# gtCLIP — Gene-set Text CLIP

A CLIP-style model that aligns gene set embeddings (via [GSFM](https://huggingface.co/maayanlab/gsfm)) with biomedical text embeddings (via [S-PubMedBert-MS-MARCO](https://huggingface.co/pritamdeka/S-PubMedBert-MS-MARCO)) in a shared latent space.

## Quick Start
```bash
GIT_LFS_SKIP_SMUDGE=1 pip install git+https://huggingface.co/maayanlab/gsfm
pip install sentence-transformers huggingface_hub
```
```python
from huggingface_hub import hf_hub_download
import importlib.util

py_path = hf_hub_download("DaveLab/gtCLIP", "gtCLIP.py")
spec = importlib.util.spec_from_file_location("gtCLIP", py_path)
mod = importlib.util.module_from_spec(spec)
spec.loader.exec_module(mod)
GeneSetTextEncoder = mod.GeneSetTextEncoder

model = GeneSetTextEncoder.from_pretrained("DaveLab/gtCLIP", device="cuda")

# Encode a gene set
gene_emb = model.encode_gene_symbols(["TP53", "BRCA1", "MYC", "CDK2"])

# Encode a text description
text_emb = model.encode_text_normalized(["DNA damage repair pathway"])

# Compute similarity
similarity = (gene_emb @ text_emb.T).item()
print(f"Similarity: {similarity:.4f}")
```

## Files

| File | Description |
|------|-------------|
| `gtCLIP.py` | Model class definition |
| `gtCLIP.pt` | Trained weights + config |
| `config.json` | Architecture hyperparameters |

## Architecture

- **Gene encoder**: GSFM (frozen or partially fine-tuned) → projection MLP → shared space
- **Text encoder**: PubMedBERT (frozen or partially fine-tuned) → projection MLP → shared space
- **Training**: Contrastive CLIP loss with gene-set overlap soft targets + reconstruction auxiliary loss