Evo 2 probes detect antimicrobial resistance in metagenomic data with high accuracy

Researchers evaluated the biosecurity potential of genomic foundation models by training linear and attention probes on frozen Evo 2 layer-26 activations to screen for antimicrobial resistance (AMR) in metagenomic data. The study found that these lightweight probes can detect AMR with strong discrimination, achieving a region-level ROC-AUC of 0.977 using a single-head attention probe.

• Linear probes reached a region-level ROC-AUC of 0.888, while single-head attention probes improved this to 0.977.
• The probes resolve finer-grained AMR drug-class subcategories and separate them from unrelated functional genes.
• Bacterial virulence was also decodable, though more weakly, with a region-level ROC-AUC of 0.833.
• The AMR probe maintained comparable ranking performance on simulated short reads without retraining, achieving a read-level ROC-AUC of 0.898.
• Within SynGenome, AMR-associated prompt labels were only weakly recoverable from Evo 1.5-generated sequences.

These results position embedding-based probes as a fast, inexpensive first-pass detection layer for metagenomic biosurveillance, mapping the strengths and limits of using foundation model representations for biosecurity screening.