Influence relevance voting: an accurate and interpretable virtual high throughput screening method.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• Virtual high-throughput screening (vHTS) predicts chemical activity using computational methods.
• Existing vHTS methods face challenges in accuracy and interpretability.

Purpose of the Study:

• To develop and evaluate a novel method, the Influence Relevance Voter (IRV), for enhanced vHTS.
• To assess IRV's performance against established methods, including support vector machines (SVMs).

Main Methods:

• The IRV is a low-parameter neural network refining a k-nearest neighbor classifier.
• It nonlinearly combines neighbor influences, decomposed into relevance and vote components.
• Performance was benchmarked using data from two large, open competitions.

Main Results:

• IRV achieved state-of-the-art results on benchmark datasets.
• It demonstrated performance comparable to or significantly better than SVMs.
• IRV identified three times more active compounds in the top 1% compared to SVMs in one case.

Conclusions:

• IRV offers a powerful and efficient alternative for vHTS.
• Advantages include probabilistic predictions, interpretability, rapid training, and minimal overfitting.
• Its unique architecture facilitates easy incorporation of additional information, making it well-suited for vHTS tasks.