Publish 2026_W39_encoder

README.md

@@ -29,7 +29,10 @@ A published encoder is not the weights alone. `predict` cannot run without every
 
 - `spatial_encoder.pt` -- the trained network.
 - the context PCA volumes (`agea_vol_pca.npy`, `merfish_vol_pca.npy`) -- the anatomical context
-  it samples at each position.
+  it samples at each position: principal components of the AGEA gene-expression volume and of the
+  MERFISH subclass cell-type densities, denoised with `iblatlas.genomics.merfish.denoise_volume`
+  (NaNs filled, Gaussian-smoothed, renormalised within the brain; see `config.context.sources`
+  in `ephysatlas_model.json`).
 - `neighbor_bank.npz` -- the **neighbour bank**: the position, standardised feature vector and
   insertion id of every training channel. At inference the model gathers the nearest recorded
   channels from this bank and attends to them, so the bank *is* part of the model, the way the

agea_vol_pca.npy

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4544da5f861e76858f6de15dc1e2926da3c0f2abafd0fde966410169bca23aad
+oid sha256:648f676671127966e2df58bb3a9aa694a4ea1d35e246bd02689f9909ba34a28f
 size 31865328

checksums.json

@@ -3,23 +3,23 @@
   "files": [
     {
       "path": "agea_vol_pca.npy",
-      "hash": "63b28e9dc7b2038de5de4a13ff0f41f343ceb3f3",
+      "hash": "b710b6b26797e78bbc49762e80f2ae5ab6bb3896",
       "bytes": 31865328
     },
     {
       "path": "confidence_model.pt",
-      "hash": "845fc230fada14fbd396cbed8d53d570d011270b",
+      "hash": "7c5eed503a08c6d3afe5047702f0b3c961def869",
       "bytes": 387569
     },
     {
       "path": "ephysatlas_model.json",
-      "hash": "496b819dcf1a77d93d566758861847375956f319",
-      "bytes": 6662
+      "hash": "c854d99dd6a81a38ad8c5b5d234648e62f029031",
+      "bytes": 6992
     },
     {
       "path": "example/expected_predictions.parquet",
-      "hash": "c5d4b8aef65523d040e900505b5ada72e90b03e5",
-      "bytes": 155898
+      "hash": "f3d23df86e52e720a223595884c6ea91ac578a6e",
+      "bytes": 155858
     },
     {
       "path": "example/features_sample.parquet",
@@ -28,7 +28,7 @@
     },
     {
       "path": "merfish_vol_pca.npy",
-      "hash": "436d32acbcfd303a53b40e42a7da72e120277f7b",
+      "hash": "c8fd5484d8c0619c5c542242c9ac99876db56892",
       "bytes": 31865328
     },
     {
@@ -38,22 +38,22 @@
     },
     {
       "path": "preprocessing/channel_stats.npz",
-      "hash": "66020ce3c717f89c3851edcc2445fd26fb9b8ed3",
-      "bytes": 3402
+      "hash": "122634168092c7bc5fea981960100e7902a5c857",
+      "bytes": 3384
     },
     {
       "path": "results/confidence_validation.json",
-      "hash": "5e4ebe6a928073fd445c40fd9e608aed37d7cbaa",
+      "hash": "f04a224d8afa2f818b5c197b948497bd07c76b0d",
       "bytes": 291
     },
     {
       "path": "results/test_r2.json",
-      "hash": "6b10e9e55ce46e8ceeaa9f41121219eb33fd935a",
-      "bytes": 2021
+      "hash": "d655cf0658a0bc7ffc31332c0347afd233adcd52",
+      "bytes": 2018
     },
     {
       "path": "spatial_encoder.pt",
-      "hash": "13ade9f2714699cf3571f40a0cd28080c8ea6ac3",
+      "hash": "7d1a1f6223eb0494dfb7fb1a218022a89dfb0ba9",
       "bytes": 4625417
     },
     {

