A combined drug discovery strategy based on machine learning and molecular docking

Area of Science:

• Computational chemistry
• Medicinal chemistry
• Bioinformatics

Background:

• Machine learning (ML) is vital for accelerating drug design and discovery.
• Evaluating various ML algorithms is crucial for optimizing hit identification.

Purpose of the Study:

• To comprehensively evaluate eight ML methods for drug discovery.
• To assess the efficacy of ML combined with structure-based virtual screening (SBVS).

Main Methods:

• Eight ML algorithms were tested: decision trees, k-Nearest Neighbor, support vector machines, random forests, extremely randomized trees, AdaBoost, gradient boosting trees, and XGBoost.
• Internal and external datasets were used for rigorous cross-validation.
• A two-step virtual screening approach combining ML with SBVS was implemented.

Main Results:

• Extremely randomized trees demonstrated superior performance among the evaluated ML methods.
• The combined ML and SBVS strategy yielded desirable results in identifying potential drug candidates.
• Cross-validation confirmed the robustness of the ML models.

Conclusions:

• Extremely randomized trees are highly effective for the initial stage of virtual screening.
• Integrating ML with traditional SBVS significantly enhances the capability to discover potential hits from extensive compound libraries.
• This hybrid approach offers a powerful strategy for modern drug discovery pipelines.