RCNNricePPI: leveraging deep learning to predict rice protein-protein interactions for climate-responsive agriculture

Protein-protein interactions (PPIs) are fundamental to cellular regulation, metabolism and stress-response pathways in rice, but experimentally characterized interaction networks remain incomplete. Here, we present RCNNricePPI, a sequence-based deep learning framework for rice PPI prediction that combines word2vec-derived amino acid embedding with a Siamese recurrent convolutional neural network integrating convolutional and bidirectional gated recurrent units. Rice PPIs data were curated from publicly available interaction resources, followed by removal of self-interactions, redundant pairs and low quality sequence. To reduce protein-level information leakage, model development and evaluation were performed using a protein-aware partitioning strategy, including protein disjoint five-fold cross-validation and an independent held-out test set. At a classification threshold of 0.40, RCNNricePPI achieved an accuracy of 0.8996, balanced accuracy of 0.7451 and AUROC of 0.8295, compared with a baseline logistic regression values of 0.6459, 0.6964, 0.7820 respectively. Performance was further assessed based protein novelty, with AUROC values of 0.9259, 0.7995 and 0.7340 for both-seen, one-unseen and both-unseen protein pairs, respectively. Architectural ablation and conventional machine learning comparisons further supported the contribution of the integrated architecture. RCNNricePPI is implemented as an accessible web-based resource and provides a framework for prioritizing candidate interactions for subsequent experimental investigation. The RCNNricePPI is freely available at riceppi.biodb.in .