VenusGenome is a gene‑level masked autoencoder that combines visible‑only encoding with genome‑aware attention. During encoding the model explicitly accounts for contig boundaries, relative strand orientation, and the positional relationship of genes within a genome, preventing artificially concatenated contigs from being treated as true adjacency.\n\nThe model was pretrained on more than 1.7 million microbial genome assemblies from isolate and metagenomic sources, using frozen protein embeddings as inputs—one vector per gene. A high masking ratio (e.g., 70%) forces the network to reconstruct masked gene embeddings from a sparse genomic context.\n\nAcross genome‑level phenotype prediction and gene‑level functional tasks, the frozen representations achieved the highest mean performance on nine of eleven primary metrics. An exploratory gene‑input perturbation analysis linked representation sensitivity to experimentally measured fitness defects. Ablation studies confirmed that genome‑aware attention substantially improves results, that masked autoencoding outperforms BERT‑style masked modeling, and that higher masking ratios provide additional gains.\n\nReview: This work demonstrates the power of embedding explicit genomic structure into a self‑supervised framework, delivering biologically meaningful features that advance microbial function prediction.