Systematic mapping of the prokaryotic noncoding sequence space with a genomic language model

Abstract

Prokaryotic genomes encode a vast repertoire of noncoding elements such as ribozymes, riboswitches, and CRISPR RNAs. The sequence-level diversity of these intergenic elements precludes alignment-based clustering that would enable annotation and discovery of novel noncoding RNA families. Here, we show that a genomic language model fine-tuned on prokaryotic intergenic sequences learns representations of noncoding elements from DNA sequence. We used these embeddings to organize 376 million intergenic regions from 1.3 terabase pairs of prokaryotic genomes and metagenomes into 5.14 million clusters. This approach nominated 20,149 clusters of putative structured RNAs with a candidate protein partner. Periodicity in token entropies revealed internal architecture, recovering CRISPR, TIGR and VIPR arrays together with their cognate effector proteins from sequence alone. Targeted clustering of bridge recombinase-adjacent sequences identified bridge RNA variants with longer guide recognition sequences than any characterized ortholog. Finally, we directed teams of AI agents to mine the database under researcher-defined criteria, uncovering divergent riboswitches, ribozymes, and other structured RNAs with biological hypotheses for their function. We release a publicly available database that transforms the dark matter of prokaryotic genomes into an explorable resource for human-agent molecular discovery.