SynthMol: A Drug Safety Prediction Framework Integrating Graph Attention and Molecular Descriptors into Pre-Trained...

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning in pharmacology

Background:

• Drug safety is paramount for clinical success, influenced by diverse molecular properties.
• Accurate safety assessment is crucial for evaluating drug candidates.
• Machine learning models trained on bioactivity data offer a promising avenue for drug safety evaluation.

Purpose of the Study:

• To introduce SynthMol, a novel deep learning framework for molecular property prediction.
• To enhance the accuracy of drug safety assessments using advanced computational methods.
• To validate SynthMol's performance on established and real-world drug safety datasets.

Main Methods:

• SynthMol integrates pre-trained 3D structural features, graph attention networks, and molecular fingerprints.
• The framework was evaluated on 22 diverse datasets, including MoleculeNet, MolData, and drug safety data.
• Performance was benchmarked against state-of-the-art models for molecular property prediction.

Main Results:

• SynthMol demonstrated superior prediction accuracy across most evaluated tasks compared to existing models.
• Achieved an ROC-AUC of 0.944 on the BBBP dataset (2.61% improvement) and 0.906 on the hERG dataset (2.38% improvement).
• Validation with experimental hERG toxicity and CYP inhibition data confirmed SynthMol's utility in distinguishing functional changes relevant to drug development.

Conclusions:

• SynthMol represents a significant advancement in deep learning for drug safety assessment.
• The framework's ability to accurately predict molecular properties aids in identifying safer drug candidates.
• SynthMol provides a valuable tool for accelerating drug development by improving early-stage safety evaluations.