PromoterAtlas

Bacterial promoter prediction has been bottlenecked by labelled data. Most published tools were trained on a few hundred experimentally validated promoters from a single organism — usually Escherichia coli — and framed the task as binary classification of isolated sequences rather than genome-wide annotation. The resulting annotation deficit is stark: across the 3,371 gammaproteobacterial genomes used to build PromoterAtlas, GenBank files contained only 263 annotated promoter features against nearly 8.5 million annotated coding sequences.

PromoterAtlas, built by Lucas Coppens and Rodrigo Ledesma-Amaro at Imperial College London, takes a middle path between single-species classifiers and tree-of-life genomic foundation models such as Evo 2 and Nucleotide Transformer. Rather than modelling whole genomes across all domains of life, it is pretrained by masked-token prediction on roughly 9 million 200-nucleotide windows lying immediately upstream of coding sequences, extracted from 3,371 gammaproteobacterial species. Restricting both the taxonomic scope and the genomic window concentrates regulatory signal in the training corpus, and lets the backbone stay small: 1.8 million parameters, small enough to annotate a bacterial genome on a laptop CPU.

The work was posted as a bioRxiv preprint in July 2025 and published in Nature Communications in May 2026. Code, trained weights, and the training corpus are all public.

#Key Features

• Whole-genome promoter annotation: A segmentation head over the pretrained backbone annotates promoters directly in GenBank, GFF3, or FASTA files, writing features back with sigma-factor and box-position qualifiers. An E. coli genome takes two to three minutes on CPU.
• Sigma-factor resolution: Predictions are typed by sigma factor across five classes — σ70/σ38, σ54, σ32, σ28, and σ24 — rather than collapsing to a single promoter/non-promoter call.
• Interpretable logit plots: Plotting per-base pre-softmax logits of the pretrained model exposes which motifs it recognises, including ribosomal binding sites, terminators, stop codons, and transcription-factor binding sites for CRP, NtrC, LuxR, and FNR.
• Frozen-embedding expression prediction: A convolutional regression head on frozen embeddings predicts transcription and translation levels from sequence alone, outperforming both a one-hot baseline and Evo 2 embeddings.
• Tunable precision/recall: Annotation thresholds and motif co-occurrence rules are user-adjustable, spanning a conservative setting that covers ~15% of regulatory regions and a recall-oriented setting that covers ~70%.

#Technical Details

The DNATransformer backbone one-hot encodes DNA, projects to a hidden dimension of 128, and applies eight blocks that each combine a double convolution (window 3), a rotary attention layer, and a 128→256→128 feed-forward layer, all with residual connections and a fixed sequence length of 200. Pretraining zero-masks 20 random positions per sequence with cross-entropy loss applied only at masked positions; training used AdamW, batch size 1024, and an initial learning rate of 1e-3, converging after 264 epochs (about five days on a single NVIDIA A100).

The annotation head is three 1D convolutions (kernel 10) producing a 12-way per-position output, trained on 318 manually curated sequences (147 σ70/σ38, 50 σ54, 76 σ32, 27 σ28, 18 σ24) that deliberately excluded E. coli to leave it as held-out validation. Against E. coli ChIP-seq data and RegulonDB, PromoterAtlas predictions validated at 88% overall — 100% for σ24, σ28, σ32, and σ54, and 85% for σ70 — compared with 74% for ProPr70, 71% for SAPPHIRE, and 58% for ProPr54, all of which predicted many more promoters. For expression, an identical regression head on frozen 128-dimensional embeddings beat both one-hot input and 4096-dimensional evo2_7b embeddings across four transcription datasets (5,392, 4,350, 10,898, and 1,494 sequences), with the best model matching or exceeding Promoter Calculator. Protein-level prediction reached a Pearson correlation of 0.79, near the 0.82 obtained by an ANOVA model using explicit RNA–protein interaction terms.

#Applications

The annotation pipeline fills a practical gap for anyone characterising a newly sequenced gammaproteobacterial genome: promoter features can be added to standard annotation files in minutes without per-genome retraining, and the sigma-factor labels support downstream questions about stress response, motility, or nitrogen regulation. Synthetic biologists can use the expression heads to rank candidate promoter and RBS sequences before construction, and the embedding space offers a route to clustering co-regulated genes across species.

#Impact

PromoterAtlas is an argument for taxonomically focused pretraining: a 1.8M-parameter model trained only on regulatory windows outperformed a 7B-parameter general genomic foundation model on the regulatory tasks tested, at a fraction of the compute. Its limitations are equally clear. Coverage is confined to Gammaproteobacteria and to 200 nucleotides upstream of coding sequences, so distal regulation is out of scope; the annotation head was trained on 318 hand-curated examples; and the expression heads were fit on E. coli datasets, leaving cross-species expression prediction untested. The code, model weights, and the full training corpus are released under an MIT license.