Initial release: trained PLASMA heads (21 task x backbone variants) + model card

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+assets/visual_abstract.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,184 @@
+---
+license: mit
+library_name: plasma-protein-local-alignment
+pipeline_tag: feature-extraction
+tags:
+  - protein
+  - protein-language-model
+  - alignment
+  - optimal-transport
+  - sinkhorn
+  - bioinformatics
+  - biology
+---
+
+# PLASMA: Pluggable Local Alignment via Sinkhorn MAtrix
+
+[![arXiv](https://img.shields.io/badge/arXiv-2510.11752-b31b1b.svg)](https://arxiv.org/abs/2510.11752)
+[![ICLR 2026](https://img.shields.io/badge/ICLR-2026-blue.svg)](https://arxiv.org/abs/2510.11752)
+[![GitHub stars](https://img.shields.io/github/stars/ZW471/PLASMA-Protein-Local-Alignment?style=social)](https://github.com/ZW471/PLASMA-Protein-Local-Alignment)
+
+![Visual abstract](assets/visual_abstract.png)
+
+**PLASMA** is a tiny, pluggable head that turns any frozen protein-language-model
+(PLM) into a residue-level *local* aligner. It reformulates protein substructure
+alignment as a regularised optimal transport problem and runs ~50× faster than
+structure-based aligners (TM-Align, Foldseek) by operating on pre-computed
+embeddings.
+
+This repository hosts the trained **PLASMA** heads for every (task, backbone)
+combination from the paper, plus instructions for the parameter-free
+**PLASMA-PF** baseline (which has no learned weights). PLASMA was published at
+**ICLR 2026**.
+
+- **Paper:** <https://arxiv.org/abs/2510.11752> (ICLR 2026)
+- **Code:** <https://github.com/ZW471/PLASMA-Protein-Local-Alignment>
+- **License:** MIT
+
+## What's in this repo
+
+Each variant lives in its own subfolder and is loaded by the `load_plasma`
+helper from the GitHub package:
+
+```
+weights/
+  active_site/
+    prot_bert/                 # config.json + model.safetensors + metadata.json
+    ankh-base/
+    TM-Vec/
+    ProstT5/
+    prot_t5_xl_half_uniref50-enc/
+    esm2_t33_650M_UR50D/
+    ProtSSN/
+  binding_site/
+    ...
+  motif/
+    ...
+```
+
+All heads share the same architecture: a small `LRL` non-linearity
+(`LazyLinear → ReLU → Linear → LayerNorm`, hidden dim 512) followed by a
+parameter-free Sinkhorn iteration (`temperature=0.1`, `n_iters=20`). The
+checkpoint files are ~3 MB each.
+
+## Quickstart
+
+Install the PLASMA package from source (the model class is shipped with the
+GitHub repo):
+
+```bash
+git clone https://github.com/ZW471/PLASMA-Protein-Local-Alignment
+cd PLASMA-Protein-Local-Alignment
+uv sync
+```
+
+Then load any trained head with the high-level helper:
+
+```python
+import torch
+from alignment import load_plasma
+
+model = load_plasma(task="active_site", backbone="prot_bert")
+model.eval()
+
+# Feed pre-computed AA-level embeddings from the matching backbone.
+# H_q / H_c are residue-level embeddings; batch_q / batch_c assign each
+# residue to a sample (use zeros if you only have one pair).
+H_q = torch.randn(120, 1024)            # query: 120 residues, ProtBERT dim
+H_c = torch.randn(180, 1024)            # candidate: 180 residues
+batch_q = torch.zeros(120, dtype=torch.long)
+batch_c = torch.zeros(180, dtype=torch.long)
+
+with torch.no_grad():
+    alignment_matrix = model(H_q, H_c, batch_q, batch_c)  # (120, 180)
+```
+
+The output is a doubly-stochastic transport plan describing the residue-level
+correspondence between the two substructures. To reduce it to a similarity
+score, reuse `utils.alignment_score` from the GitHub repo (it applies the
+diagonal convolution + threshold described in the paper).
+
+## PLASMA-PF (parameter-free)
+
+PLASMA-PF is a hinge / Sinkhorn baseline with **no learned weights**. There is
+nothing to download — just instantiate it from the same `Alignment` class:
+
+```python
+from alignment import load_plasma_pf
+
+model = load_plasma_pf()  # Alignment(eta='hinge', omega='sinkhorn', ...)
+```
+
+It accepts the same forward signature as the trained heads above.
+
+## Available variants & evaluation results
+
+Numbers below are 3-seed averages (mean ± std) reported in the paper. The seven
+backbone columns correspond to the seven subfolders under each task.
+
+### Interpolation (in-distribution test split)
+
+| Task | Metric | Ankh | ESM-2 | ProstT5 | ProtBERT | ProtSSN | ProtT5 | TM-Vec |
+| --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| **Motif** | ROC-AUC | .925 ± .002 | .933 ± .005 | .954 ± .002 | .854 ± .003 | .922 ± .002 | **.972 ± .001** | .910 ± .003 |
