#5 · Genomics

Models infer microbial and viral traits from large genomic and metagenomic datasets

The common thread is inference of microbial and viral traits from sequence data rather than direct measurement. Mean-pooled ESM-2 whole-proteome embeddings of 21,657 foodborne bacterial genomes recover taxonomy and support geometry-based triage with confidence signals. PanBART, a species-specific transformer trained on gene content and gene order of E. coli and S. pneumoniae, supports genomic surveillance tasks such as lineage assignment. V- and VL-scores provide annotation-free virus-likeness for protein families and genomes. Phageome reconstruction from 9,141 human gut metagenomes yields 3,794 modular endolysins. The STEC study of 3,418 isolates across 11 European countries is descriptive rather than inferential. What is new is the pairing of learned representations with explicit assessment of the inference step itself.