#11 · ML / AI methods

Transformer and foundation models are learning microbial traits from genomes and…

The dominant thread is the use of pretrained or fine-tuned transformer models to predict microbial traits from sequence and text. MicroGenomer, a 470M-parameter foundation model, is pre-trained on 234.5 Gbp of microbial genomic data and tuned with GTDB marker genes for ecophysiological trait prediction. A separate study fine-tunes transformer language models on 3.83M articles to extract a bacterial strain-phenotype network and links those traits to predicted genes. HELM-BERT, a DeBERTa-based encoder pretrained on HELM notation for 39,079 peptides, applies similar encoder pretraining to peptide property prediction and outperforms SMILES language models on cyclic peptide permeability. What is new is the pairing of literature-scale text mining with genome-level representation learning, aimed at phenotype and biosynthetic capacity. The items differ in validation targets and output granularity, so the shared element is methodological rather than a common experimental result.