Probabilistic multimodal learning for drug–target interaction prediction

Deep learning has driven substantial progress in drug-target interaction (DTI) prediction. However, existing methods often ignore representation uncertainty arising from feature noise in the text modality and protein structure prediction errors in the structure modality. Such uncertainties are further amplified during multimodal fusion, thereby limits model generalization. To address this issue, this study proposes a probabilistic multimodal learning framework for drug-target interaction prediction (PML-DTI) with a dual-branch architecture to extract text and structure features. A pocket-guided interaction module highlights functional binding regions, while a collaborative bidirectional Mamba module refines intra-modal context and supports preliminary cross-modal modeling. PML-DTI embeds multimodal features into a hypersphere by the von Mises-Fisher distribution and estimates their representation reliability using concentration parameters. It facilitates adaptive suppression of unreliable modalities through a reliability-aware fusion mechanism. Extensive experiments on multiple benchmark datasets demonstrate that PML-DTI surpasses seven state-of-the-art methods under both in-domain and cross-domain settings. PML-DTI provides a robust and interpretable solution for DTI prediction, with potential to accelerate drug discovery.