Abstract
Abstract
Soil ecosystems represent a vast, largely uncharacterised reservoir of microbial diversity. Metagenomic assembly has unlocked access to this resource, and recent advances in long-read sequencing have improved the recovery and quality of microbial genomes from complex samples. While genomic language models have proven highly effective at capturing biological concepts from large sequence datasets, they have predominantly been trained on reference genomes or short-read assemblies. Here, we present LOAM, a family of decoder-only genomic language models ranging from 25 to 624 million parameters and trained exclusively on Oxford Nanopore long-read environmental metagenomes. LOAM model performance scales predictably with model size and training-token budget. Despite a relatively small training sequence corpus, LOAM models out-performed comparably sized models across biological benchmarks, and achieved performance competitive with substantially larger state-of-the-art models trained on much larger datasets. Context-intervention experiments further showed that LOAM models use genomic information over several kilobases, highlighting the potential value of increased contiguity provided by long-read metagenome-assembled genomes. For probe-based benchmark tasks, we systematically evaluated representations across hidden layers and revealed that task-relevant biological information was frequently more linearly accessible from intermediate than final model layers. Finally, we observed that variation in zero-shot variant-effect prediction was strongly associated with the presence of homologous target sequences in the pre-training corpus. Together, these results establish long-read environmental metagenomes as a viable foundation for training competitive genomic language models and demonstrate the importance of both model scale and training-corpus composition for biological generalisation.