+|  | F1-Max | .885 ± .002 | .877 ± .005 | .885 ± .003 | .784 ± .002 | .866 ± .002 | **.918 ± .003** | .853 ± .003 |
+|  | PR-AUC | .921 ± .002 | .931 ± .004 | .953 ± .003 | .872 ± .003 | .920 ± .002 | **.971 ± .002** | .914 ± .003 |
+|  | Label Match Score | .921 ± .004 | .890 ± .008 | .929 ± .001 | .746 ± .007 | .767 ± .008 | **.937 ± .001** | .792 ± .008 |
+| **Binding Site** | ROC-AUC | **.995 ± .000** | .992 ± .000 | .993 ± .001 | .981 ± .001 | .992 ± .001 | .993 ± .000 | .980 ± .001 |
+|  | F1-Max | .987 ± .001 | .986 ± .001 | .983 ± .001 | .948 ± .002 | .982 ± .001 | **.988 ± .001** | .970 ± .001 |
+|  | PR-AUC | **.996 ± .001** | .994 ± .001 | .995 ± .001 | .985 ± .001 | .993 ± .001 | .995 ± .000 | .984 ± .001 |
+|  | Label Match Score | **.951 ± .002** | .950 ± .002 | **.951 ± .002** | .880 ± .008 | .872 ± .005 | **.951 ± .001** | .900 ± .004 |
+| **Active Site** | ROC-AUC | **.994 ± .001** | .991 ± .001 | .993 ± .001 | .986 ± .001 | .992 ± .001 | **.994 ± .001** | .991 ± .001 |
+|  | F1-Max | **.989 ± .001** | .985 ± .001 | .987 ± .001 | .967 ± .001 | .987 ± .001 | .987 ± .001 | .982 ± .001 |
+|  | PR-AUC | **.994 ± .001** | .992 ± .001 | **.994 ± .001** | .988 ± .001 | **.994 ± .001** | **.994 ± .001** | .992 ± .001 |
+|  | Label Match Score | **.975 ± .001** | .969 ± .002 | **.975 ± .001** | .904 ± .003 | .885 ± .013 | .972 ± .001 | .938 ± .001 |
+
+### Extrapolation (held-out hard test split)
+
+| Task | Metric | Ankh | ESM-2 | ProstT5 | ProtBERT | ProtSSN | ProtT5 | TM-Vec |
+| --- | --- | ---: | ---: | ---: | ---: | ---: | ---: | ---: |
+| **Motif** | ROC-AUC | .960 ± .011 | .972 ± .010 | **.975 ± .009** | .870 ± .030 | .949 ± .013 | .968 ± .012 | .954 ± .013 |
+|  | F1-Max | .915 ± .021 | **.931 ± .016** | .926 ± .020 | .799 ± .039 | .896 ± .023 | .922 ± .023 | .903 ± .026 |
+|  | PR-AUC | .948 ± .020 | **.970 ± .010** | .969 ± .016 | .873 ± .036 | .940 ± .020 | .962 ± .018 | .944 ± .022 |
+|  | Label Match Score | **.842 ± .025** | .786 ± .032 | .801 ± .022 | .541 ± .060 | .537 ± .025 | .738 ± .028 | .704 ± .020 |
+| **Binding Site** | ROC-AUC | .995 ± .005 | **.999 ± .001** | .993 ± .005 | .951 ± .014 | **.999 ± .001** | **.999 ± .001** | .990 ± .008 |
+|  | F1-Max | .992 ± .005 | .991 ± .005 | .985 ± .009 | .896 ± .019 | .988 ± .006 | **.996 ± .003** | .983 ± .011 |
+|  | PR-AUC | .997 ± .003 | **.999 ± .001** | .995 ± .003 | .958 ± .012 | .998 ± .001 | **.999 ± .000** | .992 ± .006 |
+|  | Label Match Score | .894 ± .026 | .851 ± .031 | .891 ± .029 | .603 ± .041 | .753 ± .041 | **.902 ± .019** | .824 ± .031 |
+| **Active Site** | ROC-AUC | .995 ± .002 | .996 ± .003 | .996 ± .003 | .980 ± .004 | .997 ± .001 | **.999 ± .000** | .995 ± .002 |
+|  | F1-Max | **.992 ± .002** | .986 ± .004 | .991 ± .004 | .950 ± .005 | .991 ± .003 | .991 ± .002 | .985 ± .003 |
+|  | PR-AUC | .995 ± .003 | .997 ± .002 | .997 ± .002 | .984 ± .003 | .998 ± .001 | **.999 ± .000** | .996 ± .002 |
+|  | Label Match Score | **.938 ± .014** | .882 ± .027 | .931 ± .026 | .697 ± .019 | .737 ± .011 | .893 ± .017 | .880 ± .023 |
+
+Each subfolder also contains a `metadata.json` with the full hyperparameter
+config in machine-readable form.
+
+## Training details
+
+- **Architecture:** `Alignment(eta='lrl', omega='sinkhorn',
+  eta_kwargs={'hidden_dim': 512},
+  omega_kwargs={'temperature': 0.1, 'n_iters': 20})`.
+- **Score head:** `K=10`, `threshold=0.5` (used by
+  `utils.alignment_score` to reduce the transport plan to a scalar).
+- **Optimiser / loss:** Adam (`lr=1e-4`), `BCEWithLogitsLoss` on the alignment
+  score plus a label-match auxiliary loss (`target_loss_weight=1.0`).
+- **Data:** the InterPro-derived motif / binding-site / active-site datasets
+  shipped under `data/raw/` in the GitHub repo, split into train / validation /
+  test / test-hard with `dataset_fraction=0.1` (default sweep) and
+  `dataset_fraction=1.0` (full sweep — checkpoints here are from the full
+  sweep).
+- **Selection metric:** validation loss (early stopping, `patience=3`).
+
+## Citation
+
+If you use these weights, please cite the PLASMA paper:
+
+```bibtex
+@inproceedings{wang2026plasma,
+  title     = {Fast and Interpretable Protein Substructure Alignment via Optimal Transport},
+  author    = {Wang, Zhiyu and Zhou, Bingxin and Wang, Jing and Tan, Yang and Zhao, Weishu and Li{\`o}, Pietro and Hong, Liang},
+  booktitle = {International Conference on Learning Representations (ICLR)},
+  year      = {2026},
+  url       = {https://arxiv.org/abs/2510.11752},
+}
+```

