as shown in Fig. 4c.  
+
+<|ref|>text<|/ref|><|det|>[[117, 472, 880, 510]]<|/det|>
+4. I would suggest to show the nucleotide exchange data for with the addition of agonists and inverse agonists for the mutant (maybe just in the sup material)  
+
+<|ref|>text<|/ref|><|det|>[[118, 528, 879, 602]]<|/det|>
+Thank you for the suggestion. We have indeed performed nucleotide exchange experiments with the addition of both agonists and inverse agonists for the mutant. The data has been included in the revised manuscript and can be found in the supplementary material (Extended Data Fig. 3).  
+
+<|ref|>text<|/ref|><|det|>[[117, 620, 879, 694]]<|/det|>
+5. There are some more modelling errors. The unmodelled loop between 148 and 166 in the receptor will most likely have the N-terminus pointing in the wrong direction. As its currently modelled the only way for this loop to be completed is by forming a protein knot through ECL3.  
+
+<|ref|>text<|/ref|><|det|>[[118, 713, 879, 787]]<|/det|>
+Thank you for your detailed inspection. We were able to model V164, A165 into the density map, which provided a clear direction for the loop between residues 147- 164. Additionally, I corrected the fitting errors for F70 and C71. The updated model now clearly shows that the loop is not forming a knot through ECL3.  
+
+<|ref|>text<|/ref|><|det|>[[117, 805, 880, 899]]<|/det|>
+6. This reviewer has major issues with the use of DeepEMhancer for modelling. The map that the authors originally provided is a prime example of Al hallucination. The Al model started to turn the detergent micelle into protein looking density and should NOT be trusted. The map I calculated from the half maps and the authors traditionally sharpened map is more than suitable for PDB modelling. EM enhanced maps also
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 878, 120]]<|/det|>
+SHOULD NOT be the main map deposited in the PDB, as is taints the PDB and EMDB with data that is not suitable for interpretation.  
+
+<|ref|>text<|/ref|><|det|>[[118, 140, 880, 251]]<|/det|>
+We used DeepEMhancer map solely for illustration and initial modeling purposes. However, we have confirmed that all model refinements in Phenix were conducted using the raw map without the use of DeepEMhancer. For each EMDB and PDB entry, we have deposit both the sharpened map (sharpened by - 10Å in cisTEM) and the raw unsharpened map. In the revised manuscript, we have removed all references to the DeepEMhancer map.  
+
+<|ref|>text<|/ref|><|det|>[[118, 287, 690, 305]]<|/det|>
+Matthew Belousoff, Monash School of Pharmaceutical Sciences  
+
+<|ref|>text<|/ref|><|det|>[[118, 343, 456, 361]]<|/det|>
+Reviewer #3 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[118, 380, 880, 529]]<|/det|>
+This manuscript reports the structure and function of a ligand- free GPCR- Gasβy intermediate complex. The authors determined the cryo- EM structure of an intermediate A2AR- mini- Gasβy complex. They presented experimental and computational evidences that the intermediate complex (S4- Gasβy) initiates a rate- limited nucleotide exchange without progressing to the fully activated complex (S5- Gasβy), in which the α- helical domain (AHD) of the Gas is partially open engaged by a second nucleotide. Based on these, they proposed a mechanistic model for the rate- limited nucleotide exchange in the intermediate GPCR- Gasβy complex.  
+
+<|ref|>text<|/ref|><|det|>[[118, 548, 558, 564]]<|/det|>
+Thanks for the commendation on our manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[118, 585, 880, 657]]<|/det|>
+The experimental method presented in this paper is suitable and the MD simulations are well- designed. The approach can be applied to characterize transient intermediate states of other GPCRs. However, there are a few short comings in this study. I recommend that this work is publishable with revision as follows:  
+
+<|ref|>text<|/ref|><|det|>[[118, 659, 880, 750]]<|/det|>
+1. The authors determined the cryo-EM structure of an intermediate A2AR-mini- Gasβy complex. Has there any other intermediate state of other GPCRs reported in other papers? What's the similarity and difference between this intermediate structure and the others? It would be better to discuss this intermediate A2AR-mini-Gasβy structure in the context of the literature.  
+
+<|ref|>text<|/ref|><|det|>[[118, 770, 880, 861]]<|/det|>
+Thank you for the thoughtful suggestion. While no intermediate complex of A2AR has been reported so far, a partially active structure has been reported with the binding of LUF5833 and LUF5834<sup>4,5</sup>. In response to the reviewer's suggestion, we have added a discussion of other GPCR intermediate states in the literature to contextualize our findings, please refer to page 11, lines 318- 328.  
+
+<|ref|>text<|/ref|><|det|>[[116, 881, 878, 899]]<|/det|>
+Additionally, we compare our structure determined in this manuscript with all resolved
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 880, 159]]<|/det|>
+receptor- Gs complex for family A GPCR with the assistance of MD simulation (Fig.4c and 4d). Our results clearly indicate the uniqueness of our structure representing an intermediate position from the G protein perspective on both Ras- like domain (resolved structure) and AHD domain (simulated structure).  
+
+<|ref|>text<|/ref|><|det|>[[118, 195, 880, 251]]<|/det|>
+2. In the determined cryo-EM structure of the A2AR-mini-Gasβγ complex, the AHD of Gαs was missing. The mini-Gasβγ is different from the full-length Gs protein, will this influence the binding of Gs and A2AR?  
+
+<|ref|>text<|/ref|><|det|>[[118, 270, 880, 362]]<|/det|>
+We agree that using the mini- Gasβγ instead of the full- length Gs protein may influence binding. However, obtaining an intermediate GPCR- G protein complex with the full- length Gs protein remains challenging. We believe that protein- engineering of the G protein, similar to what we did for the receptor, would help stabilize an intermediate G protein conformation, and this is part of our future plans.  
+
+<|ref|>text<|/ref|><|det|>[[118, 380, 880, 473]]<|/det|>
+3. It is seen from Fig.1a that the most obvious difference between the intermediate complex (S4-Gasβγ) and other complex (S1-Gasβγ, S2-Gasβγ, S3-Gasβγ and S5-Gasβγ) is whether the AHD of Gs is partially open and whether there is GDP bound in Gs. It would be nice if the A2AR bound by the full-length Gs protein (with GDP bound in Gs) is characterized by cryo-EM.  
+
+<|ref|>text<|/ref|><|det|>[[118, 492, 880, 565]]<|/det|>
+Thank you for the suggestion. As mentioned above, we agree that characterizing the full- length Gs protein in complex with A2AR would provide more insight. This is part of our ongoing research, and we plan to explore this by engineering a stable full- length G protein in future studies.  
+
+<|ref|>text<|/ref|><|det|>[[118, 584, 880, 695]]<|/det|>
+4. The authors proposed two nucleotide-bound sites (Fig. 5, site 1 and site 2) in the transition process from the intermediate to the fully activated GPCR-Gasβγ complex. They performed additional GaMD simulations to examine the GDP release process in the cS4-Gasβγ system. GDP bound in site 2 of Gasβγ in the starting structure of GaMD simulations of GDP release. In the simulation of the GDP release process, can the GDP be observed to make a short stay at site 1?  
+
+<|ref|>text<|/ref|><|det|>[[118, 714, 880, 899]]<|/det|>
+Thank you for your detailed analysis. As shown in Fig. 5c, GDP rapidly transitions from the bound to the unbound state, with no intermediate state observed during this process. Similarly, as indicated in Extended Data Fig. 11e, no other low- energy states were detected. Additionally, the initial state of the G protein starts in a fully open conformation, and during GDP release, it remains largely open rather than adopting a partially open state. This could be the caveat of the MD here to capture the nucleotide binding to the site 2 (we redefine the original site 2 as the site 1 and the original site 1 as site 2) that was observed in our bench experiments. We will modify our model in the future or waiting for the structure of an intermediate A2AR- full length G protein resolved to revisit this phenomenon.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 121, 880, 195]]<|/det|>
+5. Free energy calculations indicated a "GDP Released" state in the S5-Gasβγ system and only a "Partially Released" state in the S4-Gasβγ system (Extended Data Fig. 8d and 8e). In this "Partially Released" state, is the GDP still binds in Gasβγ? Is there a possibility that it binds in the proposed "site 1"?  
+
+<|ref|>text<|/ref|><|det|>[[117, 214, 880, 344]]<|/det|>
+The representative structure of the 'Partially Released' state from the current research is shown below (Response Fig. 4). In this state, the phosphate group remains bound within the binding site, interacting with residues G52, S54, T55, and K293. Currently, we lack the data to confirm whether this corresponds to the proposed 'site 2.' Further studies will be conducted to investigate this hypothesis as mentioned above till we have the intermediate A2A-R-full length G protein complex and using it as a starting model for simulation.  
+
+<|ref|>image<|/ref|><|det|>[[323, 410, 663, 613]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[118, 620, 866, 671]]<|/det|>
+<center>Response Figure 4. Representative conformation of the 'Partially Released' state is depicted. Key interacting residues are shown in stick representation, with red dotted lines indicating their interactions. </center>  
+
+<|ref|>text<|/ref|><|det|>[[117, 722, 880, 777]]<|/det|>
+6. Page 8 Lines 246: "Extended Data Figs. 8a-c". It should be Extended Data Figs. 8d-e. Page 8 Lines 244-249: "Extended Data Figs. 8d-e". It should be Extended Data Figs. 8a-c.  
+
+<|ref|>text<|/ref|><|det|>[[117, 779, 880, 815]]<|/det|>
+Thank you for pointing out these discrepancies. We have corrected the references to the figures (Page 10, line 286 and line 289).  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 872, 234, 888]]<|/det|>
+## References:
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[113, 84, 883, 306]]<|/det|>
+receptor. Nat. Chem. Biol 18, 256- 263 (2022). https://doi.org/10.1038/s41589- 021- 00945- w2 Lin, X. et al. Structural basis of ligand recognition and self- activation of orphan GPR52. Nature 579, 152- 157 (2020). https://doi.org/10.1038/s41586- 020- 2019- 03Wang, X., Neale, C., Kim, S.- K., Goddard, W. A. & Ye, L. Intermediate- state- trapped mutants pinpoint G protein- coupled receptor conformational allostery. Nat. Commun. 14, 1325 (2023). https://doi.org/10.1038/s41467- 023- 36971- 64Claff, T. et al. Structural Insights into Partial Activation of the Prototypic G Protein- Coupled Adenosine A2A Receptor. ACS Pharmacol. Transl. Sci.. 7, 1415- 1425 (2024). https://doi.org/10.1021/acsptsci.4c00051Amelia, T. et al. Crystal Structure and Subsequent Ligand Design of a Nonriboside Partial Agonist Bound to the Adenosine A2A Receptor. J. Med. Chem. 64, 3827- 3842 (2021). https://doi.org/10.1021/acs.jmedchem.0c01856
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[114, 89, 880, 128]]<|/det|>
+We would like to thank all reviewers for their positive responses to our previous revision. In this final revision, we address all the remaining comments.  
+
+<|ref|>text<|/ref|><|det|>[[115, 169, 445, 188]]<|/det|>
+Reviewer #1 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[115, 210, 807, 250]]<|/det|>
+The authors have addressed all my concerns and have updated the manuscript accordingly.  
+
+<|ref|>text<|/ref|><|det|>[[116, 271, 330, 289]]<|/det|>
+Some minor corrections:  
+
+<|ref|>text<|/ref|><|det|>[[115, 310, 872, 370]]<|/det|>
+1. Extended Data Figure 2, page 33: The figure caption seems to be mislabeled. I think the red data points represent the line width of S1 resonance and the teal data point represent the line width of the S4 resonance.  
+
+<|ref|>text<|/ref|><|det|>[[115, 381, 742, 401]]<|/det|>
+Thanks for pointing out this and we have corrected all these accordingly.  
+
+<|ref|>text<|/ref|><|det|>[[115, 421, 620, 441]]<|/det|>
+2. Page 5 lines 134-135: Omit the word domain after AHD  
+
+<|ref|>text<|/ref|><|det|>[[115, 452, 857, 491]]<|/det|>
+We have omitted the word "domain" after AHD and corrected all these throughout the manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[115, 512, 847, 552]]<|/det|>
+3. Page 5 line 134-135: formed by Ras-like domain and AHD resulting from a partial opening of AHD.  
+
+<|ref|>text<|/ref|><|det|>[[115, 553, 631, 572]]<|/det|>
+We have changed the sentence as the reviewer suggested.  
+
+<|ref|>text<|/ref|><|det|>[[115, 613, 445, 632]]<|/det|>
+Reviewer #2 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[115, 653, 878, 732]]<|/det|>
+Firstly, I wish to commend the authors on their significant efforts to reanalyse the cryoEM data. I am of the opinion that this has vastly improved the authors manuscript. I do not fundamentally disagree with the overall experimental design and overall analysis but I have a comment that the authors may wish to consider.  
+
+<|ref|>text<|/ref|><|det|>[[115, 753, 880, 894]]<|/det|>
+Our research group has performed 3DVA on over 300 GPCR ternary complexes and we observe incredibly similar dynamic sampling of both G- protein engagement angles and G- protein conformations. However, does this mean their interpretation is incorrect? I do not think so, it is more likely that they have found an appropriate experimental system to more fully explore and interpret the dynamics and 3DVA data that are commonly observed across all active state GPCR structures. Put another way, essentially all reported structures of active state GPCRs are an average of the S4- S5 state, but the
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[115, 88, 866, 148]]<|/det|>
+authors in this manuscript have perhaps skewed the overall population towards the S4 state and provided a great hypothesis for interpretation of the observed dynamic landscape in active state GPCR structures.  
+
+<|ref|>text<|/ref|><|det|>[[115, 179, 841, 259]]<|/det|>
+We appreciate this comment. Indeed, it is likely that our approach of combining 19F NMR and point mutation to bias the conformation towards S4 enabled us to characterize this intermediate state more thoroughly. We revised the discussion to reflect this point.  
+
+<|ref|>text<|/ref|><|det|>[[115, 290, 861, 330]]<|/det|>
+I would recommend this paper for publication as it is my opinion that the authors have appropriately addressed the reviewers concerns.  
+
+<|ref|>text<|/ref|><|det|>[[115, 340, 864, 380]]<|/det|>
+We thank Dr. Belousoff for his support in publishing our manuscript and appreciate his comments, which helped us improve it.  
+
+<|ref|>text<|/ref|><|det|>[[115, 401, 680, 420]]<|/det|>
+Matthew Belousoff, Monash Institute of Pharmaceutical Sciences  
+
+<|ref|>text<|/ref|><|det|>[[115, 461, 445, 481]]<|/det|>
+Reviewer #3 (Remarks to the Author):  
+
+<|ref|>text<|/ref|><|det|>[[115, 502, 879, 581]]<|/det|>
+The authors have addressed all my previous concerns with detailed explanations and appropriate revisions. The revisions have improved the overall quality of the manuscript. I have no further issues with the current version of the manuscript, and I believe it can be published.  
