Paper: Screening Biosecurity Features in Metagenomic Data with Evo 2 Probes

Forum topic · 小凯 · 2026-07-17

Summary

This arXiv paper (2607.14070) by Jeremy Guntoro, Alexander Dack, Dylan Danno, Michaela Jančovičová, Križan Jurinović, and Vanessa Smilansky explores whether biosecurity-relevant signals are linearly accessible in genomic foundation model representations. The authors train lightweight linear and single-head attention probes on frozen Evo 2 layer-26 activations, without fine-tuning the underlying model. On held-out metagenomic test sets, probes detect antimicrobial resistance (AMR) with strong discrimination: a linear probe reaches region-level ROC-AUC of 0.888 (mean-pool), rising to 0.977 with an attention probe. Probes resolve finer-grained AMR drug-class subcategories and separate them from unrelated functional genes, suggesting the learned signal is not explained solely by generic functional-gene state. Bacterial virulence is also decodable, though more weakly (ROC-AUC 0.833). Notably, AMR probes retain comparable read-level performance (ROC-AUC 0.898) on simulated short reads without retraining, enabling pre-assembly screening where assembly is costly or unreliable. Complementary sparse autoencoder analysis recovers interpretable resistance-related features, but with less consistency than supervised probes. The results position lightweight embedding probes as a fast, inexpensive first-pass screening layer for metagenomic biosecurity monitoring.

Screening Biosecurity Features in Metagenomic Data with Evo 2 Probes

Paper Overview

Field: Machine Learning Authors: Jeremy Guntoro, Alexander Dack, Dylan Danno, Michaela Jančovičová, Križan Jurinović, Vanessa Smilansky Published: 2026-07-15 arXiv: 2607.14070

Summary

Genomic foundation models such as Evo 2 learn rich sequence representations, but their value for biosecurity screening is largely unexplored. The authors ask how much biosecurity-relevant signal is linearly accessible in these representations by training minimal linear and attention probes on frozen Evo 2 layer-26 activations, without fine-tuning the underlying model.

Key findings

Strong AMR detection: Across held-out metagenomic test sets, probes detect antimicrobial resistance (AMR) with strong discrimination — a linear probe reaches region-level ROC-AUC of 0.888 (mean-pool), rising to 0.977 with a single-head attention probe.
Fine-grained resolution: Probes resolve finer-grained AMR drug-class subcategories and separate them from unrelated functional genes, providing evidence that the learned signal is not explained solely by generic functional-gene state.
Virulence decodable, though weaker: Bacterial virulence is also decodable with region-level ROC-AUC of 0.833.
Pre-assembly screening: AMR probes retain comparable ranking performance on simulated short reads without retraining (read-level ROC-AUC 0.898, mean-pool), making them usable before assembly in environments where assembly is computationally expensive or unreliable.
Generative label caveat: Within SynGenome, AMR-related prompt labels are only weakly recovered from Evo 1.5-generated sequences; these prompt-derived labels do not establish functionality of generated responsive sequences.
SAE comparison: Complementary sparse autoencoder analysis recovers interpretable resistance-related features, but proves less consistent than supervised probes.

Conclusion

These results position lightweight embedding probes as a fast, inexpensive first-pass screening layer for metagenomic biosurveillance, while mapping both the strengths and current limitations of this approach.

--- *Auto-collected on 2026-07-17*

Paper: Screening Biosecurity Features in Metagenomic Data with Evo 2 Probes

Summary

Screening Biosecurity Features in Metagenomic Data with Evo 2 Probes

Paper Overview

Summary

Key findings

Conclusion

Tags