weights/active_site/ProstT5/config.json

@@ -0,0 +1,11 @@
+{
+  "eta": "lrl",
+  "eta_kwargs": {
+    "hidden_dim": 512
+  },
+  "omega": "sinkhorn",
+  "omega_kwargs": {
+    "n_iters": 20,
+    "temperature": 0.1
+  }
+}

weights/active_site/ProstT5/metadata.json

@@ -0,0 +1,33 @@
+{
+  "task": "active_site",
+  "backbone": "ProstT5",
+  "split": 0,
+  "eta_config": {
+    "type": "lrl",
+    "hidden_dim": 512
+  },
+  "omega_config": {
+    "type": "sinkhorn",
+    "temperature": 0.1,
+    "n_iters": 20
+  },
+  "score_config": {
+    "K": 10,
+    "threshold": 0.5
+  },
+  "metrics": {
+    "test_frequent": {
+      "rocauc": 0.992344856262207,
+      "f1_max": 0.9862671660424469,
+      "pr_auc": 0.9944091439247131,
+      "label_match_score": 0.9720931127164464
+    },
+    "test_hard": {
+      "rocauc": 0.9999887347221375,
+      "f1_max": 0.9987496874218554,
+      "pr_auc": 0.9999887943267822,
+      "label_match_score": 0.9361866737141609
+    }
+  },
+  "source_checkpoint": "sweeps/train/all/active_site/ProstT5/27/active_site_split0_best.pt"
+}

weights/active_site/ProstT5/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e65cdc9a5beb16a5f25be6f4dbb788481d16acece4a83056ed59827bb21a938c
+size 3154416

weights/active_site/ProtSSN/config.json

@@ -0,0 +1,11 @@
+{
+  "eta": "lrl",
+  "eta_kwargs": {
+    "hidden_dim": 512
+  },
+  "omega": "sinkhorn",
+  "omega_kwargs": {
+    "n_iters": 20,
+    "temperature": 0.1
+  }
+}

weights/active_site/ProtSSN/metadata.json

@@ -0,0 +1,33 @@
+{
+  "task": "active_site",
+  "backbone": "ProtSSN",
+  "split": 0,
+  "eta_config": {
+    "type": "lrl",
+    "hidden_dim": 512
+  },
+  "omega_config": {
+    "type": "sinkhorn",
+    "temperature": 0.1,
+    "n_iters": 20
+  },
+  "score_config": {
+    "K": 10,
+    "threshold": 0.5
+  },
+  "metrics": {
+    "test_frequent": {
+      "rocauc": 0.9907435178756714,
+      "f1_max": 0.9846695149535059,
+      "pr_auc": 0.9934981465339661,
+      "label_match_score": 0.8660282360271119
+    },
+    "test_hard": {
+      "rocauc": 0.9999032616615295,
+      "f1_max": 0.9965,
+      "pr_auc": 0.9999053478240967,
+      "label_match_score": 0.7096276976610534
+    }
+  },
+  "source_checkpoint": "sweeps/train/all/active_site/ProtSSN/54/active_site_split0_best.pt"
+}

