Data Scaling and Generalization Insights for Medicinal Chemistry Deep Learning Models.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Pharmacology

Background:

• Predictive models accelerate the discovery of safer and more effective therapeutics.
• Understanding and enhancing small-molecule predictive model performance is crucial for drug discovery.
• Both deep learning and traditional machine learning approaches are employed.

Purpose of the Study:

• To compare the performance of deep learning and traditional machine learning models for small-molecule drug discovery.
• To identify factors contributing to model performance differences.
• To develop a predictive scaling relationship for model performance.

Main Methods:

• Experiments using deep learning (graph neural networks) and traditional machine learning (XGBoost, random forest).
• Leveraging large internal and public datasets.
• Assessing model performance on random, temporal, and reverse-temporal data ablation tasks, and extrapolation tasks.

Main Results:

• Graph neural networks demonstrated superior performance compared to traditional methods.
• A developed scaling relationship explained 81% of the variance in model performance across various assays and data regimes.
• Identified key factors influencing model performance.

Conclusions:

• Deep learning, particularly graph neural networks, offers significant advantages for small-molecule predictive modeling in drug discovery.
• The established scaling relationship provides a valuable tool for estimating model performance and guiding future development.
• Findings offer practical guidance for improving predictive model efficacy in drug discovery pipelines.