Profile-QSAR 2.0: Kinase Virtual Screening Accuracy Comparable to Four-Concentration IC₅₀s for Realistically Novel...

Area of Science:

• Computational chemistry
• cheminformatics
• drug discovery

Background:

• Random forest regression (RFR) models show high accuracy on random test sets but fail in real-world virtual screening.
• Realistic test sets, mirroring chemical novelty, reveal poor predictive power of conventional RFR models.
• Previous Profile-QSAR (pQSAR) methods improved applicability but still lacked accuracy on realistic test sets.

Purpose of the Study:

• To develop an improved virtual screening method (pQSAR 2.0) that accurately predicts compound activity on realistic test sets.
• To enhance the predictive power of quantitative structure-activity relationship (QSAR) models for drug discovery.
• To enable more effective hitlist triaging and virtual screening panel development.

Main Methods:

• Developed a cluster-based "realistic" training/test set split to mimic real virtual screening scenarios.
• Implemented the pQSAR 2.0 method, replacing categorical activity probabilities with predicted IC50 values from RFR models.
• Excluded the RFR model for the target assay from the independent variable profile to achieve high accuracy.

Main Results:

• pQSAR 2.0 achieved statistically comparable accuracy to medium-throughput IC50 measurements on realistic test sets.
• The new method overcomes the limitations of conventional RFR models on chemically novel compounds.
• Predicted IC50 values enable semiquantitative assessments of potency, efficiency, and selectivity.

Conclusions:

• pQSAR 2.0 offers a significant advancement in virtual screening accuracy and applicability.
• The method facilitates more informed decision-making in early-stage drug discovery.
• pQSAR 2.0 supports the development of advanced virtual screening panels for toxicity and promiscuity prediction.