weights/active_site/ProtSSN/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:21388d6626e4faa3bc260f1f6439cd66dbc03710052aff49b2111e62a1444bdb
+size 3678704

weights/active_site/TM-Vec/config.json

@@ -0,0 +1,11 @@
+{
+  "eta": "lrl",
+  "eta_kwargs": {
+    "hidden_dim": 512
+  },
+  "omega": "sinkhorn",
+  "omega_kwargs": {
+    "n_iters": 20,
+    "temperature": 0.1
+  }
+}

weights/active_site/TM-Vec/metadata.json

@@ -0,0 +1,33 @@
+{
+  "task": "active_site",
+  "backbone": "TM-Vec",
+  "split": 0,
+  "eta_config": {
+    "type": "lrl",
+    "hidden_dim": 512
+  },
+  "omega_config": {
+    "type": "sinkhorn",
+    "temperature": 0.1,
+    "n_iters": 20
+  },
+  "score_config": {
+    "K": 10,
+    "threshold": 0.5
+  },
+  "metrics": {
+    "test_frequent": {
+      "rocauc": 0.9918658137321472,
+      "f1_max": 0.9819729594391587,
+      "pr_auc": 0.9937556982040405,
+      "label_match_score": 0.9401966694747831
+    },
+    "test_hard": {
+      "rocauc": 0.9987612366676331,
+      "f1_max": 0.9860865165696939,
+      "pr_auc": 0.9988466501235962,
+      "label_match_score": 0.854244963310197
+    }
+  },
+  "source_checkpoint": "sweeps/train/all/active_site/TM-Vec/18/active_site_split0_best.pt"
+}

weights/active_site/TM-Vec/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5dc025d05e15cd99328dfd341064756f01efa5ff8aa331e7622bf3657e9e4c6d
+size 3154416

weights/active_site/ankh-base/config.json

@@ -0,0 +1,11 @@
+{
+  "eta": "lrl",
+  "eta_kwargs": {
+    "hidden_dim": 512
+  },
+  "omega": "sinkhorn",
+  "omega_kwargs": {
+    "n_iters": 20,
+    "temperature": 0.1
+  }
+}

weights/active_site/ankh-base/metadata.json

@@ -0,0 +1,33 @@
+{
+  "task": "active_site",
+  "backbone": "ankh-base",
+  "split": 0,
+  "eta_config": {
+    "type": "lrl",
+    "hidden_dim": 512
+  },
+  "omega_config": {
+    "type": "sinkhorn",
+    "temperature": 0.1,
+    "n_iters": 20
+  },
+  "score_config": {
+    "K": 10,
+    "threshold": 0.5
+  },
+  "metrics": {
+    "test_frequent": {
+      "rocauc": 0.9938057661056519,
+      "f1_max": 0.9892041174993723,
+      "pr_auc": 0.9950824975967407,
+      "label_match_score": 0.9734372955750135
+    },
+    "test_hard": {
+      "rocauc": 0.9999867081642151,
+      "f1_max": 0.9987503124218945,
+      "pr_auc": 0.9999868869781494,
+      "label_match_score": 0.9452198707803406
+    }
+  },
+  "source_checkpoint": "sweeps/train/all/active_site/ankh-base/9/active_site_split0_best.pt"
+}

weights/active_site/ankh-base/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:acf83dac2b7ed05b521bb659492aa192914222424f96107cd64a6f3a467cd182
+size 2630128

weights/active_site/esm2_t33_650M_UR50D/config.json

@@ -0,0 +1,11 @@
+{
+  "eta": "lrl",
+  "eta_kwargs": {
+    "hidden_dim": 512
+  },
+  "omega": "sinkhorn",
+  "omega_kwargs": {
+    "n_iters": 20,
+    "temperature": 0.1
+  }
+}

