Release 10 SimpleP2G single-patch checkpoints (fold 0, seed 42)

LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2025 SichenZhu
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

README.md

@@ -0,0 +1,81 @@
+---
+license: mit
+language:
+- en
+library_name: pytorch
+datasets:
+- MahmoodLab/hest
+tags:
+- biology
+- spatial-transcriptomics
+- computational-pathology
+- gene-expression
+- safetensors
+---
+
+# SimpleP2G pretrained prediction heads
+
+This repository contains **10 SimpleP2G checkpoints**: five HEST-1K datasets × two 200-gene panels × 1 setting(s), using **fold 0 and seed 42**.
+Each checkpoint was selected by validation-slide PCC-200 after a 100-epoch training run. The released weights are the EMA weights used for the saved test predictions.
+
+The models consume frozen 2,560-dimensional Virchow2 embeddings and predict gene expression in `log2(raw_count + 1)` units. **Virchow2 weights, image embeddings, tissue images, expression data, optimizer states and training logs are not included.** Obtain the image encoder separately from [paige-ai/Virchow2](https://huggingface.co/paige-ai/Virchow2).
+
+## Checkpoints
+
+| Setting | Dataset | Panel | Test slide | Selected epoch | Test PCC-200 |
+| --- | --- | --- | --- | ---: | ---: |
+| single-patch | Brain | [HMHVG-200](https://huggingface.co/king159/SimpleP2G/tree/main/models/single-patch/brain/hmhvg) | NCBI667 | 12 | 0.3803 |
+| single-patch | Brain | [HVG-200](https://huggingface.co/king159/SimpleP2G/tree/main/models/single-patch/brain/hvg) | NCBI667 | 16 | 0.1690 |
+| single-patch | ccRCC | [HMHVG-200](https://huggingface.co/king159/SimpleP2G/tree/main/models/single-patch/ccrcc/hmhvg) | INT2 | 19 | 0.2704 |
+| single-patch | ccRCC | [HVG-200](https://huggingface.co/king159/SimpleP2G/tree/main/models/single-patch/ccrcc/hvg) | INT2 | 7 | 0.1098 |
+| single-patch | HER2ST | [HMHVG-200](https://huggingface.co/king159/SimpleP2G/tree/main/models/single-patch/her2st/hmhvg) | SPA148 | 21 | 0.7089 |
+| single-patch | HER2ST | [HVG-200](https://huggingface.co/king159/SimpleP2G/tree/main/models/single-patch/her2st/hvg) | SPA148 | 37 | 0.3798 |
+| single-patch | Lung | [HMHVG-200](https://huggingface.co/king159/SimpleP2G/tree/main/models/single-patch/lung/hmhvg) | NCBI856 | 7 | 0.8717 |
+| single-patch | Lung | [HVG-200](https://huggingface.co/king159/SimpleP2G/tree/main/models/single-patch/lung/hvg) | NCBI856 | 8 | 0.4954 |
+| single-patch | PRAD | [HMHVG-200](https://huggingface.co/king159/SimpleP2G/tree/main/models/single-patch/prad/hmhvg) | MEND145 | 9 | 0.5561 |
+| single-patch | PRAD | [HVG-200](https://huggingface.co/king159/SimpleP2G/tree/main/models/single-patch/prad/hvg) | MEND145 | 12 | 0.2199 |
+
+These are the scores of the released fold only, not means across the three evaluation folds. Panels are selected using training slides and can contain different genes across datasets. Keep the exact output-column order in each checkpoint’s `genes.json`.
+
+## Download
+
+```python
+from huggingface_hub import snapshot_download
+
+snapshot_download("king159/SimpleP2G", local_dir="simplep2g-checkpoints")
+```
+
+## Load and predict
+
+The repository includes the small prediction-head implementation and a loader. From the downloaded repository directory (or an up-to-date SimpleP2G source checkout), install `torch`, `safetensors` and `huggingface_hub`, then:
+
+```python
+import torch
+from src.pretrained import load_pretrained
+
+model, genes, metadata = load_pretrained(
+    "king159/SimpleP2G", dataset="brain", panel="hmhvg", setting="single-patch", device="cpu"
+)
+
+# Supply your precomputed image embeddings: shape (batch, 2560).
+embeddings = torch.load("target_embeddings.pt", map_location="cpu", weights_only=True).float()
+with torch.inference_mode():
+    mean, log_variance = model(embeddings[:16])
+# mean[:, i] predicts genes[i] in log2(count + 1) units.
+```
+
+To load offline, pass the downloaded repository directory as the first argument. To pin a particular release, pass its Hub commit hash as `revision=`.
+
+The image feature convention is the CLS token concatenated with the mean of patch tokens (register tokens excluded), using the encoder’s prescribed preprocessing. Each model’s `config.json` records architecture, input/output conventions, split IDs, selection rule and metrics.
+
+## Files
+
+`models/<setting>/<dataset>/<panel>/` contains `model.safetensors`, `config.json`, and `genes.json`. `model_index.json` lists all checkpoints; `checksums.json` records SHA-256 hashes. Weights are exported losslessly and checked against the training implementation using strict state-dictionary loading and CPU forward comparisons.
+
+## Scope
+
+These are research models for predicting measured expression from image features. The fixed gene panels and tissue cohorts define their evaluated scope; the release does not include a clinical validation or cell-type classifier.
+
+## License
+
+This release follows the project’s MIT license; see [LICENSE](LICENSE). The separately obtained image encoder and source datasets retain their own terms.

checksums.json

@@ -0,0 +1,40 @@
+{
+  "LICENSE": "bc54668eb751e259d8cba32c737cc18713f1d8559a2215dbc70bb0841f8d7f92",
+  "README.md": "1fcb8245c14139fab21fcf130f538ca1eddb60153a3549b6b3a33ff0d2eff32a",
+  "model_index.json": "50ac11c642bbfb80229de50504b2f8802eb3c5389b3c8cdc15d24fdedf46e15b",
+  "models/single-patch/brain/hmhvg/config.json": "f5e8ea3c5c366ba9f2678992cbb8d0f77cb5f83c9b65413af6fde5e319bff9c0",
+  "models/single-patch/brain/hmhvg/genes.json": "2d5649cf126a6910d2671651d95491e44c23a9c9730d58ff70d5a1dfcafa8814",
+  "models/single-patch/brain/hmhvg/model.safetensors": "30e10f1adc24365ff8ae5d3e7851e3336dfdd1775b0a8640ba94fd46630633de",
+  "models/single-patch/brain/hvg/config.json": "4902b72f8d6a95f111c41915ec3a2926f9837db3bb63f7c75ab6cfa64a8e31fc",
+  "models/single-patch/brain/hvg/genes.json": "9f1b13b2b3be4158e50f8dede0259b880a9e789b30a88612f4da899b22744eb7",
+  "models/single-patch/brain/hvg/model.safetensors": "a671becb75830a258455a6136e1be92f5f45ae822c6875290f61edb0f90a9eaa",
+  "models/single-patch/ccrcc/hmhvg/config.json": "ee9418e61c755b41c390541c2cad5fe7ef86fb2e8164fcbc76e2342f3429d4bc",
+  "models/single-patch/ccrcc/hmhvg/genes.json": "ff93267954d744e7e3db035f7d920e281b41cca703413841d6d52b27e1dc4815",
+  "models/single-patch/ccrcc/hmhvg/model.safetensors": "c913908cfee0f4712e5680140c1e3d5ee86baac1278bf65a7da97f24985f8aa2",
+  "models/single-patch/ccrcc/hvg/config.json": "0c7f5c18b7c5c120c87a3b1f784cae8115b48d6ec711c17e138cd77c42d28b66",
+  "models/single-patch/ccrcc/hvg/genes.json": "fc7c6641ec375172b1e19d671f6b5c9d016b7678b4fc49d5116f81c87dbc1922",
+  "models/single-patch/ccrcc/hvg/model.safetensors": "19a643e778694859ee2e25614c7839a5aa237436d575e0c2d1034501311480fb",
+  "models/single-patch/her2st/hmhvg/config.json": "53c6ada123c0df0add70da1723727122503ee399abc4c1278c9ecd4a0dfaf1b8",
+  "models/single-patch/her2st/hmhvg/genes.json": "280e7e0593a090cce56a3288bddb6faf65be9d7ad24a1e1122e24bd82bc6e77f",
+  "models/single-patch/her2st/hmhvg/model.safetensors": "0022fa86b49afcfaa4c9151d0eaf5c106425dc7cf428c2c19c349455dbdef4dd",
+  "models/single-patch/her2st/hvg/config.json": "9ee562a453368b8417446dff676b746e98462dacbac5f3e283854b771a4db152",
+  "models/single-patch/her2st/hvg/genes.json": "1807423efcd65ee9358dd4a26fe2c3164dbcb80c0141878674e623689023ba7b",
+  "models/single-patch/her2st/hvg/model.safetensors": "254fc5a129245998360cdcac313dd3d1c84b2a3feefbf4ce33689a81509c740d",
+  "models/single-patch/lung/hmhvg/config.json": "42276a98dcd0db5323082add208743170033080beacba623befd680fd0f02bc9",
+  "models/single-patch/lung/hmhvg/genes.json": "8a3a986d63531aeff3abb02d59014533ff3893738961e3d4e15735c9c0bf53ca",
+  "models/single-patch/lung/hmhvg/model.safetensors": "dcbde320a1df51a876a4c0797983d160f5deb3e4bae0f402af6af97b52ccc2f5",
+  "models/single-patch/lung/hvg/config.json": "5b4d6426c97c17e1876d8ef9ba1e09689c423da409d86fc6bd6e39839b65971c",
+  "models/single-patch/lung/hvg/genes.json": "b8661251c651801dad15bda1ceaaad59736790fdb3d110d4adb5f8c3aeff0026",
+  "models/single-patch/lung/hvg/model.safetensors": "f41735851688f4b7368d890c84527a2ba93a620c3df10095029149a5f0d5db6d",
+  "models/single-patch/prad/hmhvg/config.json": "19ddd3b57706c2ea759d1a3066428b35596fec2d19ccfd74498e689681566fc8",
+  "models/single-patch/prad/hmhvg/genes.json": "d08cd995114d1fa0bff9574535425e8f8e1f1dfb27b490ff0ac135ed695c93fa",
+  "models/single-patch/prad/hmhvg/model.safetensors": "ce7d28e32b4a84c0f0cba6f0561b22a997de7e9202c970968e767b2e71257142",
+  "models/single-patch/prad/hvg/config.json": "5900eebd65c4d4a7c55f5a18abaa9cdd51a1a1362ed214b288a3a7527f351409",
+  "models/single-patch/prad/hvg/genes.json": "7ec83111bb5c60fbe0cc4d7a524e63a604bddc377fb99a719f13618680445e2d",
+  "models/single-patch/prad/hvg/model.safetensors": "4f297f40eab72aa2e7706892106ee4428584e7604d0ef91a12272acde8aa60cf",
+  "requirements.txt": "98762691a6e741f153f12f9d4f741f710158d7c38903017e3ec166588f7a5499",
+  "src/__init__.py": "ef6563e43196efa4bff258400d603b6dea99dab3de9960df39872f7b9d17f8ca",
+  "src/pretrained.py": "7e6fd794dd9a88f7ebe7c76a679fc62558398af6e44fcb56198de6b6426c4386",
+  "src/simpleP2G.py": "8d7b01abc5318cfe92142a56a39ce9c1a31f2e9276392e78f9bac8998065cdd0",
+  "src/simpleP2G_neigh.py": "3db1bfade62c5a884011671476ac294f526c8b46a6ad79e2261857a32b374929"
+}

