TejasviTomar
October 2026 · Preprint

A simple distance rule beats genomic language models at operon prediction

Tejasvi Singh Tomar

In plain words

In bacteria, neighbouring genes that point the same way along the DNA are often switched on together, as a group called an operon. Language models trained on genomes have been proposed as general tools for biology, and benchmarks test them on tasks such as finding operons. The paper re-ran the twelve protein and genome language models tested in one such benchmark, scoring their operon predictions with the benchmark's own method. All twelve lose to an idea over 25 years old: a simple rule that uses only the direction and spacing of genes. Once gene position is accounted for, the models add almost nothing to finding operons.

Abstract

Language models trained on bacterial genomes are evaluated on operon identification and on protein–protein interaction (PPI) prediction, and good scores are read as evidence that they learn genome organisation. Yet genes in one operon lie on the same strand with short gaps between them, and interacting proteins are often encoded by neighbouring genes. We re-ran the 12 protein and genomic language models of the BacBench benchmark and asked how much of their performance gene position (strand, intergenic distance and adjacency) explains. An untuned rule that uses only strand and intergenic distance exceeds every model's BacBench score on long-read operon maps (0.826 against at most 0.776 AUROC), and it outperforms the operon predictor of the genomic language model gLM on gLM's own evaluation set. At fixed position, the models' embedding similarity carries little or no operon signal, so scores on these operon benchmarks leave open whether the models have learned operon structure. For PPI, the labels from STRING are themselves built partly from genome organisation, and position alone outperforms most models zero-shot, although five of the six models we re-ran for this task keep significant AUROC signal beyond it. In supervised probes, the advantage of the genomic language model Bacformer Large over the protein language model it builds on disappears once position features are added. Benchmarks of genomic language models should report position baselines and what the models add to position.

Gene arrows of different lengths along a horizontal black line, those above it pointing right and those below pointing left. Four runs of neighbouring arrows are filled teal, orange, magenta and yellow, and the rest are white. Arcs in the colour of each run join the neighbours inside it, rising above the upper arrows and hanging below the lower ones. Two shorter grey arcs each join a white arrow to the nearest arrow of the coloured run beside it, and the other neighbours have no arc.
Illustration of every protein-coding gene of Escherichia coli K-12 MG1655 from lysC to plsB, a stretch that contains two maltose operons. The genes are drawn to scale as arrows pointing in the direction in which each gene is read, with plus-strand genes above the line and minus-strand genes below it. Genes that RegulonDB places in the same operon share a colour (yjbE, yjbF, yjbG and yjbH; malG, malF and malE; malK, lamB and malM; ubiC and ubiA). Each white gene forms a single-gene operon. The height of the arc between two neighbours is their score under the distance rule: 1[same strand] × exp(−gap/500), or zero for gaps above 500 bp. The arcs are tall within operons, shorter and grey between them, and absent where neighbours lie on opposite strands or more than 500 bp apart. In this stretch, every same-operon pair scores higher than any other pair. Across all 4,114 RegulonDB-labelled neighbour pairs in the genome, the rule reaches an AUROC of 0.954; the 12 BacBench models reach 0.778 to 0.857 when scored with the benchmark's formula, (cosine + 1)/2 × 1[same strand].