Genomic foundation models like Evo 2 contain detectable biosecurity signals that can be extracted with minimal computational cost using frozen-layer probes, enabling fast screening of metagenomic data for antimicrobial resistance and virulence without model fine-tuning.
This paper shows that Evo 2, a genomic foundation model, encodes biosecurity-relevant information in its learned representations that can be detected using simple linear probes without retraining.