confidence_model.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:efeed1d0b1a02ccc26863204a2960071957b6285410c238ea382cdfaba1ebbb6
+oid sha256:61383c56e54c04d3d8e5b78f1d9e116a42352033199a8336b72212507144b253
 size 387569

ephysatlas_model.json

@@ -9,7 +9,7 @@
     "validation_size": 81,
     "testing_size": 162,
     "best_epoch": 3,
-    "mean_test_r2": 0.39225515723228455
+    "mean_test_r2": 0.3942371606826782
   },
   "environment": {
     "python": "3.11.9",
@@ -117,7 +117,18 @@
     "context": {
       "n_cell_pcs": 50,
       "n_gene_pcs": 50,
-      "voxel_um": 200
+      "voxel_um": 200,
+      "sources": {
+        "merfish": {
+          "level": "subclass",
+          "volume": "iblatlas.genomics.merfish.denoise_volume",
+          "n_drop_non_neuronal": 0,
+          "sigma_voxels": 0.5
+        },
+        "agea": {
+          "volume": "iblatlas.genomics.agea.load(label=\"processed\")"
+        }
+      }
     },
     "preprocessing": {
       "clip_percentiles": [

example/expected_predictions.parquet

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:748b30b51a5bc3c9b1fb23c088fb309af1ccafb52f024c231844c88329cf9e6d
-size 155898
+oid sha256:ff49e53f4cdb187d53edb32f24d47d11abe418f01f2f154873d07bcd53ae6e7a
+size 155858

merfish_vol_pca.npy

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c9e02fa3ce79fb97f1e1a8d4808ddd6cd37934b037bb65d87803532c6a13cfa3
+oid sha256:95fcc9a5d829e0485c372e0ac9e3511acbebc89c56f6861e2ef8b4b6f34c5643
 size 31865328

preprocessing/channel_stats.npz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:81606c3d1a5f4a547c2963e426a7e59d900209da1c4c23f17666f5800baa0b09
-size 3402
+oid sha256:b0548b3e1872120a85b7ec0da03071fb24dafdf9d224932f25f179693ba0342e
+size 3384

results/confidence_validation.json

@@ -1,14 +1,14 @@
 {
   "split": "validation (synthetic misalignments)",
-  "accuracy": 0.7697482638888888,
+  "accuracy": 0.7644595550411523,
   "confusion_matrix": [
     [
-      145968,
-      44237
+      145775,
+      44430
     ],
     [
-      13057,
-      45570
+      14180,
+      44447
     ]
   ],
   "classes": [

