Screening of Biosecurity Features in Metagenomic Data with Evo 2 Probes
Merged summary
TL;DR — This work probes frozen internal activations of the genomic foundation model Evo 2 to test how much biosecurity-relevant signal (antimicrobial resistance, virulence) is linearly accessible, positioning lightweight embedding probes as a fast first-pass layer for metagenomic biosurveillance.
- Minimal linear/attention probes on frozen Evo 2 layer-26 activations (no fine-tuning) detect AMR strongly: linear mean-pool ROC-AUC 0.888, rising to 0.977 with a single-head attention probe; virulence is weaker (0.833).
- Probes resolve finer AMR drug-class subcategories and separate them from unrelated functional genes, suggesting the signal isn't merely generic functional-gene status.
- The AMR probe transfers to simulated short reads without retraining (read-level ROC-AUC 0.898), enabling pre-assembly evaluation when assembly is costly or unreliable.
- A complementary sparse-autoencoder analysis recovers interpretable resistance features but is less consistent than the supervised probes; SynGenome prompt labels were only weakly recoverable from Evo 1.5-generated sequences. (From AIxBio Hackathon 2026.)
Sources (1)
Screening of Biosecurity Features in Metagenomic Data with Evo 2 Probes
TL;DR — This work probes frozen internal activations of the genomic foundation model Evo 2 to test how much biosecurity-relevant signal (antimicrobial resistance, virulence) is linearly accessible, positioning lightweight embedding probes as a fast first-pass layer for metagenomic biosurveillance.
- Minimal linear/attention probes on frozen Evo 2 layer-26 activations (no fine-tuning) detect AMR strongly: linear mean-pool ROC-AUC 0.888, rising to 0.977 with a single-head attention probe; virulence is weaker (0.833).
- Probes resolve finer AMR drug-class subcategories and separate them from unrelated functional genes, suggesting the signal isn't merely generic functional-gene status.
- The AMR probe transfers to simulated short reads without retraining (read-level ROC-AUC 0.898), enabling pre-assembly evaluation when assembly is costly or unreliable.
- A complementary sparse-autoencoder analysis recovers interpretable resistance features but is less consistent than the supervised probes; SynGenome prompt labels were only weakly recoverable from Evo 1.5-generated sequences. (From AIxBio Hackathon 2026.)