Add LogiCA pretrained backbones (8M/35M/150M/650M) + model card

README.md

@@ -0,0 +1,65 @@
+---
+license: mit
+library_name: pytorch
+tags:
+  - protein-ligand
+  - drug-target-interaction
+  - contrastive-learning
+  - esm2
+  - selformer
+  - logica
+---
+
+# LogiCA — pretrained protein–ligand backbones
+
+Pretrained checkpoints for **LogiCA** (*Contextualizing Biological Language Models
+across Modalities via Logit-Space Contrastive Alignment*). Each checkpoint is a
+bidirectional ESM-2 (protein) + SELFormer (ligand) model with a cross-attention
+adapter, pretrained on BindingDB with the logit-space contrastive objective
+(`--mode logicl`, 40-epoch `logicl40_clean` track).
+
+Code: **https://github.com/Yale-CompBio/logica**
+
+| Size | Protein backbone | File | Weights | Epoch | Val loss |
+| --- | --- | --- | --- | --- | --- |
+| 8M | `esm2_t6_8M_UR50D` | `checkpoints/8m/best.pt` | 19 MB | 100 | 1.8036 |
+| 35M | `esm2_t12_35M_UR50D` | `checkpoints/35m/best.pt` | 38 MB | 99 | 1.6101 |
+| 150M | `esm2_t30_150M_UR50D` | `checkpoints/150m/best.pt` | 64 MB | 98 | 1.3963 |
+| 650M | `esm2_t33_650M_UR50D` | `checkpoints/650m/best.pt` | 97 MB | 92 | 1.0994 |
+
+Ligand encoder: `HUBioDataLab/SELFormer` for all sizes.
+
+## Contents
+
+Checkpoints are **weights-only** (`trainable_only=True`): a dict with
+`model_state_dict` (+ `epoch`, `metrics`, `rng_state`). They load into the
+`BiDirectionalDrugProteinModel` with `strict=False`.
+
+## Usage
+
+```python
+from huggingface_hub import hf_hub_download
+ckpt = hf_hub_download("Yale-CompBio/logica", "checkpoints/150m/best.pt")
+```
+
+Then fine-tune with the released code (the entry points take `--pretrained_checkpoint`):
+
+```bash
+python dti.py --method logica --dataset DAVIS \
+  --esm_model esm2_t30_150M_UR50D --pretrained_checkpoint $ckpt --output_dir runs/dti_davis
+python variant.py --objective pairwise --held_out EGFR --split_strategy lopo \
+  --esm_model esm2_t30_150M_UR50D --pretrained_checkpoint $ckpt \
+  --data_csv splits/data.csv --fasta_path data/proteins.fasta --drugs_selfies data/drugs.selfies \
+  --use_lora --output_dir runs/variant_egfr
+```
+
+## Citation
+
+```
+@inproceedings{logica2026,
+  title  = {Contextualizing Biological Language Models across Modalities via Logit-Space Contrastive Alignment},
+  author = {anonymous},
+  booktitle = {NeurIPS},
+  year   = {2026}
+}
+```

checkpoints/150m/best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b6f256484aeb614c1fd0ad1ede84cf05c139ec7d2e8d5a66fc1ece6f4db2e98d
+size 63838684

checkpoints/35m/best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4a09c115bd35f266817e3e34844b6ca03a8ca36830d705304a5845996a486dfe
+size 37724235

checkpoints/650m/best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e509b6df98da09a28645e94e4b51fb1384a532d8768ea657640ea5be840b28fe
+size 97234134

checkpoints/8m/best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e20c9c907a55e0c0c2846895c1c15c69f33f99fd0de8047f7a8359a96bfb434c
+size 18783965