Back to Feed
Safety & Alignment / Efficiency & Inference

Evo 2 Probes for Biosecurity Screening

Original: Screening of Biosecurity Features in Metagenomic Data with Evo 2 Probes
By Jeremy Guntoro, Alexander Dack, Dylan Danno, Michaela Jančovičová, Križan Jurinović, Vanessa Smilansky Published 2026-07-15

Listen to the summary

Uses a voice available on your device

Audio options
On this page 4 sections
Related concepts 2 concepts

Key Takeaways

  • The paper shows that a simple attention probe, trained on Evo 2's internal representations, achieved a high ROC-AUC of 0.977 for detecting antimicrobial resistance at the region level.
  • The approach is practical for real-world scenarios, as the binary antimicrobial resistance probe maintained a strong ROC-AUC of 0.898 on simulated short reads without requiring retraining.
  • Beyond antimicrobial resistance, the method also demonstrated potential for broader biosecurity applications by achieving a region-level ROC-AUC of 0.833 for detecting bacterial virulence.
  • The use of frozen intermediate layers from a large genomic foundation model like Evo 2-7B-262k allows for efficient adaptation to new detection tasks with minimal probe training.
  • While effective, some limitations include challenges with specific data labels and inconsistencies in complementary analysis methods like sparse autoencoders.

Summary & Methodology Analysis

This research explores the use of genomic foundation models, specifically Evo 2, to enhance biosecurity screening by detecting critical genetic features like antimicrobial resistance (AMR) and bacterial virulence in metagenomic data. The core methodology involves extracting internal residual-stream activations from layer-26 of the Evo 2-7B-262k model, which acts as a powerful feature extractor for input DNA sequences. These activations, which are vector representations of the DNA, capture rich contextual information without requiring direct training on the specific detection task.

To perform the detection, the paper trains minimal single-layer linear and attention probes on these extracted activations. Genomic regions, defined as either coding sequences (CDS) or CDS with flanking contexts from the target dataset, are then scored by these probes. Linear probes aggregate per-token logits, which are raw model outputs before normalization, using mean-pooling, a simple averaging technique. Attention probes, on the other hand, use learned query-based softmax attention, a mechanism that weighs different parts of the input sequence based on their relevance to the query, to aggregate per-token logits into region-level scores. This allows the attention probe to focus on the most informative parts of the sequence for making a prediction. A complementary approach, sparse autoencoder (SAE) analysis, was also implemented to identify latent dimensions, which are underlying patterns, associated with genomic annotations.

The probes demonstrated strong performance, with an attention probe for binary AMR detection achieving a region-level ROC-AUC of 0.977. A linear probe for binary AMR detection achieved a region-level ROC-AUC of 0.888 using mean-pooling. The bacterial virulence probe reached a region-level ROC-AUC of 0.833. Furthermore, the binary AMR probe achieved a read-level ROC-AUC of 0.898 on simulated short reads without retraining, indicating its robustness. However, limitations include that AMR-associated prompt labels in the SynGenome dataset were only weakly recoverable. The sparse autoencoder analysis was less consistent than the supervised probes and showed significant reconstruction loss on short reads. The beta-lactam probe also showed weaker separation against bacterial virulence, possibly due to limited data or biological overlap.

Interactive System Flowchart

Click diagram to expand and zoom

Cross-Examination & FAQs

A deeper dive clarifying mechanics, constraints, and baseline evaluations.

Q1. What is the main problem this paper addresses?

The paper investigates how effectively genomic foundation model representations from Evo 2 can be used to detect biosecurity features like antimicrobial resistance and bacterial virulence in metagenomic data.

Q2. What kind of data does this method work with?

The method is designed to work with metagenomic data, which consists of DNA sequences, to identify specific genomic regions related to antimicrobial resistance and bacterial virulence.

Q3. What were the most significant results?

An attention probe achieved a region-level ROC-AUC of 0.977 for binary antimicrobial resistance detection, and it achieved a read-level ROC-AUC of 0.898 on simulated short reads without retraining.

Q4. Which specific model was primarily used for feature extraction?

The primary model used for feature extraction was the Evo 2-7B-262k model, specifically its internal residual-stream activations from layer-26.

