Sequence architecture of RNA reveals key molecular signatures driving RNA phase separation

Abstract

Biomolecular condensates are dynamic membraneless assemblies that organize various cellular processes through liquid-liquid phase separation (LLPS). RNA is an active regulator of biomolecular condensates and influence their formation through sequence-encoded features. Despite its established role in condensate formation, no computational framework exists to identify underlying signatures to predict LLPS-associated RNAs. We develop an interpretable hybrid framework to distinguish reviewed, experimentally supported LLPS-associated RNAs from background RNAs. We uncover biologically meaningful sequence properties associated with condensate participation. Each RNA is represented by 167 handcrafted descriptors capturing composition, k-mer usage, sequence complexity, repeat organization and dinucleotide autocorrelation together with 640-dimensional contextual embeddings from RNA-FM. These representations are integrated through cross-validated stacking with neural-network meta-learner. The framework achieves an AUC-ROC of 0.935, AUC-PR of 0.944 and MCC of 0.782. Feature statistics, ablation analysis and SHAP interpretation reveal a distributed sequence architecture dominated by altered local k-mer usage with additional contributions from low-complexity burden, repeat-associated organization and long-range dinucleotide pattern. Biological case studies including MALAT1 and TMPO-AS1 further connect these patterns with experimentally characterized modes of multivalent RNA-protein assembly. Together, the results show how interpretable machine learning can move beyond prediction to identify sequence features that generate mechanistic hypotheses for RNA-mediated condensation.