Pill Picker Pipeline 💊

Tired of your drugs only treating what they were designed for?

This repo includes a drug repurposing pipeline that teaches molecular models like ChemBERTa and ChemBERTa-2 (hopefully MolFormer in the future) to guess how clingy a molecule will be with your protein of choice. Trained on ChEMBL's bioactivity data, the model predicts log(IC50) values—aka how tightly a compound binds—and screens the DrugBank universe to suggest new potential meds.

Now featuring fewer lab coats and more Python.

---

Features

• Train ChemBERTa or ChemBERTa-2 on ChEMBL targets with IC50 data
• Predict binding affinities (log(IC50)) on DrugBank molecules
• Rank and export top candidate drugs
• Clean CLI interface using argument parsing
• Easily extendable for other transformer-based models

---

Usage

1. Fetch ChEMBL Data

python src/fetch_chembl_ic50.py --target CHEMBL1234

2. Train Model

python src/train_affinity_model.py --target CHEMBL1234 --model chemberta

3. Repurpose from DrugBank

python src/repurpose.py --model chemberta --target CHEMBL1234 --input data/filtered_drugbank_smiles.csv

Directory Structure

.
├── data/                # Training and inference data
├── models/              # Saved trained model weights
├── notebooks/           # Notebooks to play around and test stuff
├── results/             # Saved predictions
├── src/
│   ├── models/          # Model definitions and tokenizers
│   ├── train_affinity_model.py
│   ├── repurpose.py
│   └── fetch_chembl_ic50.py