Conformalized Graph\nLearning for Molecular ADMET Property\nPrediction and Reliable Uncertainty Quantification

Drug discovery and development is a complex and costly\nprocess,\nwith a substantial portion of the expense dedicated to characterizing\nthe absorption, distribution, metabolism, excretion, and toxicity\n(ADMET) properties of new drug candidates. While the advent of deep\nlearning and molecular graph neural networks (GNNs) has significantly\nenhanced in silico ADMET prediction capabilities, reliably quantifying\nprediction uncertainty remains a critical challenge. The performance\nof GNNs is influenced by both the volume and the quality of the data.\nHence, determining the reliability and extent of a prediction is as\ncrucial as achieving accurate predictions, especially for out-of-domain\n(OoD) compounds. This paper introduces a novel GNN model called conformalized\nfusion regression (CFR). CFR combined a GNN model with a joint mean–quantile\nregression loss and an ensemble-based conformal prediction (CP) method.\nThrough rigorous evaluation across various ADMET tasks, we demonstrate\nthat our framework provides accurate predictions, reliable probability\ncalibration, and high-quality prediction intervals, outperforming\nexisting uncertainty quantification methods.