model_index.json

@@ -0,0 +1,276 @@
+{
+  "format_version": 1,
+  "repo_id": "king159/SimpleP2G",
+  "models": [
+    {
+      "setting": "single-patch",
+      "dataset": "brain",
+      "panel": "hmhvg",
+      "directory": "models/single-patch/brain/hmhvg",
+      "fold": 0,
+      "seed": 42,
+      "test_slide": "NCBI667",
+      "validation_slide": "NCBI671",
+      "checkpoint_epoch": 12,
+      "parameters": 45971074,
+      "test_pcc200": 0.3803240847090443,
+      "files": {
+        "model.safetensors": {
+          "sha256": "30e10f1adc24365ff8ae5d3e7851e3336dfdd1775b0a8640ba94fd46630633de",
+          "bytes": 183903592
+        },
+        "config.json": {
+          "sha256": "f5e8ea3c5c366ba9f2678992cbb8d0f77cb5f83c9b65413af6fde5e319bff9c0",
+          "bytes": 1899
+        },
+        "genes.json": {
+          "sha256": "2d5649cf126a6910d2671651d95491e44c23a9c9730d58ff70d5a1dfcafa8814",
+          "bytes": 2214
+        }
+      }
+    },
+    {
+      "setting": "single-patch",
+      "dataset": "brain",
+      "panel": "hvg",
+      "directory": "models/single-patch/brain/hvg",
+      "fold": 0,
+      "seed": 42,
+      "test_slide": "NCBI667",
+      "validation_slide": "NCBI671",
+      "checkpoint_epoch": 16,
+      "parameters": 45971074,
+      "test_pcc200": 0.1689760419646251,
+      "files": {
+        "model.safetensors": {
+          "sha256": "a671becb75830a258455a6136e1be92f5f45ae822c6875290f61edb0f90a9eaa",
+          "bytes": 183903592
+        },
+        "config.json": {
+          "sha256": "4902b72f8d6a95f111c41915ec3a2926f9837db3bb63f7c75ab6cfa64a8e31fc",
+          "bytes": 1904
+        },
+        "genes.json": {
+          "sha256": "9f1b13b2b3be4158e50f8dede0259b880a9e789b30a88612f4da899b22744eb7",
+          "bytes": 2279
+        }
+      }
+    },
+    {
+      "setting": "single-patch",
+      "dataset": "ccrcc",
+      "panel": "hmhvg",
+      "directory": "models/single-patch/ccrcc/hmhvg",
+      "fold": 0,
+      "seed": 42,
+      "test_slide": "INT2",
+      "validation_slide": "INT9",
+      "checkpoint_epoch": 19,
+      "parameters": 45971074,
+      "test_pcc200": 0.27036171397452496,
+      "files": {
+        "model.safetensors": {
+          "sha256": "c913908cfee0f4712e5680140c1e3d5ee86baac1278bf65a7da97f24985f8aa2",
+          "bytes": 183903592
+        },
+        "config.json": {
+          "sha256": "ee9418e61c755b41c390541c2cad5fe7ef86fb2e8164fcbc76e2342f3429d4bc",
+          "bytes": 2008
+        },
+        "genes.json": {
+          "sha256": "ff93267954d744e7e3db035f7d920e281b41cca703413841d6d52b27e1dc4815",
+          "bytes": 2197
+        }
+      }
+    },
+    {
+      "setting": "single-patch",
+      "dataset": "ccrcc",
+      "panel": "hvg",
+      "directory": "models/single-patch/ccrcc/hvg",
+      "fold": 0,
+      "seed": 42,
+      "test_slide": "INT2",
+      "validation_slide": "INT9",
+      "checkpoint_epoch": 7,
+      "parameters": 45971074,
+      "test_pcc200": 0.10978539095367261,
+      "files": {
+        "model.safetensors": {
+          "sha256": "19a643e778694859ee2e25614c7839a5aa237436d575e0c2d1034501311480fb",
+          "bytes": 183903592
+        },
+        "config.json": {
+          "sha256": "0c7f5c18b7c5c120c87a3b1f784cae8115b48d6ec711c17e138cd77c42d28b66",
+          "bytes": 2010
+        },
+        "genes.json": {
+          "sha256": "fc7c6641ec375172b1e19d671f6b5c9d016b7678b4fc49d5116f81c87dbc1922",
+          "bytes": 2188
+        }
+      }
+    },
+    {
+      "setting": "single-patch",
+      "dataset": "her2st",
+      "panel": "hmhvg",
+      "directory": "models/single-patch/her2st/hmhvg",
+      "fold": 0,
+      "seed": 42,
+      "test_slide": "SPA148",
+      "validation_slide": "SPA127",
+      "checkpoint_epoch": 21,
+      "parameters": 45971074,
+      "test_pcc200": 0.70892135019349,
+      "files": {
+        "model.safetensors": {
+          "sha256": "0022fa86b49afcfaa4c9151d0eaf5c106425dc7cf428c2c19c349455dbdef4dd",
+          "bytes": 183903592
+        },
+        "config.json": {
+          "sha256": "53c6ada123c0df0add70da1723727122503ee399abc4c1278c9ecd4a0dfaf1b8",
+          "bytes": 2236
+        },
+        "genes.json": {
+          "sha256": "280e7e0593a090cce56a3288bddb6faf65be9d7ad24a1e1122e24bd82bc6e77f",
+          "bytes": 2197
+        }
+      }
+    },
+    {
+      "setting": "single-patch",
+      "dataset": "her2st",
+      "panel": "hvg",
+      "directory": "models/single-patch/her2st/hvg",
+      "fold": 0,
+      "seed": 42,
+      "test_slide": "SPA148",
+      "validation_slide": "SPA127",
+      "checkpoint_epoch": 37,
+      "parameters": 45971074,
+      "test_pcc200": 0.3798230854871934,
+      "files": {
+        "model.safetensors": {
+          "sha256": "254fc5a129245998360cdcac313dd3d1c84b2a3feefbf4ce33689a81509c740d",
+          "bytes": 183903592
+        },
+        "config.json": {
+          "sha256": "9ee562a453368b8417446dff676b746e98462dacbac5f3e283854b771a4db152",
+          "bytes": 2238
+        },
+        "genes.json": {
+          "sha256": "1807423efcd65ee9358dd4a26fe2c3164dbcb80c0141878674e623689023ba7b",
+          "bytes": 2223
+        }
+      }
+    },
+    {
+      "setting": "single-patch",
+      "dataset": "lung",
+      "panel": "hmhvg",
+      "directory": "models/single-patch/lung/hmhvg",
+      "fold": 0,
+      "seed": 42,
+      "test_slide": "NCBI856",
+      "validation_slide": "NCBI860",
+      "checkpoint_epoch": 7,
+      "parameters": 45971074,
+      "test_pcc200": 0.8717158056511088,
+      "files": {
+        "model.safetensors": {
+          "sha256": "dcbde320a1df51a876a4c0797983d160f5deb3e4bae0f402af6af97b52ccc2f5",
+          "bytes": 183903592
+        },
+        "config.json": {
+          "sha256": "42276a98dcd0db5323082add208743170033080beacba623befd680fd0f02bc9",
+          "bytes": 1997
+        },
+        "genes.json": {
+          "sha256": "8a3a986d63531aeff3abb02d59014533ff3893738961e3d4e15735c9c0bf53ca",
+          "bytes": 2196
+        }
+      }
+    },
+    {
+      "setting": "single-patch",
+      "dataset": "lung",
+      "panel": "hvg",
+      "directory": "models/single-patch/lung/hvg",
+      "fold": 0,
+      "seed": 42,
+      "test_slide": "NCBI856",
+      "validation_slide": "NCBI860",
+      "checkpoint_epoch": 8,
+      "parameters": 45971074,
+      "test_pcc200": 0.4954411997870103,
+      "files": {
+        "model.safetensors": {
+          "sha256": "f41735851688f4b7368d890c84527a2ba93a620c3df10095029149a5f0d5db6d",
+          "bytes": 183903592
+        },
+        "config.json": {
+          "sha256": "5b4d6426c97c17e1876d8ef9ba1e09689c423da409d86fc6bd6e39839b65971c",
+          "bytes": 1996
+        },
+        "genes.json": {
+          "sha256": "b8661251c651801dad15bda1ceaaad59736790fdb3d110d4adb5f8c3aeff0026",
+          "bytes": 4216
+        }
+      }
+    },
+    {
+      "setting": "single-patch",
+      "dataset": "prad",
+      "panel": "hmhvg",
+      "directory": "models/single-patch/prad/hmhvg",
+      "fold": 0,
+      "seed": 42,
+      "test_slide": "MEND145",
+      "validation_slide": "MEND142",
+      "checkpoint_epoch": 9,
+      "parameters": 45971074,
+      "test_pcc200": 0.556121254362732,
+      "files": {
+        "model.safetensors": {
+          "sha256": "ce7d28e32b4a84c0f0cba6f0561b22a997de7e9202c970968e767b2e71257142",
+          "bytes": 183903592
+        },
+        "config.json": {
+          "sha256": "19ddd3b57706c2ea759d1a3066428b35596fec2d19ccfd74498e689681566fc8",
+          "bytes": 2045
+        },
+        "genes.json": {
+          "sha256": "d08cd995114d1fa0bff9574535425e8f8e1f1dfb27b490ff0ac135ed695c93fa",
+          "bytes": 2188
+        }
+      }
+    },
+    {
+      "setting": "single-patch",
+      "dataset": "prad",
+      "panel": "hvg",
+      "directory": "models/single-patch/prad/hvg",
+      "fold": 0,
+      "seed": 42,
+      "test_slide": "MEND145",
+      "validation_slide": "MEND142",
+      "checkpoint_epoch": 12,
+      "parameters": 45971074,
+      "test_pcc200": 0.2199021112938024,
+      "files": {
+        "model.safetensors": {
+          "sha256": "4f297f40eab72aa2e7706892106ee4428584e7604d0ef91a12272acde8aa60cf",
+          "bytes": 183903592
+        },
+        "config.json": {
+          "sha256": "5900eebd65c4d4a7c55f5a18abaa9cdd51a1a1362ed214b288a3a7527f351409",
+          "bytes": 2051
+        },
+        "genes.json": {
+          "sha256": "7ec83111bb5c60fbe0cc4d7a524e63a604bddc377fb99a719f13618680445e2d",
+          "bytes": 2242
+        }
+      }
+    }
+  ]
+}

models/single-patch/brain/hmhvg/config.json

@@ -0,0 +1,87 @@
+{
+  "format_version": 1,
+  "architecture": "SimpleP2G",
+  "setting": "single-patch",
+  "dataset": "brain",
+  "panel": "hmhvg",
+  "fold": 0,
+  "seed": 42,
+  "model_config": {
+    "decoder_depth": 12,
+    "num_genes": 200,
+    "hidden_size": 384,
+    "cond_size": 2560,
+    "num_cond_tokens": 4,
+    "num_heads": 6,
+    "loss_name": "gaussian_nll",
+    "dropout": 0.05
+  },
+  "input": {
+    "representation": "frozen Virchow2 CLS token concatenated with mean patch tokens",
+    "embedding_dimension": 2560,
+    "image_encoder_weights_included": false
+  },
+  "output": {
+    "num_genes": 200,
+    "gene_order": "genes.json",
+    "expression_scale": "log2(raw_count + 1)",
+    "returns": [
+      "expression_mean",
+      "log_variance"
+    ]
+  },
+  "checkpoint": {
+    "epoch": 12,
+    "weights": "ema",
+    "selection": "highest validation-slide PCC-200; earliest epoch breaks ties"
+  },
+  "training": {
+    "epochs": 100,
+    "batch_size": 256,
+    "optimizer": {
+      "name": "adamw",
+      "lr": 5e-05,
+      "weight_decay": 0.05,
+      "betas": [
+        0.9,
+        0.999
+      ]
+    },
+    "loss": "gaussian_nll",
+    "precision": "fp16",
+    "ema_decay": 0.999,
+    "sampling": "pooled spots"
+  },
+  "split": {
+    "training_slides": [
+      "NCBI658",
+      "NCBI659",
+      "NCBI660",
+      "NCBI661",
+      "NCBI662",
+      "NCBI663",
+      "NCBI664",
+      "NCBI665",
+      "NCBI666",
+      "NCBI668",
+      "NCBI669",
+      "NCBI670"
+    ],
+    "validation_slide": "NCBI671",
+    "test_slide": "NCBI667"
+  },
+  "validation_metrics": {
+    "PCC10": 0.6837345136688756,
+    "PCC50": 0.5748279101783605,
+    "PCC200": 0.4576825647353083,
+    "MSE": 0.7553753852844238,
+    "MAE": 0.6992919445037842
+  },
+  "test_metrics": {
+    "PCC10": 0.5845393964585056,
+    "PCC50": 0.5154498406392661,
+    "PCC200": 0.3803240847090443,
+    "MSE": 0.9970911145210266,
+    "MAE": 0.7918735146522522
+  }
+}

models/single-patch/brain/hmhvg/genes.json

@@ -0,0 +1,202 @@
+[
+  "1110008P14Rik",
+  "6330403K07Rik",
+  "Apod",
+  "Apoe",
+  "App",
+  "Arpp19",
+  "Arpp21",
+  "Atp1a1",
+  "Atp1a2",
+  "Atp1b1",
+  "Atp2a2",
+  "Atp2b2",
+  "Atp5b",
+  "Atp5e",
+  "Atp5g1",
+  "Atp5o",
+  "Atp6v1e1",
+  "Baiap2",
+  "Basp1",
+  "Bc1",
+  "Bex2",
+  "Bsg",
+  "Calm2",
+  "Calm3",
+  "Camk2a",
+  "Camk2n1",
+  "Camkv",
+  "Cartpt",
+  "Cck",
+  "Cd81",
+  "Cfl1",
+  "Chgb",
+  "Chn1",
+  "Chst1",
+  "Ckb",
+  "Clstn1",
+  "Cnp",
+  "Cox5a",
+  "Cox7b",
+  "Cox7c",
+  "Cox8a",
+  "Cplx1",
+  "Cplx2",
+  "Cryab",
+  "Cst3",
+  "Ctsd",
+  "Ctxn1",
+  "Dbi",
+  "Dclk1",
+  "Dnm1",
+  "Dpysl2",
+  "Dynll1",
+  "Dynll2",
+  "Eef1a1",
+  "Enc1",
+  "Eno1",
+  "Eno2",
+  "Fis1",
+  "Fkbp1a",
+  "Fkbp8",
+  "Fth1",
+  "Gaa",
+  "Gabra1",
+  "Gad1",
+  "Gdi1",
+  "Gm42418",
+  "Gnao1",
+  "Gnas",
+  "Gng3",
+  "Gpi1",
+  "Gpm6a",
+  "Gpr88",
+  "Gprasp1",
+  "Hba-a1",
+  "Hba-a2",
+  "Hbb-bs",
+  "Hint1",
+  "Hpca",
+  "Hspa8",
+  "Kif1a",
+  "Kif5a",
+  "Lrrc17",
+  "Ly6h",
+  "Maged1",
+  "Malat1",
+  "Mbp",
+  "Mdh2",
+  "Meg3",
+  "Mlf2",
+  "Mobp",
+  "Mrfap1",
+  "Mt1",
+  "Myl6",
+  "Naca",
+  "Nap1l5",
+  "Ncdn",
+  "Ndfip1",
+  "Ndrg2",
+  "Ndufa12",
+  "Ndufa3",
+  "Ndufa4",
+  "Ndufb4",
+  "Ndufb9",
+  "Ndufc1",
+  "Nisch",
+  "Nnat",
+  "Nptxr",
+  "Nrgn",
+  "Nsf",
+  "Oaz1",
+  "Ociad1",
+  "Olfm1",
+  "Pcp4",
+  "Pde1b",
+  "Pea15a",
+  "Penk",
+  "Pfdn5",
+  "Pja2",
+  "Plp1",
+  "Ppp1r1b",
+  "Ppp3ca",
+  "Prkar1b",
+  "Ptgds",
+  "Ptma",
+  "Ptprn",
+  "Pvalb",
+  "Rab3a",
+  "Rac1",
+  "Rpl10",
+  "Rpl13",
+  "Rpl14",
+  "Rpl18",
+  "Rpl18a",
+  "Rpl19",
+  "Rpl22l1",
+  "Rpl23a",
+  "Rpl27",
+  "Rpl32",
+  "Rpl34",
+  "Rpl36a",
+  "Rpl37",
+  "Rpl39",
+  "Rpl4",
+  "Rpl5",
+  "Rpl6",
+  "Rpl7",
+  "Rpl9",
+  "Rps12",
+  "Rps15a",
+  "Rps17",
+  "Rps18",
+  "Rps19",
+  "Rps2",
+  "Rps23",
+  "Rps24",
+  "Rps27a",
+  "Rps3",
+  "Rps4x",
+  "Rps6",
+  "Rps7",
+  "Rtn4",
+  "Scd2",
+  "Scn1b",
+  "Selenow",
+  "Serf2",
+  "Sez6l2",
+  "Slc17a7",
+  "Slc1a2",
+  "Slc22a17",
+  "Slc25a5",
+  "Snap25",
+  "Snap47",
+  "Snca",
+  "Snrpn",
+  "Sod1",
+  "Sparcl1",
+  "Sst",
+  "Stmn1",
+  "Stmn3",
+  "Stmn4",
+  "Sub1",
+  "Syn1",
+  "Syn2",
+  "Syngr1",
+  "Syt11",
+  "Tmsb10",
+  "Tmsb4x",
+  "Tpi1",
+  "Trf",
+  "Tuba1a",
+  "Tubb2a",
+  "Tubb4a",
+  "Tubb5",
+  "Ubl5",
+  "Uchl1",
+  "Uqcrh",
+  "Vsnl1",
+  "Wbp2",
+  "Ywhag",
+  "Ywhah"
+]

models/single-patch/brain/hmhvg/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30e10f1adc24365ff8ae5d3e7851e3336dfdd1775b0a8640ba94fd46630633de
+size 183903592

