Discovery of microbial intergenic features with genomic language modeling and multimodal search

Abstract

Systematic characterization of microbial noncoding regions is limited by two distinct challenges: discovery of conserved sequence features without predefined motifs and functional interpretation of newly identified elements. We address these challenges by training a sparse autoencoder on genomic language model (gLM2) representations to identify intergenic sequence features without prior annotation, and by implementing multimodal search to generate functional hypotheses from conserved associations with neighboring proteins, RNA families, and genomic organization. This framework uncovered divergent, previously uncharacterized noncoding elements, including candidate regulatory DNA sequences and structured RNAs not captured by existing annotation models. gLM2-derived intergenic features can be explored through SeqHub's multimodal search, freely available for academic use at seqhub.org.