+
+<|ref|>text<|/ref|><|det|>[[115, 593, 488, 611]]<|/det|>
+We thank the Reviewer for his/her support.
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__731ba13fea195e66a184a6160440c66e587c835c6bd2a8366ca608b373eec757/images_list.json

@@ -0,0 +1 @@
+[]

peer_reviews/supplementary_0_Transparent Peer Review file__731ba13fea195e66a184a6160440c66e587c835c6bd2a8366ca608b373eec757/supplementary_0_Transparent Peer Review file__731ba13fea195e66a184a6160440c66e587c835c6bd2a8366ca608b373eec757.mmd

@@ -0,0 +1,30 @@
+
+# nature portfolio  
+
+Peer Review File  
+
+# CryoPROS: Correcting Misalignment Caused by Preferred Orientation Using AI-generated Auxiliary Particles  
+
+Corresponding Author: Dr Chenglong Bao  
+
+This manuscript has been previously reviewed at another journal. This document only contains information relating to versions considered at Nature Communications.  
+
+Version 0:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author) This is resubmission of the revised manuscript that I previously reviewed. Authors have clearly addressed all of my comments, suggestions and concerns.  
+
+(Remarks on code availability) I have not read the code.  
+
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source. The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons
+
+<--- Page Split --->
+
+license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__731ba13fea195e66a184a6160440c66e587c835c6bd2a8366ca608b373eec757/supplementary_0_Transparent Peer Review file__731ba13fea195e66a184a6160440c66e587c835c6bd2a8366ca608b373eec757_det.mmd

@@ -0,0 +1,41 @@
+<|ref|>title<|/ref|><|det|>[[72, 50, 296, 80]]<|/det|>
+# nature portfolio  
+
+<|ref|>text<|/ref|><|det|>[[74, 96, 297, 119]]<|/det|>
+Peer Review File  
+
+<|ref|>title<|/ref|><|det|>[[73, 160, 875, 211]]<|/det|>
+# CryoPROS: Correcting Misalignment Caused by Preferred Orientation Using AI-generated Auxiliary Particles  
+
+<|ref|>text<|/ref|><|det|>[[73, 224, 433, 241]]<|/det|>
+Corresponding Author: Dr Chenglong Bao  
+
+<|ref|>text<|/ref|><|det|>[[70, 274, 875, 302]]<|/det|>
+This manuscript has been previously reviewed at another journal. This document only contains information relating to versions considered at Nature Communications.  
+
+<|ref|>text<|/ref|><|det|>[[73, 340, 144, 353]]<|/det|>
+Version 0:  
+
+<|ref|>text<|/ref|><|det|>[[73, 367, 220, 380]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[73, 392, 160, 405]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[73, 418, 857, 459]]<|/det|>
+(Remarks to the Author) This is resubmission of the revised manuscript that I previously reviewed. Authors have clearly addressed all of my comments, suggestions and concerns.  
+
+<|ref|>text<|/ref|><|det|>[[73, 471, 283, 497]]<|/det|>
+(Remarks on code availability) I have not read the code.  
+
+<|ref|>text<|/ref|><|det|>[[72, 850, 916, 904]]<|/det|>
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+<|ref|>text<|/ref|><|det|>[[72, 904, 894, 944]]<|/det|>
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source. The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[70, 46, 910, 88]]<|/det|>
+license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__7329c365184db0d565099581b1cfbb6f0161d689a0c9d63c319a0b43f6297141/images_list.json

@@ -0,0 +1,40 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_4.jpg",
+    "caption": "Reviewer's Figure R1 Correlation between no-scaling with the 8-component WSoR model for the CNS data of Figure 4.",
+    "footnote": [],
+    "bbox": [
+      [
+        272,
+        94,
+        710,
+        374
+      ]
+    ],
+    "page_idx": 7
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_0.jpg",
+    "caption": "Reviewer's Figure R2: Individual images of major FAs analysed in this study.",
+    "footnote": [],
+    "bbox": [],
+    "page_idx": 8
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_1.jpg",
+    "caption": "Reviewer's Figure R3: Ratio images with and without C14 FAs.",
+    "footnote": [],
+    "bbox": [
+      [
+        117,
+        81,
+        666,
+        300
+      ]
+    ],
+    "page_idx": 10
+  }
+]

peer_reviews/supplementary_0_Transparent Peer Review file__7329c365184db0d565099581b1cfbb6f0161d689a0c9d63c319a0b43f6297141/supplementary_0_Transparent Peer Review file__7329c365184db0d565099581b1cfbb6f0161d689a0c9d63c319a0b43f6297141.mmd

@@ -0,0 +1,268 @@
+
+# nature portfolio  
+
+Peer Review File  
+
+# Orbitrap noise structure and method for noise unbiased multivariate analysis  
+
+Corresponding Author: Professor Ian Gilmore  
+
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+Version 0:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author)  
+
+This publication presents a new scaling method for orbitrap data generated using a secondary ion mass spectrometer (SIMS). The untargeted nature of SIMS analysis provides a wealth of data that is not straightforward to interpret, especially on complex organic samples. The combination with a high resolution mass analyzer such as an orbitrap aids in identifying the mass signals but even increases the data load, further complicating manual data analysis in order to identify significant changes in the data. Multivariate analysis has become an invaluable tool to do just that but due to the way the MS analysis is performed, data scaling is necessary to gain information on higher mass, biologically relevant molecules. Several scaling methods are available but it is not at all obvious which one is most suitable for any given dataset, especially novel ones such as generated by an OrbiSIMS. Therefore, studies like this are very important to facilitate biological discoveries hidden in the dataset.  
+
+Major comments:  
+
+This publication is heavy on the introduction and establishing the mathematical foundation for their newly developed scaling method. At the beginning in the abstract the authors mention life sciences which lets one assume that this will be the focus of the publication and/or the intended application of the algorithm. At the end they do come around to show "life- science"- results but while it seems to work excellently to simulate inorganic data, they fail to demonstrating major benefits of, or improvements on existing scaling methods for the biological samples shown. So, while this work is by no means uninteresting, the results are currently lackluster. My recommendation is to focus solely on the theoretical background and the inorganic data in this publication, while further exploring biological applications and publish those at a later date, provided that significant improvements are found. Alternatively, provide an outlook for further developments to improve biological data analysis.  
+
+Minor comments:  
+
+Page 16, line 429: Please list the papers, also I assume the most cited publications are older publications, maybe more recent studies contain more information on data scaling. Page 20, line 508: Rephrase, do you mean: all scaling and log transformation are effective... Line 513: "Pareto scaling.... may be less robust for complex, biological OrbiSIMS imaging datasets." Is this assumption supported by your experimental? Figure 5j: The figure is too small to be useful, add labels for the highest loadings or add a bigger version to the supplementary information, the loadings in the SI are too small as well. Line 518 ff: The refences to the Figure 5 b- h are mixed up in the entire paragraph. Page 20 line 540: "ether- linked Pes" change to PEs  
+
+(Remarks on code availability)  
+
+Reviewer #2  
+
+(Remarks to the Author)  
+
+Keenan et al. propose a theoretical model of the noise generated by the Orbitrap mass analyzers in the low intensity region.
+
+<--- Page Split --->
+
+For modeling, they use data generated by SIMS which provides a consistent input ion current with known noise properties. Moreover, they propose a new method (WSoR) for data rescaling that capitalizes on the knowledge of the noise model.  
+
+The manuscript strengths are in the theoretical work to mathematically model the noise. The manuscript weaknesses are in the validation and in demonstrating the practical impact of the proposed theory.  
+
+1. A key concern is about the validation of the proposed WSoR method. The authors do it in three ways: through data simulation (Figure 4), through discovery of the Cs+ contamination in first N principal components, and through application to an imaging dataset. However, the simulation is still a theoretical example and the practical use and impact for discovering the Cs+ contamination is not clear. Re application to an imaging dataset, Figure 5i shows no improvement for WSoR compared to "no scaling" in terms of discovering spatial patterns. For each spatial pattern (what the authors call "CNS pattern") shown in PCA scores for WSoR-transformed data, there is a similar spatial pattern visible in the PCA scores in "no scaling" results.  
+
+2. It's not quite clear what's the basis of the molecular interpretations in Figure 5 because of two factors. First, there are no details on how the molecular identification was done. The authors should briefly comment on the methodology used for the reported molecular assignments. If they used the accurate m/z matching, they should share the details on the database (if any), the m/z tolerance used, and the m/z delta between the theoretical m/z and the observed one. It may additionally be worth mentioning that annotations of plasmalogen lipids and nucleotides are commonly the result of fragmentation reactions rather than intact detection. Moreover, the reported PE-Cer(34:2) is isobaric to PA(34:1) which can be a fragment of larger lipids. Second, the claimed correspondence between the mass spectrometry signals and cell types is not supported with any evidence.  
+
+3. A brief inspection of the raw data behind Figure 5 raises some questions about the interpretation (lines 501-553), especially the fatty acid analysis (subfigure 5b). Looking at individual images and intensity ratios, the statement about polyunsaturated vs saturated fatty acids does not appear to hold up. The pattern shown in 5b is caused mainly by the fact that palmitic acid (16:0) has a very different distribution from the other peaks. The inclusion of C14 fatty acids is somewhat redundant as only 14:0 and 14:1 are present, and a full order of magnitude below the corresponding C16 and C18 signals. Similarly, 16:2 does not contribute meaningfully to the sum, meaning that the image is indistinguishable from 18:2 / (18:0+16:0). In this set, I would argue that 18:2 and 18:0 are strongly correlated, which is the opposite observation from what the authors discuss! This also is readily apparent from a quick manual inspection, and does not really provide a good example of multivariate analysis used for hypothesis generation.  
+
+4. Within the imaging dataset analysis (lines 501-553), figure panels are referred to with the wrong figure number, e.g. references to 5b as 5d on line 519; 5e as 5g on line 530; 5g as 5h on line 539. The data in 5h (ratio of C18 to C16 fatty acids) is not referred to anywhere.  
+
+5. The testis analysis in supplementary note 9 is not mentioned in the main text, but is listed as the main individual contribution of two of the co-authors. The authors may want to consider including a short statement about that analysis and the value it adds to the manuscript.  
+
+(Remarks on code availability) There are no installation instructions provided and no README file  
+
+Reviewer #3  
+
+(Remarks to the Author) I co-reviewed this manuscript with one of the reviewers who provided the listed reports. This is part of the Nature Communications initiative to facilitate training in peer review and to provide appropriate recognition for Early Career Researchers who co-review manuscripts.  
+
+(Remarks on code availability)  
+
+Version 1:  
+
+Reviewer comments:  
+
+Reviewer #1  
+
+(Remarks to the Author)  
+
+I appreciate the extensive responses to my comments and the additional experiments. My concerns have been addressed and the significance of the model for biological data-sets is way clearer now. In my opinion, no additional experiments are needed.  
+
+I only have a few minor comments:  
+
+Figure 5a WSoR PC6 seems to highlight a feature captured by no other scaling method. Is this anything of interest? Line 572: why not test on the full data-set? Is it too computationally demanding? The 2 other bio data-sets use 100- 500 peaks.
+
+<--- Page Split --->
+
+Line 573ff: What do mean with group C/D, this is not explained anywhere in the text or SI. Is there a group A/B? This section needs to be slightly reworked.  
+
+(Remarks on code availability)  
+
+Reviewer #3  
+
+(Remarks to the Author) I co- reviewed this manuscript with one of the reviewers who provided the listed reports. This is part of the Nature Communications initiative to facilitate training in peer review and to provide appropriate recognition for Early Career Researchers who co- review manuscripts.  
+
+(Remarks on code availability)  
+
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source.  
+
+The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.  
+
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+
+## Orbitrap noise structure and method for noise-unbiased multivariate analysis  
+
+We are very grateful for the reviewers' comments and for the opportunity to submit a major revision. We are sorry for the delay but it took time to do a substantial revision including an additional biological data set using desorption electrospray ionisation (DESI) as the ion source.  
+
+We have taken into account the reviewer's request to focus more on biological data to demonstrate the effects of the different scaling methods. We have now removed Figure 4 using silver data and focus on three biological data sets (1) drosophila CNS, (2) mouse testis - significantly expanding our previous analysis and adding a new main figure and (3) DESI image of rat liver detailed in (A. Dannhorn et al, Anal. Chem. 2020, 92, 16, 11080- 11088) including a new main figure. The three data sets cover examples of high intensities, low intensities and high to low intensities. We have developed a much clearer way to show how no- scaling and the commonly used scaling methods variance, Pareto, root- mean scaling (RMS) upweight and down- weight ions, which introduces bias into principal component analysis (PCA). The objective of PCA is to parsimoniously capture the chemical components with separation from noise components. We clearly show that our Weighted Sum of Ricians (WSoR) based on our theoretical framework always provides the most efficient solution, whereas the other scaling methods perform variably being sometimes good or poor depending on the statistical detail of the data.  
+
+Furthermore, the addition of the DESI data set demonstrates the generality of our WSoR scaling method to a completely different ionisation source and is also data from an independent laboratory. From our understanding of the literature and discussions with the community there is, at present, no principled way to select a scaling method and so it is a matter of luck if they work well or not. With the growing use of computational methods on ever larger data sets, we therefore think that our WSoR method is important for improving the reliability and effectiveness of Orbitrap data analysis.  
+
+Recently, we have been conducting an international interlaboratory study to look at noise in OrbiSIMS instruments and determined their WSoR functions. We received data for 8 instruments including a newer Orbitrap Exploris design and all exhibited excellent agreement with our WSoR model. This will be reported separately later but we are happy to share a plot in confidence, if requested.  
+
+In the following we provide a point- by- point response to the reviewers.
+
+<--- Page Split --->
+
+## Reviewer #1  
+
+This publication presents a new scaling method for Orbitrap data generated using a secondary ion mass spectrometer (SIMS). The untargeted nature of SIMS analysis provides a wealth of data that is not straightforward to interpret, especially on complex organic samples. The combination with a high resolution mass analyzer such as an orbitrap aids in identifying the mass signals but even increases the data load, further complicating manual data analysis in order to identify significant changes in the data. Multivariate analysis has become an invaluable tool to do just that but due to the way the MS analysis is performed, data scaling is necessary to gain information on higher mass, biologically relevant molecules. Several scaling methods are available but it is not at all obvious which one is most suitable for any given dataset, especially novel ones such as generated by an OrbiSIMS. Therefore, studies like this are very important to facilitate biological discoveries hidden in the dataset.  