weights/active_site/esm2_t33_650M_UR50D/metadata.json

@@ -0,0 +1,33 @@
+{
+  "task": "active_site",
+  "backbone": "esm2_t33_650M_UR50D",
+  "split": 0,
+  "eta_config": {
+    "type": "lrl",
+    "hidden_dim": 512
+  },
+  "omega_config": {
+    "type": "sinkhorn",
+    "temperature": 0.1,
+    "n_iters": 20
+  },
+  "score_config": {
+    "K": 10,
+    "threshold": 0.5
+  },
+  "metrics": {
+    "test_frequent": {
+      "rocauc": 0.9898462891578674,
+      "f1_max": 0.985478217325989,
+      "pr_auc": 0.9927905797958374,
+      "label_match_score": 0.966968783440344
+    },
+    "test_hard": {
+      "rocauc": 0.9998506903648376,
+      "f1_max": 0.9959939909864797,
+      "pr_auc": 0.9998576045036316,
+      "label_match_score": 0.8552783124974398
+    }
+  },
+  "source_checkpoint": "sweeps/train/all/active_site/esm2_t33_650M_UR50D/45/active_site_split0_best.pt"
+}

weights/active_site/esm2_t33_650M_UR50D/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:543c89f2bcd6b30fbadc7acdbfef207ef285f73aa90ef48bff4db6b4b871b5f0
+size 3678704

weights/active_site/prot_bert/config.json

@@ -0,0 +1,11 @@
+{
+  "eta": "lrl",
+  "eta_kwargs": {
+    "hidden_dim": 512
+  },
+  "omega": "sinkhorn",
+  "omega_kwargs": {
+    "n_iters": 20,
+    "temperature": 0.1
+  }
+}

weights/active_site/prot_bert/metadata.json

@@ -0,0 +1,33 @@
+{
+  "task": "active_site",
+  "backbone": "prot_bert",
+  "split": 0,
+  "eta_config": {
+    "type": "lrl",
+    "hidden_dim": 512
+  },
+  "omega_config": {
+    "type": "sinkhorn",
+    "temperature": 0.1,
+    "n_iters": 20
+  },
+  "score_config": {
+    "K": 10,
+    "threshold": 0.5
+  },
+  "metrics": {
+    "test_frequent": {
+      "rocauc": 0.9855157732963562,
+      "f1_max": 0.9628886659979939,
+      "pr_auc": 0.9884096384048462,
+      "label_match_score": 0.910069686397606
+    },
+    "test_hard": {
+      "rocauc": 0.9875518083572388,
+      "f1_max": 0.9543965734441925,
+      "pr_auc": 0.9901764988899231,
+      "label_match_score": 0.6242303718695857
+    }
+  },
+  "source_checkpoint": "sweeps/train/all/active_site/prot_bert/0/active_site_split0_best.pt"
+}

weights/active_site/prot_bert/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6192f5a73807ff7721cfda218a7d8e88b56772cd4476868ed9781e73305c343
+size 3154416

weights/active_site/prot_t5_xl_half_uniref50-enc/config.json

@@ -0,0 +1,11 @@
+{
+  "eta": "lrl",
+  "eta_kwargs": {
+    "hidden_dim": 512
+  },
+  "omega": "sinkhorn",
+  "omega_kwargs": {
+    "n_iters": 20,
+    "temperature": 0.1
+  }
+}

weights/active_site/prot_t5_xl_half_uniref50-enc/metadata.json

@@ -0,0 +1,33 @@
+{
+  "task": "active_site",
+  "backbone": "prot_t5_xl_half_uniref50-enc",
+  "split": 0,
+  "eta_config": {
+    "type": "lrl",
+    "hidden_dim": 512
+  },
+  "omega_config": {
+    "type": "sinkhorn",
+    "temperature": 0.1,
+    "n_iters": 20
+  },
+  "score_config": {
+    "K": 10,
+    "threshold": 0.5
+  },
+  "metrics": {
+    "test_frequent": {
+      "rocauc": 0.9920676946640015,
+      "f1_max": 0.984769038701623,
+      "pr_auc": 0.9936845302581787,
+      "label_match_score": 0.9688430294838205
+    },
+    "test_hard": {
+      "rocauc": 0.9998884797096252,
+      "f1_max": 0.9960059910134798,
+      "pr_auc": 0.999886691570282,
+      "label_match_score": 0.8415421822864932
+    }
+  },
+  "source_checkpoint": "sweeps/train/all/active_site/prot_t5_xl_half_uniref50-enc/36/active_site_split0_best.pt"
+}

weights/active_site/prot_t5_xl_half_uniref50-enc/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4f629be09a89aee0e13bea4d420878e017e68415b386e70eb25114b2ed70d72
+size 3154416

weights/binding_site/ProstT5/config.json

@@ -0,0 +1,11 @@
+{
+  "eta": "lrl",
+  "eta_kwargs": {
+    "hidden_dim": 512
+  },
+  "omega": "sinkhorn",
+  "omega_kwargs": {
+    "n_iters": 20,
+    "temperature": 0.1
+  }
+}

