Accurate prediction of drug-target interactions (DTIs) is essential for drug discovery and repurposing. Despite recent advances, deep learning models often exhibit limited generalization under realistic cold-start scenarios and suffer from poor interpretability. To address these challenges, we present BioFusionDTI, a multimodal deep learning framework that integrates graph-based and sequence-based representations of drugs and proteins. BioFusionDTI employs graph convolutional networks (GCNs) to extract structural features from molecular and protein graphs, and convolutional neural networks (CNNs) to process sequence embeddings derived from pretrained biomolecular language models. A bilinear attention network (BAN) is further introduced to capture fine-grained cross-modal interactions, thereby enhancing predictive accuracy and interpretability. Extensive experiments on three benchmark data sets (SNAP, DRH, and Kinase) under warm, cold-drug, and cold-protein settings demonstrate that BioFusionDTI consistently outperforms state-of-the-art baselines. Ablation studies highlight the effectiveness of the fusion strategy, with the BAN module making a substantial contribution to performance gains. Moreover, attention visualizations reveal biologically plausible interaction sites, in agreement with molecular docking results. The present study suggests that BioFusionDTI is a robust and interpretable tool for drug discovery and repositioning. The source code is available at https://github.com/QiufenChen/BioFusionDTI.