Accuracy or novelty: what can we gain from target-specific machine-learning-based scoring functions in virtual...

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Machine-learning scoring functions (MLSFs) are emerging for protein-ligand binding affinity prediction and virtual screening.
• Doubts persist regarding the benefits and performance of these novel scoring functions.

Purpose of the Study:

• To benchmark target-specific MLSFs against classical scoring functions and QSAR models using an unbiased dataset.
• To evaluate MLSFs based on prediction accuracy and hit novelty in virtual screening.

Main Methods:

• Assessed MLSFs trained on three protein-ligand interaction representations using the LIT-PCBA dataset.
• Compared MLSF performance with Glide SP scoring function and ligand-based quantitative structure-activity relationship (QSAR) models.
• Systematically explored prediction accuracy and hit novelty.

Main Results:

• Target-specific MLSFs generally outperformed the classical Glide SP scoring function.
• MLSFs did not consistently outperform 2D fingerprint-based QSAR models, even with integrated features.
• MLSFs identified different hits based on featurization strategies, suggesting they are more akin to QSAR models than traditional scoring functions.

Conclusions:

• Target-specific MLSFs show potential but are not yet superior to established QSAR models for virtual screening.
• MLSFs can identify diverse hits but lack the intrinsic attributes of traditional scoring functions.
• MLSFs can be viewed as advanced ligand-based QSAR models, offering valuable insights for future development.