Evaluating Scalable Supervised Learning for Synthesize-on-Demand Chemical Libraries.

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Traditional drug discovery is slow and expensive, with high-throughput screening examining only a small chemical space.
• Machine learning (ML) offers powerful virtual screening (VS) capabilities for large chemical libraries, but lacks extensive experimental validation.

Purpose of the Study:

• To prospectively evaluate ligand-based virtual screening using ML for identifying active compounds against the bacterial protein-protein interaction PriA-SSB.
• To compare the performance of different ML models and validate findings on large commercial and custom chemical libraries.

Main Methods:

• Cross-validation was used to compare supervised learning models, selecting a random forest (RF) classifier as optimal.
• The RF model screened over 8 million compounds from Aldrich Market Select and one billion from the Enamine REAL database.
• Experimental validation was performed on selected compounds from both libraries.

Main Results:

• The RF model significantly outperformed a structure similarity baseline, with 48% of 701 selected compounds showing activity against PriA-SSB.
• Testing 68 diverse top predictions from Enamine REAL yielded 31 hits (46%), including one with a 1.3 μM IC50.
• The ML approach demonstrated scalability and effectiveness in identifying novel active compounds.

Conclusions:

• Ligand-based virtual screening powered by ML, specifically RF, is a highly effective strategy for discovering active small molecules from vast chemical libraries.
• This validated ML approach accelerates the identification of drug candidates, overcoming limitations of traditional screening methods.
• The study highlights the practical utility of ML-driven VS in real-world drug discovery programs.