MSCMF-DTB: a multi-scale cross-modal fusion framework for drug–target binding prediction

Predicting drug–target binding remains a central challenge in computational drug discovery, particularly due to the need for models that jointly capture molecular topology, chemical substructures, and protein sequence dependencies. We propose MSCMF-DTB, an end-to-end deep learning framework supporting both drug–target interaction (DTI) classification and drug–target affinity (DTA) regression. On the drug side, molecular graphs generated with RDKit are encoded using a DenseGCN module, while a parallel fingerprint channel captures fragment-level and compositional features. On the protein side, contextualized embeddings from TAPE-BERT are processed through a multi-scale 1D CNN to extract local sequence patterns. Cross-modal drug–protein relationships are modeled using cross-attention mechanism coupled with a tensor network for higher-order feature interaction. The fused representations are fed into an MLP for final prediction. Extensive experiments demonstrate that MSCMF-DTB achieves competitive and consistent performance across small- and large-scale datasets (Human, C. elegans, GPCR, BioSNAP, and DrugBank for DTI, and DAVIS and KIBA for DTA). Notably, on the large-scale DrugBank dataset for DTI prediction, MSCMF-DTB improved AUC and Recall by up to 3.2% and 6.1%, respectively, compared with the second-best model (DrugBAN). For DTA prediction, the model achieved stable performance on the large and heterogeneous KIBA dataset, with an MSE of 0.146, a Concordance Index of 0.886, and an rm² of 0.765. Attention-based interpretability further shows that the model learns biologically meaningful interaction regions. Finally, a cold-start case study indicates that MSCMF-DTB successfully identifies experimentally validated inhibitors to AKT1, illustrating its practical utility in virtual screening and drug repurposing.