Uncertainty quantification enables reliable deep learning for protein-ligand binding affinity prediction.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Deep learning (DL) models are crucial for predicting protein-ligand binding affinity in drug design.
• Current models often lack generalization and confidence estimates, hindering reliable decision-making.

Purpose of the Study:

• To compare five uncertainty quantification methods for DL models predicting protein-ligand binding affinity.
• To introduce and evaluate Bayes by Backprop for this application.

Main Methods:

• Comparison of Deep Ensemble, Monte Carlo Dropout, Laplace approximation, Bayes by Backprop, and Evidential Neural Networks.
• Utilized the Leak-Proof PDBBind dataset for unbiased evaluation.
• Employed feed-forward neural networks (FFNN) with extended connectivity interaction features (ECIF).

Main Results:

• The FFNN with ECIF and Bayes by Backprop demonstrated superior predictive performance.
• Bayes by Backprop provided highly reliable uncertainty quantification and excellent calibration.
• This method excelled across multiple evaluation metrics without recalibration.

Conclusions:

• Bayes by Backprop is a promising method for uncertainty quantification in protein-ligand binding affinity prediction.
• This approach enhances the reliability and reproducibility of DL models in drug discovery.
• Findings support active learning-driven model development for more robust drug design.