A Reproducible Hierarchical Virtual Screening Framework Integrating Scaffold-Aware Machine Learning, Ensemble...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in pharmacology

Background:

• Indoleamine 2,3-dioxygenase 1 (IDO1) is a key enzyme in cancer immune evasion, making it a significant therapeutic target.
• Despite its potential, clinical development of IDO1 inhibitors has faced challenges, necessitating improved drug discovery strategies.

Purpose of the Study:

• To establish a reproducible, hierarchical virtual screening framework for prioritizing potent IDO1 inhibitors.
• To integrate scaffold-aware machine learning with ensemble docking and molecular dynamics for enhanced prediction accuracy.

Main Methods:

• A curated dataset of IDO1 inhibitors was processed and used to train machine learning models (Random Forest, XGBoost, SVM) with scaffold-based splitting.
• Ensemble docking with GNINA and CNN rescoring was employed for prospective screening of FDA-approved drugs.
• Molecular dynamics simulations were performed on top-ranked candidates to assess binding stability.

Main Results:

• Machine learning models achieved high predictive performance (ROC-AUC ≈0.88-0.89), with applicability domain filtering ensuring reliability.
• Prospective screening identified 39 potential IDO1 inhibitors among FDA-approved drugs.
• Ensemble docking and molecular dynamics confirmed stable binding modes for top candidates, validating the screening framework's efficacy.

Conclusions:

• The hierarchical integration of scaffold-aware machine learning and structure-based ensemble methods significantly improves the robustness of virtual screening.
• This generalizable workflow effectively reduces false positives and supports reproducible candidate prioritization in computational drug discovery.
• The study provides a fully executable Jupyter notebook for implementing the described virtual screening pipeline.