PMPIHGLL: predicting metabolite-protein interactions using dual hypergraph convolutional networks and large language models.

Metabolite-protein interactions (MPIs) play pivotal roles in regulating cellular processes and metabolic pathways and have important implications for systems biology and drug discovery. Although several computational models have been proposed, alternative approaches are needed to broaden the methodological toolkit for MPI prediction. In this study, we developed a new deep-learning model for predicting MPIs. Chemical and protein large language models were used to generate informative representations of metabolites and proteins. For each entity type, we constructed two K-nearest-neighbor hypergraphs using different values of K, thereby capturing higher-order relationships at different scales and expanding the feature-learning space. A dual hypergraph convolutional network (HGCN) learned from these relationships. The resulting high-level features were further processed using a channel-wise attention mechanism and a one-dimensional convolutional neural network (1D-CNN) to obtain integrated metabolite and protein representations. Finally, a fully connected layer generated the predictions. The model was evaluated on four MPI datasets using global five-fold cross-validation; AUC and AUPR exceeded 0.9 on most datasets. Local five-fold cross-validation and an independent test further indicated that the model could predict MPIs involving previously unobserved metabolites or proteins. The model outperformed several existing MPI prediction models and related models adapted to this task. Ablation studies supported the rationale for its design. Short abstract In this study, we focused on the metabolite-protein interactions (MPIs), which play pivotal roles in regulating cellular processes and metabolic pathways, and designed an effective deep-learning model for predicting MPIs. The model received the chemical and protein features yielded by large language models and refined them using a dual hypergraph convolutional network with two K-nearest-neighbor hypergraphs using different values of K. The resulting high-level features were further processed by a channel-wise attention mechanism and a one-dimensional convolutional neural network. A fully connected layer finally generated the predictions. The model shown competitive performance on four MPI datasets and outperformed several existing MPI prediction models.