An Interpretable Deep Learning Framework Leveraging RNA Foundation Model and Capsule Networks for Accurate Prediction of RNA 2′-O-Methylation Sites

RNA 2'-O-methylation (2OMe) is a widespread post-transcriptional modification that influences RNA stability, translation, and immune recognition. Yet, accurate computational identification of 2OMe sites remains challenging because conventional sequence encodings incompletely capture contextual dependencies and higher-order sequence patterns. Here, we present Caps-2OMe, an interpretable multimodal deep learning framework that combines a Chaos Game Representation (CGR) branch to encode positional and compositional sequence patterns with an RNA-FM branch to capture context-aware sequence representations and long-range dependencies. The two feature streams are adaptively fused and decoded by a capsule-based architecture for robust 2OMe site prediction. Caps-2OMe achieved an accuracy of 0.936 and an AUC of 0.987 on the independent test set, outperforming existing computational predictors. Capsule-level analyses further revealed sparse and class-specific routing patterns, while digit capsule length provided a meaningful estimate of prediction confidence. Motif analysis further revealed biologically meaningful sequence signatures, including an AGAUC-like dominant motif and a CU-enriched local sequence context. Cross-nucleotide evaluation showed that the general model remained stable across four subsets, whereas subtype-specific models exhibited limited transferability. These results establish Caps-2OMe as an accurate and interpretable framework for 2OMe site prediction and provide a useful framework for advancing the computational study of RNA modifications and their regulatory roles.