Distribution-preserved sampling (DPS) for smarter machine learning assisted ultra-large-scale virtual screening.

Area of Science:

• Computational Chemistry
• Drug Discovery
• Bioinformatics

Background:

• Ultra-large-scale structure-based virtual screening (SBVS) faces computational challenges due to massive chemical libraries.
• High-performance computing resources are often inaccessible to smaller research labs.
• Managing large-scale docking data requires specialized pipelines and expertise.

Purpose of the Study:

• To develop a data-driven pipeline for scalable virtual screening.
• To reduce the computational burden of identifying novel bioactive compounds.
• To establish a benchmark dataset for machine learning in drug discovery.

Main Methods:

• Leveraged a common scaffold (2-phenylthiazole moiety) to reduce the chemical search space.
• Selected over 400,000 2-phenylthiazole-containing molecules from the zinc database.
• Trained a random forest regression model on a sampled subset (1%) to predict binding scores for the entire library using KMeans clustering and binning for distribution preservation.

Main Results:

• Successfully reduced the chemical search space significantly.
• The random forest model accurately predicted binding scores, preserving the original score distribution.
• Statistical fidelity was validated using KS, Wasserstein, JS, and KL divergence metrics.

Conclusions:

• Data-driven strategies are effective for scalable virtual screening.
• The presented pipeline overcomes computational limitations in drug discovery.
• This work provides a valuable benchmark dataset for machine learning applications in the field.