---
license: mit
library_name: pytorch
tags:
  - protein-ligand
  - drug-target-interaction
  - contrastive-learning
---

# LogiCA — pretrained protein-ligand backbones (anonymous, under review)

Anonymous artifact for double-blind review. Pretrained checkpoints for a
logit-space contrastive alignment model: a bidirectional ESM-2 (protein) +
SELFormer (ligand) backbone with a cross-attention adapter, pretrained on
BindingDB with a token-likelihood contrastive objective (40 epochs).

| Size | Protein backbone | File | Weights |
| --- | --- | --- | --- |
| 8M   | `esm2_t6_8M_UR50D`    | `checkpoints/8m/best.pt`   | 19 MB |
| 35M  | `esm2_t12_35M_UR50D`  | `checkpoints/35m/best.pt`  | 38 MB |
| 150M | `esm2_t30_150M_UR50D` | `checkpoints/150m/best.pt` | 64 MB |
| 650M | `esm2_t33_650M_UR50D` | `checkpoints/650m/best.pt` | 97 MB |

Ligand encoder: `HUBioDataLab/SELFormer` (all sizes).

Checkpoints are weights-only (`model_state_dict` + `epoch` + `metrics`); load with
`strict=False` into the bidirectional model. Download the size you need and pass it
to the training code as the pretrained checkpoint:

```python
from huggingface_hub import hf_hub_download
ckpt = hf_hub_download(REPO_ID, "checkpoints/150m/best.pt")  # REPO_ID withheld for review
```

Training and fine-tuning code is provided as an anonymized supplement with the
submission. Repository, authors, and citation are withheld for double-blind review.