Conformalized Graph Learning for Molecular ADMET Property Prediction and Reliable Uncertainty Quantification.

Area of Science:

• Computational chemistry
• Pharmacology
• Machine learning

Background:

• Drug discovery is expensive, with significant costs in characterizing ADMET properties.
• Graph neural networks (GNNs) improve in silico ADMET prediction but struggle with uncertainty quantification.
• Data volume and quality impact GNN performance, making uncertainty estimation crucial, especially for out-of-domain compounds.

Purpose of the Study:

• To introduce a novel GNN model, Conformalized Fusion Regression (CFR), for robust ADMET property prediction and uncertainty quantification.
• To address the critical challenge of reliably quantifying prediction uncertainty in GNN-based ADMET models.

Main Methods:

• Developed Conformalized Fusion Regression (CFR), integrating a GNN with joint mean-quantile regression loss.
• Employed an ensemble-based conformal prediction (CP) method for uncertainty quantification.
• Evaluated the framework across diverse ADMET prediction tasks.

Main Results:

• CFR achieved accurate ADMET predictions.
• The framework demonstrated reliable probability calibration.
• High-quality prediction intervals were generated, outperforming existing methods for uncertainty quantification.

Conclusions:

• CFR offers a promising approach for reliable in silico ADMET prediction and uncertainty quantification.
• The model enhances the trustworthiness of predictions, particularly for novel or out-of-domain drug candidates.
• This work advances the application of GNNs in drug discovery by providing robust uncertainty estimates.