BioGraphX-RNA: A Universal Physicochemical Graph Encoding for Interpretable RNA Subcellular Localization Prediction

Abstract RNA subcellular localization is a critical determinant of cellular function. However, current computational approaches often operate as "black boxes," overlooking the complex interplay among sequence, structure, and physicochemical interactions that govern RNA localization. Building upon BioGraphX, originally developed for proteins, we introduce BioGraphX-RNA, a universal physicochemical graph-encoding framework that provides structure-informed encoding by translating primary nucleotide sequences into multi-scale interaction graphs using explicit biophysical rules. When combined with frozen RiNALMo embeddings through an interpretable gated fusion layer, BioGraphX-RNA outperforms DeepLocRNA and, uniquely, quantifies the relative contributions of sequence and structure for each RNA type, achieving macro-AUROC improvements of 0.0172 for mRNA, 0.0545 for miRNA, and 0.0422 for lncRNA on human datasets. In a blind cross-species prediction task on mouse data, the model demonstrates promising zero-shot transfer performance, suggesting that biophysical localization cues are evolutionarily conserved. Notably, the BioGraphX graph-only model outperforms RNAfold-derived secondary-structure graphs for miRNA (macro-AUROC 0.9482 vs. 0.8787), validating the structural proxy hypothesis under the most stringent possible conditions. Explainability analyses further reveal RNA-type-specific structural dependencies. In particular, miRNA exhibits a near-equilibrium balance between sequence and structure. SHAP-based interpretations provide mechanistic insights, identifying patterned GC content as a potential nuclear retention signal and an anti-structure profile as indicative of exosome-mediated targeting. These advances are achieved with only 2.05 million trainable parameters, aligning with Green AI principles. BioGraphX-RNA therefore demonstrates that explicitly integrating biophysical constraints into graph-based encodings enables accurate, generalizable, and interpretable predictions, advancing structure-aware RNA biology and laying a foundation for precision medicine. Competing Interest Statement The authors have declared no competing interest. Footnotes Adding RNAfold validation and ablation study.