Language model generates cis-regulatory elements across prokaryotes

Abstract

Deep learning had succeeded in designing Cis-regulatory elements (CREs) for certain species, but necessitated training data derived from experiments. Here, we present Promoter-Factory, a protocol that leverages language models (LM) to design CREs for prokaryotes without experimental prior. Millions of sequences were drawn from thousands of prokaryotic genomes to train a suite of language models, named PromoGen2, and achieved the highest zero-shot promoter strength prediction accuracy among tested LMs. Artificial CREs designed with Promoter-Factory achieved a 100% success rate to express gene in Escherichia coli, Bacillus subtilis, and Bacillus licheniformis. Furthermore, most of the promoters designed targeting Jejubacter sp. L23, a halophilic bacterium without available CREs, were active and successfully drove lycopene overproduction. The generation of 2 million putative promoters across 1,757 prokaryotic genera, along with the Promoter-Factory protocol, will significantly expand the sequence space and facilitate the development of an extensive repertoire of prokaryotic CREs.