models/single-patch/brain/hvg/config.json

@@ -0,0 +1,87 @@
+{
+  "format_version": 1,
+  "architecture": "SimpleP2G",
+  "setting": "single-patch",
+  "dataset": "brain",
+  "panel": "hvg",
+  "fold": 0,
+  "seed": 42,
+  "model_config": {
+    "decoder_depth": 12,
+    "num_genes": 200,
+    "hidden_size": 384,
+    "cond_size": 2560,
+    "num_cond_tokens": 4,
+    "num_heads": 6,
+    "loss_name": "gaussian_nll",
+    "dropout": 0.05
+  },
+  "input": {
+    "representation": "frozen Virchow2 CLS token concatenated with mean patch tokens",
+    "embedding_dimension": 2560,
+    "image_encoder_weights_included": false
+  },
+  "output": {
+    "num_genes": 200,
+    "gene_order": "genes.json",
+    "expression_scale": "log2(raw_count + 1)",
+    "returns": [
+      "expression_mean",
+      "log_variance"
+    ]
+  },
+  "checkpoint": {
+    "epoch": 16,
+    "weights": "ema",
+    "selection": "highest validation-slide PCC-200; earliest epoch breaks ties"
+  },
+  "training": {
+    "epochs": 100,
+    "batch_size": 256,
+    "optimizer": {
+      "name": "adamw",
+      "lr": 5e-05,
+      "weight_decay": 0.05,
+      "betas": [
+        0.9,
+        0.999
+      ]
+    },
+    "loss": "gaussian_nll",
+    "precision": "fp16",
+    "ema_decay": 0.999,
+    "sampling": "pooled spots"
+  },
+  "split": {
+    "training_slides": [
+      "NCBI658",
+      "NCBI659",
+      "NCBI660",
+      "NCBI661",
+      "NCBI662",
+      "NCBI663",
+      "NCBI664",
+      "NCBI665",
+      "NCBI666",
+      "NCBI668",
+      "NCBI669",
+      "NCBI670"
+    ],
+    "validation_slide": "NCBI671",
+    "test_slide": "NCBI667"
+  },
+  "validation_metrics": {
+    "PCC10": 0.5519836169944861,
+    "PCC50": 0.41539179241866847,
+    "PCC200": 0.21614730186065806,
+    "MSE": 0.20960445702075958,
+    "MAE": 0.2335881143808365
+  },
+  "test_metrics": {
+    "PCC10": 0.47346420145244456,
+    "PCC50": 0.34249158576801364,
+    "PCC200": 0.1689760419646251,
+    "MSE": 0.22083990275859833,
+    "MAE": 0.24713166058063507
+  }
+}

models/single-patch/brain/hvg/genes.json

@@ -0,0 +1,202 @@
+[
+  "1700052K11Rik",
+  "1810010H24Rik",
+  "2810468N07Rik",
+  "4932438A13Rik",
+  "9530059O14Rik",
+  "A930004D18Rik",
+  "Ackr1",
+  "Acta2",
+  "Aebp1",
+  "Ahnak",
+  "Aldh1a1",
+  "Anapc16",
+  "Anln",
+  "Apod",
+  "Arid4a",
+  "As3mt",
+  "Ascc2",
+  "Ascl1",
+  "Atg4c",
+  "B930036N10Rik",
+  "Bbs2",
+  "Bcas1",
+  "Bfsp2",
+  "C1qc",
+  "C1ql1",
+  "C3",
+  "C4b",
+  "C78859",
+  "Cadm1",
+  "Calb2",
+  "Car2",
+  "Card19",
+  "Cartpt",
+  "Ccdc153",
+  "Ccl24",
+  "Cd2ap",
+  "Cd74",
+  "Cd9",
+  "Cdr1os",
+  "Celsr3",
+  "Cldn11",
+  "Cnp",
+  "Coro1b",
+  "Cort",
+  "Cplx3",
+  "Cryab",
+  "Cryl1",
+  "Csrp1",
+  "Dbndd2",
+  "Dexi",
+  "Enpp2",
+  "Enpp6",
+  "Eps8l2",
+  "Ermn",
+  "Fa2h",
+  "Fabp3",
+  "Fam53b",
+  "Fcgr2b",
+  "Fkbp9",
+  "Fnbp1",
+  "Foxc1",
+  "Fuca1",
+  "Fxyd1",
+  "Gal3st1",
+  "Gask1b",
+  "Gatm",
+  "Gfap",
+  "Gjc2",
+  "Gjc3",
+  "Gm16323",
+  "Gm20594",
+  "Gm30906",
+  "Gm33195",
+  "Gnrh1",
+  "Gpr37",
+  "Gpx3",
+  "Grb14",
+  "Gsn",
+  "Hba-a1",
+  "Hba-a2",
+  "Hbb-bs",
+  "Hbb-bt",
+  "Iars",
+  "Igf2",
+  "Il17rc",
+  "Kap",
+  "Kcna4",
+  "Kcnmb4",
+  "Kif21b",
+  "Kif7",
+  "Larp4",
+  "Lars2",
+  "Lcn2",
+  "Litaf",
+  "Lpar1",
+  "Lyz2",
+  "Lzic",
+  "Mag",
+  "Mal",
+  "Map6d1",
+  "Mdm2",
+  "Metrn",
+  "Mfsd14b",
+  "Mgp",
+  "Mobp",
+  "Mog",
+  "Myoc",
+  "N4bp2l1",
+  "Nab1",
+  "Ndrg1",
+  "Neat1",
+  "Nif3l1",
+  "Nit1",
+  "Nnat",
+  "Npy",
+  "Nr3c1",
+  "Nyap1",
+  "Opalin",
+  "Ostf1",
+  "Pacs2",
+  "Padi2",
+  "Pcdhga7",
+  "Pcsk6",
+  "Pdlim2",
+  "Pgd",
+  "Phf6",
+  "Phldb1",
+  "Pisd",
+  "Plekhb1",
+  "Pllp",
+  "Plp1",
+  "Poglut1",
+  "Ppard",
+  "Ppp1r14a",
+  "Prkcd",
+  "Prr18",
+  "Ptgds",
+  "Pth2",
+  "Pvalb",
+  "Pxn",
+  "Rab7b",
+  "Rad1",
+  "Rarres2",
+  "Rhog",
+  "Rnf122",
+  "Rpap3",
+  "Rps19bp1",
+  "S100a6",
+  "S100a8",
+  "S100a9",
+  "Sac3d1",
+  "Samd1",
+  "Scgb1a1",
+  "Sepsecs",
+  "Sept4",
+  "Sgcb",
+  "Shtn1",
+  "Sirt4",
+  "Slain1",
+  "Slc17a7",
+  "Slc44a1",
+  "Slc6a13",
+  "Slc6a3",
+  "Smg6",
+  "Smim1",
+  "Sox10",
+  "Sox2ot",
+  "Sspo",
+  "Sst",
+  "Steap4",
+  "Tap1",
+  "Tgif1",
+  "Th",
+  "Thbs4",
+  "Timp1",
+  "Tm7sf3",
+  "Tmbim1",
+  "Tmem178b",
+  "Tmem189",
+  "Tmem212",
+  "Tmem63a",
+  "Tmem87b",
+  "Tmem88b",
+  "Tppp3",
+  "Trf",
+  "Trim12c",
+  "Trim59",
+  "Tspan2",
+  "Tyrobp",
+  "Ucn",
+  "Ugt8a",
+  "Usp4",
+  "Vim",
+  "Vip",
+  "Vkorc1",
+  "Vmac",
+  "Vtn",
+  "Zfp101",
+  "Zfp606",
+  "Zfp950"
+]

models/single-patch/brain/hvg/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a671becb75830a258455a6136e1be92f5f45ae822c6875290f61edb0f90a9eaa
+size 183903592

models/single-patch/ccrcc/hmhvg/config.json

@@ -0,0 +1,97 @@
+{
+  "format_version": 1,
+  "architecture": "SimpleP2G",
+  "setting": "single-patch",
+  "dataset": "ccrcc",
+  "panel": "hmhvg",
+  "fold": 0,
+  "seed": 42,
+  "model_config": {
+    "decoder_depth": 12,
+    "num_genes": 200,
+    "hidden_size": 384,
+    "cond_size": 2560,
+    "num_cond_tokens": 4,
+    "num_heads": 6,
+    "loss_name": "gaussian_nll",
+    "dropout": 0.1
+  },
+  "input": {
+    "representation": "frozen Virchow2 CLS token concatenated with mean patch tokens",
+    "embedding_dimension": 2560,
+    "image_encoder_weights_included": false
+  },
+  "output": {
+    "num_genes": 200,
+    "gene_order": "genes.json",
+    "expression_scale": "log2(raw_count + 1)",
+    "returns": [
+      "expression_mean",
+      "log_variance"
+    ]
+  },
+  "checkpoint": {
+    "epoch": 19,
+    "weights": "ema",
+    "selection": "highest validation-slide PCC-200; earliest epoch breaks ties"
+  },
+  "training": {
+    "epochs": 100,
+    "batch_size": 256,
+    "optimizer": {
+      "name": "adamw",
+      "lr": 3e-05,
+      "weight_decay": 0.1,
+      "betas": [
+        0.9,
+        0.999
+      ]
+    },
+    "loss": "gaussian_nll",
+    "precision": "fp16",
+    "ema_decay": 0.999,
+    "sampling": "pooled spots"
+  },
+  "split": {
+    "training_slides": [
+      "INT1",
+      "INT10",
+      "INT11",
+      "INT12",
+      "INT13",
+      "INT14",
+      "INT15",
+      "INT16",
+      "INT17",
+      "INT18",
+      "INT19",
+      "INT20",
+      "INT21",
+      "INT22",
+      "INT23",
+      "INT24",
+      "INT3",
+      "INT4",
+      "INT5",
+      "INT6",
+      "INT7",
+      "INT8"
+    ],
+    "validation_slide": "INT9",
+    "test_slide": "INT2"
+  },
+  "validation_metrics": {
+    "PCC10": 0.6926724980557916,
+    "PCC50": 0.6391818018883983,
+    "PCC200": 0.42965969649900465,
+    "MSE": 2.185840368270874,
+    "MAE": 1.17231023311615
+  },
+  "test_metrics": {
+    "PCC10": 0.502469291003792,
+    "PCC50": 0.4121004550758145,
+    "PCC200": 0.27036171397452496,
+    "MSE": 1.6523619890213013,
+    "MAE": 0.9979413747787476
+  }
+}

models/single-patch/ccrcc/hmhvg/genes.json

@@ -0,0 +1,202 @@
+[
+  "A2M",
+  "ACTA2",
+  "ACTB",
+  "ADIRF",
+  "AEBP1",
+  "ANGPTL4",
+  "ANPEP",
+  "ANXA2",
+  "APOC1",
+  "APOE",
+  "APOL1",
+  "APP",
+  "ARF1",
+  "ARPC1B",
+  "ARPC2",
+  "ATP1A1",
+  "ATP1B1",
+  "ATP5F1B",
+  "ATP5MC2",
+  "ATP5ME",
+  "B2M",
+  "BGN",
+  "BIRC3",
+  "BSG",
+  "BST2",
+  "C19orf33",
+  "C1QA",
+  "C1QB",
+  "C1QC",
+  "C1R",
+  "C1S",
+  "C3",
+  "CA12",
+  "CALD1",
+  "CANX",
+  "CCDC91",
+  "CCN2",
+  "CCNI",
+  "CD24",
+  "CD44",
+  "CD63",
+  "CD74",
+  "CD81",
+  "CD99",
+  "CEBPD",
+  "CHCHD2",
+  "CIRBP",
+  "CLU",
+  "COL18A1",
+  "COL1A1",
+  "COL1A2",
+  "COL3A1",
+  "COL4A1",
+  "COL4A2",
+  "COL6A1",
+  "COL6A2",
+  "COL6A3",
+  "CP",
+  "CPE",
+  "CRYAB",
+  "CST3",
+  "CSTB",
+  "CTSB",
+  "CTSD",
+  "CTSZ",
+  "CYB5A",
+  "CYB5R3",
+  "DCN",
+  "DDIT4",
+  "DDX17",
+  "DEPP1",
+  "DUSP1",
+  "EEF1G",
+  "EIF1",
+  "EIF4G2",
+  "EIF4H",
+  "ENPP3",
+  "FABP7",
+  "FGB",
+  "FKBP5",
+  "FLNA",
+  "FN1",
+  "FOS",
+  "FTH1",
+  "FTL",
+  "FXYD2",
+  "GABARAP",
+  "GLUL",
+  "GPX3",
+  "GSN",
+  "GSTP1",
+  "HINT1",
+  "HIST1H4C",
+  "HLA-A",
+  "HLA-DPA1",
+  "HLA-DPB1",
+  "HLA-DQB1",
+  "HLA-DRA",
+  "HLA-DRB1",
+  "HNRNPA2B1",
+  "HNRNPA3",
+  "HSP90AA1",
+  "HSP90B1",
+  "HSPA8",
+  "HSPD1",
+  "HTRA1",
+  "IFI27",
+  "IFI30",
+  "IFI6",
+  "IFITM2",
+  "IGFBP3",
+  "IGFBP4",
+  "IGFBP5",
+  "IGFBP7",
+  "IGHA1",
+  "IGHG1",
+  "IGHG2",
+  "IGHG3",
+  "IGHG4",
+  "IGHM",
+  "IGKC",
+  "IGLC1",
+  "IL32",
+  "ITGA3",
+  "ITGB1",
+  "JCHAIN",
+  "KRT8",
+  "LAPTM5",
+  "LGALS1",
+  "LGALS3BP",
+  "LY6E",
+  "LYZ",
+  "MGP",
+  "MIF",
+  "MMP7",
+  "MMP9",
+  "MYH9",
+  "MYL9",
+  "NDRG1",
+  "NDUFA4L2",
+  "NME2",
+  "NPC2",
+  "NUPR1",
+  "P4HB",
+  "PARK7",
+  "PCBP1",
+  "PCBP2",
+  "PDIA6",
+  "PDK4",
+  "PDZK1IP1",
+  "PFDN5",
+  "PFKP",
+  "PGAM1",
+  "PGF",
+  "PLEC",
+  "PLIN2",
+  "PLOD2",
+  "PPDPF",
+  "PRDX6",
+  "PRR13",
+  "PTMA",
+  "PTMS",
+  "PTTG1IP",
+  "RARRES2",
+  "RGS5",
+  "RHOB",
+  "RNASET2",
+  "S100A11",
+  "S100A6",
+  "SAT1",
+  "SCD",
+  "SERPINA1",
+  "SERPINE1",
+  "SERPING1",
+  "SOD2",
+  "SPARC",
+  "SPP1",
+  "SQSTM1",
+  "SRRM2",
+  "SSR4",
+  "TAGLN",
+  "TAGLN2",
+  "TGFBI",
+  "TGM2",
+  "TIMP1",
+  "TIMP3",
+  "TMBIM6",
+  "TMSB4X",
+  "TOMM7",
+  "TPM1",
+  "TPM2",
+  "TSC22D3",
+  "TUBA1B",
+  "TXNIP",
+  "TYROBP",
+  "UBC",
+  "UQCRQ",
+  "VEGFA",
+  "VIM",
+  "VWF"
+]

models/single-patch/ccrcc/hmhvg/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c913908cfee0f4712e5680140c1e3d5ee86baac1278bf65a7da97f24985f8aa2
+size 183903592

