MolBO: Bayesian Optimization for Multi-Objective Drug Design

---

Pareto-optimal molecule discovery via qNEHVI Bayesian optimization over large molecular libraries.

---

Motivation

Drug discovery demands simultaneous optimization of conflicting properties — high binding affinity must be balanced against metabolic stability, low toxicity, aqueous solubility, and synthetic accessibility. Classical single-objective optimization fails here: optimizing one property almost always degrades another.

MolBO addresses this by framing molecular design as a multi-objective Bayesian optimization (MOBO) problem. Building on the framework of Fromer & Coley (2024), we implement the q-Noisy Expected Hypervolume Improvement (qNEHVI) acquisition function from BoTorch over molecular fingerprint spaces, enabling sample-efficient discovery of the full Pareto front across any set of molecular objectives.

---

Features

• 🧬 qNEHVI acquisition — batch-parallel, noisy multi-objective BO with state-of-the-art sample efficiency
• ⚗️ RDKit integration — Morgan fingerprints, property oracles (logP, QED, SA score, Tanimoto similarity)
• 📈 Pareto front tracking — per-iteration hypervolume computation and front visualization
• 🔌 Modular oracle API — plug in any scoring function (docking scores, ADMET predictors, ML surrogates)
• 📓 Jupyter walkthrough — end-to-end notebook with visualizations and ablations
• 🧪 Full test suite — pytest coverage for all core modules
• 🏎️ GPU-ready — automatic CUDA detection via PyTorch

---

Repository Structure

MolBO/
├── data/
│   ├── raw/                  # Input molecular libraries (SMILES files)
│   └── processed/            # Featurized fingerprint tensors
├── models/
│   ├── __init__.py
│   ├── gp_model.py           # Batched GP surrogate (BoTorch)
│   └── fingerprints.py       # Morgan fingerprint featurizer (RDKit)
├── notebooks/
│   └── 01_quickstart.ipynb   # End-to-end walkthrough
├── tests/
│   ├── test_fingerprints.py
│   ├── test_gp_model.py
│   ├── test_optimize.py
│   └── test_evaluate.py
├── optimize.py               # Main Pareto optimization loop (qNEHVI)
├── evaluate.py               # Oracle definitions & Pareto metrics
├── requirements.txt
├── setup.py
├── .gitignore
└── README.md

---

Installation

1. Clone

git clone https://github.com/Islamomar-1/MolBO.git
cd MolBO

2. Create environment

conda create -n molbo python=3.10
conda activate molbo

3. Install PyTorch (CUDA 11.8 example)

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu118

4. Install RDKit

conda install -c conda-forge rdkit

5. Install remaining dependencies

pip install -e .

---

Quick Start

CLI — optimize a SMILES library

python optimize.py \
    --library data/raw/zinc_10k.smi \
    --objectives qed logp sa \
    --n_init 50 \
    --n_iter 20 \
    --batch_size 5 \
    --output results/pareto_run1.json

Python API

from models.fingerprints import MorganFeaturizer
from evaluate import ObjectiveOracle
from optimize import ParetoOptimizer

# Load a SMILES library
smiles_list = open("data/raw/zinc_10k.smi").read().splitlines()

# Define objectives (any callable returning a float, higher = better)
oracle = ObjectiveOracle(objectives=["qed", "logp", "sa_score"])

# Featurize
featurizer = MorganFeaturizer(radius=2, n_bits=2048)
X = featurizer.transform(smiles_list)         # (N, 2048) tensor
Y = oracle.evaluate_batch(smiles_list)         # (N, n_obj) tensor

# Run Pareto optimization
optimizer = ParetoOptimizer(
    X=X,
    Y=Y,
    smiles=smiles_list,
    n_iter=20,
    batch_size=5,
)
pareto_smiles, pareto_scores = optimizer.run()

print(f"Pareto front size: {len(pareto_smiles)}")

---

Benchmark

Results on the ZINC 10k subset, 20 BO iterations × batch size 5 (100 oracle calls total). Hypervolume (HV) normalized by the analytic maximum.

Method	HV (↑)	Pareto Set Size	Oracle Calls
Random	0.41	12	100
NSGA-II	0.59	18	100
Single-obj BO (QED)	0.52	9	100
MolBO (qNEHVI)	0.78	31	100
MolBO (qEHVI)	0.74	27	100
MolBO (qParEGO)	0.69	22	100

Objectives: QED, logP ∈ [2, 5], SA Score (inverted). Mean ± std over 5 seeds.

---

Technology Stack

Component	Library / Tool
Bayesian Optimization	BoTorch
GP & Tensors	PyTorch
Molecular Features	RDKit
Acquisition Fn	qNEHVI (BoTorch built-in)
Property Oracles	RDKit, SA Score (Ertl)
Visualization	Matplotlib, Plotly
Testing	pytest

---

Contributing

Contributions are welcome! Please follow these steps:

1. Fork the repository
2. Create a feature branch: git checkout -b feature/my-feature
3. Write tests for any new functionality in tests/
4. Run the test suite: pytest tests/ -v
5. Submit a pull request with a clear description

Adding a new objective

Implement the BaseOracle interface in evaluate.py:

class MyOracle(BaseOracle):
    name = "my_property"

    def __call__(self, smiles: str) -> float:
        mol = Chem.MolFromSmiles(smiles)
        # ... compute and return float (higher = better)
        return score

Then pass "my_property" to ObjectiveOracle(objectives=[..., "my_property"]).

---

Citation

If you use MolBO in your research, please cite:

@article{fromer2024computer,
  title   = {Computer-aided multi-objective optimization in small molecule discovery},
  author  = {Fromer, Jenna C and Coley, Connor W},
  journal = {Nature Computational Science},
  volume  = {4},
  pages   = {22--33},
  year    = {2024},
  doi     = {10.1038/s43588-023-00601-0}
}

And the BoTorch qNEHVI implementation:

@inproceedings{daulton2021parallel,
  title     = {Parallel Bayesian Optimization of Multiple Noisy Objectives with Expected Hypervolume Improvement},
  author    = {Daulton, Samuel and Balandat, Maximilian and Bakshy, Eytan},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2021}
}

---

License

MIT © Islam Omar

MolBO