MARRI: Mechanism–Aware Learning of RNA–RNA Interactions with Foundation Models
Year: 2026
Type: preprint
Abstract: RNA–RNA interactions (RRIs) are central to post-transcriptional regulation, yet their prediction requires accounting for structural accessibility and intermolecular base pairing beyond sequence complementarity. Pretrained RNA foundation models provide representations of individual RNAs, but translating these representations into interaction predictions requires mechanistic knowledge of interaction recognition. Here, we introduce MARRI, a mechanism-aware multimodal framework that couples pretrained representations with domain knowledge by embedding learned opening penalties and hybridization gains into bidirectional cross-RNA attention. MARRI improves ROC--AUC and PR--AUC over the strongest baseline by 12.20\% and 10.25\%, respectively, and maintains strong discrimination for unseen RNA partnerships and interaction regions. Perturbation and repeat-context analyses identify prediction-relevant sequence regions. Integrating MARRI-derived interaction propensities with disease-associated expression profiles links molecular prediction to systems-level models of disease progression and network remodeling. MARRI thus illustrates a knowledge-guided strategy for coupling molecular foundation models with domain knowledge to support grounded prediction, interpretation and systems-level hypothesis generation.
Language: English
Subfield: Molecular Biology
Domain: Life Sciences
Open Access status: green
Funder: American Heart Association
OpenAlex ID: W7236388068