models/single-patch/ccrcc/hvg/config.json

@@ -0,0 +1,97 @@
+{
+  "format_version": 1,
+  "architecture": "SimpleP2G",
+  "setting": "single-patch",
+  "dataset": "ccrcc",
+  "panel": "hvg",
+  "fold": 0,
+  "seed": 42,
+  "model_config": {
+    "decoder_depth": 12,
+    "num_genes": 200,
+    "hidden_size": 384,
+    "cond_size": 2560,
+    "num_cond_tokens": 4,
+    "num_heads": 6,
+    "loss_name": "gaussian_nll",
+    "dropout": 0.1
+  },
+  "input": {
+    "representation": "frozen Virchow2 CLS token concatenated with mean patch tokens",
+    "embedding_dimension": 2560,
+    "image_encoder_weights_included": false
+  },
+  "output": {
+    "num_genes": 200,
+    "gene_order": "genes.json",
+    "expression_scale": "log2(raw_count + 1)",
+    "returns": [
+      "expression_mean",
+      "log_variance"
+    ]
+  },
+  "checkpoint": {
+    "epoch": 7,
+    "weights": "ema",
+    "selection": "highest validation-slide PCC-200; earliest epoch breaks ties"
+  },
+  "training": {
+    "epochs": 100,
+    "batch_size": 256,
+    "optimizer": {
+      "name": "adamw",
+      "lr": 3e-05,
+      "weight_decay": 0.1,
+      "betas": [
+        0.9,
+        0.999
+      ]
+    },
+    "loss": "gaussian_nll",
+    "precision": "fp16",
+    "ema_decay": 0.999,
+    "sampling": "pooled spots"
+  },
+  "split": {
+    "training_slides": [
+      "INT1",
+      "INT10",
+      "INT11",
+      "INT12",
+      "INT13",
+      "INT14",
+      "INT15",
+      "INT16",
+      "INT17",
+      "INT18",
+      "INT19",
+      "INT20",
+      "INT21",
+      "INT22",
+      "INT23",
+      "INT24",
+      "INT3",
+      "INT4",
+      "INT5",
+      "INT6",
+      "INT7",
+      "INT8"
+    ],
+    "validation_slide": "INT9",
+    "test_slide": "INT2"
+  },
+  "validation_metrics": {
+    "PCC10": 0.5082438063563466,
+    "PCC50": 0.2864946574885263,
+    "PCC200": 0.11651308085853335,
+    "MSE": 0.5486487746238708,
+    "MAE": 0.4507647454738617
+  },
+  "test_metrics": {
+    "PCC10": 0.3779823666623128,
+    "PCC50": 0.24105083678052494,
+    "PCC200": 0.10978539095367261,
+    "MSE": 0.5488066077232361,
+    "MAE": 0.4156813621520996
+  }
+}

models/single-patch/ccrcc/hvg/genes.json

@@ -0,0 +1,202 @@
+[
+  "ACKR1",
+  "ACTA2",
+  "ACTC1",
+  "ACTG2",
+  "AEBP1",
+  "AIF1L",
+  "ALDOB",
+  "ANKRD1",
+  "APOD",
+  "AQP1",
+  "ARPC3",
+  "BANK1",
+  "BLK",
+  "BRIX1",
+  "C11orf96",
+  "C1orf198",
+  "C7",
+  "CAMP",
+  "CBL",
+  "CCDC80",
+  "CCDC91",
+  "CCL14",
+  "CCL18",
+  "CCL19",
+  "CCL2",
+  "CCL21",
+  "CCL5",
+  "CCL8",
+  "CCN2",
+  "CCN3",
+  "CD209",
+  "CD79A",
+  "CD8A",
+  "CFAP97D1",
+  "CHIT1",
+  "CLDN5",
+  "CLIC6",
+  "CLU",
+  "CNN1",
+  "COL1A1",
+  "COL1A2",
+  "CR2",
+  "CRIP1",
+  "CRYBG3",
+  "CTSK",
+  "CTSS",
+  "CXCL10",
+  "CXCL11",
+  "CXCL12",
+  "CXCL13",
+  "CXCL8",
+  "CXCL9",
+  "DCN",
+  "DERL3",
+  "DES",
+  "DGKA",
+  "DSG4",
+  "ELMO1",
+  "ELN",
+  "ENPP2",
+  "F13A1",
+  "FABP7",
+  "FAM102A",
+  "FBLN1",
+  "FBLN2",
+  "FBLN5",
+  "FBXW7",
+  "FCRL1",
+  "FDCSP",
+  "FN1",
+  "FOS",
+  "FOSB",
+  "FRZB",
+  "G0S2",
+  "GART",
+  "GJA9",
+  "GNLY",
+  "GPNMB",
+  "GREM1",
+  "GUCY2D",
+  "GZMB",
+  "HBA1",
+  "HBA2",
+  "HP",
+  "HSPA1A",
+  "HSPA1B",
+  "HSPA6",
+  "HSPB6",
+  "IGF1",
+  "IGHA1",
+  "IGHA2",
+  "IGHD",
+  "IGHG1",
+  "IGHG2",
+  "IGHG3",
+  "IGHG4",
+  "IGHJ2",
+  "IGHJ6",
+  "IGHM",
+  "IGHV1-46",
+  "IGHV5-10-1",
+  "IGHV6-1",
+  "IGKC",
+  "IGKV4-1",
+  "IGKV5-2",
+  "IGLC1",
+  "IGLC7",
+  "IGLV3-1",
+  "IGLV6-57",
+  "IGLV9-49",
+  "IL2RG",
+  "IL7R",
+  "INHBA",
+  "INMT",
+  "ISG15",
+  "ITGB6",
+  "JCHAIN",
+  "KCNK6",
+  "KLK10",
+  "KRT7",
+  "LCN2",
+  "LCP1",
+  "LHFPL6",
+  "LMOD1",
+  "LTB",
+  "LTBP2",
+  "LTF",
+  "LUM",
+  "MFAP4",
+  "MFGE8",
+  "MFSD11",
+  "MGP",
+  "MMP11",
+  "MMP2",
+  "MMP7",
+  "MMP9",
+  "MT1G",
+  "MT1H",
+  "MT1X",
+  "MTRNR2L8",
+  "MUSTN1",
+  "MYH10",
+  "MYLK",
+  "MYLPF",
+  "MZB1",
+  "NDUFAF4",
+  "NR4A1",
+  "PCDH7",
+  "PIGR",
+  "PIM2",
+  "PIP5K1C",
+  "PLN",
+  "POSTN",
+  "POU2AF1",
+  "PPP1R14A",
+  "PRELP",
+  "PRKCB",
+  "PSMC3",
+  "PTCHD4",
+  "PTGDS",
+  "PTGIS",
+  "PTPRC",
+  "REN",
+  "RGCC",
+  "RNASE1",
+  "S100A2",
+  "S100A8",
+  "S100A9",
+  "SAA1",
+  "SAA2",
+  "SETDB1",
+  "SFRP1",
+  "SFRP2",
+  "SFRP4",
+  "SLC11A1",
+  "SLPI",
+  "SLTM",
+  "SOCS3",
+  "SOD3",
+  "SORBS1",
+  "SPP1",
+  "SULF1",
+  "SVEP1",
+  "SYNPO2",
+  "TACSTD2",
+  "TAGLN",
+  "TENT5C",
+  "THBS1",
+  "THBS2",
+  "THBS4",
+  "TIMP3",
+  "TNC",
+  "TNXB",
+  "TPM2",
+  "TPSB2",
+  "TRBC2",
+  "TRPM2",
+  "TTC27",
+  "VWF",
+  "YEATS2"
+]

models/single-patch/ccrcc/hvg/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19a643e778694859ee2e25614c7839a5aa237436d575e0c2d1034501311480fb
+size 183903592

models/single-patch/her2st/hmhvg/config.json

@@ -0,0 +1,109 @@
+{
+  "format_version": 1,
+  "architecture": "SimpleP2G",
+  "setting": "single-patch",
+  "dataset": "her2st",
+  "panel": "hmhvg",
+  "fold": 0,
+  "seed": 42,
+  "model_config": {
+    "decoder_depth": 12,
+    "num_genes": 200,
+    "hidden_size": 384,
+    "cond_size": 2560,
+    "num_cond_tokens": 4,
+    "num_heads": 6,
+    "loss_name": "gaussian_nll",
+    "dropout": 0.0
+  },
+  "input": {
+    "representation": "frozen Virchow2 CLS token concatenated with mean patch tokens",
+    "embedding_dimension": 2560,
+    "image_encoder_weights_included": false
+  },
+  "output": {
+    "num_genes": 200,
+    "gene_order": "genes.json",
+    "expression_scale": "log2(raw_count + 1)",
+    "returns": [
+      "expression_mean",
+      "log_variance"
+    ]
+  },
+  "checkpoint": {
+    "epoch": 21,
+    "weights": "ema",
+    "selection": "highest validation-slide PCC-200; earliest epoch breaks ties"
+  },
+  "training": {
+    "epochs": 100,
+    "batch_size": 256,
+    "optimizer": {
+      "name": "adamw",
+      "lr": 0.0001,
+      "weight_decay": 0.05,
+      "betas": [
+        0.9,
+        0.999
+      ]
+    },
+    "loss": "gaussian_nll",
+    "precision": "fp16",
+    "ema_decay": 0.999,
+    "sampling": "pooled spots"
+  },
+  "split": {
+    "training_slides": [
+      "SPA119",
+      "SPA120",
+      "SPA121",
+      "SPA122",
+      "SPA123",
+      "SPA124",
+      "SPA125",
+      "SPA126",
+      "SPA128",
+      "SPA129",
+      "SPA130",
+      "SPA131",
+      "SPA132",
+      "SPA133",
+      "SPA134",
+      "SPA135",
+      "SPA136",
+      "SPA137",
+      "SPA138",
+      "SPA139",
+      "SPA140",
+      "SPA141",
+      "SPA142",
+      "SPA143",
+      "SPA144",
+      "SPA145",
+      "SPA146",
+      "SPA147",
+      "SPA149",
+      "SPA150",
+      "SPA151",
+      "SPA152",
+      "SPA153",
+      "SPA154"
+    ],
+    "validation_slide": "SPA127",
+    "test_slide": "SPA148"
+  },
+  "validation_metrics": {
+    "PCC10": 0.456390537005097,
+    "PCC50": 0.30044265527694825,
+    "PCC200": 0.15609376481639806,
+    "MSE": 0.7690321803092957,
+    "MAE": 0.6974937915802002
+  },
+  "test_metrics": {
+    "PCC10": 0.8710780333678818,
+    "PCC50": 0.822812513630561,
+    "PCC200": 0.70892135019349,
+    "MSE": 0.8401367664337158,
+    "MAE": 0.7023175954818726
+  }
+}

models/single-patch/her2st/hmhvg/genes.json

@@ -0,0 +1,202 @@
+[
+  "A2M",
+  "ACTB",
+  "ACTG1",
+  "ACTN4",
+  "ADAM15",
+  "AEBP1",
+  "AES",
+  "ALDOA",
+  "AP000769.1",
+  "APOC1",
+  "APOE",
+  "ARHGDIA",
+  "ATG10",
+  "ATP5B",
+  "ATP5E",
+  "ATP6V0B",
+  "AZGP1",
+  "B2M",
+  "BEST1",
+  "BGN",
+  "BSG",
+  "BST2",
+  "C12orf57",
+  "C1QA",
+  "C3",
+  "CALM2",
+  "CALML5",
+  "CALR",
+  "CCT3",
+  "CD24",
+  "CD63",
+  "CD74",
+  "CFL1",
+  "CHCHD2",
+  "CHPF",
+  "CIB1",
+  "CLDN3",
+  "CLDN4",
+  "CNN3",
+  "COL18A1",
+  "COL1A1",
+  "COL1A2",
+  "COL3A1",
+  "COL6A2",
+  "COMP",
+  "COPE",
+  "COPS9",
+  "COX4I1",
+  "COX5B",
+  "COX6B1",
+  "COX6C",
+  "COX7C",
+  "CRIP2",
+  "CST3",
+  "CTSB",
+  "CTSD",
+  "CYBA",
+  "DDX5",
+  "DHCR24",
+  "EDF1",
+  "EEF1D",
+  "EEF2",
+  "EIF4G1",
+  "ELOVL1",
+  "ENO1",
+  "ERBB2",
+  "FASN",
+  "FAU",
+  "FLNA",
+  "FN1",
+  "FNBP1L",
+  "FTH1",
+  "FTL",
+  "GAPDH",
+  "GNAS",
+  "GPX4",
+  "GRB7",
+  "GRINA",
+  "GUK1",
+  "H2AFJ",
+  "HLA-A",
+  "HLA-B",
+  "HLA-C",
+  "HLA-DRA",
+  "HLA-E",
+  "HNRNPA2B1",
+  "HSP90AA1",
+  "HSP90AB1",
+  "HSP90B1",
+  "HSPA8",
+  "HSPB1",
+  "IDH2",
+  "IFI27",
+  "IGFBP2",
+  "IGHA1",
+  "IGHG1",
+  "IGHG3",
+  "IGHG4",
+  "IGHM",
+  "IGKC",
+  "IGLC2",
+  "IGLC3",
+  "INTS1",
+  "ISG15",
+  "JTB",
+  "KDELR1",
+  "KRT18",
+  "KRT19",
+  "KRT7",
+  "KRT81",
+  "LAPTM4A",
+  "LAPTM5",
+  "LASP1",
+  "LGALS1",
+  "LGALS3",
+  "LGALS3BP",
+  "LMAN2",
+  "LMNA",
+  "LUM",
+  "LY6E",
+  "MAPKAPK2",
+  "MDK",
+  "MGP",
+  "MIDN",
+  "MIEN1",
+  "MLLT6",
+  "MMACHC",
+  "MMP14",
+  "MUC1",
+  "MUCL1",
+  "MYL6",
+  "MYL9",
+  "MZT2B",
+  "NBL1",
+  "NDUFB9",
+  "NUCKS1",
+  "NUPR1",
+  "ORMDL3",
+  "P4HB",
+  "PCGF2",
+  "PEBP1",
+  "PERP",
+  "PFDN5",
+  "PGAP3",
+  "PHB",
+  "PIP4K2B",
+  "PLD3",
+  "PNMT",
+  "POSTN",
+  "PPDPF",
+  "PPP1CA",
+  "PPP1R1B",
+  "PRDX1",
+  "PRRC2A",
+  "PRSS8",
+  "PSMB3",
+  "PSMB4",
+  "PSMD3",
+  "PTMA",
+  "PTMS",
+  "PTPRF",
+  "RACK1",
+  "S100A14",
+  "S100A6",
+  "S100A8",
+  "S100A9",
+  "SCAND1",
+  "SCD",
+  "SDC1",
+  "SEC61A1",
+  "SEPW1",
+  "SERF2",
+  "SF3B5",
+  "SH3BGRL3",
+  "SLC9A3R1",
+  "SNRPB",
+  "SPARC",
+  "SPDEF",
+  "SPINT2",
+  "SSR2",
+  "SSR4",
+  "STARD10",
+  "STARD3",
+  "SUPT6H",
+  "SYNGR2",
+  "TAGLN",
+  "TAPBP",
+  "TFF3",
+  "TIMP1",
+  "TMSB10",
+  "TPT1",
+  "TSPO",
+  "TUBB",
+  "TXNIP",
+  "UBA52",
+  "UBC",
+  "UBE2M",
+  "UQCRQ",
+  "VIM",
+  "ZYX"
+]

