IonFusion: Multimodal Fusion for Hierarchical Classification and Functional Annotation of Ion Channels and Transporters

Ion channels and transporters are diverse membrane proteins that regulate ion flux across biological membranes and play essential roles in cellular signaling, homeostasis, and physiological processes. Accurate classification and functional annotation of these proteins are important for characterizing membrane protein function, interpreting pathogenic mutations, and identifying therapeutic targets. However, classification of ion channels and transporters is complicated by similar fine-grained classes, hierarchical labels, and functional determinants spanning sequence motifs, domains, and structural features. We present IonFusion, which uses sequence-conditioned cross-attention to aggregate domain-derived and structure-aware representations. To model the hierarchical organization of classes of ion channels and transporters, IonFusion employs a coarse-to-fine autoregressive classifier linking five coarse-grained classes to nine fine-grained classes. For GO annotation, IonFusion combines multimodal predictions with DIAMOND homology scores. Through comprehensive benchmarks, IonFusion achieved 0.988 accuracy and 0.860 macro-F1, outperforming eight baselines. Across three GO ontologies and sequence-identity thresholds of ≤10%, ≤30%, and ≤50%, it achieved the highest Fmax in all nine settings (mean 0.732). Ablations supported contributions from each modality and both prediction components. Attention-based and targeted perturbation analyses highlighted a conserved, class-associated motif in potassium channels. IonFusion provides a multimodal computational framework for fine-grained classification of ion channels and transporters and GO annotation within the evaluated label spaces. By integrating complementary multimodal representations with hierarchical prediction and homology-based fusion, IonFusion improves the discrimination of closely related ion channel and transporter classes and enhances GO annotation performance across different sequence-identity thresholds.