+
+We are very grateful for the reviewer's supportive comments on the value of our work, especially selecting the most suitable scaling method for biological data. We agree that it is not obvious which scaling method to use and we show that our WSoR method works best in all cases.  
+
+## Major comments:  
+
+This publication is heavy on the introduction and establishing the mathematical foundation for their newly developed scaling method. At the beginning in the abstract the authors mention life sciences which lets one assume that this will be the focus of the publication and/or the intended application of the algorithm. At the end they do come around to show "life- science"- results but while it seems to work excellently to simulate inorganic data, they fail to demonstrating major benefits of, or improvements on existing scaling methods for the biological samples shown. So, while this work is by no means uninteresting, the results are currently lackluster. My recommendation is to focus solely on the theoretical background and the inorganic data in this publication, while further exploring biological applications and publish those at a later date, provided that significant improvements are found. Alternatively, provide an outlook for further developments to improve biological data analysis.  
+
+The statistics of the Orbitrap measurement process are complex and it has been a long and difficult task to work it all out. So, we are therefore very grateful for the reviewer's appreciation of the mathematical foundations that we have developed.  
+
+This understanding allowed us to develop the Weighted Sum of Rician's (WSoR) scaling method for noise unbiased multivariate analysis. On reflection, we agree with the reviewer that we did not clearly present and compare the performance with other scaling methods. We have now extended the analysis of the mouse testis data as a main figure. We have also included an additional biological imaging data set of rat liver using desorption electrospray ionisation (DESI) Orbitrap MS (A. Dannhorn et al, Anal. Chem. 2020, 92, 16, 11080- 11088). To make the effects of noise bias clearer for each biological example, we show how the variance for no- scaling, variance, Pareto, RMS and Probabilistic Factor Analysis (PFA) used for the covariance matrix for scaling compares to the fundamental WSoR model. In every case, the WSoR model strongly agrees with the machine learned variance using probabilistic factor analysis (PFA) validating our method using 3 biological data sets.
+
+<--- Page Split --->
+
+These plots highlight how the different models introduce bias by down- weighting or upweighting peaks. The effects of bias on the multivariate scores images are now clearer to understand. For example, in the drosophila CNS example, there are 8 chemical components that are significant with respect to the noise level. Variance scaling performs poorly down- weighting intense ions so that chemical components are found in high principal components amongst noise. In the testis data example, the intensities are lower and Pareto scaling now performs poorly, down- weighting most ions and features found in component 4 for WSoR scaling are not found until component 34. The DESI data set demonstrates the generality of our method to a completely different ionisation source as well as independently generated data. In this example, eigenvalue analysis finds 7 distinct components. Variance, Pareto and root- mean scaling all introduce a bias that leads to chemical components being relegated to high principal components defeating the purpose of multivariate analysis.  
+
+In each example, we find that the WSoR provides noise unbiased scaling allowing the chemical information to be captured efficiently in the leading Principal Components and separated from the noise. In contrast, no- scaling, variance scaling, Pareto scaling and RMS are sometimes good or poor on a case- by- case basis depending on the distribution of ion intensities. This leads to inefficient capture of the chemical components with some mixed between noise components.  
+
+We hope that our revisions more clearly show the biases introduced by the commonly used scaling method adopted by the community with a rationale and how the efficiency of multivariate analysis is affected by them.  
+
+## Minor comments:  
+
+Page 16, line 429: Please list the papers, also I assume the most cited publications are older publications, maybe more recent studies contain more information on data scaling. Thank you, we now provide this in Supplementary Table 2. Page 20, line 508: Rephrase, do you mean: all scaling and log transformation are effective... Line 513: "Pareto scaling... ...may be less robust for complex, biological OrbiSIMS imaging datasets." Is this assumption supported by your experimental? Log- transformation is always very poor. We have substantially revised the text and figure and hope that this is now clearer. Figure 5j: The figure is too small to be useful, add labels for the highest loadings or add a bigger version to the supplementary information, the loadings in the SI are too small as well. Thank you, we have revised the figure. Line 518 ff: The refences to the Figure 5 b- h are mixed up in the entire paragraph. We have substantially revised the text and figure to correct this. Page 20 line 540: "ether- linked Pes" change to PEs As part of the major revisions this text has been deleted.  
+
+## Reviewer #2  
+
+Keenan et al. propose a theoretical model of the noise generated by the Orbitrap mass analyzers in the low intensity region. For modeling, they use data generated by SIMS which provides a consistent
+
+<--- Page Split --->
+
+input ion current with known noise properties. Moreover, they propose a new method (WSoR) for data rescaling that capitalizes on the knowledge of the noise model.  
+
+The manuscript strengths are in the theoretical work to mathematically model the noise. The manuscript weaknesses are in the validation and in demonstrating the practical impact of the proposed theory.  
+
+We are very grateful to Reviewers 2 and 3 for their positive comments on the mathematical model and we hope that the additional work that we have done including more analysis on the testis data and analysis of an additional DESI Orbitrap data demonstrates more clearly the practical impact of our method. We provide further details on these additions in the response to Reviewer 1.  
+
+A key concern is about the validation of the proposed WSoR method. The authors do it in three ways: through data simulation (Figure 4), through discovery of the Cs+ contamination in first N principal components, and through application to an imaging dataset. However, the simulation is still a theoretical example and the practical use and impact for discovering the Cs+ contamination is not clear.  
+
+To address the reviewers' concern we have removed the simulations and have clarified the use of the silver data as a simple data set acting as a control. We have now developed a method to clearly show the bias that is introduced for principal component analysis with different scaling methods and if no scaling is used and validate this using three biological data sets. For further details, please see our response to Reviewer 1  
+
+Re application to an imaging dataset, Figure 5i shows no improvement for WSoR compared to "no scaling" in terms of discovering spatial patterns. For each spatial pattern (what the authors call "CNS pattern") shown in PCA scores for WSoR- transformed data, there is a similar spatial pattern visible in the PCA scores in "no scaling" results.  
+
+The no- scaling data set is poor at capturing the information in the leading principal components and this is demonstrated by noise beginning to dominate the scores images beyond the \(7^{\text{th}}\) PC. The WSoR method is optimally parsimonious containing the biological information with the first 8 PCs. To demonstrate the difference we show in Reviewer's Figure R1 the correlation between no- scaling and the 8 component WSoR model, analogous to new Fig 4b for variance scaling. This shows that with unscaled PCA over 50 components would be required to describe the data with the same fidelity as the 8 component WSoR model. This is because chemical information is relegated to higher PCs with no- scaling.
+
+<--- Page Split --->
+![](images/Figure_4.jpg)
+
+<center>Reviewer's Figure R1 Correlation between no-scaling with the 8-component WSoR model for the CNS data of Figure 4. </center>  
+
+2. It's not quite clear what's the basis of the molecular interpretations in Figure 5 because of two factors. First, there are no details on how the molecular identification was done. The authors should briefly comment on the methodology used for the reported molecular assignments. If they used the accurate m/z matching, they should share the details on the database (if any), the m/z tolerance used, and the m/z delta between the theoretical m/z and the observed one.  
+
+We thank the reviewer for pointing out this omission in our methods section. We have added a sentence describing that molecular assignments were performed by accurate matching using the CEU mass tool (Gil- de- la- Fuente et al., 2019), which compares data against the HMDB, LipidMaps, Metlin, Kegg and in- house CEU mass libraries. The m/z tolerance was set at 2 ppm during ID searching however, as shown in the table below, all the putative IDs we have made have a mass accuracy \(< 1\) ppm. We have added this information into the supplementary data.  
+
+<table><tr><td>m/z</td><td>Formula (experimental)</td><td>Putative I.D.</td><td>Adduct</td><td>Mass deviation (ppm)</td></tr><tr><td>134.047</td><td>C5H4N5-</td><td>Adenine</td><td>[M-H]-</td><td>0.2</td></tr><tr><td>150.042</td><td>C5H4N5O-</td><td>Guanine</td><td>[M-H]-</td><td>-0.2</td></tr><tr><td>111.020</td><td>C4H3N2O2-</td><td>Uracil</td><td>[M-H]-</td><td>-0.1</td></tr><tr><td>255.233</td><td>C16H31O2-</td><td>C16:0</td><td>[M-H]-</td><td>-0.3</td></tr><tr><td>253.217</td><td>C16H31O2-</td><td>C16:1</td><td>[M-H]-</td><td>-0.3</td></tr><tr><td>277.217</td><td>C18H29O2-</td><td>C18:3</td><td>[M-H]-</td><td>-0.7</td></tr><tr><td>279.233</td><td>C18H31O2-</td><td>C18:2</td><td>[M-H]-</td><td>-0.7</td></tr><tr><td>281.248</td><td>C18H33O2-</td><td>C18:1</td><td>[M-H]-</td><td>-0.6</td></tr><tr><td>283.264</td><td>C18H35O2-</td><td>C18:0</td><td>[M-H]-</td><td>-0.6</td></tr><tr><td>657.498</td><td>C36H70N2O6P-</td><td>PE-Cer(34:2)</td><td>[M-H]-</td><td>0.4</td></tr><tr><td>702.545</td><td>C39H77NO7P-</td><td>PE(O-34:1)</td><td>[M-H]-</td><td>0.4</td></tr></table>
+
+<--- Page Split --->
+
+It may additionally be worth mentioning that annotations of plasmalogen lipids and nucleotides are commonly the result of fragmentation reactions rather than intact detection. Moreover, the reported PE- Cer(34:2) is isobaric to PA(34:1) which can be a fragment of larger lipids.  
+
+We appreciate that this interpretation may not have been so clear in our original text, so we have clarified that the nucleobases we have putatively identified as (adenine, guanine and uracil) are detected as a result of fragmentation of DNA/RNA by adding an additional sentence describing this to line 524 on page 23.  
+
+We have putatively identified the plasmalogen lipids as [M- H]- ions, as in our experience, lipids ionized with the OrbiSIMS argon GCIB form primarily intact molecular ions, with additional fragments that originate from loss of the entire headgroup or entire fatty acid chains (as opposed to the high degree of fragmentation induced by the bismuth LMIG) (Passarelli et al, 2017; Newell et al, 2020). Therefore, these fragments do not resemble other large lipid classes.  
+
+PE- Cer(34:2) and PA(34:1) are not isobaric as [M- H]- ions, which is how the majority of intact lipids ionize with the OrbiSIMS argon GCIB. PA(34:1) has a very different \(m / z\) value \((C_{37}H_{71}O_{8}P, m / z\) 673.481) compared to PE- Cer(34:2) \((C_{36}H_{71}N_{2}O_{6}P, m / z\) 657.498). The reviewer may be referring to PA(O- 34:2) which is the second search result on LIPID MAPS after PE- Cer(34:2). However, the mass deviation for PE- Cer(34:2) is 0.416 ppm whereas PA(O- 34:2) is 17.491 ppm. The instrument calibration result on the day of image acquisition (with silver) at \(m / z\) 538.52534 was 0.105 ppm and at \(m / z\) 754.33519 was 0.114 ppm. Therefore, we are confident that this ion is not PA(O- 34:2) due to the high mass deviation for this assignment.  
+
+Second, the claimed correspondence between the mass spectrometry signals and cell types is not supported with any evidence.  
+
+The claims about mass spectrometry signals and cell types are stated as hypotheses based on existing literature, and the location of the cell types referenced is well established in the literature which has been cited accordingly. The Drosophila larval CNS has a very distinct and reproducible pattern of cell- type distribution (we have the updated the anatomical schematic in Figure 5a and shown below), which makes it an ideal test- application for this method.  
+
+![](images/Figure_unknown_0.jpg)
+
+
+<--- Page Split --->
+
+3. A brief inspection of the raw data behind Figure 5 raises some questions about the interpretation (lines 501-553), especially the fatty acid analysis (subfigure 5b). Looking at individual images and intensity ratios, the statement about polyunsaturated vs saturated fatty acids does not appear to hold up. The pattern shown in 5b is caused mainly by the fact that palmitic acid (16:0) has a very different distribution from the other peaks. The inclusion of C14 fatty acids is somewhat redundant as only 14:0 and 14:1 are present, and a full order of magnitude below the corresponding C16 and C18 signals. Similarly, 16:2 does not contribute meaningfully to the sum, meaning that the image is indistinguishable from 18:2 / (18:0+16:0). In this set, I would argue that 18:2 and 18:0 are strongly correlated, which is the opposite observation from what the authors discuss! This also is readily apparent from a quick manual inspection, and does not really provide a good example of multivariate analysis used for hypothesis generation.  
+
+Thank you for these important points. In our images, all the major FAs we have examined (regardless of whether they are polyunsaturated or saturated) gave a higher signal in the neuropil compared to the cortex (Reviewer's Figure R2). This also applies to 16:0 although as the reviewer suggests the signal does appear somewhat higher in the cortex than with other FAs. In line with these results, the ratios of polyunsaturated (C18:2, C16:2) to saturated (C18:0, C16:0) FAs are also higher in the neuropil than the cortex.
+
+<--- Page Split --->
+![](images/Figure_unknown_1.jpg)
+
+<center>Reviewer's Figure R2: Individual images of major FAs analysed in this study. </center>  
+
+The top three chain- lengths detected in the Drosophila CNS are from highest to lowest signal: C18, C16 and C14 fatty acids. C14 FAs are still considerably above background level thus they are included. Nevertheless, in line with Reviewer's request, we have repeated the analysis of PUFA:SFA ratios without including C14 FAs (Reviewer's Figure R3). This analysis shows a broadly similar results to the original PUFA:SFA ratios that included C14 FAs.
+
+<--- Page Split --->
+![PLACEHOLDER_11_0]
+
+<center>Reviewer's Figure R3: Ratio images with and without C14 FAs. </center>  
+
+We agree entirely with the Reviewer that the ratios shown are readily apparent from manual inspection of individual images. This is a good thing because it provides a proof- of- principle sanity check on our automated analysis. Moving forward, manual inspection for many large datasets would be very time- consuming and challenging to identify patterns. The algorithm provided in this paper offers an unbiased and automated method for detecting patterns from mass spectrometry imaging, with significant advantages for analysing large image datasets.  
+
+4. Within the imaging dataset analysis (lines 501-553), figure panels are referred to with the wrong figure number, e.g. references to 5b as 5d on line 519; 5e as 5g on line 530; 5g as 5h on line 539. The data in 5h (ratio of C18 to C16 fatty acids) is not referred to anywhere.  
+
+Thank you, we have completely revised the text and figures and this is now corrected.  
+
+5. The testis analysis in supplementary note 9 is not mentioned in the main text, but is listed as the main individual contribution of two of the co-authors. The authors may want to consider including a short statement about that analysis and the value it adds to the manuscript.  