results/test_r2.json

@@ -1,50 +1,50 @@
 {
   "split": "test",
   "n_test_insertions": 162,
-  "mean_r2": 0.39225515723228455,
+  "mean_r2": 0.3942371606826782,
   "per_feature": {
-    "rms_lf": 0.6568270921707153,
-    "psd_delta": 0.6041296720504761,
-    "psd_theta": 0.7125661373138428,
-    "psd_alpha": 0.7067911624908447,
-    "psd_beta": 0.6807988882064819,
-    "psd_gamma": 0.6520721912384033,
-    "psd_lfp": 0.6709305047988892,
-    "rms_lf_csd_diff1": 0.3776247501373291,
-    "psd_delta_csd_diff1": 0.34319454431533813,
-    "psd_theta_csd_diff1": 0.49149876832962036,
-    "psd_alpha_csd_diff1": 0.48544758558273315,
-    "psd_beta_csd_diff1": 0.43134593963623047,
-    "psd_gamma_csd_diff1": 0.3505679965019226,
-    "psd_lfp_csd_diff1": 0.38819313049316406,
-    "psd_residual_delta": 0.38726121187210083,
-    "psd_residual_theta": 0.40748900175094604,
-    "psd_residual_alpha": 0.3592684268951416,
-    "psd_residual_beta": 0.2527044415473938,
-    "psd_residual_gamma": 0.20339089632034302,
-    "psd_residual_lfp": 0.15038365125656128,
-    "aperiodic_offset": 0.6621072292327881,
-    "aperiodic_exponent": 0.3824368715286255,
-    "decay_fit_error": 0.26942986249923706,
-    "decay_fit_r_squared": 0.3475661277770996,
-    "decay_n_peaks": 0.09780853986740112,
-    "cor_ratio": 0.590096116065979,
-    "rms_ap": 0.5781934261322021,
-    "alpha_mean": 0.36733919382095337,
-    "alpha_std": 0.32872676849365234,
-    "spike_count": 0.24967080354690552,
-    "depolarisation_slope": 0.41250598430633545,
-    "polarity": 0.2455672025680542,
-    "recovery_slope": 0.3471274971961975,
-    "repolarisation_slope": 0.36749958992004395,
-    "slowness_s_per_m": 0.09602916240692139,
-    "slowness_s_per_m_std": 0.4324660897254944,
-    "spatial_spread_um": 0.2887027859687805,
-    "spatial_spread_um_std": 0.13961225748062134,
-    "tip_val": 0.2865339517593384,
-    "spike_width_secs": 0.4519430994987488,
-    "predepolarisation_width_secs": 0.2852904200553894,
-    "spike_amplitude": 0.3234935998916626,
-    "peak_to_trough_ratio_log": 0.004339396953582764
+    "rms_lf": 0.6550701260566711,
+    "psd_delta": 0.5976907014846802,
+    "psd_theta": 0.7105530500411987,
+    "psd_alpha": 0.7033395767211914,
+    "psd_beta": 0.6790828704833984,
+    "psd_gamma": 0.6526632308959961,
+    "psd_lfp": 0.6669749021530151,
+    "rms_lf_csd_diff1": 0.372251033782959,
+    "psd_delta_csd_diff1": 0.33365315198898315,
+    "psd_theta_csd_diff1": 0.4834861755371094,
+    "psd_alpha_csd_diff1": 0.47833847999572754,
+    "psd_beta_csd_diff1": 0.4298723340034485,
+    "psd_gamma_csd_diff1": 0.34356796741485596,
+    "psd_lfp_csd_diff1": 0.382626473903656,
+    "psd_residual_delta": 0.3934798240661621,
+    "psd_residual_theta": 0.4087831377983093,
+    "psd_residual_alpha": 0.36000263690948486,
+    "psd_residual_beta": 0.2635628581047058,
+    "psd_residual_gamma": 0.2118266224861145,
+    "psd_residual_lfp": 0.16072648763656616,
+    "aperiodic_offset": 0.6590027809143066,
+    "aperiodic_exponent": 0.38344305753707886,
+    "decay_fit_error": 0.2743866443634033,
+    "decay_fit_r_squared": 0.3550477623939514,
+    "decay_n_peaks": 0.09816223382949829,
+    "cor_ratio": 0.5910254716873169,
+    "rms_ap": 0.5826618671417236,
+    "alpha_mean": 0.38318175077438354,
+    "alpha_std": 0.3446558713912964,
+    "spike_count": 0.24994319677352905,
+    "depolarisation_slope": 0.41396862268447876,
+    "polarity": 0.24988126754760742,
+    "recovery_slope": 0.3518065810203552,
+    "repolarisation_slope": 0.3700753450393677,
+    "slowness_s_per_m": 0.0968065857887268,
+    "slowness_s_per_m_std": 0.4426867365837097,
+    "spatial_spread_um": 0.29377901554107666,
+    "spatial_spread_um_std": 0.14359837770462036,
+    "tip_val": 0.29089581966400146,
+    "spike_width_secs": 0.45722121000289917,
+    "predepolarisation_width_secs": 0.28831756114959717,
+    "spike_amplitude": 0.32823026180267334,
+    "peak_to_trough_ratio_log": 0.015869319438934326
   }
 }

spatial_encoder.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:ab0e4a71f7a5b72120ed9a40fe98002beada9500c8a89651a6d999371f9730e8
+oid sha256:0f8c9347db81656c1e2e4d7dde60758ee8e6f15bbf2f67a65477f8aa1d7621d0
 size 4625417