Add model card

README.md

@@ -0,0 +1,75 @@
+---
+language: en
+license: mit
+tags:
+  - protein-protein-interaction
+  - interpretability
+  - explainability
+  - prochlorococcus
+  - cyanobacteria
+  - deeplift
+  - captum
+  - pytorch
+library_name: pytorch
+pipeline_tag: other
+---
+
+# ppiGPT — Prochlorococcus MED4 Protein-Protein Interaction Model
+
+## Model Description
+
+ppiGPT is a GPT-2 architecture trained from scratch to predict protein-protein interactions in *Prochlorococcus* MED4. The model takes two protein sequences as input and predicts whether they interact.
+
+| Parameter | Value |
+|-----------|-------|
+| Architecture | GPT-2 (decoder-only transformer) |
+| Layers | 12 |
+| Attention heads | 12 |
+| Embedding dimension | 768 |
+| Total parameters | ~84.98M |
+| Vocabulary | 29 tokens (20 amino acids + 9 special tokens) |
+| Max sequence length | 1024 tokens |
+| Training epochs | 3 |
+
+## Files
+
+| File | Description | Size |
+|------|-------------|------|
+| `model/out_3e/ckpt.pt` | Trained model checkpoint | 1.0 GB |
+| `model/data/meta.pkl` | Character-level tokenizer metadata | 343 B |
+| `results/deeplift_motif_analysis_results.pkl` | DeepLift attributions for all 2,168 protein pairs | 78 MB |
+| `results/integrated_gradients_random_ppi_per_token_attributions.csv` | Per-token Integrated Gradients attributions | 174 MB |
+
+## Training Data
+
+The model was trained on *Prochlorococcus* MED4 protein sequences with interaction labels derived from yeast two-hybrid (Y2H) experiments:
+
+- **Positive Reference Set (PRS)**: 1,084 experimentally validated Y2H interactions
+- **Random Reference Set (RRS)**: 1,084 randomly paired MED4 proteins
+
+## Input Format
+
+Protein pairs are encoded as character-level sequences:
+
+```
+<ps1>,PROTEIN_SEQUENCE_1,<ps2>,PROTEIN_SEQUENCE_2,<
+```
+
+The model predicts interaction probability via softmax over output tokens at positions 25-26.
+
+## Intended Use
+
+This model and its associated interpretability outputs accompany a manuscript on *Prochlorococcus* MED4 interactome analysis. The repository provides tools for understanding what sequence features drive the model's interaction predictions, including DeepLift attribution, Integrated Gradients, attention analysis, and alanine substitution scanning.
+
+## Code Repository
+
+Analysis scripts and source data: [github.com/olympus-terminal/Prochlorococcus_interactome_model_explainability](https://github.com/olympus-terminal/Prochlorococcus_interactome_model_explainability)
+
+## Framework
+
+- PyTorch >= 2.0.0
+- Captum (for DeepLift and Integrated Gradients)
+
+## License
+
+MIT