This paper tests whether a large DNA language model called Evo2 already encodes signals that point to important biosecurity features. The authors show that simple, fast classifiers — called probes — trained on Evo2’s internal activations can detect genes linked to antimicrobial resistance (AMR) and, to a lesser extent, bacterial virulence, directly from metagenomic data.

The team extracted the model’s layer-26 activations for each nucleotide in genomic regions. They kept Evo2 frozen and trained tiny downstream models on those activations. One probe was a linear map that averages per-nucleotide scores (mean-pooling). Another probe learned attention weights that focus on the most informative nucleotides. Probes were trained on labeled regions drawn from three data sources: MGnify metagenome-assembled genomes (816 chicken-gut MAGs and 255 human-skin MAGs spanning 901 bacterial species), the Virulence Factor Database (34 species), and SynGenome, a set of Evo1.5-generated sequences up to 5 kilobases long.

On held-out metagenomic test sets the approach worked well for AMR. A linear probe achieved a region-level ROC-AUC of 0.888 using mean-pooling. A single-head attention probe improved that to 0.977, indicating that attention can pick out key subsequences. The AMR probe also ranked simulated short reads effectively without retraining, reaching a read-level ROC-AUC of 0.898. Bacterial virulence signals were weaker but still detectable (region-level ROC-AUC 0.833). A separate sparse autoencoder analysis recovered some interpretable resistance-associated features but was less consistent than the supervised probes.

Why this matters: traditional AMR screening in metagenomes often depends on aligning reads to curated resistance databases and assembling genomes. The embedding-based probes operate on Evo2 activations and can flag likely resistance signals early, before costly assembly. That makes them a potentially fast, inexpensive first-pass layer for biosurveillance in complex or low-coverage samples where assembly is difficult.