Bayesian Optimization with MM/GBSA Scoring

This repository accompanies the paper:

Bayesian Optimization for Structure-Based Drug Design with High-Fidelity MM/GBSA Scoring https://doi.org/10.1101/2025.06.22.660936

We study active learning / Bayesian optimization (BO) for structure-based drug design, using MM/GBSA as a high-fidelity oracle in place of fast docking scores. The main finding is that BO with molecular language model embeddings efficiently recovers top-scoring compounds with far fewer oracle evaluations than random screening.

For a guide on running the optimizer on your own data or with a custom scoring oracle, see USAGE.md.

---

Reproducing Results

Setup

All steps except docking use the base conda environment:

conda env create -f docker/environment.yml
conda activate base

The docking step requires a separate environment due to Python version incompatibilities with the vina and easydock packages:

conda env create -f docker/environment-docking.yml

Docker images are also provided (docker/Dockerfile.base, docker/Dockerfile.docking) for fully reproducible containerized execution.

Prepare the data

Run the preprocessing notebook to produce the required CSV files:

data/process_data.ipynb

Expected output files in data/:

• MCL1-mmgbsa.csv — MCL1 MM/GBSA scores
• MCL1-vina.csv — MCL1 Vina docking scores
• Enamine10k_scores.csv
• Enamine50k_scores.csv

Run Bayesian Optimization

bash run_bayesian_optimization.sh

Results are written to bayesian_optimization_results/. To run ablations:

bash run_seeds.sh           # Initialization robustness
bash run_batch_size.sh      # batch acquisition size
bash run_diverse_init.sh    # diverse initialization
bash run_ucb_experiment.sh  # UCB kappa sweep
bash run_ucb_kappa.sh       # UCB acquisition function for specific kappa

Generate figures

Open the notebooks in analysis/ corresponding to the experiment of interest (e.g. analysis/mcl1_mmgbsa.ipynb for the main MM/GBSA results). Figures are saved to analysis/figures/.

---

Full Pipeline (from scratch)

If you want to reproduce the full pipeline including data collection:

1. Screening — ZINC22 similarity screening; see screening/README.md
2. Docking — AutoDock Vina via easydock (docking environment); see docking/README.md
3. MD simulation + MM/GBSA — GROMACS + gmx_MMPBSA; see simulation_configuration/README.md
4. Preprocessing — data/process_data.ipynb
5. BO experiments — run_bayesian_optimization.sh
6. Analysis — notebooks in analysis/

---

Repository Structure

bayesian-optimization-mmgbsa/
│
├── docker/                      # Conda environment files and Docker images
│   ├── environment.yml          # Base environment (BO + analysis)
│   ├── environment-docking.yml  # Docking environment (separate due to version constraints)
│   ├── Dockerfile.base          # Docker image for base environment
│   └── Dockerfile.docking       # Docker image for docking environment
│
├── screening/                   # Step 1: ZINC22 similarity screening
│   ├── similarity.py            # MCS-based screening script
│   └── README.md
│
├── docking/                     # Step 2: AutoDock Vina docking via easydock
│   ├── start_docking            # Docking launcher script
│   ├── extract_mols.py          # Post-processing: extract mol blocks from .db files
│   └── README.md
│
├── simulation_configuration/    # Step 3: GROMACS + gmx_MMPBSA setup
│   ├── *.mdp                    # GROMACS MD parameter files
│   ├── mmpbsa.in                # gmx_MMPBSA input
│   └── README.md
│
├── data/                        # Step 4: Data preprocessing
│   ├── process_data.ipynb       # Deduplication and formatting
│   └── README.md
│
├── bayesian_optimization/       # Step 5: Active learning experiments
│   ├── run.py                   # Main entry point
│   └── ligbo/        # Core BO library (surrogate models, embeddings, samplers)
│
├── bayesian_optimization_results/  # Output: BO run results (populated at runtime)
│
├── analysis/                    # Step 6: Analysis notebooks and figures
│   ├── figures/                 # Publication figures
│   ├── mcl1_mmgbsa.ipynb        # Main MCL1 MM/GBSA results
│   ├── mcl1_vina.ipynb          # MCL1 Vina results
│   ├── mcl1_vina_docking_in_one.ipynb  # Combined Vina analysis
│   ├── mcl1_batch_size.ipynb    # Effect of batch acquisition size
│   ├── mcl1_medoid_effect.ipynb # Effect of initialization strategy (medoid)
│   ├── mcl1_mmgbsa_diverse_init.ipynb  # Diverse initialization ablation
│   ├── mcl1_mmgbsa_ucb.ipynb    # UCB acquisition function experiments
│   ├── enamine10k.ipynb         # Enamine 10k benchmark
│   ├── enamine50k.ipynb         # Enamine 50k benchmark
│   ├── seeds_error_bars.ipynb   # Initialization robustness / error bars
│   ├── rogi_curves.ipynb        # ROGI landscape analysis
│   ├── utils.py                 # Shared plotting utilities
│   └── exploratory/             # Exploratory dataset analysis (score distributions, embeddings)
│
├── run_bayesian_optimization.sh # Runs all main BO experiments
├── run_seeds.sh                 # Initialization robustness sweep
├── run_batch_size.sh            # Batch size ablation
├── run_diverse_init.sh          # Diverse initialization ablation
├── run_ucb_experiment.sh        # UCB kappa hyperparameter sweep
└── run_ucb_kappa.sh             # UCB acquisition ablation