The forecasting of drug-target interactions (DTIs) is a crucial element in the domain of drug repositioning. Current methodologies, primarily based on dual-branch architectures or graph neural networks (GNNs), typically model binary associations─specifically drug-target or target-disease relationships─thereby overlooking the directional dependencies and synergistic mechanisms intrinsic to tripartite drug-target-disease (GTD) interactions. To address this disparity, we present MTGNN (Multimodal Transformer Graph Neural Network), a comprehensive prediction framework designed to model GTD triplets directly. MTGNN specifically constructs a heterogeneous graph that incorporates direction-aware metapaths to capture biologically significant directional dependencies (e.g., drug → target → disease) and utilizes a dual-path Transformer architecture to integrate both the topological structure and semantic features of biomedical entities (drugs, targets, and diseases). A cross-attention technique is also implemented to dynamically align graph-based and modality-specific semantic representations, promoting improved cross-modal interaction. Comprehensive tests performed validate the effectiveness of MTGNN in precisely inferring GTD connections, exhibiting enhanced performance and generalization capacities. These findings highlight the efficacy of MTGNN as a formidable computational instrument for medication repositioning.