MicroGenomer: A Foundation Model for Transferable Microbial Genome Representations Enabling Multi-scale Genomic Understanding and Ecophysiological Trait Prediction

Abstract

Microorganisms underpin global biogeochemical cycles and represent a vast, underexplored reservoir for sustainable biotechnology and human health. Their diversity, arising from the interplay of numerous genes within holistic genomic contexts, dictates complex ecophysiological traits across varied environments. To bridge the gap between complex genomic sequences and associated biological functions and phenotypic outcomes, we present MicroGenomer, a foundation model for transferable microbial genome representations enabling multi-scale genomic understanding and ecophysiological trait prediction. MicroGenomer leverages a hierarchical training strategy comprising pre-training on large-scale genomic sequences (234.5 billion base pairs), domain-specific mid-training using the GTDB-curated marker gene set, and task-specific post-training. Despite a streamlined architecture of 470 million parameters, MicroGenomer achieves performance on key tasks competitive with models nearly 85 times its size. Advancing beyond established gene-scale encoders, MicroGenomer generates robust embeddings at the genome scale for downstream modeling. Extensive evaluations demonstrate that MicroGenomer effectively captures phylogenetic structures in species space, excelling in geneand genome-scale understanding as well as ecophysiological trait prediction. The practical utility of these capabilities is further demonstrated by targeted wet-lab validation on newly isolated strains, indicating that MicroGenomers predictions provide reliable guidance for biological experiments. Collectively, MicroGenomer offers a high-performance, resource-efficient framework that transforms raw sequence data into actionable biological insights, providing a powerful foundation for microbiome research and biotechnology.