models/single-patch/her2st/hmhvg/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0022fa86b49afcfaa4c9151d0eaf5c106425dc7cf428c2c19c349455dbdef4dd
+size 183903592

models/single-patch/her2st/hvg/config.json

@@ -0,0 +1,109 @@
+{
+  "format_version": 1,
+  "architecture": "SimpleP2G",
+  "setting": "single-patch",
+  "dataset": "her2st",
+  "panel": "hvg",
+  "fold": 0,
+  "seed": 42,
+  "model_config": {
+    "decoder_depth": 12,
+    "num_genes": 200,
+    "hidden_size": 384,
+    "cond_size": 2560,
+    "num_cond_tokens": 4,
+    "num_heads": 6,
+    "loss_name": "gaussian_nll",
+    "dropout": 0.0
+  },
+  "input": {
+    "representation": "frozen Virchow2 CLS token concatenated with mean patch tokens",
+    "embedding_dimension": 2560,
+    "image_encoder_weights_included": false
+  },
+  "output": {
+    "num_genes": 200,
+    "gene_order": "genes.json",
+    "expression_scale": "log2(raw_count + 1)",
+    "returns": [
+      "expression_mean",
+      "log_variance"
+    ]
+  },
+  "checkpoint": {
+    "epoch": 37,
+    "weights": "ema",
+    "selection": "highest validation-slide PCC-200; earliest epoch breaks ties"
+  },
+  "training": {
+    "epochs": 100,
+    "batch_size": 256,
+    "optimizer": {
+      "name": "adamw",
+      "lr": 0.0001,
+      "weight_decay": 0.05,
+      "betas": [
+        0.9,
+        0.999
+      ]
+    },
+    "loss": "gaussian_nll",
+    "precision": "fp16",
+    "ema_decay": 0.999,
+    "sampling": "pooled spots"
+  },
+  "split": {
+    "training_slides": [
+      "SPA119",
+      "SPA120",
+      "SPA121",
+      "SPA122",
+      "SPA123",
+      "SPA124",
+      "SPA125",
+      "SPA126",
+      "SPA128",
+      "SPA129",
+      "SPA130",
+      "SPA131",
+      "SPA132",
+      "SPA133",
+      "SPA134",
+      "SPA135",
+      "SPA136",
+      "SPA137",
+      "SPA138",
+      "SPA139",
+      "SPA140",
+      "SPA141",
+      "SPA142",
+      "SPA143",
+      "SPA144",
+      "SPA145",
+      "SPA146",
+      "SPA147",
+      "SPA149",
+      "SPA150",
+      "SPA151",
+      "SPA152",
+      "SPA153",
+      "SPA154"
+    ],
+    "validation_slide": "SPA127",
+    "test_slide": "SPA148"
+  },
+  "validation_metrics": {
+    "PCC10": 0.43139689241917323,
+    "PCC50": 0.20696065692979265,
+    "PCC200": 0.0662583081649352,
+    "MSE": 0.2628522515296936,
+    "MAE": 0.31125587224960327
+  },
+  "test_metrics": {
+    "PCC10": 0.751352123780151,
+    "PCC50": 0.6407706227872064,
+    "PCC200": 0.3798230854871934,
+    "MSE": 0.3085944652557373,
+    "MAE": 0.2869880795478821
+  }
+}

models/single-patch/her2st/hvg/genes.json

@@ -0,0 +1,202 @@
+[
+  "A2M",
+  "ACKR1",
+  "ADIRF",
+  "AGR2",
+  "AKAP9",
+  "AMDHD2",
+  "ANKRD36",
+  "ANKRD36C",
+  "APOC1",
+  "APOD",
+  "ARID4A",
+  "ARMC9",
+  "ATP5D",
+  "BAG6",
+  "BRCA1",
+  "BRD4",
+  "C3",
+  "C4orf48",
+  "CACNA2D1",
+  "CCDC34",
+  "CCL19",
+  "CCL21",
+  "CD3D",
+  "CD7",
+  "CDK5RAP3",
+  "CDR1",
+  "CEP135",
+  "CEP162",
+  "CEP290",
+  "CFD",
+  "CHCHD6",
+  "CLDN5",
+  "CPSF3",
+  "CRABP2",
+  "CST6",
+  "CTNNBL1",
+  "CWF19L2",
+  "CX3CL1",
+  "CXCL14",
+  "CXCL9",
+  "DDX60L",
+  "DECR1",
+  "DNAJC15",
+  "DNAJC2",
+  "DNTTIP2",
+  "DOCK11",
+  "DYNC1H1",
+  "EIF2B3",
+  "ENPP1",
+  "ENY2",
+  "EPB41L1",
+  "ERCC1",
+  "ERICH1",
+  "EXOSC10",
+  "FKBP11",
+  "FMNL2",
+  "FRA10AC1",
+  "FRMD4B",
+  "G0S2",
+  "GBP2",
+  "GDPD3",
+  "GLTSCR2",
+  "GMFG",
+  "GNL2",
+  "GNL3L",
+  "GNLY",
+  "GPX3",
+  "GTF2I",
+  "HLA-DPB1",
+  "HLA-F",
+  "HOOK3",
+  "HSPB11",
+  "IFI27L2",
+  "IFT81",
+  "IFT88",
+  "IGHA1",
+  "IGHG1",
+  "IGHG3",
+  "IGHG4",
+  "IGHM",
+  "IGKC",
+  "IGLC2",
+  "IGLC3",
+  "IGLC7",
+  "ITGB1",
+  "ITPR2",
+  "JCHAIN",
+  "KDM7A",
+  "KMT2E",
+  "KNTC1",
+  "KRT15",
+  "KRT17",
+  "LRRC29",
+  "LSM6",
+  "LTA4H",
+  "LTF",
+  "LYZ",
+  "MAP4K1",
+  "MAP7D3",
+  "MCUR1",
+  "MGP",
+  "MICAL3",
+  "MMP9",
+  "MPND",
+  "MRPL21",
+  "MRPL52",
+  "MSRB2",
+  "MUCL1",
+  "MUTYH",
+  "MYH3",
+  "NAP1L1",
+  "NDUFA1",
+  "NELFE",
+  "NIT2",
+  "NPW",
+  "NRDC",
+  "NTN4",
+  "NUP214",
+  "NUP37",
+  "OCEL1",
+  "OCIAD2",
+  "OFD1",
+  "OPTN",
+  "PDCD5",
+  "PDE3B",
+  "PHF14",
+  "PIP",
+  "PITX1",
+  "PLAC8",
+  "PLAC9",
+  "PLIN1",
+  "PLIN4",
+  "PLK2",
+  "PNMT",
+  "PPIH",
+  "PPP1R1A",
+  "PPP1R1B",
+  "PREX2",
+  "PRMT7",
+  "PSMD1",
+  "PTGDS",
+  "PTPN6",
+  "R3HDM2",
+  "RAD50",
+  "RBP1",
+  "RBP7",
+  "RFC1",
+  "RNF181",
+  "RNPC3",
+  "ROCK1",
+  "SAA1",
+  "SCAPER",
+  "SCFD1",
+  "SCGB3A1",
+  "SFRP4",
+  "SFSWAP",
+  "SLF1",
+  "SLIRP",
+  "SLPI",
+  "SMCHD1",
+  "SNRNP48",
+  "SOD1",
+  "SP100",
+  "SRSF11",
+  "SRSF7",
+  "STAT4",
+  "SUGP1",
+  "SVIL",
+  "SYCP2",
+  "TEX10",
+  "TFF3",
+  "TFPT",
+  "THOC5",
+  "TIMM50",
+  "TJP3",
+  "TMEM258",
+  "TMEM40",
+  "TNNI2",
+  "TPR",
+  "TRAF3IP1",
+  "TRAP1",
+  "TRBC2",
+  "TRPV2",
+  "TXN",
+  "TXNDC17",
+  "UACA",
+  "UPF2",
+  "UPP1",
+  "USE1",
+  "USP48",
+  "UTRN",
+  "UXT",
+  "VARS",
+  "VWF",
+  "WDR60",
+  "WWP1",
+  "ZBTB20",
+  "ZNF638",
+  "ZNF880",
+  "ZRANB2"
+]

models/single-patch/her2st/hvg/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:254fc5a129245998360cdcac313dd3d1c84b2a3feefbf4ce33689a81509c740d
+size 183903592

models/single-patch/lung/hmhvg/config.json

@@ -0,0 +1,93 @@
+{
+  "format_version": 1,
+  "architecture": "SimpleP2G",
+  "setting": "single-patch",
+  "dataset": "lung",
+  "panel": "hmhvg",
+  "fold": 0,
+  "seed": 42,
+  "model_config": {
+    "decoder_depth": 12,
+    "num_genes": 200,
+    "hidden_size": 384,
+    "cond_size": 2560,
+    "num_cond_tokens": 4,
+    "num_heads": 6,
+    "loss_name": "gaussian_nll",
+    "dropout": 0.0
+  },
+  "input": {
+    "representation": "frozen Virchow2 CLS token concatenated with mean patch tokens",
+    "embedding_dimension": 2560,
+    "image_encoder_weights_included": false
+  },
+  "output": {
+    "num_genes": 200,
+    "gene_order": "genes.json",
+    "expression_scale": "log2(raw_count + 1)",
+    "returns": [
+      "expression_mean",
+      "log_variance"
+    ]
+  },
+  "checkpoint": {
+    "epoch": 7,
+    "weights": "ema",
+    "selection": "highest validation-slide PCC-200; earliest epoch breaks ties"
+  },
+  "training": {
+    "epochs": 100,
+    "batch_size": 256,
+    "optimizer": {
+      "name": "adamw",
+      "lr": 5e-05,
+      "weight_decay": 0.05,
+      "betas": [
+        0.9,
+        0.999
+      ]
+    },
+    "loss": "gaussian_nll",
+    "precision": "fp16",
+    "ema_decay": 0.999,
+    "sampling": "pooled spots"
+  },
+  "split": {
+    "training_slides": [
+      "NCBI857",
+      "NCBI858",
+      "NCBI859",
+      "NCBI861",
+      "NCBI864",
+      "NCBI865",
+      "NCBI866",
+      "NCBI867",
+      "NCBI870",
+      "NCBI873",
+      "NCBI875",
+      "NCBI876",
+      "NCBI879",
+      "NCBI880",
+      "NCBI881",
+      "NCBI882",
+      "NCBI883",
+      "NCBI884"
+    ],
+    "validation_slide": "NCBI860",
+    "test_slide": "NCBI856"
+  },
+  "validation_metrics": {
+    "PCC10": 0.9362276377737937,
+    "PCC50": 0.9226602827904639,
+    "PCC200": 0.8183084345389446,
+    "MSE": 1.1610033512115479,
+    "MAE": 0.7474100589752197
+  },
+  "test_metrics": {
+    "PCC10": 0.970405946201667,
+    "PCC50": 0.9624952765742016,
+    "PCC200": 0.8717158056511088,
+    "MSE": 1.2483162879943848,
+    "MAE": 0.8023096919059753
+  }
+}

models/single-patch/lung/hmhvg/genes.json

@@ -0,0 +1,202 @@
+[
+  "ACKR1",
+  "ACTA2",
+  "AGER",
+  "AGR3",
+  "AIF1",
+  "AKR1C1",
+  "AKR1C2",
+  "ANKRD28",
+  "ATF3",
+  "ATF4",
+  "ATG7",
+  "ATP2A3",
+  "AXIN2",
+  "AXL",
+  "BANK1",
+  "BCL2L1",
+  "BCL2L11",
+  "BMPR2",
+  "BPIFA1",
+  "C20orf85",
+  "CCL18",
+  "CCL2",
+  "CCL21",
+  "CCN2",
+  "CD14",
+  "CD2",
+  "CD34",
+  "CD3E",
+  "CD4",
+  "CD44",
+  "CD52",
+  "CD68",
+  "CD69",
+  "CD79A",
+  "CD8A",
+  "CEACAM5",
+  "CEACAM6",
+  "COL15A1",
+  "COL1A1",
+  "COL1A2",
+  "COL3A1",
+  "CPA3",
+  "CST3",
+  "CTHRC1",
+  "CTNNB1",
+  "CXCL13",
+  "CXCL14",
+  "CXCL9",
+  "CXCR4",
+  "DCN",
+  "DMBT1",
+  "DNAJB9",
+  "DUOX1",
+  "EGFR",
+  "EHMT1",
+  "EPAS1",
+  "EPCAM",
+  "ERLEC1",
+  "FABP4",
+  "FAS",
+  "FCER1G",
+  "FCGR3A",
+  "FCN3",
+  "FGF7",
+  "FKBP11",
+  "FN1",
+  "FOXJ1",
+  "GCLM",
+  "GNG11",
+  "GPR183",
+  "GSR",
+  "GZMA",
+  "HAVCR2",
+  "HERPUD1",
+  "HES1",
+  "HIF1A",
+  "HIST1H1C",
+  "HLA-DQA1",
+  "HLA-DQB1",
+  "HLA-DRA",
+  "HMGA1",
+  "HMOX1",
+  "HSPA5",
+  "HYOU1",
+  "ICAM1",
+  "IDH1",
+  "IFIT1",
+  "IFIT2",
+  "IFIT3",
+  "IL7R",
+  "IRF1",
+  "ITGA3",
+  "ITGAM",
+  "ITGAV",
+  "ITGAX",
+  "ITGB1",
+  "ITGB6",
+  "ITM2C",
+  "JCHAIN",
+  "KDR",
+  "KIT",
+  "KRT15",
+  "KRT17",
+  "KRT18",
+  "KRT5",
+  "KRT6A",
+  "KRT8",
+  "LAMP3",
+  "LGALS1",
+  "LMAN1",
+  "LTF",
+  "LUM",
+  "LYZ",
+  "MANF",
+  "MARCO",
+  "MCEMP1",
+  "MEG3",
+  "MGST1",
+  "MMP7",
+  "MRC1",
+  "MS4A1",
+  "MS4A7",
+  "MSLN",
+  "MUC5B",
+  "MYC",
+  "MYDGF",
+  "NAPSA",
+  "NFKB1",
+  "NKX2-1",
+  "NUCB2",
+  "NUTF2",
+  "OAS2",
+  "OAS3",
+  "PCNA",
+  "PDGFRA",
+  "PDGFRB",
+  "PDIA3",
+  "PDIA4",
+  "PDIA6",
+  "PECAM1",
+  "PGC",
+  "PIM2",
+  "PKM",
+  "PLIN2",
+  "PLVAP",
+  "POSTN",
+  "PPARG",
+  "PRDX4",
+  "PTGDS",
+  "PTPRC",
+  "RHOA",
+  "RNASE1",
+  "RTKN2",
+  "S100A12",
+  "S100A2",
+  "S100A8",
+  "S100A9",
+  "SAA2",
+  "SCGB1A1",
+  "SCGB3A2",
+  "SEC11C",
+  "SELENOS",
+  "SFRP2",
+  "SFRP4",
+  "SFTA2",
+  "SFTPC",
+  "SFTPD",
+  "SLC1A3",
+  "SLC25A4",
+  "SMAD4",
+  "SNAI2",
+  "SOD2",
+  "SOX4",
+  "SOX9",
+  "SPARCL1",
+  "SPCS2",
+  "SPCS3",
+  "SPP1",
+  "SPRY2",
+  "SSR3",
+  "STAT1",
+  "STAT6",
+  "TGFB1",
+  "TGFB3",
+  "TNFRSF13C",
+  "TOP1",
+  "TPSAB1",
+  "TRAC",
+  "TTC19",
+  "UBE2J1",
+  "UCHL3",
+  "UGDH",
+  "VEGFA",
+  "VIM",
+  "WFDC2",
+  "WNT5A",
+  "WWTR1",
+  "XBP1",
+  "YAP1",
+  "ZEB1"
+]

