Beware of the generic machine learning-based scoring functions in structure-based virtual screening.

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Machine learning-based scoring functions (MLSFs) are emerging as promising tools for structure-based virtual screening (SBVS).
• A key challenge is evaluating the generalizability of MLSFs trained on diverse datasets for specific drug targets.

Purpose of the Study:

• To systematically assess the effectiveness of 14 reported MLSFs in virtual screening (VS).
• To determine if generic MLSFs can consistently outperform classical scoring functions and achieve satisfactory results across various targets.

Main Methods:

• Evaluated 14 MLSFs across multiple VS datasets.
• Compared MLSF performance against classical scoring functions like Glide SP.
• Investigated the impact of using top-three docking poses and retraining models with updated datasets.

Main Results:

• Most MLSFs demonstrated poor performance and did not outperform classical scoring functions.
• RFscore-VS, trained on the Directory of Useful Decoys-Enhanced dataset, showed superiority for some targets.
• Performance significantly decreased for targets dissimilar to those in the training sets, indicating limited generalization.

Conclusions:

• Generic MLSFs often exhibit poor generalization capabilities for real-world virtual screening campaigns.
• Caution is advised when employing generic MLSFs for VS due to their potential limitations.
• Further research may be needed to develop MLSFs with improved target-specific applicability.