Real-World Assessment of Machine-Learned Docking Using Bioassay-Derived Benchmarks

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in cheminformatics

Background:

• The expansion of compound libraries and chemical space presents challenges for traditional drug discovery methods.
• Machine learning (ML) docking offers potential speed and scalability for virtual screening.
• Existing ML benchmarks may use biased datasets, overestimating real-world performance.

Purpose of the Study:

• To systematically evaluate the performance of the ML-based DiffDock-Pocket method.
• To compare ML docking against traditional physics-based approaches using realistic high-throughput screening (HTS) data.
• To provide a clearer understanding of ML model reliability in practical drug discovery scenarios.

Main Methods:

• Utilized high-throughput screening (HTS) datasets from the PubChem BioAssay database.
• Evaluated the performance of DiffDock-Pocket, a popular ML-based docking tool.
• Compared ML docking results against traditional physics-based docking methods.

Main Results:

• The study provides a realistic assessment of ML docking performance on HTS data.
• Identified specific strengths and limitations of ML-based docking methods in practical applications.
• Highlighted discrepancies between ML performance on benchmark versus real-world datasets.

Conclusions:

• ML docking methods show potential but require careful validation on realistic data.
• Current ML models may not fully capture the complexities of real-world virtual screening.
• Further research is needed to improve the reliability and applicability of ML in drug discovery.