CUE: A Chemical Uncertainty-Aware Embedding Framework for Multimodal Drug Selectivity Prediction

Accurate prediction of drug selectivity is critical for prioritizing candidate compounds with reduced off-target effects in AI-based drug discovery. Although selectivity can be inferred indirectly from drug-target affinity (DTA) prediction, cumulative errors from affinity predictions across multiple targets can reduce reliability, motivating the development of methods that directly predict compound-level selectivity. We propose a Chemical Uncertainty-aware Embedding (CUE) framework that integrates molecular fingerprints and 2D molecular structure image embeddings. The two modalities are combined via Loss Trajectory Analysis for Uncertainty (LTAU)-based weighted fusion, which adaptively reweights features according to their predictive reliability. Across eight benchmark data sets, CUE achieved RMSE values ranging from 0.163 to 1.691 and outperformed affinity-based and quantitative structure-activity relationship (QSAR) baseline models. Furthermore, a virtual screening case study for EGFR(T790M/C797S) inhibitors further demonstrated its utility for identifying selective hits. The source code is available at https://github.com/jjjabcd/CUE.