models/single-patch/lung/hmhvg/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dcbde320a1df51a876a4c0797983d160f5deb3e4bae0f402af6af97b52ccc2f5
+size 183903592

models/single-patch/lung/hvg/config.json

@@ -0,0 +1,93 @@
+{
+  "format_version": 1,
+  "architecture": "SimpleP2G",
+  "setting": "single-patch",
+  "dataset": "lung",
+  "panel": "hvg",
+  "fold": 0,
+  "seed": 42,
+  "model_config": {
+    "decoder_depth": 12,
+    "num_genes": 200,
+    "hidden_size": 384,
+    "cond_size": 2560,
+    "num_cond_tokens": 4,
+    "num_heads": 6,
+    "loss_name": "gaussian_nll",
+    "dropout": 0.0
+  },
+  "input": {
+    "representation": "frozen Virchow2 CLS token concatenated with mean patch tokens",
+    "embedding_dimension": 2560,
+    "image_encoder_weights_included": false
+  },
+  "output": {
+    "num_genes": 200,
+    "gene_order": "genes.json",
+    "expression_scale": "log2(raw_count + 1)",
+    "returns": [
+      "expression_mean",
+      "log_variance"
+    ]
+  },
+  "checkpoint": {
+    "epoch": 8,
+    "weights": "ema",
+    "selection": "highest validation-slide PCC-200; earliest epoch breaks ties"
+  },
+  "training": {
+    "epochs": 100,
+    "batch_size": 256,
+    "optimizer": {
+      "name": "adamw",
+      "lr": 5e-05,
+      "weight_decay": 0.05,
+      "betas": [
+        0.9,
+        0.999
+      ]
+    },
+    "loss": "gaussian_nll",
+    "precision": "fp16",
+    "ema_decay": 0.999,
+    "sampling": "pooled spots"
+  },
+  "split": {
+    "training_slides": [
+      "NCBI857",
+      "NCBI858",
+      "NCBI859",
+      "NCBI861",
+      "NCBI864",
+      "NCBI865",
+      "NCBI866",
+      "NCBI867",
+      "NCBI870",
+      "NCBI873",
+      "NCBI875",
+      "NCBI876",
+      "NCBI879",
+      "NCBI880",
+      "NCBI881",
+      "NCBI882",
+      "NCBI883",
+      "NCBI884"
+    ],
+    "validation_slide": "NCBI860",
+    "test_slide": "NCBI856"
+  },
+  "validation_metrics": {
+    "PCC10": 0.9030630133389671,
+    "PCC50": 0.8287717950429099,
+    "PCC200": 0.4339199710155945,
+    "MSE": 0.606461763381958,
+    "MAE": 0.42725834250450134
+  },
+  "test_metrics": {
+    "PCC10": 0.9472604745563193,
+    "PCC50": 0.8581366731430391,
+    "PCC200": 0.4954411997870103,
+    "MSE": 0.7060865759849548,
+    "MAE": 0.4992806017398834
+  }
+}

models/single-patch/lung/hvg/genes.json

@@ -0,0 +1,202 @@
+[
+  "ACKR1",
+  "ACTA2",
+  "AGR3",
+  "AIF1",
+  "APLN",
+  "ASCL1",
+  "BPIFA1",
+  "C20orf85",
+  "CCL18",
+  "CCL21",
+  "CCL22",
+  "CCNA1",
+  "CD1A",
+  "CD52",
+  "CD68",
+  "CD8A",
+  "CDKN2A",
+  "CEACAM6",
+  "COL1A1",
+  "COL3A1",
+  "CXCL13",
+  "CXCL14",
+  "CXCL9",
+  "CXCR5",
+  "DMBT1",
+  "EHMT1",
+  "EREG",
+  "FABP4",
+  "FGF2",
+  "FGFBP2",
+  "FOXJ1",
+  "GDF15",
+  "GZMA",
+  "HAS1",
+  "HLA-DRA",
+  "HMGA1",
+  "HMOX1",
+  "IFIT1",
+  "IL7R",
+  "IRF1",
+  "IRF7",
+  "ITGAV",
+  "JCHAIN",
+  "KRT14",
+  "KRT15",
+  "KRT17",
+  "KRT18",
+  "KRT8",
+  "LAMP3",
+  "LTB",
+  "LTF",
+  "LY6D",
+  "LYZ",
+  "MARCO",
+  "MCEMP1",
+  "MFAP5",
+  "MMP7",
+  "MUC5AC",
+  "MUC5B",
+  "NAPSA",
+  "NegControlCodeword_0501",
+  "NegControlCodeword_0502",
+  "NegControlCodeword_0504",
+  "NegControlCodeword_0508",
+  "NegControlCodeword_0511",
+  "NegControlCodeword_0512",
+  "NegControlCodeword_0514",
+  "NegControlCodeword_0515",
+  "NegControlCodeword_0516",
+  "NegControlCodeword_0517",
+  "NegControlCodeword_0518",
+  "NegControlCodeword_0520",
+  "NegControlCodeword_0524",
+  "NegControlCodeword_0526",
+  "NegControlCodeword_0527",
+  "NegControlCodeword_0528",
+  "NegControlCodeword_0530",
+  "NegControlCodeword_0534",
+  "NegControlCodeword_0536",
+  "NegControlCodeword_0537",
+  "NegControlCodeword_0540",
+  "NegControlProbe_00002",
+  "NegControlProbe_00003",
+  "NegControlProbe_00004",
+  "NegControlProbe_00012",
+  "NegControlProbe_00013",
+  "NegControlProbe_00014",
+  "NegControlProbe_00016",
+  "NegControlProbe_00017",
+  "NegControlProbe_00019",
+  "NegControlProbe_00022",
+  "NegControlProbe_00024",
+  "NegControlProbe_00025",
+  "NegControlProbe_00031",
+  "NegControlProbe_00033",
+  "NegControlProbe_00034",
+  "NegControlProbe_00039",
+  "NegControlProbe_00041",
+  "PCNA",
+  "PDCD1",
+  "PGC",
+  "PLIN2",
+  "POSTN",
+  "PPARG",
+  "PTPRC",
+  "RETN",
+  "S100A12",
+  "S100A2",
+  "S100A8",
+  "S100A9",
+  "SAA2",
+  "SCGB1A1",
+  "SCGB3A2",
+  "SFRP2",
+  "SFRP4",
+  "SFTPC",
+  "SFTPD",
+  "SLC25A4",
+  "SPP1",
+  "TERT",
+  "UGDH",
+  "UnassignedCodeword_0000",
+  "UnassignedCodeword_0001",
+  "UnassignedCodeword_0011",
+  "UnassignedCodeword_0014",
+  "UnassignedCodeword_0028",
+  "UnassignedCodeword_0034",
+  "UnassignedCodeword_0043",
+  "UnassignedCodeword_0044",
+  "UnassignedCodeword_0045",
+  "UnassignedCodeword_0048",
+  "UnassignedCodeword_0050",
+  "UnassignedCodeword_0052",
+  "UnassignedCodeword_0055",
+  "UnassignedCodeword_0057",
+  "UnassignedCodeword_0063",
+  "UnassignedCodeword_0066",
+  "UnassignedCodeword_0068",
+  "UnassignedCodeword_0071",
+  "UnassignedCodeword_0082",
+  "UnassignedCodeword_0085",
+  "UnassignedCodeword_0095",
+  "UnassignedCodeword_0104",
+  "UnassignedCodeword_0107",
+  "UnassignedCodeword_0112",
+  "UnassignedCodeword_0118",
+  "UnassignedCodeword_0122",
+  "UnassignedCodeword_0123",
+  "UnassignedCodeword_0127",
+  "UnassignedCodeword_0129",
+  "UnassignedCodeword_0131",
+  "UnassignedCodeword_0139",
+  "UnassignedCodeword_0148",
+  "UnassignedCodeword_0150",
+  "UnassignedCodeword_0155",
+  "UnassignedCodeword_0162",
+  "UnassignedCodeword_0166",
+  "UnassignedCodeword_0168",
+  "UnassignedCodeword_0172",
+  "UnassignedCodeword_0174",
+  "UnassignedCodeword_0185",
+  "UnassignedCodeword_0188",
+  "UnassignedCodeword_0198",
+  "UnassignedCodeword_0201",
+  "UnassignedCodeword_0205",
+  "UnassignedCodeword_0210",
+  "UnassignedCodeword_0213",
+  "UnassignedCodeword_0214",
+  "UnassignedCodeword_0219",
+  "UnassignedCodeword_0224",
+  "UnassignedCodeword_0236",
+  "UnassignedCodeword_0241",
+  "UnassignedCodeword_0245",
+  "UnassignedCodeword_0252",
+  "UnassignedCodeword_0258",
+  "UnassignedCodeword_0259",
+  "UnassignedCodeword_0260",
+  "UnassignedCodeword_0261",
+  "UnassignedCodeword_0266",
+  "UnassignedCodeword_0275",
+  "UnassignedCodeword_0280",
+  "UnassignedCodeword_0282",
+  "UnassignedCodeword_0285",
+  "UnassignedCodeword_0290",
+  "UnassignedCodeword_0292",
+  "UnassignedCodeword_0299",
+  "UnassignedCodeword_0300",
+  "UnassignedCodeword_0305",
+  "UnassignedCodeword_0312",
+  "UnassignedCodeword_0322",
+  "UnassignedCodeword_0323",
+  "UnassignedCodeword_0324",
+  "UnassignedCodeword_0327",
+  "UnassignedCodeword_0328",
+  "UnassignedCodeword_0330",
+  "UnassignedCodeword_0332",
+  "UnassignedCodeword_0413",
+  "VIM",
+  "WFDC2",
+  "WNT3A"
+]

models/single-patch/lung/hvg/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f41735851688f4b7368d890c84527a2ba93a620c3df10095029149a5f0d5db6d
+size 183903592

models/single-patch/prad/hmhvg/config.json

@@ -0,0 +1,96 @@
+{
+  "format_version": 1,
+  "architecture": "SimpleP2G",
+  "setting": "single-patch",
+  "dataset": "prad",
+  "panel": "hmhvg",
+  "fold": 0,
+  "seed": 42,
+  "model_config": {
+    "decoder_depth": 12,
+    "num_genes": 200,
+    "hidden_size": 384,
+    "cond_size": 2560,
+    "num_cond_tokens": 4,
+    "num_heads": 6,
+    "loss_name": "gaussian_nll",
+    "dropout": 0.1
+  },
+  "input": {
+    "representation": "frozen Virchow2 CLS token concatenated with mean patch tokens",
+    "embedding_dimension": 2560,
+    "image_encoder_weights_included": false
+  },
+  "output": {
+    "num_genes": 200,
+    "gene_order": "genes.json",
+    "expression_scale": "log2(raw_count + 1)",
+    "returns": [
+      "expression_mean",
+      "log_variance"
+    ]
+  },
+  "checkpoint": {
+    "epoch": 9,
+    "weights": "ema",
+    "selection": "highest validation-slide PCC-200; earliest epoch breaks ties"
+  },
+  "training": {
+    "epochs": 100,
+    "batch_size": 256,
+    "optimizer": {
+      "name": "adamw",
+      "lr": 3e-05,
+      "weight_decay": 0.1,
+      "betas": [
+        0.9,
+        0.999
+      ]
+    },
+    "loss": "gaussian_nll",
+    "precision": "fp16",
+    "ema_decay": 0.999,
+    "sampling": "pooled spots"
+  },
+  "split": {
+    "training_slides": [
+      "MEND139",
+      "MEND140",
+      "MEND141",
+      "MEND143",
+      "MEND144",
+      "MEND146",
+      "MEND147",
+      "MEND148",
+      "MEND149",
+      "MEND150",
+      "MEND151",
+      "MEND152",
+      "MEND153",
+      "MEND154",
+      "MEND156",
+      "MEND157",
+      "MEND158",
+      "MEND159",
+      "MEND160",
+      "MEND161",
+      "MEND162"
+    ],
+    "validation_slide": "MEND142",
+    "test_slide": "MEND145"
+  },
+  "validation_metrics": {
+    "PCC10": 0.735454567647998,
+    "PCC50": 0.6725363606552629,
+    "PCC200": 0.5097357986212281,
+    "MSE": 0.9369765520095825,
+    "MAE": 0.7383893728256226
+  },
+  "test_metrics": {
+    "PCC10": 0.7297544680010257,
+    "PCC50": 0.681781272744159,
+    "PCC200": 0.556121254362732,
+    "MSE": 1.0525509119033813,
+    "MAE": 0.7671422362327576
+  }
+}