weights/binding_site/ProstT5/metadata.json

@@ -0,0 +1,33 @@
+{
+  "task": "binding_site",
+  "backbone": "ProstT5",
+  "split": 0,
+  "eta_config": {
+    "type": "lrl",
+    "hidden_dim": 512
+  },
+  "omega_config": {
+    "type": "sinkhorn",
+    "temperature": 0.1,
+    "n_iters": 20
+  },
+  "score_config": {
+    "K": 10,
+    "threshold": 0.5
+  },
+  "metrics": {
+    "test_frequent": {
+      "rocauc": 0.9942386150360107,
+      "f1_max": 0.9826327712056381,
+      "pr_auc": 0.9957546591758728,
+      "label_match_score": 0.9467616829861532
+    },
+    "test_hard": {
+      "rocauc": 0.9979713559150696,
+      "f1_max": 0.9945945945945946,
+      "pr_auc": 0.9983985424041748,
+      "label_match_score": 0.9251244336550384
+    }
+  },
+  "source_checkpoint": "sweeps/train/all/binding_site/ProstT5/30/binding_site_split0_best.pt"
+}

weights/binding_site/ProstT5/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0af354b1a2e012a204766eacbc66e708dc0978b23814d95e04dea87396f8724
+size 3154416

weights/binding_site/ProtSSN/config.json

@@ -0,0 +1,11 @@
+{
+  "eta": "lrl",
+  "eta_kwargs": {
+    "hidden_dim": 512
+  },
+  "omega": "sinkhorn",
+  "omega_kwargs": {
+    "n_iters": 20,
+    "temperature": 0.1
+  }
+}

weights/binding_site/ProtSSN/metadata.json

@@ -0,0 +1,33 @@
+{
+  "task": "binding_site",
+  "backbone": "ProtSSN",
+  "split": 0,
+  "eta_config": {
+    "type": "lrl",
+    "hidden_dim": 512
+  },
+  "omega_config": {
+    "type": "sinkhorn",
+    "temperature": 0.1,
+    "n_iters": 20
+  },
+  "score_config": {
+    "K": 10,
+    "threshold": 0.5
+  },
+  "metrics": {
+    "test_frequent": {
+      "rocauc": 0.9934399724006653,
+      "f1_max": 0.9803625377643505,
+      "pr_auc": 0.9950706362724304,
+      "label_match_score": 0.8582753864622367
+    },
+    "test_hard": {
+      "rocauc": 0.9999507665634155,
+      "f1_max": 0.9971223021582734,
+      "pr_auc": 0.9999310374259949,
+      "label_match_score": 0.785670741250131
+    }
+  },
+  "source_checkpoint": "sweeps/train/all/binding_site/ProtSSN/57/binding_site_split0_best.pt"
+}

weights/binding_site/ProtSSN/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5600eb5599169bfbb0214c707e68a20e9377b1244f6e03d1a5b2a21f8879c9c1
+size 3678704

weights/binding_site/TM-Vec/config.json

@@ -0,0 +1,11 @@
+{
+  "eta": "lrl",
+  "eta_kwargs": {
+    "hidden_dim": 512
+  },
+  "omega": "sinkhorn",
+  "omega_kwargs": {
+    "n_iters": 20,
+    "temperature": 0.1
+  }
+}

weights/binding_site/TM-Vec/metadata.json

@@ -0,0 +1,33 @@
+{
+  "task": "binding_site",
+  "backbone": "TM-Vec",
+  "split": 0,
+  "eta_config": {
+    "type": "lrl",
+    "hidden_dim": 512
+  },
+  "omega_config": {
+    "type": "sinkhorn",
+    "temperature": 0.1,
+    "n_iters": 20
+  },
+  "score_config": {
+    "K": 10,
+    "threshold": 0.5
+  },
+  "metrics": {
+    "test_frequent": {
+      "rocauc": 0.9777182936668396,
+      "f1_max": 0.9658886894075404,
+      "pr_auc": 0.9825307726860046,
+      "label_match_score": 0.8937117450104289
+    },
+    "test_hard": {
+      "rocauc": 0.9969605207443237,
+      "f1_max": 0.9967567567567568,
+      "pr_auc": 0.997014045715332,
+      "label_match_score": 0.8630872951235824
+    }
+  },
+  "source_checkpoint": "sweeps/train/all/binding_site/TM-Vec/21/binding_site_split0_best.pt"
+}

weights/binding_site/TM-Vec/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7c955e1c5e2b290f849037b2557a55ebcfc425b864e98a077646788c0bf0505
+size 3154416

weights/binding_site/ankh-base/config.json

@@ -0,0 +1,11 @@
+{
+  "eta": "lrl",
+  "eta_kwargs": {
+    "hidden_dim": 512
+  },
+  "omega": "sinkhorn",
+  "omega_kwargs": {
+    "n_iters": 20,
+    "temperature": 0.1
+  }
+}