+
+We thank the reviewer for this suggestion. We have done additional analysis on the testis data and added a new main Figure. This data set is an interesting one as the signal intensities are low providing a good complement to the CNS and DESI data sets. In this case, no scaling and Pareto scaling perform poorly since low intensity ions are down- weighted. With WSoR scaling, two localisations relating to \(\mathrm{C_4H_8N_3O_2}\) are discovered in PC4, which are not discovered until PC39 with no scaling and PC34 with Pareto scaling. Further details on the variable performance of scaling methods are given in the response to Reviewer 1.  
+
+Reviewer #2 (Remarks on code availability):  
+
+There are no installation instructions provided and no README file Details have now been included in the program header.
+
+<--- Page Split --->
+
+## Reviewer #3  
+
+I co- reviewed this manuscript with one of the reviewers who provided the listed reports. This is part of the Nature Communications initiative to facilitate training in peer review and to provide appropriate recognition for Early Career Researchers who co- review manuscripts.  
+
+We thank the reviewer for their comments which have helped us improve the clarity of the manuscript.
+
+<--- Page Split --->
+
+## Orbitrap noise structure and method for noise unbiased multivariate analysis  
+
+We are very grateful for the reviewers' comments. In the following we provide a point- by- point response to the reviewers.  
+
+## Reviewer #1  
+
+I appreciate the extensive responses to my comments and the additional experiments. My concerns have been addressed and the significance of the model for biological data- sets is way clearer now. In my opinion, no additional experiments are needed.  
+
+We are glad the revised version addressed all of the reviewer's concerns.  
+
+I only have a few minor comments:  
+
+Figure 5a WSoR PC6 seems to highlight a feature captured by no other scaling method. Is this anything of interest?  
+
+We thank the reviewer for pointing this out. We have looked into the details and WSoR PC6 is describing an artefact of the Orbitrap, which is important information for studies using isotopic labelling strategies. It results from the Orbitrap software censoring data below a threshold. Isotopologues, or other strongly correlated peaks, should all scale together linearly however those with very low intensity, for example the \(^{13}\mathrm{C}_4\mathrm{C}_9\mathrm{H}_{79}\mathrm{O}_{12}\mathrm{S}\) isotopologue, fall below the threshold in some pixels and are censored, giving rise to a non- linear effect. We have added a sentence to the main text to describe this. This artefact as well as Orbitrap linearity will be discussed in detail in a future paper.  
+
+Line 572: why not test on the full data- set? Is it too computationally demanding? The 2 other bio data- sets use 100- 500 peaks.  
+
+The DESI data set is very large and complex. We selected a sub- set of peaks with distinct spatial distributions to that provide a more visual example of the differences between no- scaling and the different scaling methods.
+
+<--- Page Split --->
+
+Line 573ff: What do mean with group C/D, this is not explained anywhere in the text or SI. Is there a group A/B? This section needs to be slightly reworked.  
+
+We thank the reviewer for pointing out this source of confusion. In fact, there are only groups C and D, where C stands for "clustered" and D for "distributed". To avoid this confusion we have now renamed those groups A and B in the text and Figure.  
+
+## Reviewer #2 (Re-addressing of comment from previous round of revisions)  
+
+3. A brief inspection of the raw data behind Figure 5 raises some questions about the interpretation (lines 501-553), especially the fatty acid analysis (subfigure 5b). Looking at individual images and intensity ratios, the statement about polyunsaturated vs saturated fatty acids does not appear to hold up. The pattern shown in 5b is caused mainly by the fact that palmitic acid (16:0) has a very different distribution from the other peaks. The inclusion of C14 fatty acids is somewhat redundant as only 14:0 and 14:1 are present, and a full order of magnitude below the corresponding C16 and C18 signals. Similarly, 16:2 does not contribute meaningfully to the sum, meaning that the image is indistinguishable from 18:2 / (18:0+16:0). In this set, I would argue that 18:2 and 18:0 are strongly correlated, which is the opposite observation from what the authors discuss! This also is readily apparent from a quick manual inspection, and does not really provide a good example of multivariate analysis used for hypothesis generation.  
+
+We agree with the reviewer that a ratio of the sums is not appropriate owing to the low intensities for some of the ions and our previous general statement is not supported. The reviewer's comment stimulated us to investigate the spatial distribution of unsaturated and saturated fatty acids in more detail. PCA (with WSoR scaling) of a subset of data containing ten C16 and C18 fatty acid peaks (Reviewer's Figure R1) shows that fatty acids are predominantly in the neuropil (PC 1) but that the distribution is nuanced (PC 2). From PC 2, we see there is no correlation of C16:0 and C18:0 or C18:1 and C16:1. However, we find that cortex/neuropil variation does reflect differences in saturation with the strongest anticorrelation between C16:0 and C18:1. Comparison of the PC2 loading with WSoR PC7 in the paper shows they are nearly the same. In other words, the subtle variation in saturation is captured by WSoR in a PC that is not found without scaling, highlighting the effect of bias if appropriate scaling is not used.  
+
+In the revision we have removed the general statement about the spatial distribution of saturated and unsaturated fatty acids and have replaced old Fig 5f with the specific example for C16 and C18 (new Fig 4f,g). We have also replaced the previous text with the following (lines 520- 524) "WSoR PC 7 (Figure 4a,d) reveals that C16:0 and C18:1 are anticorrelated with different spatial distributions. We find that the signal intensity ratio of saturated to unsaturated fatty acids C16:1 to C16:0 (Figure 4f) is enhanced in the neuropil whereas the ratio C18:1 to C18:0 (Figure 4g) is stronger at the periphery of the cortex along the blood- brain barrier." We hope that this is now clearer and thank the reviewer for pointing out the error.
+
+<--- Page Split --->
+![PLACEHOLDER_15_0]
+  
+
+Reviewer's Figure R1: Principal component analysis (with WSoR scaling) of a subset of the CNS dataset containing only 10 fatty acid peaks.
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__7329c365184db0d565099581b1cfbb6f0161d689a0c9d63c319a0b43f6297141/supplementary_0_Transparent Peer Review file__7329c365184db0d565099581b1cfbb6f0161d689a0c9d63c319a0b43f6297141_det.mmd

@@ -0,0 +1,366 @@
+<|ref|>title<|/ref|><|det|>[[72, 50, 295, 78]]<|/det|>
+# nature portfolio  
+
+<|ref|>text<|/ref|><|det|>[[75, 95, 297, 118]]<|/det|>
+Peer Review File  
+
+<|ref|>title<|/ref|><|det|>[[73, 161, 844, 209]]<|/det|>
+# Orbitrap noise structure and method for noise unbiased multivariate analysis  
+
+<|ref|>text<|/ref|><|det|>[[73, 224, 465, 240]]<|/det|>
+Corresponding Author: Professor Ian Gilmore  
+
+<|ref|>text<|/ref|><|det|>[[70, 274, 864, 289]]<|/det|>
+This file contains all reviewer reports in order by version, followed by all author rebuttals in order by version.  
+
+<|ref|>text<|/ref|><|det|>[[73, 326, 144, 339]]<|/det|>
+Version 0:  
+
+<|ref|>text<|/ref|><|det|>[[73, 353, 219, 366]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[73, 379, 160, 393]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[73, 405, 238, 418]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 419, 916, 536]]<|/det|>
+This publication presents a new scaling method for orbitrap data generated using a secondary ion mass spectrometer (SIMS). The untargeted nature of SIMS analysis provides a wealth of data that is not straightforward to interpret, especially on complex organic samples. The combination with a high resolution mass analyzer such as an orbitrap aids in identifying the mass signals but even increases the data load, further complicating manual data analysis in order to identify significant changes in the data. Multivariate analysis has become an invaluable tool to do just that but due to the way the MS analysis is performed, data scaling is necessary to gain information on higher mass, biologically relevant molecules. Several scaling methods are available but it is not at all obvious which one is most suitable for any given dataset, especially novel ones such as generated by an OrbiSIMS. Therefore, studies like this are very important to facilitate biological discoveries hidden in the dataset.  
+
+<|ref|>text<|/ref|><|det|>[[73, 550, 192, 562]]<|/det|>
+Major comments:  
+
+<|ref|>text<|/ref|><|det|>[[72, 562, 923, 679]]<|/det|>
+This publication is heavy on the introduction and establishing the mathematical foundation for their newly developed scaling method. At the beginning in the abstract the authors mention life sciences which lets one assume that this will be the focus of the publication and/or the intended application of the algorithm. At the end they do come around to show "life- science"- results but while it seems to work excellently to simulate inorganic data, they fail to demonstrating major benefits of, or improvements on existing scaling methods for the biological samples shown. So, while this work is by no means uninteresting, the results are currently lackluster. My recommendation is to focus solely on the theoretical background and the inorganic data in this publication, while further exploring biological applications and publish those at a later date, provided that significant improvements are found. Alternatively, provide an outlook for further developments to improve biological data analysis.  
+
+<|ref|>text<|/ref|><|det|>[[73, 692, 192, 704]]<|/det|>
+Minor comments:  
+
+<|ref|>text<|/ref|><|det|>[[70, 704, 900, 822]]<|/det|>
+Page 16, line 429: Please list the papers, also I assume the most cited publications are older publications, maybe more recent studies contain more information on data scaling. Page 20, line 508: Rephrase, do you mean: all scaling and log transformation are effective... Line 513: "Pareto scaling.... may be less robust for complex, biological OrbiSIMS imaging datasets." Is this assumption supported by your experimental? Figure 5j: The figure is too small to be useful, add labels for the highest loadings or add a bigger version to the supplementary information, the loadings in the SI are too small as well. Line 518 ff: The refences to the Figure 5 b- h are mixed up in the entire paragraph. Page 20 line 540: "ether- linked Pes" change to PEs  
+
+<|ref|>text<|/ref|><|det|>[[73, 847, 282, 861]]<|/det|>
+(Remarks on code availability)  
+
+<|ref|>text<|/ref|><|det|>[[73, 886, 161, 899]]<|/det|>
+Reviewer #2  
+
+<|ref|>text<|/ref|><|det|>[[73, 912, 238, 925]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[70, 925, 914, 939]]<|/det|>
+Keenan et al. propose a theoretical model of the noise generated by the Orbitrap mass analyzers in the low intensity region.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[72, 46, 910, 75]]<|/det|>
+For modeling, they use data generated by SIMS which provides a consistent input ion current with known noise properties. Moreover, they propose a new method (WSoR) for data rescaling that capitalizes on the knowledge of the noise model.  
+
+<|ref|>text<|/ref|><|det|>[[72, 85, 911, 114]]<|/det|>
+The manuscript strengths are in the theoretical work to mathematically model the noise. The manuscript weaknesses are in the validation and in demonstrating the practical impact of the proposed theory.  
+
+<|ref|>text<|/ref|><|det|>[[72, 125, 917, 218]]<|/det|>
+1. A key concern is about the validation of the proposed WSoR method. The authors do it in three ways: through data simulation (Figure 4), through discovery of the Cs+ contamination in first N principal components, and through application to an imaging dataset. However, the simulation is still a theoretical example and the practical use and impact for discovering the Cs+ contamination is not clear. Re application to an imaging dataset, Figure 5i shows no improvement for WSoR compared to "no scaling" in terms of discovering spatial patterns. For each spatial pattern (what the authors call "CNS pattern") shown in PCA scores for WSoR-transformed data, there is a similar spatial pattern visible in the PCA scores in "no scaling" results.  
+
+<|ref|>text<|/ref|><|det|>[[72, 228, 921, 334]]<|/det|>
+2. It's not quite clear what's the basis of the molecular interpretations in Figure 5 because of two factors. First, there are no details on how the molecular identification was done. The authors should briefly comment on the methodology used for the reported molecular assignments. If they used the accurate m/z matching, they should share the details on the database (if any), the m/z tolerance used, and the m/z delta between the theoretical m/z and the observed one. It may additionally be worth mentioning that annotations of plasmalogen lipids and nucleotides are commonly the result of fragmentation reactions rather than intact detection. Moreover, the reported PE-Cer(34:2) is isobaric to PA(34:1) which can be a fragment of larger lipids. Second, the claimed correspondence between the mass spectrometry signals and cell types is not supported with any evidence.  
+
+<|ref|>text<|/ref|><|det|>[[72, 345, 918, 464]]<|/det|>
+3. A brief inspection of the raw data behind Figure 5 raises some questions about the interpretation (lines 501-553), especially the fatty acid analysis (subfigure 5b). Looking at individual images and intensity ratios, the statement about polyunsaturated vs saturated fatty acids does not appear to hold up. The pattern shown in 5b is caused mainly by the fact that palmitic acid (16:0) has a very different distribution from the other peaks. The inclusion of C14 fatty acids is somewhat redundant as only 14:0 and 14:1 are present, and a full order of magnitude below the corresponding C16 and C18 signals. Similarly, 16:2 does not contribute meaningfully to the sum, meaning that the image is indistinguishable from 18:2 / (18:0+16:0). In this set, I would argue that 18:2 and 18:0 are strongly correlated, which is the opposite observation from what the authors discuss! This also is readily apparent from a quick manual inspection, and does not really provide a good example of multivariate analysis used for hypothesis generation.  
+
+<|ref|>text<|/ref|><|det|>[[72, 475, 923, 517]]<|/det|>
+4. Within the imaging dataset analysis (lines 501-553), figure panels are referred to with the wrong figure number, e.g. references to 5b as 5d on line 519; 5e as 5g on line 530; 5g as 5h on line 539. The data in 5h (ratio of C18 to C16 fatty acids) is not referred to anywhere.  
+
+<|ref|>text<|/ref|><|det|>[[72, 528, 911, 570]]<|/det|>
+5. The testis analysis in supplementary note 9 is not mentioned in the main text, but is listed as the main individual contribution of two of the co-authors. The authors may want to consider including a short statement about that analysis and the value it adds to the manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[72, 581, 535, 608]]<|/det|>
+(Remarks on code availability) There are no installation instructions provided and no README file  
+
+<|ref|>text<|/ref|><|det|>[[72, 620, 161, 633]]<|/det|>
+Reviewer #3  
+
+<|ref|>text<|/ref|><|det|>[[72, 646, 850, 700]]<|/det|>
+(Remarks to the Author) I co-reviewed this manuscript with one of the reviewers who provided the listed reports. This is part of the Nature Communications initiative to facilitate training in peer review and to provide appropriate recognition for Early Career Researchers who co-review manuscripts.  