models/single-patch/prad/hmhvg/genes.json

@@ -0,0 +1,202 @@
+[
+  "A2M",
+  "ACPP",
+  "ACTA2",
+  "ACTB",
+  "ACTG1",
+  "ACTG2",
+  "ADIRF",
+  "AGR2",
+  "AMD1",
+  "ANPEP",
+  "APLP2",
+  "ATP5F1E",
+  "ATP5MD",
+  "ATP5MF",
+  "ATP5MPL",
+  "AZGP1",
+  "B2M",
+  "BTF3",
+  "C12orf57",
+  "C19orf48",
+  "CALM2",
+  "CALR",
+  "CD63",
+  "CD74",
+  "CD81",
+  "CD9",
+  "CD99",
+  "CFD",
+  "CFL1",
+  "CHCHD2",
+  "CIRBP",
+  "CKB",
+  "CLU",
+  "CNN1",
+  "COMMD6",
+  "COX4I1",
+  "COX5B",
+  "COX6A1",
+  "COX6C",
+  "COX7A2",
+  "COX7B",
+  "COX7C",
+  "COX8A",
+  "CPE",
+  "CSRP1",
+  "CST3",
+  "DBI",
+  "DDT",
+  "DES",
+  "DHRS7",
+  "DSTN",
+  "DUSP1",
+  "EDF1",
+  "EEF1A1",
+  "EEF1B2",
+  "EEF2",
+  "EIF1",
+  "EIF3L",
+  "ELOB",
+  "FABP5",
+  "FASN",
+  "FAU",
+  "FBLN1",
+  "FLNA",
+  "FOS",
+  "FTH1",
+  "FTL",
+  "FXYD3",
+  "GAPDH",
+  "GPX4",
+  "H2AFJ",
+  "H3F3A",
+  "H3F3B",
+  "HERPUD1",
+  "HINT1",
+  "HLA-B",
+  "HLA-C",
+  "HLA-DRA",
+  "HMGN2",
+  "HNRNPA1",
+  "HOXB13",
+  "HSPA5",
+  "HSPA8",
+  "HSPB1",
+  "IGHA1",
+  "IGKC",
+  "IGLC2",
+  "ITM2B",
+  "KDELR2",
+  "KLK2",
+  "KLK3",
+  "KLK4",
+  "KRT18",
+  "KRT8",
+  "LGALS1",
+  "LTF",
+  "MALAT1",
+  "MDK",
+  "MGP",
+  "MIF",
+  "MINOS1",
+  "MPC2",
+  "MSMB",
+  "MYH11",
+  "MYL6",
+  "MYL9",
+  "MYLK",
+  "MZT2B",
+  "NACA",
+  "NBL1",
+  "NCAPD3",
+  "NDRG1",
+  "NDUFB1",
+  "NDUFB11",
+  "NDUFB4",
+  "NDUFS5",
+  "NEAT1",
+  "NEFH",
+  "NKX3-1",
+  "NME4",
+  "NPM1",
+  "NPY",
+  "NR4A1",
+  "NUPR1",
+  "OAZ1",
+  "OST4",
+  "PABPC1",
+  "PARK7",
+  "PCAT4",
+  "PDLIM5",
+  "PFDN5",
+  "PLA2G2A",
+  "PLPP1",
+  "PMEPA1",
+  "POLR2L",
+  "PPDPF",
+  "PPIA",
+  "PRAC1",
+  "PRDX2",
+  "PRDX6",
+  "PTGDS",
+  "PTMA",
+  "RACK1",
+  "RDH11",
+  "ROMO1",
+  "S100A11",
+  "S100A6",
+  "SARAF",
+  "SAT1",
+  "SEC11C",
+  "SEC61B",
+  "SEC61G",
+  "SELENOP",
+  "SELENOW",
+  "SERF2",
+  "SERP1",
+  "SKP1",
+  "SLC25A6",
+  "SLC45A3",
+  "SNHG19",
+  "SNHG25",
+  "SNHG8",
+  "SNRPD2",
+  "SORD",
+  "SPDEF",
+  "SPINT2",
+  "SPON2",
+  "SRP14",
+  "SSR4",
+  "STEAP2",
+  "TAGLN",
+  "TFF3",
+  "TIMP1",
+  "TMBIM6",
+  "TMEFF2",
+  "TMEM141",
+  "TMEM258",
+  "TMEM59",
+  "TMPRSS2",
+  "TMSB10",
+  "TMSB4X",
+  "TOMM7",
+  "TPM1",
+  "TPM2",
+  "TPT1",
+  "TRPM4",
+  "TSC22D3",
+  "TSPAN1",
+  "TSTD1",
+  "TXN",
+  "UBA52",
+  "UBB",
+  "UBL5",
+  "UQCR10",
+  "UQCRB",
+  "UQCRH",
+  "UQCRQ",
+  "VEGFA",
+  "VIM",
+  "ZFAS1"
+]

models/single-patch/prad/hmhvg/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ce7d28e32b4a84c0f0cba6f0561b22a997de7e9202c970968e767b2e71257142
+size 183903592

models/single-patch/prad/hvg/config.json

@@ -0,0 +1,96 @@
+{
+  "format_version": 1,
+  "architecture": "SimpleP2G",
+  "setting": "single-patch",
+  "dataset": "prad",
+  "panel": "hvg",
+  "fold": 0,
+  "seed": 42,
+  "model_config": {
+    "decoder_depth": 12,
+    "num_genes": 200,
+    "hidden_size": 384,
+    "cond_size": 2560,
+    "num_cond_tokens": 4,
+    "num_heads": 6,
+    "loss_name": "gaussian_nll",
+    "dropout": 0.1
+  },
+  "input": {
+    "representation": "frozen Virchow2 CLS token concatenated with mean patch tokens",
+    "embedding_dimension": 2560,
+    "image_encoder_weights_included": false
+  },
+  "output": {
+    "num_genes": 200,
+    "gene_order": "genes.json",
+    "expression_scale": "log2(raw_count + 1)",
+    "returns": [
+      "expression_mean",
+      "log_variance"
+    ]
+  },
+  "checkpoint": {
+    "epoch": 12,
+    "weights": "ema",
+    "selection": "highest validation-slide PCC-200; earliest epoch breaks ties"
+  },
+  "training": {
+    "epochs": 100,
+    "batch_size": 256,
+    "optimizer": {
+      "name": "adamw",
+      "lr": 3e-05,
+      "weight_decay": 0.1,
+      "betas": [
+        0.9,
+        0.999
+      ]
+    },
+    "loss": "gaussian_nll",
+    "precision": "fp16",
+    "ema_decay": 0.999,
+    "sampling": "pooled spots"
+  },
+  "split": {
+    "training_slides": [
+      "MEND139",
+      "MEND140",
+      "MEND141",
+      "MEND143",
+      "MEND144",
+      "MEND146",
+      "MEND147",
+      "MEND148",
+      "MEND149",
+      "MEND150",
+      "MEND151",
+      "MEND152",
+      "MEND153",
+      "MEND154",
+      "MEND156",
+      "MEND157",
+      "MEND158",
+      "MEND159",
+      "MEND160",
+      "MEND161",
+      "MEND162"
+    ],
+    "validation_slide": "MEND142",
+    "test_slide": "MEND145"
+  },
+  "validation_metrics": {
+    "PCC10": 0.7139566340145835,
+    "PCC50": 0.5177275273410492,
+    "PCC200": 0.20855663886632447,
+    "MSE": 0.4681999087333679,
+    "MAE": 0.32622313499450684
+  },
+  "test_metrics": {
+    "PCC10": 0.6859417787277124,
+    "PCC50": 0.4673065945925007,
+    "PCC200": 0.2199021112938024,
+    "MSE": 0.29881447553634644,
+    "MAE": 0.22787906229496002
+  }
+}

models/single-patch/prad/hvg/genes.json

@@ -0,0 +1,202 @@
+[
+  "A2M",
+  "ABCB8",
+  "ACKR1",
+  "ACPP",
+  "ACTG2",
+  "ADAP2",
+  "ADCY5",
+  "AGBL1",
+  "AGR2",
+  "AKTIP",
+  "AL133492.1",
+  "AL355075.4",
+  "AL359220.1",
+  "ANGPTL7",
+  "ANKRD42",
+  "AP4E1",
+  "APOC1",
+  "APOD",
+  "AQP1",
+  "ASF1B",
+  "ATF3",
+  "ATP10D",
+  "AZGP1",
+  "C19orf33",
+  "C1orf52",
+  "CAMK1D",
+  "CCDC144A",
+  "CCDC180",
+  "CCL19",
+  "CCL21",
+  "CCL4",
+  "CCL4L2",
+  "CCL5",
+  "CCM2L",
+  "CDA",
+  "CDH19",
+  "CDH3",
+  "CDK11A",
+  "CFD",
+  "CFHR5",
+  "CH25H",
+  "CHGA",
+  "CKAP2",
+  "CLDN5",
+  "CNOT4",
+  "COL15A1",
+  "COL6A1",
+  "CPXM2",
+  "CST1",
+  "CXCL10",
+  "CXCL9",
+  "DDX20",
+  "DEPP1",
+  "DES",
+  "DNM3OS",
+  "EIF5A2",
+  "ELF3",
+  "ELMO1",
+  "EMILIN2",
+  "EPM2A",
+  "ETV6",
+  "FILIP1L",
+  "FOSB",
+  "FOSL1",
+  "FYCO1",
+  "GLI4",
+  "GRIN2A",
+  "GZMA",
+  "HBA1",
+  "HBA2",
+  "HBB",
+  "HIVEP2",
+  "ICA1L",
+  "IFT52",
+  "IGF2BP2",
+  "IGFBP3",
+  "IGHA1",
+  "IGHG1",
+  "IGHG2",
+  "IGHG3",
+  "IGHG4",
+  "IGHM",
+  "IGKC",
+  "IGLC2",
+  "IGLC3",
+  "IGLL5",
+  "JAM2",
+  "JCHAIN",
+  "KANSL1",
+  "KCNG4",
+  "KCNH2",
+  "KHNYN",
+  "KLHL33",
+  "KLK2",
+  "KLK3",
+  "KRBA1",
+  "KRT13",
+  "KRT15",
+  "KRT17",
+  "KRT19",
+  "KRT5",
+  "LACTB",
+  "LAIR1",
+  "LCN2",
+  "LEPR",
+  "LMO7DN",
+  "LOX",
+  "LRRC23",
+  "LTF",
+  "LYZ",
+  "MCUB",
+  "MFSD11",
+  "MGP",
+  "MHENCR",
+  "MMP7",
+  "MSMB",
+  "MSRB3",
+  "MT1G",
+  "MT1H",
+  "MYH11",
+  "MYOM1",
+  "NECAB1",
+  "NEFH",
+  "NPR3",
+  "NPY",
+  "NSUN2",
+  "NTRK2",
+  "OLFM4",
+  "OR6S1",
+  "P3H3",
+  "PARP12",
+  "PCSK7",
+  "PDCD7",
+  "PDK1",
+  "PIGR",
+  "PLA2G2A",
+  "PLCD3",
+  "POU2F2",
+  "PRDM5",
+  "PRKRIP1",
+  "PRMT9",
+  "PRPF18",
+  "PRPF4",
+  "PSCA",
+  "PTGDS",
+  "PTGER3",
+  "PTRH1",
+  "RAB30",
+  "RASL11B",
+  "RDH11",
+  "RGS5",
+  "RIC3",
+  "RNASE1",
+  "RND2",
+  "RNF144A",
+  "RNFT2",
+  "RPAP3",
+  "RPGRIP1L",
+  "RUBCNL",
+  "S100P",
+  "SAA1",
+  "SCGB1A1",
+  "SCGB3A1",
+  "SCUBE3",
+  "SIX3",
+  "SLC14A1",
+  "SLC16A1-AS1",
+  "SLC19A2",
+  "SLC25A13",
+  "SLC26A4",
+  "SNAP47",
+  "SORD",
+  "SPAG16",
+  "SPC24",
+  "SPP1",
+  "SSTR2",
+  "ST3GAL2",
+  "SULF2",
+  "SYK",
+  "TFRC",
+  "TGM4",
+  "THAP6",
+  "THBD",
+  "TMEFF2",
+  "TMEM260",
+  "TPSAB1",
+  "TPSB2",
+  "TSPYL4",
+  "TTC27",
+  "UTP25",
+  "VHL",
+  "VIPR2",
+  "VTI1A",
+  "ZKSCAN5",
+  "ZNF2",
+  "ZNF266",
+  "ZNF385A",
+  "ZNF557",
+  "ZNF620",
+  "ZNF654"
+]

models/single-patch/prad/hvg/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f297f40eab72aa2e7706892106ee4428584e7604d0ef91a12272acde8aa60cf
+size 183903592

requirements.txt

@@ -0,0 +1,3 @@
+torch==2.10.0
+safetensors==0.8.0
+huggingface-hub==0.36.0

src/__init__.py

@@ -0,0 +1 @@
+"""SimpleP2G pretrained prediction heads."""

src/pretrained.py

@@ -0,0 +1,74 @@
+"""Load released SimpleP2G prediction heads without an image encoder."""
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from types import SimpleNamespace
+
+import torch
+from safetensors.torch import load_file
+
+from src.simpleP2G import SimpleP2G
+
+
+DATASETS = ("brain", "prad", "her2st", "lung", "ccrcc")
+PANELS = ("hmhvg", "hvg")
+SETTINGS = ("single-patch", "cross-patch")
+
+
+def load_pretrained(
+    repo_id_or_path="king159/SimpleP2G",
+    *,
+    dataset="brain",
+    panel="hmhvg",
+    setting="single-patch",
+    device="cpu",
+    revision=None,
+    cache_dir=None,
+    local_files_only=False,
+):
+    """Return an eval-mode model, ordered gene names and checkpoint metadata.
+
+    ``repo_id_or_path`` is a Hub model ID or a downloaded repository directory.
+    Inputs are precomputed image embeddings; no image-encoder weights are loaded.
+    Predictions follow the order in ``genes`` and use log2(count + 1) units.
+    """
+    dataset, panel = dataset.lower(), panel.lower()
+    if dataset not in DATASETS or panel not in PANELS or setting not in SETTINGS:
+        raise ValueError(f"Unsupported model selection: {dataset}/{panel}/{setting}")
+    folder = f"models/{setting}/{dataset}/{panel}"
+    local = Path(repo_id_or_path).expanduser()
+
+    def resolve(filename):
+        if local.is_dir():
+            path = local / folder / filename
+            if not path.is_file():
+                raise FileNotFoundError(path)
+            return path
+        from huggingface_hub import hf_hub_download
+
+        return Path(hf_hub_download(
+            repo_id=str(repo_id_or_path), filename=f"{folder}/{filename}",
+            revision=revision, cache_dir=cache_dir, local_files_only=local_files_only,
+        ))
+
+    metadata = json.loads(resolve("config.json").read_text())
+    if metadata.get("format_version") != 1:
+        raise ValueError("Unsupported checkpoint metadata version")
+    expected = {"dataset": dataset, "panel": panel, "setting": setting}
+    if any(metadata.get(key) != value for key, value in expected.items()):
+        raise ValueError("Checkpoint metadata does not match the requested model")
+    genes = json.loads(resolve("genes.json").read_text())
+    config = SimpleNamespace(**metadata["model_config"])
+    if len(genes) != config.num_genes or len(set(genes)) != len(genes):
+        raise ValueError("The ordered gene list must contain one unique name per output")
+    if setting == "single-patch":
+        model = SimpleP2G(config)
+    else:
+        from src.simpleP2G_neigh import SimpleP2GNeigh
+
+        model = SimpleP2GNeigh(config)
+    state = load_file(str(resolve("model.safetensors")), device="cpu")
+    model.load_state_dict(state, strict=True)
+    model.to(device=device).eval().requires_grad_(False)
+    return model, genes, metadata