weights/binding_site/ankh-base/metadata.json

@@ -0,0 +1,33 @@
+{
+  "task": "binding_site",
+  "backbone": "ankh-base",
+  "split": 0,
+  "eta_config": {
+    "type": "lrl",
+    "hidden_dim": 512
+  },
+  "omega_config": {
+    "type": "sinkhorn",
+    "temperature": 0.1,
+    "n_iters": 20
+  },
+  "score_config": {
+    "K": 10,
+    "threshold": 0.5
+  },
+  "metrics": {
+    "test_frequent": {
+      "rocauc": 0.9952087998390198,
+      "f1_max": 0.9896803423105965,
+      "pr_auc": 0.9964244365692139,
+      "label_match_score": 0.9444196989256388
+    },
+    "test_hard": {
+      "rocauc": 0.9999386072158813,
+      "f1_max": 0.9978432782171099,
+      "pr_auc": 0.9999148845672607,
+      "label_match_score": 0.938885366370608
+    }
+  },
+  "source_checkpoint": "sweeps/train/all/binding_site/ankh-base/12/binding_site_split0_best.pt"
+}

weights/binding_site/ankh-base/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9eb81c0ada0185276791992bc1e66a558970a009b64e78ca7dbd664532a51d93
+size 2630128

weights/binding_site/esm2_t33_650M_UR50D/config.json

@@ -0,0 +1,11 @@
+{
+  "eta": "lrl",
+  "eta_kwargs": {
+    "hidden_dim": 512
+  },
+  "omega": "sinkhorn",
+  "omega_kwargs": {
+    "n_iters": 20,
+    "temperature": 0.1
+  }
+}

weights/binding_site/esm2_t33_650M_UR50D/metadata.json

@@ -0,0 +1,33 @@
+{
+  "task": "binding_site",
+  "backbone": "esm2_t33_650M_UR50D",
+  "split": 0,
+  "eta_config": {
+    "type": "lrl",
+    "hidden_dim": 512
+  },
+  "omega_config": {
+    "type": "sinkhorn",
+    "temperature": 0.1,
+    "n_iters": 20
+  },
+  "score_config": {
+    "K": 10,
+    "threshold": 0.5
+  },
+  "metrics": {
+    "test_frequent": {
+      "rocauc": 0.9934264421463013,
+      "f1_max": 0.9860865165696939,
+      "pr_auc": 0.9950663447380066,
+      "label_match_score": 0.949677944527228
+    },
+    "test_hard": {
+      "rocauc": 0.999835193157196,
+      "f1_max": 0.9935344827586207,
+      "pr_auc": 0.9997715950012207,
+      "label_match_score": 0.8993046989344214
+    }
+  },
+  "source_checkpoint": "sweeps/train/all/binding_site/esm2_t33_650M_UR50D/48/binding_site_split0_best.pt"
+}

weights/binding_site/esm2_t33_650M_UR50D/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:56f464b1d41e8922f75ae8e7a727a32cb7bcc086f95dec46d7f13213d61c014d
+size 3678704

weights/binding_site/prot_bert/config.json

@@ -0,0 +1,11 @@
+{
+  "eta": "lrl",
+  "eta_kwargs": {
+    "hidden_dim": 512
+  },
+  "omega": "sinkhorn",
+  "omega_kwargs": {
+    "n_iters": 20,
+    "temperature": 0.1
+  }
+}

weights/binding_site/prot_bert/metadata.json

@@ -0,0 +1,33 @@
+{
+  "task": "binding_site",
+  "backbone": "prot_bert",
+  "split": 0,
+  "eta_config": {
+    "type": "lrl",
+    "hidden_dim": 512
+  },
+  "omega_config": {
+    "type": "sinkhorn",
+    "temperature": 0.1,
+    "n_iters": 20
+  },
+  "score_config": {
+    "K": 10,
+    "threshold": 0.5
+  },
+  "metrics": {
+    "test_frequent": {
+      "rocauc": 0.9826457500457764,
+      "f1_max": 0.9522342064714946,
+      "pr_auc": 0.9867449402809143,
+      "label_match_score": 0.887023501667951
+    },
+    "test_hard": {
+      "rocauc": 0.9887794256210327,
+      "f1_max": 0.9510791366906475,
+      "pr_auc": 0.9874331951141357,
+      "label_match_score": 0.7587943521090054
+    }
+  },
+  "source_checkpoint": "sweeps/train/all/binding_site/prot_bert/3/binding_site_split0_best.pt"
+}

weights/binding_site/prot_bert/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:69692d6e35fd438bdcf63c710b130636da2e495f015568e8d38820d7d86e6c06
+size 3154416

