Screening of Biosecurity Features in Metagenomic Data with Evo 2 Probes

This study investigates the utility of genomic foundation models, specifically Evo 2, for biosecurity screening. Rather than fine-tuning the entire model, the researchers extracted hidden-state activations from layer-26 of the frozen 7-billion-parameter Evo 2 model. They trained lightweight linear and attention-based probes to classify sequences based on biosecurity-relevant labels, including antimicrobial resistance (AMR) and bacterial virulence. To ensure the probes were learning meaningful biological signals rather than generic functional-gene status or taxonomic artifacts, the researchers used strictly matched negative controls, ensuring that positive and negative training examples were balanced for GC content, sequence length, and taxonomic origin.

The researchers found that Evo 2 embeddings contain highly accessible signals for AMR. A simple linear probe achieved a region-level ROC-AUC of 0.888, which improved to 0.977 when using a single-head attention probe. These probes successfully distinguished AMR genes from unrelated functional genes and could resolve finer-grained drug-class subcategories. Furthermore, the probes remained effective when applied to simulated short-read data, achieving a read-level ROC-AUC of 0.898. This suggests that the approach is robust enough for real-time biosurveillance in complex, low-coverage metagenomic environments where full genome assembly is computationally prohibitive or unreliable. Bacterial virulence was also detectable, though with lower performance (ROC-AUC 0.833).

Traditional biosecurity screening relies heavily on sequence homology against curated databases, which can miss novel or engineered threats. This study demonstrates that genomic foundation models provide a scalable, fast, and inexpensive alternative for first-pass detection. By leveraging pre-trained representations, researchers can identify potential risks in raw metagenomic data without the need for intensive assembly or alignment. This work establishes a framework for embedding-based biosurveillance that could be integrated into automated pipelines to flag emerging resistance elements in environmental or clinical samples.