VenusGenome: Genome-Aware Masked Autoencoding for Microbial Genome Representation Learning
Abstract
Genomic-context models have shown that contextualizing protein embeddings can improve predictions of microbial function. We observe that a gene's protein embedding is partially predictable from the embeddings of neighboring genes in the genome, motivating high-ratio masking that requires reconstruction from sparser genomic context. Genomic modeling must also distinguish biological relationships from assembly conventions, since the ordering of separate contigs does not imply adjacency and strand labels depend on the reference orientation. We therefore introduce VenusGenome, a gene-resolution masked autoencoder that combines visible-only encoding with genome-aware attention incorporating contig boundaries, relative strand orientations, and gene positions. We pretrain this model on a curated corpus of over 1.7 million microbial genome assemblies from isolate and metagenomic sources, using frozen protein embeddings as gene-level inputs. Across genome-level phenotype prediction and gene-level tasks, the resulting frozen representations achieve the highest mean performance among the evaluated models on nine of eleven primary metrics. An exploratory gene-input perturbation analysis further links representation sensitivity to experimentally measured fitness defects. Ablations support the contribution of genome-aware attention and show the benefits of masked autoencoding over BERT-style masked modeling, with further gains from higher masking ratios.
Related
- Genome-wide conservation and locus-specific coding variation in Teratoramularia rumicicola, a candidate microbial herbicide for Rumex weeds
- Bayesian adaptive experimental design for efficient microbial genome-wide association studies
- Genome-scale metabolic models for a synthetic soil microbial community as a path for understanding community functioning
- HyperTaxa: Multiscale Genome Representations for Generalized Microbial Taxonomic Inference
- Distinct kinetics and mechanisms of microbial inactivation of enteric virus revealed by capsid and genome integrity
BibTeX
@misc{liang2026venusgenome,
title = {{VenusGenome: Genome-Aware Masked Autoencoding for Microbial Genome Representation Learning}},
author = {Zhang Liang and Zhang Yukun and Tan Yang and Jiang Fan and Zhong Bozitao and Cheng Xiaoran and Hong Liang and Li Mingchen and Yu Yuanxi},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2611.05405},
eprint = {2611.05405},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2611.05405},
}
RIS
TY - UNPB
TI - VenusGenome: Genome-Aware Masked Autoencoding for Microbial Genome Representation Learning
AU - Zhang Liang
AU - Zhang Yukun
AU - Tan Yang
AU - Jiang Fan
AU - Zhong Bozitao
AU - Cheng Xiaoran
AU - Hong Liang
AU - Li Mingchen
AU - Yu Yuanxi
PY - 2026
DA - 2026/11/01
PB - arXiv
DO - 10.48550/arxiv.2611.05405
UR - https://arxiv.org/abs/2611.05405
AB - Genomic-context models have shown that contextualizing protein embeddings can improve predictions of microbial function. We observe that a gene's protein embedding is partially predictable from the embeddings of neighboring genes in the genome, motivating high-ratio masking that requires reconstruction from sparser genomic context. Genomic modeling must also distinguish biological relationships from assembly conventions, since the ordering of separate contigs does not imply adjacency and strand labels depend on the reference orientation. We therefore introduce VenusGenome, a gene-resolution masked autoencoder that combines visible-only encoding with genome-aware attention incorporating contig boundaries, relative strand orientations, and gene positions. We pretrain this model on a curated corpus of over 1.7 million microbial genome assemblies from isolate and metagenomic sources, using frozen protein embeddings as gene-level inputs. Across genome-level phenotype prediction and gene-level tasks, the resulting frozen representations achieve the highest mean performance among the evaluated models on nine of eleven primary metrics. An exploratory gene-input perturbation analysis further links representation sensitivity to experimentally measured fitness defects. Ablations support the contribution of genome-aware attention and show the benefits of masked autoencoding over BERT-style masked modeling, with further gains from higher masking ratios.
ER -