+
+<|ref|>text<|/ref|><|det|>[[72, 710, 283, 724]]<|/det|>
+(Remarks on code availability)  
+
+<|ref|>text<|/ref|><|det|>[[72, 749, 144, 762]]<|/det|>
+Version 1:  
+
+<|ref|>text<|/ref|><|det|>[[72, 775, 219, 789]]<|/det|>
+Reviewer comments:  
+
+<|ref|>text<|/ref|><|det|>[[72, 802, 160, 815]]<|/det|>
+Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[72, 828, 238, 841]]<|/det|>
+(Remarks to the Author)  
+
+<|ref|>text<|/ref|><|det|>[[72, 841, 905, 880]]<|/det|>
+I appreciate the extensive responses to my comments and the additional experiments. My concerns have been addressed and the significance of the model for biological data-sets is way clearer now. In my opinion, no additional experiments are needed.  
+
+<|ref|>text<|/ref|><|det|>[[72, 893, 310, 906]]<|/det|>
+I only have a few minor comments:  
+
+<|ref|>text<|/ref|><|det|>[[72, 906, 884, 947]]<|/det|>
+Figure 5a WSoR PC6 seems to highlight a feature captured by no other scaling method. Is this anything of interest? Line 572: why not test on the full data-set? Is it too computationally demanding? The 2 other bio data-sets use 100- 500 peaks.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[70, 46, 920, 75]]<|/det|>
+Line 573ff: What do mean with group C/D, this is not explained anywhere in the text or SI. Is there a group A/B? This section needs to be slightly reworked.  
+
+<|ref|>text<|/ref|><|det|>[[73, 99, 283, 113]]<|/det|>
+(Remarks on code availability)  
+
+<|ref|>text<|/ref|><|det|>[[73, 138, 162, 152]]<|/det|>
+Reviewer #3  
+
+<|ref|>text<|/ref|><|det|>[[73, 164, 864, 217]]<|/det|>
+(Remarks to the Author) I co- reviewed this manuscript with one of the reviewers who provided the listed reports. This is part of the Nature Communications initiative to facilitate training in peer review and to provide appropriate recognition for Early Career Researchers who co- review manuscripts.  
+
+<|ref|>text<|/ref|><|det|>[[73, 230, 283, 244]]<|/det|>
+(Remarks on code availability)  
+
+<|ref|>text<|/ref|><|det|>[[72, 608, 916, 660]]<|/det|>
+Open Access This Peer Review File is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made.  
+
+<|ref|>text<|/ref|><|det|>[[72, 660, 796, 675]]<|/det|>
+In cases where reviewers are anonymous, credit should be given to 'Anonymous Referee' and the source.  
+
+<|ref|>text<|/ref|><|det|>[[72, 674, 910, 727]]<|/det|>
+The images or other third party material in this Peer Review File are included in the article's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.  
+
+<|ref|>text<|/ref|><|det|>[[72, 727, 618, 741]]<|/det|>
+To view a copy of this license, visit https://creativecommons.org/licenses/by/4.0/
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[118, 260, 655, 293]]<|/det|>
+## Orbitrap noise structure and method for noise-unbiased multivariate analysis  
+
+<|ref|>text<|/ref|><|det|>[[118, 306, 828, 361]]<|/det|>
+We are very grateful for the reviewers' comments and for the opportunity to submit a major revision. We are sorry for the delay but it took time to do a substantial revision including an additional biological data set using desorption electrospray ionisation (DESI) as the ion source.  
+
+<|ref|>text<|/ref|><|det|>[[116, 379, 877, 617]]<|/det|>
+We have taken into account the reviewer's request to focus more on biological data to demonstrate the effects of the different scaling methods. We have now removed Figure 4 using silver data and focus on three biological data sets (1) drosophila CNS, (2) mouse testis - significantly expanding our previous analysis and adding a new main figure and (3) DESI image of rat liver detailed in (A. Dannhorn et al, Anal. Chem. 2020, 92, 16, 11080- 11088) including a new main figure. The three data sets cover examples of high intensities, low intensities and high to low intensities. We have developed a much clearer way to show how no- scaling and the commonly used scaling methods variance, Pareto, root- mean scaling (RMS) upweight and down- weight ions, which introduces bias into principal component analysis (PCA). The objective of PCA is to parsimoniously capture the chemical components with separation from noise components. We clearly show that our Weighted Sum of Ricians (WSoR) based on our theoretical framework always provides the most efficient solution, whereas the other scaling methods perform variably being sometimes good or poor depending on the statistical detail of the data.  
+
+<|ref|>text<|/ref|><|det|>[[118, 636, 880, 745]]<|/det|>
+Furthermore, the addition of the DESI data set demonstrates the generality of our WSoR scaling method to a completely different ionisation source and is also data from an independent laboratory. From our understanding of the literature and discussions with the community there is, at present, no principled way to select a scaling method and so it is a matter of luck if they work well or not. With the growing use of computational methods on ever larger data sets, we therefore think that our WSoR method is important for improving the reliability and effectiveness of Orbitrap data analysis.  
+
+<|ref|>text<|/ref|><|det|>[[118, 764, 850, 855]]<|/det|>
+Recently, we have been conducting an international interlaboratory study to look at noise in OrbiSIMS instruments and determined their WSoR functions. We received data for 8 instruments including a newer Orbitrap Exploris design and all exhibited excellent agreement with our WSoR model. This will be reported separately later but we are happy to share a plot in confidence, if requested.  
+
+<|ref|>text<|/ref|><|det|>[[118, 875, 655, 891]]<|/det|>
+In the following we provide a point- by- point response to the reviewers.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[118, 85, 215, 100]]<|/det|>
+## Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[117, 120, 872, 321]]<|/det|>
+This publication presents a new scaling method for Orbitrap data generated using a secondary ion mass spectrometer (SIMS). The untargeted nature of SIMS analysis provides a wealth of data that is not straightforward to interpret, especially on complex organic samples. The combination with a high resolution mass analyzer such as an orbitrap aids in identifying the mass signals but even increases the data load, further complicating manual data analysis in order to identify significant changes in the data. Multivariate analysis has become an invaluable tool to do just that but due to the way the MS analysis is performed, data scaling is necessary to gain information on higher mass, biologically relevant molecules. Several scaling methods are available but it is not at all obvious which one is most suitable for any given dataset, especially novel ones such as generated by an OrbiSIMS. Therefore, studies like this are very important to facilitate biological discoveries hidden in the dataset.  
+
+<|ref|>text<|/ref|><|det|>[[118, 341, 869, 394]]<|/det|>
+We are very grateful for the reviewer's supportive comments on the value of our work, especially selecting the most suitable scaling method for biological data. We agree that it is not obvious which scaling method to use and we show that our WSoR method works best in all cases.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 415, 254, 430]]<|/det|>
+## Major comments:  
+
+<|ref|>text<|/ref|><|det|>[[117, 432, 870, 632]]<|/det|>
+This publication is heavy on the introduction and establishing the mathematical foundation for their newly developed scaling method. At the beginning in the abstract the authors mention life sciences which lets one assume that this will be the focus of the publication and/or the intended application of the algorithm. At the end they do come around to show "life- science"- results but while it seems to work excellently to simulate inorganic data, they fail to demonstrating major benefits of, or improvements on existing scaling methods for the biological samples shown. So, while this work is by no means uninteresting, the results are currently lackluster. My recommendation is to focus solely on the theoretical background and the inorganic data in this publication, while further exploring biological applications and publish those at a later date, provided that significant improvements are found. Alternatively, provide an outlook for further developments to improve biological data analysis.  
+
+<|ref|>text<|/ref|><|det|>[[118, 652, 867, 705]]<|/det|>
+The statistics of the Orbitrap measurement process are complex and it has been a long and difficult task to work it all out. So, we are therefore very grateful for the reviewer's appreciation of the mathematical foundations that we have developed.  
+
+<|ref|>text<|/ref|><|det|>[[117, 725, 876, 907]]<|/det|>
+This understanding allowed us to develop the Weighted Sum of Rician's (WSoR) scaling method for noise unbiased multivariate analysis. On reflection, we agree with the reviewer that we did not clearly present and compare the performance with other scaling methods. We have now extended the analysis of the mouse testis data as a main figure. We have also included an additional biological imaging data set of rat liver using desorption electrospray ionisation (DESI) Orbitrap MS (A. Dannhorn et al, Anal. Chem. 2020, 92, 16, 11080- 11088). To make the effects of noise bias clearer for each biological example, we show how the variance for no- scaling, variance, Pareto, RMS and Probabilistic Factor Analysis (PFA) used for the covariance matrix for scaling compares to the fundamental WSoR model. In every case, the WSoR model strongly agrees with the machine learned variance using probabilistic factor analysis (PFA) validating our method using 3 biological data sets.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 84, 875, 303]]<|/det|>
+These plots highlight how the different models introduce bias by down- weighting or upweighting peaks. The effects of bias on the multivariate scores images are now clearer to understand. For example, in the drosophila CNS example, there are 8 chemical components that are significant with respect to the noise level. Variance scaling performs poorly down- weighting intense ions so that chemical components are found in high principal components amongst noise. In the testis data example, the intensities are lower and Pareto scaling now performs poorly, down- weighting most ions and features found in component 4 for WSoR scaling are not found until component 34. The DESI data set demonstrates the generality of our method to a completely different ionisation source as well as independently generated data. In this example, eigenvalue analysis finds 7 distinct components. Variance, Pareto and root- mean scaling all introduce a bias that leads to chemical components being relegated to high principal components defeating the purpose of multivariate analysis.  
+
+<|ref|>text<|/ref|><|det|>[[118, 323, 870, 412]]<|/det|>
+In each example, we find that the WSoR provides noise unbiased scaling allowing the chemical information to be captured efficiently in the leading Principal Components and separated from the noise. In contrast, no- scaling, variance scaling, Pareto scaling and RMS are sometimes good or poor on a case- by- case basis depending on the distribution of ion intensities. This leads to inefficient capture of the chemical components with some mixed between noise components.  
+
+<|ref|>text<|/ref|><|det|>[[118, 432, 879, 485]]<|/det|>
+We hope that our revisions more clearly show the biases introduced by the commonly used scaling method adopted by the community with a rationale and how the efficiency of multivariate analysis is affected by them.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 506, 255, 521]]<|/det|>
+## Minor comments:  
+
+<|ref|>text<|/ref|><|det|>[[115, 524, 870, 800]]<|/det|>
+Page 16, line 429: Please list the papers, also I assume the most cited publications are older publications, maybe more recent studies contain more information on data scaling. Thank you, we now provide this in Supplementary Table 2. Page 20, line 508: Rephrase, do you mean: all scaling and log transformation are effective... Line 513: "Pareto scaling... ...may be less robust for complex, biological OrbiSIMS imaging datasets." Is this assumption supported by your experimental? Log- transformation is always very poor. We have substantially revised the text and figure and hope that this is now clearer. Figure 5j: The figure is too small to be useful, add labels for the highest loadings or add a bigger version to the supplementary information, the loadings in the SI are too small as well. Thank you, we have revised the figure. Line 518 ff: The refences to the Figure 5 b- h are mixed up in the entire paragraph. We have substantially revised the text and figure to correct this. Page 20 line 540: "ether- linked Pes" change to PEs As part of the major revisions this text has been deleted.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 819, 216, 833]]<|/det|>
+## Reviewer #2  
+
+<|ref|>text<|/ref|><|det|>[[118, 855, 875, 889]]<|/det|>
+Keenan et al. propose a theoretical model of the noise generated by the Orbitrap mass analyzers in the low intensity region. For modeling, they use data generated by SIMS which provides a consistent
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 860, 119]]<|/det|>
+input ion current with known noise properties. Moreover, they propose a new method (WSoR) for data rescaling that capitalizes on the knowledge of the noise model.  
+
+<|ref|>text<|/ref|><|det|>[[118, 139, 824, 191]]<|/det|>
+The manuscript strengths are in the theoretical work to mathematically model the noise. The manuscript weaknesses are in the validation and in demonstrating the practical impact of the proposed theory.  
+
+<|ref|>text<|/ref|><|det|>[[118, 212, 865, 284]]<|/det|>
+We are very grateful to Reviewers 2 and 3 for their positive comments on the mathematical model and we hope that the additional work that we have done including more analysis on the testis data and analysis of an additional DESI Orbitrap data demonstrates more clearly the practical impact of our method. We provide further details on these additions in the response to Reviewer 1.  
+
+<|ref|>text<|/ref|><|det|>[[118, 304, 880, 393]]<|/det|>
+A key concern is about the validation of the proposed WSoR method. The authors do it in three ways: through data simulation (Figure 4), through discovery of the Cs+ contamination in first N principal components, and through application to an imaging dataset. However, the simulation is still a theoretical example and the practical use and impact for discovering the Cs+ contamination is not clear.  
+
+<|ref|>text<|/ref|><|det|>[[118, 413, 878, 503]]<|/det|>
+To address the reviewers' concern we have removed the simulations and have clarified the use of the silver data as a simple data set acting as a control. We have now developed a method to clearly show the bias that is introduced for principal component analysis with different scaling methods and if no scaling is used and validate this using three biological data sets. For further details, please see our response to Reviewer 1  
+
+<|ref|>text<|/ref|><|det|>[[118, 523, 878, 595]]<|/det|>
+Re application to an imaging dataset, Figure 5i shows no improvement for WSoR compared to "no scaling" in terms of discovering spatial patterns. For each spatial pattern (what the authors call "CNS pattern") shown in PCA scores for WSoR- transformed data, there is a similar spatial pattern visible in the PCA scores in "no scaling" results.  
+
+<|ref|>text<|/ref|><|det|>[[117, 615, 880, 761]]<|/det|>
+The no- scaling data set is poor at capturing the information in the leading principal components and this is demonstrated by noise beginning to dominate the scores images beyond the \(7^{\text{th}}\) PC. The WSoR method is optimally parsimonious containing the biological information with the first 8 PCs. To demonstrate the difference we show in Reviewer's Figure R1 the correlation between no- scaling and the 8 component WSoR model, analogous to new Fig 4b for variance scaling. This shows that with unscaled PCA over 50 components would be required to describe the data with the same fidelity as the 8 component WSoR model. This is because chemical information is relegated to higher PCs with no- scaling.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[272, 94, 710, 374]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[117, 386, 860, 420]]<|/det|>
+<center>Reviewer's Figure R1 Correlation between no-scaling with the 8-component WSoR model for the CNS data of Figure 4. </center>  
+
+<|ref|>text<|/ref|><|det|>[[117, 440, 874, 530]]<|/det|>
+2. It's not quite clear what's the basis of the molecular interpretations in Figure 5 because of two factors. First, there are no details on how the molecular identification was done. The authors should briefly comment on the methodology used for the reported molecular assignments. If they used the accurate m/z matching, they should share the details on the database (if any), the m/z tolerance used, and the m/z delta between the theoretical m/z and the observed one.  
