RiboGRAM learns multi-scale non-coding RNA grammar for structural and regulatory inference

Protein-coding sequences are read through the triplet genetic code, whereas the organization of molecular information within non-coding RNA sequences remains elusive. Here, we identify a non-coding RNA grammar defined by multi-scale motif usage and 5’-to-3’ positional gradients, and develop RiboGRAM to learn this organization. Across nine organisms, non-coding RNAs showed scale-dependent divergence from coding sequences, with directional U-rich and AU-rich gradients. Integrating multi-scale sequence encoding with contextual modeling, RiboGRAM organized RNA families, revealing motif-associated and pair-like patterns across layers. Without structural supervision, non-coding RNA pretraining improved base-pair recovery over coding or messenger RNA pretraining, with representations supporting alignment-free diffusion-based 3D modeling of short RNAs. In human RNA–protein binding profiles, RiboGRAM identified position-resolved motif organization across protein targets and cell types, with sequence-preference shifts associated with reduced cross-cell transferability. Together, these results establish a learnable grammar of non-coding RNA sequence organization for structural and regulatory inference.