VenusGenome: Genome-Aware Masked Autoencoding for Microbial Genome Representation Learning

VenusGenome learns gene-level microbial representations where frozen embeddings achieve the highest mean performance among evaluated models on nine of eleven primary metrics.

• Why it matters: Genomic-context models can improve microbial function prediction, but they must reconstruct masked genes from sparse neighboring context while separating true biology from contig and strand assembly conventions.
• What they did: The gene-resolution masked autoencoder uses genome-aware attention with contig boundaries, relative strand orientations, and gene positions, and is pretrained on over 1.7 million microbial genome assemblies using frozen protein embeddings as inputs.
• The result: Across genome-level phenotype prediction and gene-level tasks, the frozen representations reach top mean performance on nine of eleven metrics, and perturbation analysis links representation sensitivity to experimentally measured fitness defects.