Smarter, not harder: machine learning for scalable drug screening

This repository contains the proof-of-concept (POC) analysis presented in our study on distribution-preserving sampling subset selection approach for docking score prediction. The goal of this work is to demonstrate that for a scaffold-based dataset, it is possible to train predictive models on a very small fraction of the data (≈1%) while maintaining accuracy, thereby reducing computational costs by over 100× compared to training on larger fractions.

---

Repository Structure

• POC.ipynb — Jupyter notebook containing all analyses (feature extraction, distribution comparison, model training, and evaluation).
• merged_filtered_data_no_outliers.csv — Scaffold-based dataset (≈413k molecules), derived from fragment-based virtual screening and filtered for outliers.
• HSD11B1_filtered_data_no_outliers.csv — Benchmark dataset from the DOCKSTRING project, filtered for outliers (≈255k molecules).

---

Methods

• Divergence Analysis: We quantified how well feature and docking score distributions were preserved across subsets using Kolmogorov–Smirnov, Wasserstein, Jensen–Shannon, and Kullback–Leibler metrics.
• Feature Selection: Top 50 molecular descriptors were selected using regression-based mutual information.
• Subset Sampling: Random vs. distribution-preserving strategies were compared across fractions (1%–75%).
• Model Training: Random Forest regressors were trained on subsets to evaluate predictive performance.
• Runtime Comparison: Training on 1% of the scaffold dataset required ~1 min, compared to ~115 min for 75%, while maintaining similar accuracy.

---

Requirements

The analysis uses only widely available Python libraries:

• numpy
• pandas
• matplotlib
• seaborn
• scikit-learn
• scipy

Install requirements with:

pip install numpy pandas matplotlib seaborn scikit-learn scipy

Usage

1. Clone this repository:

   git clone https://github.com/csbarak/POC.git cd POC

2. Open the Jupyter notebook:

   jupyter notebook POC.ipynb

3. Run the notebook cells step by step to reproduce the full analysis.

Citation

If you use this repository, data, or methodology in your work, please cite:

Trachtenberg, A.; Spelkov A.; Akabayov B. (2025). Smarter, not harder: machine learning for scalable drug screening.

License

This project is released under the MIT License.