Combining Bayesian and Evidential Uncertainty Quantification for Improved Bioactivity Modeling.

Area of Science:

• Computational chemistry and cheminformatics
• Machine learning in drug discovery
• Uncertainty quantification in predictive modeling

Background:

• Reliable computational modeling in drug discovery necessitates robust uncertainty quantification (UQ).
• Bayesian methods (deep ensemble, MC dropout) and evidential learning offer distinct UQ capabilities and computational trade-offs.
• Existing methods present limitations in expressivity and computational demands for complex drug discovery tasks.

Purpose of the Study:

• To develop and evaluate hybrid UQ approaches combining Bayesian methods and evidential learning.
• To benchmark these hybrid models against established methods on a drug discovery dataset.
• To assess the performance and computational efficiency of novel UQ strategies for bioactivity prediction.

Main Methods:

• Implementation of hybrid models integrating Bayesian principles with evidential learning.
• Benchmarking on the Papyrus++ dataset for bioactivity endpoints (xC50, Kx) using diverse data splits.
• Evaluation using metrics such as Root Mean Squared Error (RMSE), Continuous Ranked Probability Score (CRPS), and interval scores.

Main Results:

• The ensemble of evidential models (EOE) demonstrated superior performance across all tested endpoints and split strategies.
• EOE achieved the lowest RMSE and leading CRPS and interval scores, outperforming traditional methods.
• EOE matched or surpassed large ensembles in utility metrics at significantly reduced computational cost.

Conclusions:

• Hybrid UQ approaches, particularly EOE, provide more accurate and informative uncertainties for bioactivity modeling.
• EOE presents a computationally practical and robust default for uncertainty-aware decision-making in drug discovery.
• The findings advocate for the integration of evidential and Bayesian principles for enhanced UQ in cheminformatics.