🧬 BELKA-mini: Predict Small Molecule Binding

Welcome to the BELKA-mini challenge — a scaled-down version of the BELKA competition from NeurIPS 2024. In this hackathon-style challenge, your goal is to predict whether a small molecule binds to a specific protein target based on its SMILES representation.

This repo provides everything you need to get started.

---

📁 Repository Contents

File / Folder	Description
small_belka_splits/	Cleaned and balanced train.csv, val.csv, test.csv files
small_belka_splits_900k/	A larger version of the problem; use it if you can, computationally.
starter_model.ipynb	A baseline model using ECFPs + Random Forest
belka_hint_exploration.ipynb	A deeper exploration notebook (shared later during the hackathon)
submission.csv (optional)	Example format for leaderboard submission
BELKA-mini-Overview.md	A detailed overview of the challenge
BELKA-mini-Dataset-Description.md	A detailed description of the dataset
README.md	You’re here!

---

🎯 Challenge Overview

Given:

• A small molecule's SMILES string (representing its structure)
• The protein target name

Your task is to:

• Predict whether the molecule will bind to the target.

This is a binary classification problem. Binding is highly imbalanced (less than 1% of molecules bind), so careful modeling and evaluation are key!

---

🔬 Dataset Structure

Each row contains:

• id: unique identifier
• buildingblock1/2/3_smiles: SMILES for each component
• molecule_smiles: full molecule SMILES (including triazine core)
• protein_name: one of three targets (HSA, BRD4, sEH)
• binds: 1 if the molecule binds the protein, 0 otherwise
• is_novel: (test only) 1 if the molecule includes building blocks not seen in the training set

---

🧪 Evaluation Metric

Submissions are scored using average precision (AP).
Final score = average AP per protein + novelty group (as in the original BELKA challenge).

---

🚀 How to Get Started

1. Open starter_model.ipynb
2. Run the featurization and baseline training
3. Submit predictions on the test.csv molecules

You’ll be using:

• ECFP fingerprints (via RDKit)
• Protein one-hot encoding
• A simple Random Forest

This notebook is designed to work directly on Kaggle with no extra setup.

---

💡 Hint Notebook (Advanced)

Once you’ve got a baseline, check out belka_hint_exploration.ipynb:

• Analyze class imbalance
• Explore novel vs. known molecules
• Evaluate predictions by protein or novelty
• Learn how the test set was constructed for generalization

Use this notebook to improve how you evaluate and tune your models.

---

🧠 Tips for Participants

• The test set contains molecules from new parts of chemical space
• A good model should generalize, not just memorize
• Try:
  • Graph neural networks (GNNs)
  • 3D-aware models
  • Contrastive learning on SMILES
  • Pre-trained chemical embeddings (ChemBERTa, Mol2Vec)

---

📤 Submitting Predictions

Your submission should be a CSV with:

id,binds
123456789,0.013
123456790,0.872
...

Submit to the platform’s leaderboard for evaluation!

---

🙌 Credits

This mini-challenge is derived from the original BELKA dataset by Leash Biosciences, restructured for educational use.

Special thanks to NeurIPS 2024 + BELKA organizers for sharing this awesome dataset.