+
+<|ref|>text<|/ref|><|det|>[[117, 550, 856, 658]]<|/det|>
+We thank the reviewer for pointing out this omission in our methods section. We have added a sentence describing that molecular assignments were performed by accurate matching using the CEU mass tool (Gil- de- la- Fuente et al., 2019), which compares data against the HMDB, LipidMaps, Metlin, Kegg and in- house CEU mass libraries. The m/z tolerance was set at 2 ppm during ID searching however, as shown in the table below, all the putative IDs we have made have a mass accuracy \(< 1\) ppm. We have added this information into the supplementary data.  
+
+<|ref|>table<|/ref|><|det|>[[120, 679, 657, 866]]<|/det|>
+
+<table><tr><td>m/z</td><td>Formula (experimental)</td><td>Putative I.D.</td><td>Adduct</td><td>Mass deviation (ppm)</td></tr><tr><td>134.047</td><td>C5H4N5-</td><td>Adenine</td><td>[M-H]-</td><td>0.2</td></tr><tr><td>150.042</td><td>C5H4N5O-</td><td>Guanine</td><td>[M-H]-</td><td>-0.2</td></tr><tr><td>111.020</td><td>C4H3N2O2-</td><td>Uracil</td><td>[M-H]-</td><td>-0.1</td></tr><tr><td>255.233</td><td>C16H31O2-</td><td>C16:0</td><td>[M-H]-</td><td>-0.3</td></tr><tr><td>253.217</td><td>C16H31O2-</td><td>C16:1</td><td>[M-H]-</td><td>-0.3</td></tr><tr><td>277.217</td><td>C18H29O2-</td><td>C18:3</td><td>[M-H]-</td><td>-0.7</td></tr><tr><td>279.233</td><td>C18H31O2-</td><td>C18:2</td><td>[M-H]-</td><td>-0.7</td></tr><tr><td>281.248</td><td>C18H33O2-</td><td>C18:1</td><td>[M-H]-</td><td>-0.6</td></tr><tr><td>283.264</td><td>C18H35O2-</td><td>C18:0</td><td>[M-H]-</td><td>-0.6</td></tr><tr><td>657.498</td><td>C36H70N2O6P-</td><td>PE-Cer(34:2)</td><td>[M-H]-</td><td>0.4</td></tr><tr><td>702.545</td><td>C39H77NO7P-</td><td>PE(O-34:1)</td><td>[M-H]-</td><td>0.4</td></tr></table>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 84, 868, 138]]<|/det|>
+It may additionally be worth mentioning that annotations of plasmalogen lipids and nucleotides are commonly the result of fragmentation reactions rather than intact detection. Moreover, the reported PE- Cer(34:2) is isobaric to PA(34:1) which can be a fragment of larger lipids.  
+
+<|ref|>text<|/ref|><|det|>[[118, 158, 866, 230]]<|/det|>
+We appreciate that this interpretation may not have been so clear in our original text, so we have clarified that the nucleobases we have putatively identified as (adenine, guanine and uracil) are detected as a result of fragmentation of DNA/RNA by adding an additional sentence describing this to line 524 on page 23.  
+
+<|ref|>text<|/ref|><|det|>[[118, 250, 875, 340]]<|/det|>
+We have putatively identified the plasmalogen lipids as [M- H]- ions, as in our experience, lipids ionized with the OrbiSIMS argon GCIB form primarily intact molecular ions, with additional fragments that originate from loss of the entire headgroup or entire fatty acid chains (as opposed to the high degree of fragmentation induced by the bismuth LMIG) (Passarelli et al, 2017; Newell et al, 2020). Therefore, these fragments do not resemble other large lipid classes.  
+
+<|ref|>text<|/ref|><|det|>[[117, 359, 872, 504]]<|/det|>
+PE- Cer(34:2) and PA(34:1) are not isobaric as [M- H]- ions, which is how the majority of intact lipids ionize with the OrbiSIMS argon GCIB. PA(34:1) has a very different \(m / z\) value \((C_{37}H_{71}O_{8}P, m / z\) 673.481) compared to PE- Cer(34:2) \((C_{36}H_{71}N_{2}O_{6}P, m / z\) 657.498). The reviewer may be referring to PA(O- 34:2) which is the second search result on LIPID MAPS after PE- Cer(34:2). However, the mass deviation for PE- Cer(34:2) is 0.416 ppm whereas PA(O- 34:2) is 17.491 ppm. The instrument calibration result on the day of image acquisition (with silver) at \(m / z\) 538.52534 was 0.105 ppm and at \(m / z\) 754.33519 was 0.114 ppm. Therefore, we are confident that this ion is not PA(O- 34:2) due to the high mass deviation for this assignment.  
+
+<|ref|>text<|/ref|><|det|>[[118, 516, 857, 551]]<|/det|>
+Second, the claimed correspondence between the mass spectrometry signals and cell types is not supported with any evidence.  
+
+<|ref|>text<|/ref|><|det|>[[117, 580, 860, 670]]<|/det|>
+The claims about mass spectrometry signals and cell types are stated as hypotheses based on existing literature, and the location of the cell types referenced is well established in the literature which has been cited accordingly. The Drosophila larval CNS has a very distinct and reproducible pattern of cell- type distribution (we have the updated the anatomical schematic in Figure 5a and shown below), which makes it an ideal test- application for this method.  
+
+<|ref|>image<|/ref|><|det|>[[120, 680, 386, 905]]<|/det|>
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[117, 101, 872, 303]]<|/det|>
+3. A brief inspection of the raw data behind Figure 5 raises some questions about the interpretation (lines 501-553), especially the fatty acid analysis (subfigure 5b). Looking at individual images and intensity ratios, the statement about polyunsaturated vs saturated fatty acids does not appear to hold up. The pattern shown in 5b is caused mainly by the fact that palmitic acid (16:0) has a very different distribution from the other peaks. The inclusion of C14 fatty acids is somewhat redundant as only 14:0 and 14:1 are present, and a full order of magnitude below the corresponding C16 and C18 signals. Similarly, 16:2 does not contribute meaningfully to the sum, meaning that the image is indistinguishable from 18:2 / (18:0+16:0). In this set, I would argue that 18:2 and 18:0 are strongly correlated, which is the opposite observation from what the authors discuss! This also is readily apparent from a quick manual inspection, and does not really provide a good example of multivariate analysis used for hypothesis generation.  
+
+<|ref|>text<|/ref|><|det|>[[117, 333, 880, 442]]<|/det|>
+Thank you for these important points. In our images, all the major FAs we have examined (regardless of whether they are polyunsaturated or saturated) gave a higher signal in the neuropil compared to the cortex (Reviewer's Figure R2). This also applies to 16:0 although as the reviewer suggests the signal does appear somewhat higher in the cortex than with other FAs. In line with these results, the ratios of polyunsaturated (C18:2, C16:2) to saturated (C18:0, C16:0) FAs are also higher in the neuropil than the cortex.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[115, 78, 843, 620]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[118, 623, 688, 639]]<|/det|>
+<center>Reviewer's Figure R2: Individual images of major FAs analysed in this study. </center>  
+
+<|ref|>text<|/ref|><|det|>[[118, 650, 870, 740]]<|/det|>
+The top three chain- lengths detected in the Drosophila CNS are from highest to lowest signal: C18, C16 and C14 fatty acids. C14 FAs are still considerably above background level thus they are included. Nevertheless, in line with Reviewer's request, we have repeated the analysis of PUFA:SFA ratios without including C14 FAs (Reviewer's Figure R3). This analysis shows a broadly similar results to the original PUFA:SFA ratios that included C14 FAs.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[117, 81, 666, 300]]<|/det|>
+<|ref|>image_caption<|/ref|><|det|>[[118, 303, 586, 318]]<|/det|>
+<center>Reviewer's Figure R3: Ratio images with and without C14 FAs. </center>  
+
+<|ref|>text<|/ref|><|det|>[[117, 348, 878, 456]]<|/det|>
+We agree entirely with the Reviewer that the ratios shown are readily apparent from manual inspection of individual images. This is a good thing because it provides a proof- of- principle sanity check on our automated analysis. Moving forward, manual inspection for many large datasets would be very time- consuming and challenging to identify patterns. The algorithm provided in this paper offers an unbiased and automated method for detecting patterns from mass spectrometry imaging, with significant advantages for analysing large image datasets.  
+
+<|ref|>text<|/ref|><|det|>[[118, 485, 876, 539]]<|/det|>
+4. Within the imaging dataset analysis (lines 501-553), figure panels are referred to with the wrong figure number, e.g. references to 5b as 5d on line 519; 5e as 5g on line 530; 5g as 5h on line 539. The data in 5h (ratio of C18 to C16 fatty acids) is not referred to anywhere.  
+
+<|ref|>text<|/ref|><|det|>[[118, 559, 763, 575]]<|/det|>
+Thank you, we have completely revised the text and figures and this is now corrected.  
+
+<|ref|>text<|/ref|><|det|>[[118, 595, 873, 649]]<|/det|>
+5. The testis analysis in supplementary note 9 is not mentioned in the main text, but is listed as the main individual contribution of two of the co-authors. The authors may want to consider including a short statement about that analysis and the value it adds to the manuscript.  
+
+<|ref|>text<|/ref|><|det|>[[117, 668, 876, 796]]<|/det|>
+We thank the reviewer for this suggestion. We have done additional analysis on the testis data and added a new main Figure. This data set is an interesting one as the signal intensities are low providing a good complement to the CNS and DESI data sets. In this case, no scaling and Pareto scaling perform poorly since low intensity ions are down- weighted. With WSoR scaling, two localisations relating to \(\mathrm{C_4H_8N_3O_2}\) are discovered in PC4, which are not discovered until PC39 with no scaling and PC34 with Pareto scaling. Further details on the variable performance of scaling methods are given in the response to Reviewer 1.  
+
+<|ref|>text<|/ref|><|det|>[[119, 816, 447, 832]]<|/det|>
+Reviewer #2 (Remarks on code availability):  
+
+<|ref|>text<|/ref|><|det|>[[118, 852, 622, 887]]<|/det|>
+There are no installation instructions provided and no README file Details have now been included in the program header.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[118, 103, 216, 118]]<|/det|>
+## Reviewer #3  
+
+<|ref|>text<|/ref|><|det|>[[118, 139, 872, 191]]<|/det|>
+I co- reviewed this manuscript with one of the reviewers who provided the listed reports. This is part of the Nature Communications initiative to facilitate training in peer review and to provide appropriate recognition for Early Career Researchers who co- review manuscripts.  
+
+<|ref|>text<|/ref|><|det|>[[118, 212, 805, 246]]<|/det|>
+We thank the reviewer for their comments which have helped us improve the clarity of the manuscript.
+
+<--- Page Split --->
+<|ref|>sub_title<|/ref|><|det|>[[118, 261, 715, 277]]<|/det|>
+## Orbitrap noise structure and method for noise unbiased multivariate analysis  
+
+<|ref|>text<|/ref|><|det|>[[118, 291, 838, 325]]<|/det|>
+We are very grateful for the reviewers' comments. In the following we provide a point- by- point response to the reviewers.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 346, 215, 361]]<|/det|>
+## Reviewer #1  
+
+<|ref|>text<|/ref|><|det|>[[118, 383, 876, 435]]<|/det|>
+I appreciate the extensive responses to my comments and the additional experiments. My concerns have been addressed and the significance of the model for biological data- sets is way clearer now. In my opinion, no additional experiments are needed.  
+
+<|ref|>text<|/ref|><|det|>[[118, 455, 667, 471]]<|/det|>
+We are glad the revised version addressed all of the reviewer's concerns.  
+
+<|ref|>text<|/ref|><|det|>[[118, 493, 384, 508]]<|/det|>
+I only have a few minor comments:  
+
+<|ref|>text<|/ref|><|det|>[[118, 529, 822, 564]]<|/det|>
+Figure 5a WSoR PC6 seems to highlight a feature captured by no other scaling method. Is this anything of interest?  
+
+<|ref|>text<|/ref|><|det|>[[117, 584, 860, 729]]<|/det|>
+We thank the reviewer for pointing this out. We have looked into the details and WSoR PC6 is describing an artefact of the Orbitrap, which is important information for studies using isotopic labelling strategies. It results from the Orbitrap software censoring data below a threshold. Isotopologues, or other strongly correlated peaks, should all scale together linearly however those with very low intensity, for example the \(^{13}\mathrm{C}_4\mathrm{C}_9\mathrm{H}_{79}\mathrm{O}_{12}\mathrm{S}\) isotopologue, fall below the threshold in some pixels and are censored, giving rise to a non- linear effect. We have added a sentence to the main text to describe this. This artefact as well as Orbitrap linearity will be discussed in detail in a future paper.  
+
+<|ref|>text<|/ref|><|det|>[[118, 749, 848, 783]]<|/det|>
+Line 572: why not test on the full data- set? Is it too computationally demanding? The 2 other bio data- sets use 100- 500 peaks.  
+
+<|ref|>text<|/ref|><|det|>[[118, 804, 855, 856]]<|/det|>
+The DESI data set is very large and complex. We selected a sub- set of peaks with distinct spatial distributions to that provide a more visual example of the differences between no- scaling and the different scaling methods.
+
+<--- Page Split --->
+<|ref|>text<|/ref|><|det|>[[118, 102, 870, 138]]<|/det|>
+Line 573ff: What do mean with group C/D, this is not explained anywhere in the text or SI. Is there a group A/B? This section needs to be slightly reworked.  
+
+<|ref|>text<|/ref|><|det|>[[118, 157, 871, 211]]<|/det|>
+We thank the reviewer for pointing out this source of confusion. In fact, there are only groups C and D, where C stands for "clustered" and D for "distributed". To avoid this confusion we have now renamed those groups A and B in the text and Figure.  
+
+<|ref|>sub_title<|/ref|><|det|>[[118, 230, 691, 248]]<|/det|>
+## Reviewer #2 (Re-addressing of comment from previous round of revisions)  
+
+<|ref|>text<|/ref|><|det|>[[117, 266, 871, 468]]<|/det|>
+3. A brief inspection of the raw data behind Figure 5 raises some questions about the interpretation (lines 501-553), especially the fatty acid analysis (subfigure 5b). Looking at individual images and intensity ratios, the statement about polyunsaturated vs saturated fatty acids does not appear to hold up. The pattern shown in 5b is caused mainly by the fact that palmitic acid (16:0) has a very different distribution from the other peaks. The inclusion of C14 fatty acids is somewhat redundant as only 14:0 and 14:1 are present, and a full order of magnitude below the corresponding C16 and C18 signals. Similarly, 16:2 does not contribute meaningfully to the sum, meaning that the image is indistinguishable from 18:2 / (18:0+16:0). In this set, I would argue that 18:2 and 18:0 are strongly correlated, which is the opposite observation from what the authors discuss! This also is readily apparent from a quick manual inspection, and does not really provide a good example of multivariate analysis used for hypothesis generation.  
