evo-1-5-0134e6a8·1 events·first seen Aliases: Evo 1.5
Researchers trained lightweight linear and attention probes on frozen activations from the Evo 2 genomic foundation model to screen metagenomic data for biosecurity-relevant signals, without fine-tuning the base model. The probes achieve strong discrimination for antimicrobial resistance detection (ROC-AUC up to 0.977 with attention probing) and moderate performance on bacterial virulence, and retain performance on simulated short reads. The work positions embedding-based probes as a fast, low-cost first-pass layer for metagenomic biosurveillance, while also mapping current limits including weak recoverability of AMR labels from Evo 1.5-generated sequences.