Data-augmented machine learning scoring functions for virtual screening of YTHDF1 m⁶A reader protein.

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and medicinal chemistry
• Machine learning in bioinformatics

Background:

• Machine learning accelerates drug discovery, particularly structure- and ligand-based approaches.
• Limited inhibitor data necessitates data augmentation for enhanced model performance.
• YTHDF1 is a promising cancer therapeutic target, but traditional virtual screening methods struggle with its unique binding characteristics.

Purpose of the Study:

• To develop predictive machine learning models for structure-based drug discovery.
• To create YTHDF1-specific machine learning scoring functions (MLSFs) to improve structure-based virtual screening (SBVS).
• To address challenges in identifying potent YTHDF1 inhibitors due to limited data and protein binding complexities.

Main Methods:

• Trained predictive models using traditional machine learning algorithms on target and ligand dynamics-aware datasets.
• Developed YTHDF1-specific MLSFs using data augmentation, including multiple ligand and protein conformations.
• Trained 64 MLSFs with four algorithms and evaluated their predictive and ranking power on ten test sets.

Main Results:

• The artificial neural network with protein-ligand extended connectivity fingerprints (ANN-PLEC) demonstrated superior performance.
• ANN-PLEC achieved a high area under the precision-recall curve (PR-AUC) of 0.87.
• The developed MLSFs show promise for drug discovery targeting proteins with limited active molecule data.

Conclusions:

• The ANN-PLEC model offers a viable approach for structure-based drug discovery, especially for challenging targets like YTHDF1.
• This method enhances SBVS efficacy by providing specific scoring functions tailored to the target protein.
• The ANN-PLEC scoring function is publicly available on GitHub to facilitate further research.