Loading…
Prokaryotic promoter design and zero-shot promoter-strength ranking from a nucleotide language model pretrained on 17,806 microbial genomes.
No providers recorded yet. Browse all providers
A synthetic biologist who has just isolated a promising bacterium wants to dial gene expression up and down in it, and finds that no promoter in the parts registries is known to work there. Machine-learning promoter design has an uncomfortable answer: nearly every method needs a labelled expression dataset from the target organism, so the species that most need designed regulatory parts are the ones least able to supply the data.
PromoGen2 removes that step by learning promoter grammar itself — a series of decoder-only nucleotide language models trained on upstream regulatory sequence drawn broadly enough across the prokaryotic tree that the distribution transfers to organisms never seen with measured expression. The same frozen checkpoint does two jobs: it emits candidate cis-regulatory elements by autoregressive sampling, and it ranks them by mean per-token log-likelihood, a zero-shot proxy for promoter strength that requires no fitting. It was developed by Yi-Xin Huo's group at the Beijing Institute of Technology with collaborators in Beijing, posted as a preprint in November 2024 and published in Nucleic Acids Research in 2026.
Two artifacts sit on top of the pretrained series. Promoter-Factory is a Streamlit application that annotates an unannotated genome, mines putative core promoters from it, and adapts PromoGen2 to that species without the user writing code. PromoGen2-proka removes even the adaptation step: one released checkpoint conditioned on a taxonomy tag, so a bare genus name in the prompt generates promoters for that genus.
GPT2LMHeadModel checkpoints.PromoGen2 is a decoder-only Transformer with GPT-2 positional encoding and single-nucleotide tokenization over a ten-token vocabulary, trained with a next-token objective at a context length of 256. The three sizes use 6, 12 and 30 blocks with embedding dimensions of 320, 480 and 640. Pretraining data came from 17,806 NCBI RefSeq prokaryotic genomes: 160 nt upstream of every coding sequence gave 59 million sequences (the PURD dataset), MMseqs2 clustering at 90% identity left 16 million clusters, and a convolutional strength predictor retained the 1.4 million highest-scoring as PURD-core, the pretraining corpus. Every sequence was presented in both 5'→3' and 3'→5' orientations to offset the autoregressive model's one-directional view. Training ran three epochs; holdout perplexity fell from 1.195 to 1.186 across the size range.
On zero-shot strength prediction against GFP-measured datasets, PromoGen2 raised the average Spearman correlation across species to 0.50 from the 0.27 of the best baseline, while Nucleotide Transformer at 50M and 100M parameters and DNABERT-2 at 117M did not produce usable rankings. PromoGen2-proka came from fine-tuning the 149M model once on 427,477 core promoters spanning 5,800 genomes and 1,757 genera, then generating roughly 2 million putative promoters. In B. subtilis, all 50 tested designs were active and 62% beat the native P43 promoter; 48 of 50 were active in B. licheniformis; and in Jejubacter sp. L23, 95% exceeded the PgadA60 control and drove lycopene titres across a 3.68-fold range.
The model targets metabolic engineering and synthetic biology in organisms outside the handful of well-characterized chassis — newly isolated environmental strains, industrial production hosts, and any bacterium or archaeon whose genome is sequenced but whose regulatory parts are not. Typical use is to generate a graded promoter library for a target species, rank it by likelihood, synthesize only the top candidates, and use the resulting expression ladder to balance flux through a heterologous pathway, as the lycopene work demonstrates. The scorer also triages promoters mined from genomes.
PromoGen2 makes the case that a generative nucleotide language model can be a foundation for regulatory-element design rather than a per-species tool, and that its own likelihood is a serviceable activity predictor — a framing the authors propose as a benchmark for judging whether a sequence model suits cis-regulatory design at all. That a genus tag alone yields functional promoters lowers the barrier for labs in understudied organisms, where the alternative is months of characterization. Limits remain: the models predict relative rather than absolute expression, so hitting a specified strength still requires screening; the annotation path depends on external tools; the in-house convolutional strength predictor that filtered PURD-core was never released, so the pretraining corpus cannot be rebuilt; and the weights carry a non-commercial license even though the code is MIT.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.