CARINA learns transferable tissue-resolved RNA-binding grammars

RNA-binding proteins (RBPs) execute tissue-specific post-transcriptional programs, yet most quantitative binding predictors are trained and evaluated in a single experimental context, limiting controlled cross-tissue comparison, transfer and interpretation. Here we introduce CARINA, a context-calibrated topological grammar framework that learns transferable tissue-resolved RNA-binding rules from heterogeneous CLIP resources. CARINA casts tissue-resolved RBP binding as multi-task conditional grammar learning on RNA sequence-structure graphs. The model constructs transfer-stable grammar evidence through motif-field initialization and probabilistic pairing transport, and converts this evidence into tissue-conditioned predictions through context-calibrated affine reparameterization and multiscale motif energy pooling. CARINA is trained on a GTEx-guided proxy benchmark spanning 160 tissue--RBP combinations across eight human tissues and evaluated under stringent biological and technical dataset shift on 173 independent real-tissue CLIP datasets. Across these settings, CARINA preserves train-in-one-tissue, test-in-another transfer, reconstructs complete locus-level binding architectures in challenging extrapolation settings, including all 15 TARDBP sites within human brain in the NAT8L 3'UTR from a restricted training subset, and yields quantitatively coherent affinity landscapes across independent tissues. Interpretable readouts recover canonical and context-conditioned motif families, enable tissue- and species-resolved motif atlases, and support mechanism-oriented analyses at loci such as MAPT and in tissue-resolved autoregulatory circuits. CARINA thus provides a quantitative framework for comparing how transferable biochemical recognition programs are reused, modulated and deployed across tissues and species.