+
+<|ref|>text<|/ref|><|det|>[[117, 486, 876, 687]]<|/det|>
+We agree with the reviewer that a ratio of the sums is not appropriate owing to the low intensities for some of the ions and our previous general statement is not supported. The reviewer's comment stimulated us to investigate the spatial distribution of unsaturated and saturated fatty acids in more detail. PCA (with WSoR scaling) of a subset of data containing ten C16 and C18 fatty acid peaks (Reviewer's Figure R1) shows that fatty acids are predominantly in the neuropil (PC 1) but that the distribution is nuanced (PC 2). From PC 2, we see there is no correlation of C16:0 and C18:0 or C18:1 and C16:1. However, we find that cortex/neuropil variation does reflect differences in saturation with the strongest anticorrelation between C16:0 and C18:1. Comparison of the PC2 loading with WSoR PC7 in the paper shows they are nearly the same. In other words, the subtle variation in saturation is captured by WSoR in a PC that is not found without scaling, highlighting the effect of bias if appropriate scaling is not used.  
+
+<|ref|>text<|/ref|><|det|>[[117, 707, 871, 852]]<|/det|>
+In the revision we have removed the general statement about the spatial distribution of saturated and unsaturated fatty acids and have replaced old Fig 5f with the specific example for C16 and C18 (new Fig 4f,g). We have also replaced the previous text with the following (lines 520- 524) "WSoR PC 7 (Figure 4a,d) reveals that C16:0 and C18:1 are anticorrelated with different spatial distributions. We find that the signal intensity ratio of saturated to unsaturated fatty acids C16:1 to C16:0 (Figure 4f) is enhanced in the neuropil whereas the ratio C18:1 to C18:0 (Figure 4g) is stronger at the periphery of the cortex along the blood- brain barrier." We hope that this is now clearer and thank the reviewer for pointing out the error.
+
+<--- Page Split --->
+<|ref|>image<|/ref|><|det|>[[137, 88, 848, 435]]<|/det|>  
+
+<|ref|>text<|/ref|><|det|>[[117, 458, 844, 494]]<|/det|>
+Reviewer's Figure R1: Principal component analysis (with WSoR scaling) of a subset of the CNS dataset containing only 10 fatty acid peaks.
+
+<--- Page Split --->

peer_reviews/supplementary_0_Transparent Peer Review file__7449e55a7fb2473376ccf3c30b6b2a5c1de2ca5732f3d4e94009eb78eeac2ebd/images_list.json

@@ -0,0 +1,550 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_0.jpg",
+    "caption": "Fig. R1 | Comparison between traditional jammers and the STCM system. a. Physical image of the traditional jammer, with dimensions of \\(200\\mathrm{mm}*150\\mathrm{mm}*120\\mathrm{mm}\\) , weighing approximately \\(10\\mathrm{kg}\\) , featuring a",
+    "footnote": [],
+    "bbox": [
+      [
+        272,
+        654,
+        750,
+        850
+      ]
+    ],
+    "page_idx": 5
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_1.jpg",
+    "caption": "Fig. R2 | Comparison of the scattering characteristics between the RF absorber and the STCM system. a. The RF absorber can only suppress the fundamental frequency. b. The STCM system has the ability to suppress the fundamental frequency while simultaneously generating multiple harmonics. It can also provide stealth and jamming.",
+    "footnote": [],
+    "bbox": [
+      [
+        180,
+        235,
+        808,
+        411
+      ]
+    ],
+    "page_idx": 6
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_2.jpg",
+    "caption": "Fig. R3 | Advantages of STCM's stealth and jamming over RF absorbers when radar transmit power is increased.",
+    "footnote": [],
+    "bbox": [
+      [
+        130,
+        82,
+        870,
+        265
+      ]
+    ],
+    "page_idx": 7
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_3.jpg",
+    "caption": "Fig. R4 | Schematic diagram of the false target camouflage zones. a, The camouflage zones of the false target generated by TDOA localization with varying modulation frequencies. b, The camouflage zones of the false target at each receiver of MRDF with varying modulation frequencies.",
+    "footnote": [],
+    "bbox": [
+      [
+        117,
+        652,
+        876,
+        792
+      ]
+    ],
+    "page_idx": 8
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_4.jpg",
+    "caption": "Fig. R5 | STCM scattering characteristics at different incident angles from the transmitters, and the echo conditions at each Receiver. The angles of Receiver 1, Receiver 2, and Receiver 3 are -26.56°, 18.43°, and 45°, respectively. a. The transmitter incident angle is 0°. b. The transmitter incident angle is 20°. c. The transmitter incident angle is -40°.",
+    "footnote": [],
+    "bbox": [
+      [
+        118,
+        364,
+        875,
+        521
+      ]
+    ],
+    "page_idx": 10
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_5.jpg",
+    "caption": "Fig. R6 | Oblique incidence performance of the STCM unit. a, c, Programmable elements amplitude and phase spectra at \\(30^{\\circ}\\) oblique incidence. b, d, Programmable elements amplitude and phase spectra at \\(60^{\\circ}\\) oblique incidence.",
+    "footnote": [],
+    "bbox": [
+      [
+        234,
+        90,
+        757,
+        277
+      ]
+    ],
+    "page_idx": 11
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_6.jpg",
+    "caption": "Fig. R7 | Countermeasure effectiveness of the STCM system against the FDOA localization of the MSR system when multiple transmitters are present. a, d, The MSR localization results and the Doppler frequencies of signals at each receiver when Transmitter 1(T1) illuminates from -53.13°. b, e, The MSR localization results and the Doppler frequencies of signals at each receiver when Transmitter 2(T2) illuminates from 0°. c, f, The MSR localization results and the Doppler frequencies of signals at each receiver when Transmitter 3(T3) illuminates from 59.04°.",
+    "footnote": [],
+    "bbox": [
+      [
+        115,
+        535,
+        880,
+        724
+      ]
+    ],
+    "page_idx": 11
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_7.jpg",
+    "caption": "Fig. R8 | Countermeasure effectiveness of the STCM system against the TDOA localization of the MSR system when multiple transmitters are present. a, d, The MSR localization results and the signal delays at each receiver when Transmitter 1(T1) illuminates from \\(-53.13^{\\circ}\\) . b, e, The MSR localization results and the signal delays at each receiver when Transmitter 2(T2) illuminates from \\(0^{\\circ}\\) . c, f, The MSR localization results and the signal delays at each receiver when Transmitter 3(T3) illuminates from \\(59.04^{\\circ}\\) .",
+    "footnote": [],
+    "bbox": [
+      [
+        122,
+        88,
+        880,
+        263
+      ]
+    ],
+    "page_idx": 12
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_8.jpg",
+    "caption": "Fig. R9 | Outdoor dynamic experimental validation of the STCM-based anti-MSR. a, The drone is equipped with three STCMs and is powered by a small battery that supplies energy to the entire system. b, Outdoor dynamic experiment scenarios. The signal processing system is connected to a transmitting horn and four receiving horns to simulate the MSR system's transmitters and receivers. We selected three flight points to observe the echo signal characteristics at each receiver. c-e, Frequency spectra of each receiver at the flight",
+    "footnote": [],
+    "bbox": [
+      [
+        115,
+        150,
+        881,
+        812
+      ]
+    ],
+    "page_idx": 13
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_9.jpg",
+    "caption": "Fig. R10 | Outdoor dynamic experimental validation of the STCM-based anti-MSR. a, The drone is equipped with three STCMs and is powered by a small battery that supplies energy to the entire system. b, Outdoor dynamic experiment scenarios. The signal processing system is connected to a transmitting horn and",
+    "footnote": [],
+    "bbox": [
+      [
+        115,
+        180,
+        880,
+        840
+      ]
+    ],
+    "page_idx": 17
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_10.jpg",
+    "caption": "Fig. R11 | Existing studies on curved metasurface. a, Schematic of the flexible metasurface with real-time and independent control of orthogonal-polarized EM waves. b, Conformal metasurface and its static performance.",
+    "footnote": [],
+    "bbox": [
+      [
+        121,
+        545,
+        870,
+        712
+      ]
+    ],
+    "page_idx": 18
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_11.jpg",
+    "caption": "Fig. R12 | Schematic diagram of the false target camouflage zones. a, The camouflage zones of the false target generated by TDOA localization with varying modulation frequencies. b, The camouflage zones of the false target at each receiver of MRDF with varying modulation frequencies.",
+    "footnote": [],
+    "bbox": [
+      [
+        117,
+        282,
+        863,
+        420
+      ]
+    ],
+    "page_idx": 20
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_12.jpg",
+    "caption": "Fig. R13 | Comparison between traditional jammers and the STCM system. a. Physical image of the traditional jammer, with dimensions of \\(200\\mathrm{mm}*150\\mathrm{mm}*120\\mathrm{mm}\\) , weighing approximately \\(10\\mathrm{kg}\\) , featuring a single transmit channel and a single receive channel. b. Physical image of the STCM system, with dimensions of \\(224\\mathrm{mm}*250\\mathrm{mm}*1.5\\mathrm{mm}\\) , weighing \\(230\\mathrm{g}\\) , and capable of simultaneously jamming multiple Receivers with multiple beams.",
+    "footnote": [],
+    "bbox": [
+      [
+        260,
+        615,
+        777,
+        813
+      ]
+    ],
+    "page_idx": 22
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_13.jpg",
+    "caption": "Fig. R14 | Comparison of the scattering characteristics between the RF absorber and the STCM system. a. The RF absorber can only suppress the fundamental frequency. b. The STCM system has the ability to suppress the fundamental frequency while simultaneously generating multiple harmonics. It can also provide stealth and jamming.",
+    "footnote": [],
+    "bbox": [
+      [
+        118,
+        172,
+        870,
+        382
+      ]
+    ],
+    "page_idx": 23
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_14.jpg",
+    "caption": "Fig. R15 | Outdoor dynamic experimental validation of the STCM-based anti-MSR. a, The drone is",
+    "footnote": [],
+    "bbox": [
+      [
+        115,
+        216,
+        880,
+        875
+      ]
+    ],
+    "page_idx": 25
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_15.jpg",
+    "caption": "Fig. R16 | a, The drone is equipped with three STCMs, and a small battery powers the entire system. b, Flight attitude of a drone equipped with STCMs.",
+    "footnote": [],
+    "bbox": [
+      [
+        125,
+        92,
+        880,
+        279
+      ]
+    ],
+    "page_idx": 28
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_16.jpg",
+    "caption": "Fig. R17 | STCM scattering characteristics at different incident angles from the transmitters, and the echo conditions at each Receiver. The angles of Receiver 1, Receiver 2, and Receiver 3 are -26.56°, 18.43°, and 45°, respectively. a. The transmitter incident angle is 0°. b. The transmitter incident angle is 20°. c. The transmitter incident angle is -40°.",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        550,
+        870,
+        705
+      ]
+    ],
+    "page_idx": 28
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_17.jpg",
+    "caption": "Fig. R18 | Oblique incidence performance of the STCM unit. a, b, Programmable elements amplitude and phase spectra at \\(30^{\\circ}\\) oblique incidence. c, d, Programmable elements amplitude and phase spectra at \\(60^{\\circ}\\) oblique incidence.",
+    "footnote": [],
+    "bbox": [
+      [
+        230,
+        88,
+        763,
+        280
+      ]
+    ],
+    "page_idx": 29
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_18.jpg",
+    "caption": "Fig. R20 | Schematic diagram of the false target camouflage zones. a, The camouflage zones of the false target generated by TDOA localization with varying modulation frequencies. b, The camouflage zones of the false target at each receiver of MRDF with varying modulation frequencies.",
+    "footnote": [],
+    "bbox": [],
+    "page_idx": 29
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_19.jpg",
+    "caption": "Fig. R21 | Countermeasure effectiveness of the STCM system against the FDOA localization of the MSR system when multiple transmitters are present. a, d, The MSR localization results and the Doppler frequencies of signals at each receiver when Transmitter 1 (T1) illuminates from -53.13°. b, e, The MSR localization results and the Doppler frequencies of signals at each receiver when Transmitter 2 (T2) illuminates from 0°. c, f, The MSR localization results and the Doppler frequencies of signals at each receiver when Transmitter 3(T3) illuminates from 59.04°.",
+    "footnote": [],
+    "bbox": [
+      [
+        115,
+        360,
+        880,
+        550
+      ]
+    ],
+    "page_idx": 30
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_20.jpg",
+    "caption": "Fig. R22 | Countermeasure effectiveness of the STCM system against the TDOA localization of the MSR system when multiple transmitters are present. a, d, The MSR localization results and the signal",
+    "footnote": [],
+    "bbox": [
+      [
+        123,
+        679,
+        880,
+        858
+      ]
+    ],
+    "page_idx": 31
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_21.jpg",
+    "caption": "Fig. R23 | Jamming effectiveness of the STCM system when the transmitted signal is a broadband",
+    "footnote": [],
+    "bbox": [
+      [
+        300,
+        612,
+        700,
+        881
+      ]
+    ],
+    "page_idx": 31
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_22.jpg",
+    "caption": "Fig. R24 | Comparison between traditional jammers and the STCM system. a. Physical image of the traditional jammer, with dimensions of \\(200\\mathrm{mm}*150\\mathrm{mm}*120\\mathrm{mm}\\) , weighing approximately \\(10\\mathrm{kg}\\) , featuring a single transmit channel and a single receive channel. b. Physical image of the STCM system, with dimensions of \\(224\\mathrm{mm}*250\\mathrm{mm}*1.5\\mathrm{mm}\\) , weighing \\(230\\mathrm{g}\\) , and capable of simultaneously jamming multiple Receivers with multiple beams. c, The drone is equipped with three STCMs, and a small battery powers the entire STCM system.",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        578,
+        880,
+        755
+      ]
+    ],
+    "page_idx": 32
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2b.jpg",
+    "caption": "Fig. R25 | Scattering characteristics of the STCM in Figure 2b of the original manuscript, and the frequency offsets of each receiver.",
+    "footnote": [],
+    "bbox": [
+      [
