Performance of machine-learning scoring functions in structure-based virtual screening.

Area of Science:

• Computational chemistry
• Cheminformatics
• Drug discovery

Background:

• Classical scoring functions show limitations in virtual screening and binding affinity prediction.
• Machine learning (ML) scoring functions show promise but face challenges like overfitting and applicability to new targets.

Purpose of the Study:

• To develop and validate a novel, ready-to-use machine learning scoring function, RF-Score-VS, for improved virtual screening and binding affinity prediction.
• To assess RF-Score-VS performance against classical and existing ML scoring functions using extensive datasets.

Main Methods:

• RF-Score-VS was trained on a large dataset of 15,426 active and 893,897 inactive molecules across 102 targets.
• Performance was evaluated using DUD-E datasets, three docking tools, and comparisons with five classical and three ML scoring functions.
• Independent testing was conducted on the DEKOIS benchmark dataset.

Main Results:

• RF-Score-VS achieved a 55.6% hit rate in the top 1% of screened molecules, significantly outperforming Vina (16.2%).
• RF-Score-VS demonstrated superior binding affinity prediction (Pearson correlation of 0.56) compared to Vina (-0.18).
• Comparable results were observed on the independent DEKOIS benchmark, confirming robustness.

Conclusions:

• RF-Score-VS offers a substantial improvement in virtual screening efficiency and binding affinity prediction accuracy.
• The developed scoring function addresses limitations of classical methods and shows promise for novel target applications.
• The study provides open-access datasets and the RF-Score-VS tool to foster further research in computational drug discovery.