[Meta-path-enhanced multimodal heterogeneous hypergraph Transformer for herbal compound-target interaction...

Herbal compound-target interaction(CTI) prediction is a key computational task for elucidating the multi-component and multi-target mechanisms of herbal medicines and identifying potential bioactive compounds. However, existing methods still have clear limitations in integrating multimodal information and modeling high-order compound-target relationships. To address these issues, we proposed a multimodal Meta-path-aware heterogeneous hypergraph transformer model, termed MMPAHHGT, and established a unified prediction framework that integrated multimodal biological feature learning with Meta-path-driven heterogeneous hypergraph representation learning. Specifically, the model first integrated the herb's property semantics, compound SMILES-based structural representations, target protein sequence representations, and functional and disease semantic priors to generate high-quality multimodal embeddings for compounds and targets. It then constructed a multi-view heterogeneous hypergraph based on Meta-path instances. A two-level attention propagation mechanism, including intra-hyperedge and inter-hyperedge attention, was introduced. This mechanism was further combined with type-aware co-attention, multi-head attention, and residual connections to effectively capture high-order interactions in the compound-target network and reduce interference from noisy links. The model ultimately output the probability of CTI. Experiments on two public datasets show that MMPAHHGT consistently outperforms multiple state-of-the-art baseline methods in terms of AUROC, AUPR, and F1-score. Systematic ablation studies further demonstrate the critical contributions of multimodal biological representations, hypergraph attention propagation, and Meta-path semantic fusion to the performance improvement of the model. Finally, molecular docking analysis supports the biological plausibility of the predicted results by the model.