VenusGenome learns microbial gene context by hiding most of its inputs

Microbial genes do not operate in isolation. Their order, orientation and co-occurrence help organize functional groups, including operons and biosynthetic gene clusters. VenusGenome, introduced in a preprint by Liang Zhang and colleagues at Shanghai Jiao Tong University, learns from that organization without processing genomes one DNA base at a time.

The researchers report that its representations achieved the highest mean performance among the evaluated models on nine of eleven primary metrics. Those results span organism-level traits and gene-level tasks, but do not establish universal superiority or experimentally validate the model’s biological interpretations.

Learning from missing genes

VenusGenome begins with numerical representations, or embeddings, of the proteins encoded by each genome. These come from a frozen ESM2-35M protein language model. Each gene also carries information about its position, strand and contig—the assembled DNA fragment containing it.

The training strategy follows an observation about local predictability. A linear predictor using two genes on either side of a target reduced protein-embedding reconstruction error by 15.4% relative to a constant mean predictor. Four random genes from the same contig reduced it by 8.9%. Nearby genes therefore supplied useful reconstruction cues beyond the contig’s general composition.

To make those cues less readily available, VenusGenome hides 75% of gene inputs during pretraining. Its encoder processes only the remaining genes, while a smaller decoder reconstructs the missing protein features. It reconstructs embeddings, not DNA sequences or experimentally measured gene functions.

The team pretrained the model on more than 1.7 million microbial genome assemblies from isolate and metagenomic sources. The main version uses a billion-parameter encoder and windows containing up to 4,096 genes.

Respecting genome boundaries

An assembly’s separate contigs cannot simply be treated as consecutive stretches of a continuous chromosome. Their ordering does not establish biological adjacency. Likewise, individual strand labels depend on the chosen reference orientation, whereas whether two genes share a strand remains unchanged when that orientation is reversed.

VenusGenome therefore restricts direct gene-to-gene attention to genes within the same contig, incorporates their positions and learns biases for same-strand versus opposite-strand relationships. Shared global tokens relay information between contigs without inventing local adjacency.

After pretraining, the decoder is discarded and all genes become visible. The frozen encoder produces contextualized gene features; averaging these features supplies a whole-genome representation. Longer assemblies are covered with overlapping windows, but those windows are encoded independently, without direct attention between them.

What the benchmarks showed

The researchers kept representation encoders frozen and trained comparable small prediction networks for supervised tasks. Whole-genome benchmarks separated training, validation and test data by species; supervised gene-level benchmarks used genome-disjoint splits. Supervised scores were averaged over five prediction-network training seeds.

VenusGenome led five of six whole-genome tasks, covering growth traits, antimicrobial response and other phenotypes. For doubling time, Pearson correlation reached 0.336, compared with 0.254 for Evo2, the strongest competing baseline on that task. ESM2-650M remained ahead on optimal growth temperature.

At gene level, VenusGenome led four of five metrics and matched ESM2-650M’s reported mean on essential-gene prediction. Without a trained prediction head, its embedding similarities achieved a genome-averaged Spearman correlation of 0.300 with STRINGDB protein-association confidence scores, versus 0.201 for BacFormer. This measures agreement with database scores, not newly demonstrated physical interactions.

Design comparisons supported the structured attention mechanism: removing contig-aware attention reduced that correlation from 0.293 to 0.203 in the 512-million-parameter model. Higher masking improved this metric at that scale, but left the smaller 85-million-parameter model nearly unchanged. The design-analysis curves represent single pretraining runs.

Fitness signals, with limits

An exploratory analysis replaced each target gene’s protein embedding with its organism’s mean embedding, then measured changes in neighboring gene representations. This computational substitution is not a simulated biological deletion.

Across 48 organisms, the sensitivity score distinguished genes with average experimental fitness at or below −1 from other evaluated genes with a median AUROC of 0.673. An exploratory Evo2 coding-sequence-likelihood baseline scored 0.631. VenusGenome scored higher in 39 organisms.

These findings associate representation sensitivity with measured disruption effects; they do not resolve condition-specific responses. The authors also acknowledge that frozen protein inputs restrict learning to information retained by the protein encoder. The preprint thus supports a reusable microbial representation approach, while leaving condition-aware perturbation prediction as a future direction rather than an established capability.