Representation-aware conformal prediction for reliable uncertainty quantification in drug-target affinity

Quantifying predictive uncertainty is essential for computational drug discovery to reduce experimental costs and false positives. However, most drug-target affinity (DTA) models focus on improving prediction accuracy without reliable uncertainty estimates.

We propose RCP-DTA, a representation-aware conformal prediction framework that constructs adaptive prediction intervals based on feature similarity. By constructing local calibration sets, our approach generates sample-specific prediction intervals that achieve empirical coverage close to the nominal confidence level. To enhance representation quality, RCP-DTA employs a cross-modal architecture that integrates pre-trained sequence representations and structural information. Binding site context is incorporated via virtual pocket feedback to enhance residue representations, followed by gated fusion for adaptive integration of sequence and structural modalities. Evaluations across multiple DTA benchmark datasets demonstrate that RCP-DTA achieves competitive affinity prediction performance while providing reliable uncertainty quantification.

Code and data are available at: https://github.com/Lee515628/RCP-DTA.

Supplementary data are available at Bioinformatics online.