+        302,
+        653,
+        686,
+        866
+      ]
+    ],
+    "page_idx": 33
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_23.jpg",
+    "caption": "Fig. R26 | Solution of equation (4) (or the TDOA localization situation) when the number of transmitters in the MSR system increases. a, d, The MSR localization results and the signal delays at each receiver when Transmitter 1(T1) illuminates from -53.13°. b, e, The MSR localization results and the signal delays at each receiver when Transmitter 2(T2) illuminates from 0°. c, f, The MSR localization results and the signal delays at each receiver when Transmitter 3(T3) illuminates from 59.04°.",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        551,
+        883,
+        728
+      ]
+    ],
+    "page_idx": 35
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_24.jpg",
+    "caption": "Fig. R27 | Solution of equation (4) (or the TDOA localization) when there are multiple Receivers in the MSR system. a, d, The localization results of the MSR system and the signal delays at each receiver when the system consists of five receivers and STCM is not modulated. b, e, The localization results of the MSR system and the signal delays at each receiver when the system consists of five receivers and STCM is modulated.",
+    "footnote": [],
+    "bbox": [
+      [
+        160,
+        221,
+        845,
+        450
+      ]
+    ],
+    "page_idx": 36
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_25.jpg",
+    "caption": "Fig. R28 | Simulation results of the model. a. Signal spectrum. b. Signal time domain.",
+    "footnote": [],
+    "bbox": [
+      [
+        150,
+        420,
+        836,
+        564
+      ]
+    ],
+    "page_idx": 37
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_26.jpg",
+    "caption": "Fig. R29 | Comparison of the Doppler frequencies of moving target echo signals with and without STCM modulation. a. STCM without modulation. b. STCM modulation.",
+    "footnote": [],
+    "bbox": [
+      [
+        156,
+        633,
+        852,
+        860
+      ]
+    ],
+    "page_idx": 38
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_27.jpg",
+    "caption": "Fig. R1 | STCM conformal design. a-d, Various bending forms of the STCM. Through flexible conformal design, the STCM can be used to cover target surfaces of different shapes. e, Amplitude characteristics of the conformal STCM element. f, Phase characteristics of the conformal STCM element.",
+    "footnote": [],
+    "bbox": [
+      [
+        179,
+        103,
+        812,
+        448
+      ]
+    ],
+    "page_idx": 39
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_28.jpg",
+    "caption": "Fig. R2 | Design of STCM's element with dual polarization and a wide range of incident angles. a, b, STCM element structure. c, d, Phase and amplitude characteristics of the element under a wide range of incident angles. The phase and amplitude characteristics remain stable within the \\(0^{\\circ}\\) to \\(60^{\\circ}\\) incident range.",
+    "footnote": [],
+    "bbox": [
+      [
+        209,
+        519,
+        785,
+        815
+      ]
+    ],
+    "page_idx": 43
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_29.jpg",
+    "caption": "Fig. R3 | The main contribution and breakthrough of this work.",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        185,
+        870,
+        550
+      ]
+    ],
+    "page_idx": 43
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_30.jpg",
+    "caption": "Fig. R4 | STCM conformal design. a-d, Various bending forms of the STCM. Through flexible conformal design, the STCM can be used to cover target surfaces of different shapes. e, Amplitude characteristics of the conformal STCM element. f, Phase characteristics of the conformal STCM element.",
+    "footnote": [],
+    "bbox": [
+      [
+        200,
+        144,
+        789,
+        461
+      ]
+    ],
+    "page_idx": 44
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_31.jpg",
+    "caption": "Fig. R5 | Design of STCM's element with dual polarization and a wide range of incident angles. a, b, STCM element structure. c, d, Phase and amplitude characteristics of the element under a wide range of incident angles. The phase and amplitude characteristics remain stable within the \\(0^{\\circ}\\) to \\(60^{\\circ}\\) incident range.",
+    "footnote": [],
+    "bbox": [
+      [
+        210,
+        539,
+        785,
+        833
+      ]
+    ],
+    "page_idx": 48
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_32.jpg",
+    "caption": "Fig. R6 | The main contribution and breakthrough of this work.",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        185,
+        870,
+        550
+      ]
+    ],
+    "page_idx": 48
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_33.jpg",
+    "caption": "Fig. R7 | The relationship between received power and distance.",
+    "footnote": [],
+    "bbox": [
+      [
+        285,
+        216,
+        690,
+        450
+      ]
+    ],
+    "page_idx": 49
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_34.jpg",
+    "caption": "Fig. R8 | The relationship between received power and distance. a, 3-D view of the high-power element. b, High-power element structure design.",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        600,
+        856,
+        805
+      ]
+    ],
+    "page_idx": 52
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_unknown_35.jpg",
+    "caption": "Fig. R9 | Analysis of jamming effects at different signal bandwidths. a, b, Results for a signal bandwidth of 5 MHz. c, d, Results for a signal bandwidth of 20 MHz. e, f, Results for a signal bandwidth of 30 MHz.",
+    "footnote": [],
+    "bbox": [
+      [
+        240,
+        255,
+        760,
+        676
+      ]
+    ],
+    "page_idx": 52
+  }
+]

peer_reviews/supplementary_0_Transparent Peer Review file__7514694b688643c6a27efe7684b270e04cbc69dc584319e2bd9bc2f86f25d8f6/images_list.json

@@ -0,0 +1,85 @@
+[
+  {
+    "type": "image",
+    "img_path": "images/Supplementary_Figure_3.jpg",
+    "caption": "Supplementary Figure 3. Term-equivalent age topological analysis. (A) Density significantly increases, (B) modularity significantly decreases, and (C) global efficiency increases across GA at birth. (D) There is no significant change in the number of rich club connections, however (E) the average length of rich club connections significantly increases across GA at birth. The shaded area indicates 95% confidence intervals. The grey dotted line indicates the cut-off for term birth (37 weeks GA or later is term-born). \\\\*\\\\*\\\\* indicates \\(\\mathsf{p}< 0.001\\) , \\\\*\\\\* indicates \\(\\mathsf{p}< 0.01\\) , \\\\* indicates \\(\\mathsf{p}< 0.05\\)",
+    "footnote": [],
+    "bbox": [
+      [
+        120,
+        80,
+        868,
+        555
+      ]
+    ],
+    "page_idx": 12
+  },
+  {
+    "type": "image",
+    "img_path": "images/Supplementary_Figure_11.jpg",
+    "caption": "Supplementary Figure 11. Exploration of the effects of minimum fiber lengths of tractography. (A) Density (5mm: \\(F_{density,PMA} = 11.77\\) , estimated df \\(= 8.43\\) , \\(p = 2.00 \\times 10^{-16}\\) , \\(F_{density,GA}\\) at birth \\(= 7.41\\) , estimated df \\(= 3.29\\) , \\(p = 6.18 \\times 10^{-6}\\) ; 10mm: \\(F_{density,PMA} = 39.71\\) , estimated df \\(= 7.71\\) , \\(p = 2.00 \\times 10^{-16}\\) , \\(F_{density,GA}\\) at birth \\(= 4.23\\) , estimated df \\(= 1.00\\) , \\(p = 0.040\\) ; 20mm: \\(F_{density,PMA} = 15.98\\) , estimated df \\(= 10.85\\) , \\(p = 2.00 \\times 10^{-16}\\) , \\(F_{density,GA}\\) at birth \\(= 6.22\\) , estimated df \\(= 2.80\\) , \\(p = 1.55 \\times 10^{-4}\\) ; 30mm: \\(F_{density,PMA} = 37.17\\) , estimated df \\(= 7.50\\) , \\(p = 2.00 \\times 10^{-16}\\) , GA at birth: \\(p = 0.166\\) , global efficiency (5mm: \\(F_{global efficiency,PMA} = 10.89\\) , estimated df \\(= 8.69\\) , \\(p = 2.00 \\times 10^{-16}\\) , \\(F_{global efficiency,GA}\\) at birth \\(= 7.06\\) , estimated df \\(= 2.93\\) , \\(p = 3.34 \\times 10^{-5}\\) ; 10mm: \\(F_{global efficiency,PMA} = 37.84\\) , estimated df \\(= 6.87\\) , \\(p = 2.00 \\times 10^{-16}\\) , \\(F_{global efficiency,GA}\\) at birth \\(= 6.25\\) , estimated df \\(= 1.00\\) , \\(p = 0.013\\) ; 20mm: \\(F_{global efficiency,PMA} = 12.10\\) , estimated df \\(= 10.16\\) , \\(p = 2.00 \\times 10^{-16}\\) , \\(F_{global efficiency,GA}\\) at birth \\(= 6.56\\) , estimated df \\(= 2.72\\) , \\(p = 1.08 \\times 10^{-4}\\) ; 30mm: \\(F_{global efficiency,PMA} = 35.29\\) , estimated df \\(= 7.23\\) , \\(p = 2.00 \\times 10^{-16}\\) , \\(F_{global efficiency,GA}\\) at birth \\(= 4.21\\) , estimated df \\(= 1.00\\) , \\(p = 0.041\\) , and modularity (5mm: \\(F_{modularity,PMA} = 6.99\\) , estimated df \\(= 3.51\\) , \\(p = 8.83 \\times 10^{-6}\\) , \\(F_{modularity,GA}\\) at birth \\(= 19.98\\) , estimated df \\(= 1.00\\) , \\(p = 9.53 \\times 10^{-6}\\) ; 10mm: \\(F_{modularity,PMA} = 7.15\\) , estimated df \\(= 4.19\\) , \\(p = 1.43 \\times 10^{-6}\\) , \\(F_{modularity,GA}\\) at birth \\(= 9.18\\) , estimated df \\(= 1.00\\) , \\(p = 2.53 \\times 10^{-3}\\) ; 20mm: \\(PMA\\) \\(p = 0.900\\) , \\(F_{modularity,GA}\\) at birth \\(= 10.32\\) , estimated df \\(= 1.86\\) , \\(p = 2.19 \\times 10^{-5}\\) ; 30mm: \\(F_{modularity,PMA} = 6.85\\) , estimated df \\(= 2.84\\) , \\(p = 7.64 \\times 10^{-5}\\) , \\(F_{modularity,GA}\\) at birth \\(= 4.38\\) , estimated df \\(= 2.80\\) , \\(p = 3.46 \\times 10^{-3}\\) ) across GA at birth and PMA. (B) Topological fingerprints and (C) consensus networks for each tractography method. (D) Average binarized matrices of early and late PMA neonatal scans for 30mm fiber length connectomes.",
+    "footnote": [],
+    "bbox": [
+      [
+        122,
+        81,
+        875,
+        610
+      ]
+    ],
+    "page_idx": 13
+  },
+  {
+    "type": "image",
+    "img_path": "images/Supplementary_Figure_2.jpg",
+    "caption": "Supplementary Figure 2. Multiple density consensus networks. (A) Modularity decreases, (B) Global efficiency increases, and (C) characteristic path length decrease with increasingly dense connectomes. (D) Comparisons of consecutive connectomes, indicating the distribution of new connections across 7 lobes (indicated as a \\(\\%\\) change from the sparser to dense network). (E) Topological fingerprints demonstrating the correlation between local organizational measures for the \\(8\\%\\) (sparsest), \\(18\\%\\) (middle) and \\(24\\%\\) (densest) density networks. (F) Connectome matrices representing the \\(8\\%\\) , \\(18\\%\\) and \\(24\\%\\) networks. Here, the lower triangle is the matrix and for \\(18\\%\\) and \\(24\\%\\) networks the upper triangle is only comprised of connections that are present in that network that weren't present in the previous network (8-18% and 18-24% comparisons).",
+    "footnote": [],
+    "bbox": [
+      [
+        117,
+        147,
+        880,
+        650
+      ]
+    ],
+    "page_idx": 15
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_2.jpg",
+    "caption": "Figure 2. Organizational differences across PMA and GA at birth. (A) Modularity significantly decreased across PMA and GA at birth in variable density networks, but modularity significantly increased across PMA and decreased across GA at birth in controlled-density networks. Global efficiency significantly increased across both PMA and GA at birth in variable density networks. However, with density-controlled networks global efficiency significantly decreased across PMA but no significant effect was found across GA at birth. (B) The number of rich club connections significantly increased across PMA in both variable density and density-controlled networks. The number of rich club connections also significantly increased across GA at birth in variable density networks, but no significant effect was found with density-controlled networks. In addition, average length of rich club connections (Euclidean distance) significantly decreased across PMA in both variable density and density-controlled networks. The average length of rich club connections significantly increased across GA at birth in both variable density and density-controlled networks. The shaded area indicates 95% confidence intervals. The grey dotted line indicates the cut-off for term birth (37 weeks GA or later is term-born). \\\\*\\\\*\\\\* indicates \\(p< 0.001\\) , \\\\*\\\\* indicates \\(p< 0.01\\) , \\\\* indicates \\(p< 0.05\\) .",
+    "footnote": [],
+    "bbox": [
+      [
+        130,
+        80,
+        870,
+        691
+      ]
+    ],
+    "page_idx": 19
+  },
+  {
+    "type": "image",
+    "img_path": "images/Supplementary_Figure_3.jpg",
+    "caption": "Supplementary Figure 3. Term-equivalent age topological analysis. (A) Density significantly increases across GA at birth. (B) Modularity significantly decreases across GA at birth and propensity-matched analysis shows that preterm infants have significantly higher modularity compared to term infants. (C) Global efficiency significantly increases across GA at birth and propensity-matched shows preterm infants have significantly lower global efficiency compared to term infants. (D) There is no significant change in the number of rich club connections, however (E) the average length of rich club connections significantly increases across GA at birth. The shaded area indicates 95% confidence intervals. The grey dotted line indicates the cut-off for term birth (37 weeks GA or later is term-born). *** indicates \\(p < 0.001\\) , ** indicates \\(p < 0.01\\) , * indicates \\(p < 0.05\\) .",
+    "footnote": [],
+    "bbox": [
+      [
+        115,
+        81,
+        852,
+        542
+      ]
+    ],
+    "page_idx": 43
+  },
+  {
+    "type": "image",
+    "img_path": "images/Figure_1.jpg",
+    "caption": "Figure 1. Neonatal connectivity across PMA and GA at birth. (A) Tractograms highlighting the reduced connectivity between preterm and term infants who were both scanned at 42 weeks PMA. Below the tractograms, a graphic depicts the difference between PMA (orange) and GA at birth (blue) for these infants. The red bars indicate the time infants were exposed to the extra-uterine environment before being scanned. (B) Representative tractograms of four neonates born and scanned at different PMA and GA at birth. (C) Connectome density significantly increased across PMA and GA at birth (represented as predicted density from generalized additive model). The grey dotted line indicates the cut-off for term birth (37 weeks GA or later is term-born). (D) Average connectivity weighted matrix (average number of streamlines) and binarized matrix (consistency matrix), where the value indicates that proportion of participants that have that edge. *** indicates \\(p < 0.001\\) , ** indicates \\(p < 0.01\\) , * indicates \\(p < 0.05\\) .",
+    "footnote": [],
+    "bbox": [],
+    "page_idx": 44
+  }
+]