SCAD-DTA: spherical concept alignment with dynamic multi-modal fusion for drug-target affinity prediction.

Abstract

Accurate drug-target affinity (DTA) prediction is essential for virtual screening and computational drug discovery. However, existing benchmarks such as Davis and KIBA often suffer from imbalanced label distributions and potential sampling bias, which can lead to overestimated performance and limited generalization in realistic settings. In addition, integrating heterogeneous modalities such as molecular graphs, protein sequences, and chemical fingerprints remains challenging due to unstable cross-modal interactions and inconsistent representation scales. To address these issues, we first construct three curated datasets from the Therapeutic Target Database (TTD), namely TTD_IC50, TTD_EC50, and TTD_KI, using a stratified sampling and normalization strategy to improve distribution balance while maintaining biochemical diversity. These datasets are designed to better reflect real-world distribution shifts in DTA prediction. Based on these datasets, we propose SCAD-DTA, a geometry-aware multimodal learning framework that explicitly addresses instability in cross-modal fusion under distribution shift. The key idea is to stabilize multimodal interaction by jointly modeling adaptive fusion dynamics and representation geometry constraints. SCAD-DTA introduces a Dynamic Cross-Modal Attention mechanism that adaptively reweights modality contributions conditioned on sample-specific context, mitigating modality dominance. To further improve representation stability, a Spherical Constrained Projection module enforces unit-norm geometry in the latent space, reducing scale inconsistency across modalities. In addition, a Concept Alignment module maps fused representations into a learnable prototype space, enabling structured and interpretable modeling of drug-target interactions. Extensive experiments on six benchmark datasets show that SCAD-DTA achieves competitive or superior performance in most evaluation settings, with particular strength under cold-start and cross-distribution scenarios; however, its gains are less pronounced in some cold-start splits (e.g., cold-drug on KIBA, cold-target on Metz), which we discuss explicitly. The source code and datasets are publicly available at: https://github.com/xwtxbzz/SCAD-DTA.