Q5. How do the linear and attention probes differ in aggregating scores?

Linear probes use mean-pooling to aggregate per-token logits into region-level scores, while attention probes use learned query-based softmax attention for the same purpose.

Q6. What is a sparse autoencoder analysis and why was it used?

Sparse autoencoder (SAE) analysis is a complementary approach used to identify latent dimensions, or underlying patterns, associated with genomic annotations.

Q7. What datasets or models were referenced in this research?

The research referenced several models and datasets including Evo 2, Evo 2-7B-262k, MGnify, Virulence Factor Database (VFDB), SynGenome, CARD, ResFinder, AMRFinderPlus, METAGENE-1, DeepARG, DNABERT, Nucleotide Transformer, HyenaDNA, EVEE, and Goodfire Evo-2 SAE.

Q8. What were the limitations identified in the study?

Limitations include that AMR-associated prompt labels in the SynGenome dataset were only weakly recoverable, the sparse autoencoder analysis was less consistent and showed significant reconstruction loss on short reads, and the beta-lactam probe showed weaker separation against bacterial virulence.

Q9. Did the paper compare the performance against other genomic foundation models?

The paper lists several models like Evo 1.5, DNABERT, Nucleotide Transformer, HyenaDNA, and EVEE in its datasets/models section, but the extracted facts do not provide specific comparative performance metrics against these models for the tasks performed.

Q10. What is the main problem this paper addresses?

The paper investigates how effectively genomic foundation model representations from Evo 2 can be used to detect biosecurity features like antimicrobial resistance and bacterial virulence in metagenomic data.

Q11. What kind of data does this method work with?

The method is designed to work with metagenomic data, which consists of DNA sequences, to identify specific genomic regions related to antimicrobial resistance and bacterial virulence.

Q12. What were the most significant results?

An attention probe achieved a region-level ROC-AUC of 0.977 for binary antimicrobial resistance detection, and it achieved a read-level ROC-AUC of 0.898 on simulated short reads without retraining.

Q13. Which specific model was primarily used for feature extraction?

The primary model used for feature extraction was the Evo 2-7B-262k model, specifically its internal residual-stream activations from layer-26.

Q14. How do the linear and attention probes differ in aggregating scores?

Linear probes use mean-pooling to aggregate per-token logits into region-level scores, while attention probes use learned query-based softmax attention for the same purpose.

Q15. What is a sparse autoencoder analysis and why was it used?

Sparse autoencoder (SAE) analysis is a complementary approach used to identify latent dimensions, or underlying patterns, associated with genomic annotations.

Q16. What datasets or models were referenced in this research?

The research referenced several models and datasets including Evo 2, Evo 2-7B-262k, MGnify, Virulence Factor Database (VFDB), SynGenome, CARD, ResFinder, AMRFinderPlus, METAGENE-1, DeepARG, DNABERT, Nucleotide Transformer, HyenaDNA, EVEE, and Goodfire Evo-2 SAE.

Q17. What were the limitations identified in the study?

Limitations include that AMR-associated prompt labels in the SynGenome dataset were only weakly recoverable, the sparse autoencoder analysis was less consistent and showed significant reconstruction loss on short reads, and the beta-lactam probe showed weaker separation against bacterial virulence.

Q18. Did the paper compare the performance against other genomic foundation models?

The paper lists several models like Evo 1.5, DNABERT, Nucleotide Transformer, HyenaDNA, and EVEE in its datasets/models section, but the extracted facts do not provide specific comparative performance metrics against these models for the tasks performed.

Browse the archive →
Efficiency & Inference / Benchmarks & Evals Memory Efficient Optimization for Model Adaptation The paper introduces a curvature aware zeroth order optimization method that reduces memory usage during test time model adaptation without requiring backpropagation. Reinforcement Learning / Safety & Alignment Automated Cloud Threat Detection Using Reinforcement Learning The authors implement a Deep Q-Network to provide real-time, autonomous intrusion detection and threat mitigation for cloud infrastructure. Multimodal / Efficiency & Inference Enhancing Visual Understanding Without Inference Overhead The paper introduces Generation as Auxiliary Supervision, a method that uses a decoupled generation branch during training to improve Multimodal Large Language Model visual understanding while adding zero overhead at inference time.