RNA is increasingly recognized as a central regulator of diverse cellular processes, with non-coding mRNA regions encoding information that shapes subcellular localization, stability, and translation. This regulation cannot be explained by individual factors in isolation; it emerges from combinatorial interactions in which multiple RNA-binding proteins (RBPs) and RNA modifications co-occupy the same transcripts, assembling local interactomes that vary across transcript regions and cellular states. Mapping this landscape experimentally is constrained by the scale of the combinatorial space, the cost of specialized assays, and the context-specific nature of each measurement, leaving existing atlases confined to few contexts. Pretrained RNA Language Models (RNA-LMs) offer a complementary route: as shown by preliminary work in our group, even without modification labels their frozen embeddings recover modified nucleotides along with their canonical motifs, indicating they capture regulatory signal extractable from their representations.

We extend this paradigm to RBPs, treating the latent geometry of RNA-LM representations as a context-agnostic map of the regulatory landscape that encodes the implicit rules of post-transcriptional regulation. Probing this signal across model depth and sequence context quantifies how well the representation space captures known RBP binding preferences. Applied to a broad panel of RBPs profiled by CLIP-based assays across cellular contexts, the framework reveals a spectrum of recoverability. RNA-LM-derived predictions then feed transcriptome-wide non-negative matrix factorization (NMF), uncovering latent regulatory programs defined by co-occurring modifications and RBPs.

Co-authors: Vincent Jung, Raphaëlle Luisier