src/simpleP2G.py

@@ -0,0 +1,350 @@
+from typing import Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class SimpleP2GRMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+
+class SimpleP2GMLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = getattr(
+            config,
+            "intermediate_size",
+            int(config.hidden_size * getattr(config, "mlp_ratio", 4.0)),
+        )
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = nn.GELU(approximate="tanh")
+        self.dropout = nn.Dropout(getattr(config, "dropout", 0.0))
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return self.dropout(down_proj)
+
+
+class SimpleP2GAttention(nn.Module):
+    """
+    multi-head self-attention module using F.scaled_dot_product_attention.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_heads
+        self.head_dim = self.hidden_size // self.num_heads
+
+        if self.hidden_size % self.num_heads != 0:
+            raise ValueError(f"hidden_size {self.hidden_size} must be divisible by num_heads {self.num_heads}")
+
+        self.attention_dropout = getattr(config, "attention_dropout", 0.0)
+
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+
+    def forward(self, hidden_states: torch.Tensor, attention_mask: Optional[torch.Tensor] = None) -> torch.Tensor:
+        batch_size, seq_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(batch_size, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(batch_size, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(batch_size, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        # Prepare attention mask for F.scaled_dot_product_attention
+        # F.sdpa expects mask to be True for positions that should attend
+        if attention_mask is not None:
+            if attention_mask.dim() == 2:  # (B, N)
+                # Convert to (B, 1, 1, N) and invert (0 -> False, 1 -> True)
+                attention_mask = attention_mask.bool()
+                attention_mask = attention_mask.unsqueeze(1).unsqueeze(1)
+            elif attention_mask.dim() == 3:  # (B, 1, N)
+                attention_mask = attention_mask.bool()
+                attention_mask = attention_mask.unsqueeze(1)
+            elif attention_mask.dim() == 4:  # (B, 1, 1, N) or (B, H, N, N)
+                attention_mask = attention_mask.bool()
+
+        # Use F.scaled_dot_product_attention
+        attn_output = F.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            dropout_p=self.attention_dropout if self.training else 0.0,
+            is_causal=False,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(batch_size, seq_len, self.hidden_size)
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output
+
+
+class SimpleP2GCrossAttention(nn.Module):
+    """
+    Multi-head cross-attention module.
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_heads
+        self.head_dim = self.hidden_size // self.num_heads
+
+        if self.hidden_size % self.num_heads != 0:
+            raise ValueError(f"hidden_size {self.hidden_size} must be divisible by num_heads {self.num_heads}")
+
+        self.attention_dropout = getattr(config, "attention_dropout", 0.0)
+
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,  # Query
+        encoder_hidden_states: torch.Tensor,  # Key, Value
+        attention_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        batch_size, seq_len, _ = hidden_states.size()
+        encoder_seq_len = encoder_hidden_states.size(1)
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(encoder_hidden_states)
+        value_states = self.v_proj(encoder_hidden_states)
+
+        query_states = query_states.view(batch_size, seq_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(batch_size, encoder_seq_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(batch_size, encoder_seq_len, self.num_heads, self.head_dim).transpose(1, 2)
+
+        if attention_mask is not None and not (
+            attention_mask.is_floating_point() or attention_mask.dtype == torch.bool
+        ):
+            attention_mask = attention_mask.bool()
+
+        attn_output = F.scaled_dot_product_attention(
+            query_states,
+            key_states,
+            value_states,
+            attn_mask=attention_mask,
+            dropout_p=self.attention_dropout if self.training else 0.0,
+            is_causal=False,
+        )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(batch_size, seq_len, self.hidden_size)
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output
+
+
+class SimpleP2GDecoderLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        eps = getattr(config, "rms_norm_eps", 1e-6)
+
+        # Layers
+        self.self_attn = SimpleP2GAttention(config=config)
+        self.cross_attn = SimpleP2GCrossAttention(config=config)
+        self.mlp = SimpleP2GMLP(config)
+
+        # Pre-Norms (Normalization happens BEFORE the sub-layer)
+        self.norm1 = SimpleP2GRMSNorm(self.hidden_size, eps=eps)
+        self.norm2 = SimpleP2GRMSNorm(self.hidden_size, eps=eps)
+        self.norm3 = SimpleP2GRMSNorm(self.hidden_size, eps=eps)
+        self.resid_dropout = nn.Dropout(getattr(config, "dropout", 0.0))
+
+    def forward(self, x, encoder_hidden_states, attention_mask=None, cross_attention_mask=None):
+        # 1. Self-Attention (Genes interacting with each other)
+        x = x + self.resid_dropout(self.self_attn(self.norm1(x), attention_mask=attention_mask))
+
+        # 2. Cross-Attention (Genes looking at the image context)
+        x = x + self.resid_dropout(
+            self.cross_attn(self.norm2(x), encoder_hidden_states, attention_mask=cross_attention_mask)
+        )
+
+        # 3. Feed-Forward
+        x = x + self.mlp(self.norm3(x))
+
+        return x
+
+
+class SimpleP2G(nn.Module):
+    def __init__(
+        self,
+        config,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.decoder_depth = config.decoder_depth
+        self.num_genes = config.num_genes
+        self.hidden_size = config.hidden_size
+        self.gene_queries = nn.Embedding(self.num_genes, self.hidden_size)
+        self.cond_size = config.cond_size
+        self.num_cond_tokens = getattr(config, "num_cond_tokens", 1)
+        self.decoder_blocks = nn.ModuleList([SimpleP2GDecoderLayer(config) for _ in range(self.decoder_depth)])
+        self.final_norm = SimpleP2GRMSNorm(self.hidden_size)
+        self.mean_head = nn.Linear(self.hidden_size, 1)
+        if config.loss_name in ["gaussian_nll", "beta_nll"]:
+            self.log_var_head = nn.Linear(self.hidden_size, 1)
+        # if config.loss_name == "zinb":
+        #     self.pi_head = nn.Linear(self.hidden_size, 1)
+        self.gradient_checkpointing = False
+        self.emb_dropout = nn.Dropout(getattr(config, "dropout", 0.0))
+
+        # ── Additional regularization ──
+        self.gene_drop_rate = float(getattr(config, "gene_drop_rate", 0.0))
+        self.emb_noise_std = float(getattr(config, "emb_noise_std", 0.0))
+        self.stochastic_depth_rate = float(getattr(config, "stochastic_depth_rate", 0.0))
+
+        self.label_embed = nn.Sequential(
+            nn.Linear(self.cond_size, self.cond_size, bias=True),
+            nn.SiLU(),
+            nn.Linear(self.cond_size, self.hidden_size * self.num_cond_tokens, bias=True),
+        )
+
+    def forward(self, c: torch.Tensor) -> tuple[torch.Tensor, ...]:
+        """Forward pass."""
+        output = self.gene_queries.weight.unsqueeze(0).expand(c.size(0), -1, -1)  # (B, num_genes, hidden_size)
+        output = self.emb_dropout(output)
+
+        # Gene dropout: randomly zero out gene query positions during training
+        if self.training and self.gene_drop_rate > 0:
+            gene_mask = torch.rand(c.size(0), self.num_genes, 1, device=c.device) > self.gene_drop_rate
+            output = output * gene_mask  # zero out dropped genes
+
+        condition = self.label_embed(c).view(c.size(0), self.num_cond_tokens, self.hidden_size)  # (B, T, hidden_size)
+
+        # Embedding noise: add Gaussian noise to conditioning tokens during training
+        if self.training and self.emb_noise_std > 0:
+            condition = condition + torch.randn_like(condition) * self.emb_noise_std
+        for i, block in enumerate(self.decoder_blocks):
+            # Stochastic depth: randomly skip layers during training
+            if self.training and self.stochastic_depth_rate > 0:
+                # Linear scaling: deeper layers have higher drop probability
+                drop_prob = self.stochastic_depth_rate * i / max(len(self.decoder_blocks) - 1, 1)
+                if torch.rand(1).item() < drop_prob:
+                    continue
+
+            if self.gradient_checkpointing and self.training:
+                output = torch.utils.checkpoint.checkpoint(
+                    block,
+                    output,
+                    encoder_hidden_states=condition,
+                    use_reentrant=False,
+                )
+            else:
+                output = block(
+                    output,
+                    encoder_hidden_states=condition,
+                )
+        output = self.final_norm(output)
+        raw_mean = self.mean_head(output).squeeze(-1)
+        mean = F.softplus(raw_mean)  # Ensures predicted G >= 0
+        if self.config.loss_name in ["gaussian_nll", "beta_nll"]:
+            log_var = self.log_var_head(output).squeeze(-1)
+        # if self.config.loss_name == "zinb":
+        #     pi_logits = self.pi_head(output).squeeze(-1)
+        #     return mean, log_var, pi_logits
+        elif self.config.loss_name in ["mse", "l1"]:
+            log_var = torch.zeros_like(mean)
+        return mean, log_var
+
+
+def get_loss_fn(loss_type: str) -> nn.Module:
+    if loss_type == "gaussian_nll":
+        return GaussianNLLLoss()
+    elif loss_type == "beta_nll":
+        return BetaNLLLoss(beta=1.0)
+    elif loss_type == "mse":
+        return nn.MSELoss()
+    elif loss_type == "l1":
+        return nn.L1Loss()
+    else:
+        raise ValueError(f"Unsupported loss type: {loss_type}")
+
+
+class GaussianNLLLoss(nn.Module):
+    def __init__(self, logvar_clipping=(-10, 10)):
+        super().__init__()
+        self.min_logvar, self.max_logvar = logvar_clipping
+
+    def forward(self, pred_mean, pred_logvar, target):
+        # 1. Clamp logvar to prevent NaN gradients or Inf precision
+        pred_logvar = torch.clamp(pred_logvar, self.min_logvar, self.max_logvar)
+
+        # 2. Precision = exp(-logvar) is 1/sigma^2
+        precision = torch.exp(-pred_logvar)
+
+        # 3. Standard NLL formula components
+        # (target - mean)^2 / variance + log(variance)
+        diff = (target - pred_mean) ** 2
+        loss = 0.5 * (precision * diff + pred_logvar)
+
+        return loss.mean()
+
+
+class BetaNLLLoss(nn.Module):
+    """
+    Beta-Negative Log-Likelihood Loss.
+    Reference: Seitzer et al., "Properties and Benefits of Training with
+    Beta-Negative Log-Likelihood", ICLR 2022.
+    """
+
+    def __init__(self, beta=1.0, logvar_clipping=(-10, 10)):
+        super().__init__()
+        self.beta = beta
+        self.min_logvar, self.max_logvar = logvar_clipping
+
+    def forward(self, pred_mean, pred_logvar, target):
+        """
+        pred_mean: (B, num_genes)
+        pred_logvar: (B, num_genes)
+        target: (B, num_genes)
+        """
+        # 1. Clamp for stability
+        pred_logvar = torch.clamp(pred_logvar, self.min_logvar, self.max_logvar)
+
+        # 2. Calculate variance (sigma^2)
+        variance = torch.exp(pred_logvar)
+
+        # 3. Standard Gaussian NLL components
+        # (y - mu)^2 / sigma^2 + log(sigma^2)
+        squared_error = (target - pred_mean) ** 2
+        nll = (squared_error / variance) + pred_logvar
+
+        # 4. Apply Beta Weighting
+        # The 'stop_gradient' (detach) on the multiplier is optional
+        # but often used to stabilize the variance head.
+        # The paper version uses the variance gradient as well.
+        weight = variance.detach() ** self.beta
+
+        loss = 0.5 * weight * nll
+
+        return loss.mean()

src/simpleP2G_neigh.py

@@ -0,0 +1,92 @@
+"""SimpleP2G-cross: target tokens and spatially positioned neighbor tokens."""
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+from torch.utils.checkpoint import checkpoint
+
+from src.simpleP2G import SimpleP2GDecoderLayer, SimpleP2GRMSNorm
+
+
+class SimpleP2GNeigh(nn.Module):
+    """The target-plus-neighbors model used in the main comparison."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.decoder_depth = config.decoder_depth
+        self.num_genes = config.num_genes
+        self.hidden_size = config.hidden_size
+        self.cond_size = config.cond_size
+        self.num_cond_tokens = getattr(config, "num_cond_tokens", 4)
+        self.num_neighbor_tokens = int(getattr(config, "num_neighbor_tokens", 24))
+        if str(getattr(config, "neighbor_position_mode", "relative")) != "relative":
+            raise ValueError("SimpleP2G-cross requires relative neighbor positions")
+        for key in ("neighbor_cond_size", "global_cond_size"):
+            if int(getattr(config, key, self.cond_size)) != self.cond_size:
+                raise ValueError("All image embeddings must use the same encoder dimension")
+        if (
+            not getattr(config, "use_target_context", True)
+            or not getattr(config, "use_neighbor_context", True)
+            or getattr(config, "use_slide_context", False)
+            or getattr(config, "use_local_summary", False)
+            or getattr(config, "context_source_balance", 0.0) != 0
+        ):
+            raise ValueError(
+                "This model requires target + neighbor context; use the separate experiment entry point for other variants"
+            )
+        self.gene_queries = nn.Embedding(self.num_genes, self.hidden_size)
+        self.label_embed = nn.Sequential(
+            nn.Linear(self.cond_size, self.cond_size),
+            nn.SiLU(),
+            nn.Linear(self.cond_size, self.hidden_size * self.num_cond_tokens),
+        )
+        self.neighbor_proj = nn.Sequential(
+            nn.Linear(self.cond_size, self.hidden_size),
+            nn.SiLU(),
+            nn.Linear(self.hidden_size, self.hidden_size),
+        )
+        self.neighbor_pos_proj = nn.Sequential(
+            nn.Linear(2, self.hidden_size),
+            nn.SiLU(),
+            nn.Linear(self.hidden_size, self.hidden_size),
+        )
+        self.decoder_blocks = nn.ModuleList([SimpleP2GDecoderLayer(config) for _ in range(self.decoder_depth)])
+        self.final_norm = SimpleP2GRMSNorm(self.hidden_size)
+        self.mean_head = nn.Linear(self.hidden_size, 1)
+        if config.loss_name in ("gaussian_nll", "beta_nll"):
+            self.log_var_head = nn.Linear(self.hidden_size, 1)
+        self.gradient_checkpointing = False
+        self.emb_dropout = nn.Dropout(getattr(config, "dropout", 0.0))
+
+    def forward(self, c, neighbor_emb=None, neighbor_mask=None, neighbor_pos=None, global_emb=None):
+        """Predict from target (B,D), neighbors (B,K,D) and relative positions.
+
+        global_emb is accepted for the shared slide-dataset interface; the
+        main-comparison model does not consume a slide token.
+        """
+        if neighbor_emb is None or neighbor_pos is None:
+            raise ValueError("neighbor_emb and neighbor_pos are required")
+        batch_size = c.shape[0]
+        output = self.emb_dropout(self.gene_queries.weight.unsqueeze(0).expand(batch_size, -1, -1))
+        target = self.label_embed(c).view(batch_size, self.num_cond_tokens, self.hidden_size)
+        neighbors = self.neighbor_proj(neighbor_emb)
+        neighbors = neighbors + self.neighbor_pos_proj(neighbor_pos.to(dtype=neighbors.dtype))
+        condition = torch.cat((target, neighbors), dim=1)
+        if neighbor_mask is None:
+            neighbor_mask = torch.ones(neighbor_emb.shape[:2], device=c.device, dtype=torch.bool)
+        target_mask = torch.ones(batch_size, self.num_cond_tokens, device=c.device, dtype=torch.bool)
+        mask = torch.cat((target_mask, neighbor_mask.bool()), dim=1).unsqueeze(1).unsqueeze(1)
+        for block in self.decoder_blocks:
+            if self.gradient_checkpointing and self.training:
+                output = checkpoint(block, output, condition, None, mask, use_reentrant=False)
+            else:
+                output = block(output, encoder_hidden_states=condition, cross_attention_mask=mask)
+        output = self.final_norm(output)
+        mean = F.softplus(self.mean_head(output).squeeze(-1))
+        log_var = (
+            self.log_var_head(output).squeeze(-1)
+            if self.config.loss_name in ("gaussian_nll", "beta_nll")
+            else torch.zeros_like(mean)
+        )
+        return mean, log_var