weights/binding_site/prot_t5_xl_half_uniref50-enc/config.json

@@ -0,0 +1,11 @@
+{
+  "eta": "lrl",
+  "eta_kwargs": {
+    "hidden_dim": 512
+  },
+  "omega": "sinkhorn",
+  "omega_kwargs": {
+    "n_iters": 20,
+    "temperature": 0.1
+  }
+}

weights/binding_site/prot_t5_xl_half_uniref50-enc/metadata.json

@@ -0,0 +1,33 @@
+{
+  "task": "binding_site",
+  "backbone": "prot_t5_xl_half_uniref50-enc",
+  "split": 0,
+  "eta_config": {
+    "type": "lrl",
+    "hidden_dim": 512
+  },
+  "omega_config": {
+    "type": "sinkhorn",
+    "temperature": 0.1,
+    "n_iters": 20
+  },
+  "score_config": {
+    "K": 10,
+    "threshold": 0.5
+  },
+  "metrics": {
+    "test_frequent": {
+      "rocauc": 0.9949175715446472,
+      "f1_max": 0.991672975018925,
+      "pr_auc": 0.9963440895080566,
+      "label_match_score": 0.9517851873361006
+    },
+    "test_hard": {
+      "rocauc": 0.9999374151229858,
+      "f1_max": 0.9967590925459129,
+      "pr_auc": 0.9999114274978638,
+      "label_match_score": 0.8924064511657084
+    }
+  },
+  "source_checkpoint": "sweeps/train/all/binding_site/prot_t5_xl_half_uniref50-enc/39/binding_site_split0_best.pt"
+}

weights/binding_site/prot_t5_xl_half_uniref50-enc/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70c7184ded2cb29cbfba1521c4fb651514231630d5db2875c441ef64f9fa80ac
+size 3154416

weights/motif/ProstT5/config.json

@@ -0,0 +1,11 @@
+{
+  "eta": "lrl",
+  "eta_kwargs": {
+    "hidden_dim": 512
+  },
+  "omega": "sinkhorn",
+  "omega_kwargs": {
+    "n_iters": 20,
+    "temperature": 0.1
+  }
+}

weights/motif/ProstT5/metadata.json

@@ -0,0 +1,33 @@
+{
+  "task": "motif",
+  "backbone": "ProstT5",
+  "split": 0,
+  "eta_config": {
+    "type": "lrl",
+    "hidden_dim": 512
+  },
+  "omega_config": {
+    "type": "sinkhorn",
+    "temperature": 0.1,
+    "n_iters": 20
+  },
+  "score_config": {
+    "K": 10,
+    "threshold": 0.5
+  },
+  "metrics": {
+    "test_frequent": {
+      "rocauc": 0.9485077261924744,
+      "f1_max": 0.8776619845944721,
+      "pr_auc": 0.9461915493011475,
+      "label_match_score": 0.9349878941312089
+    },
+    "test_hard": {
+      "rocauc": 0.9987444877624512,
+      "f1_max": 0.9784646567012921,
+      "pr_auc": 0.9987750053405762,
+      "label_match_score": 0.7881006840105947
+    }
+  },
+  "source_checkpoint": "sweeps/train/all/motif/ProstT5/33/motif_split0_best.pt"
+}

weights/motif/ProstT5/model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ff0edb123312a3832f686a28a1dc00996af04918c7bf83c333842a1bff11407
+size 3154416

weights/motif/ProtSSN/config.json

@@ -0,0 +1,11 @@
+{
+  "eta": "lrl",
+  "eta_kwargs": {
+    "hidden_dim": 512
+  },
+  "omega": "sinkhorn",
+  "omega_kwargs": {
+    "n_iters": 20,
+    "temperature": 0.1
+  }
+}

weights/motif/ProtSSN/metadata.json

@@ -0,0 +1,33 @@
+{
+  "task": "motif",
+  "backbone": "ProtSSN",
+  "split": 0,
+  "eta_config": {
+    "type": "lrl",
+    "hidden_dim": 512
+  },
+  "omega_config": {
+    "type": "sinkhorn",
+    "temperature": 0.1,
+    "n_iters": 20
+  },
+  "score_config": {
+    "K": 10,
+    "threshold": 0.5
+  },
+  "metrics": {
+    "test_frequent": {
+      "rocauc": 0.9163562655448914,
+      "f1_max": 0.8624229979466119,
+      "pr_auc": 0.913853108882904,
+      "label_match_score": 0.7653402222577776
+    },
+    "test_hard": {
+      "rocauc": 0.9953680038452148,
+      "f1_max": 0.9669315560112791,
+      "pr_auc": 0.9959331750869751,
+      "label_match_score": 0.5649718196333852
+    }
+  },
+  "source_checkpoint": "sweeps/train/all/motif/ProtSSN/60/motif_split0_best.pt"
+}