Transformer-based Prediction of Microbial Growth Rates from Genomic Data

Abstract

Accurate estimation of microbial growth rates is essential for un- derstanding microbial life strategies, their ecological roles, and cultivation potential. Existing computational methods that infer growth rates—particularly minimum doubling times—using codon usage bias (CUB) often struggle with low accuracy for slow-growing microbes. Here, we present a novel transformer-based deep learning model called LookingGlass tailored for microbial genomes to pre- dict doubling times directly from genomic features. Fine-tuned on ri- bosomal protein-coding genes, our model consistently outperforms the widely used CUB-based estimator gRodon, achieving higher predictive accuracy and stronger correlations with experimentally measured growth rates. We further compareLookingGlass to Evo , a general-purpose DNA language model with greater context length, and show that microbial-specific pretraining yields superior perfor- mance. Analysis of 25,000+ microbial genomes from the “Genomes from Earth’s Microbiomes” (GEM) catalog reveals persistent culti- vation biases toward fast-growing taxa across environments and lineages. Our model highlights the potential of transformer archi- tectures for genome-based growth rate prediction and provides a powerful new tool for identifying microbes with high or low cultivation potential in metagenomic datasets. CCS Concepts •Applied computing → Computational genomics; Bioinfor- matics; •Computing methodologies → Information extrac- tion.