Independent analysis distinguished the similarity of 16 AI-designed phages to natural genomes from the effectiveness of the method used to find them
Independent analysis distinguished the similarity of 16 AI-designed phages to natural genomes from the effectiveness of the method used to find them
On August 12, IEEE Spectrum reported on an independent preprint, a study that has not yet undergone peer review. Its authors reassessed 302 bacteriophage ΦX174 genomes proposed by the Evo 2 model. In laboratory tests, 16 variants infected bacteria. The analysis examined two separate questions: how closely these genomes resembled natural phages and how effectively the overall procedure identified viable variants.
In a Science article published on August 6, researchers described how Evo 2 proposed variants of ΦX174, a small virus that infects bacteria. They fine-tuned the model on genomes from related phages, then applied computational constraints and filters before testing the synthetic DNA in the laboratory. Of the 302 candidate genomes, 17 could not be synthesized. Of the remaining 285, 16 were infectious. This result came from the entire process, not from the model alone.
On August 6, a comparison of cocktails containing designed phages and closely related natural phages showed that the designed phage mixture suppressed resistant E. coli, whereas the natural mixture did not. The independent analysis addresses a different question: how far the functional genomes diverged from natural variants and how effectively the overall procedure found them.
The first measure is evolutionary novelty, which indicates how far a genome lies from sequences already observed in nature. The second is design efficiency, which indicates how effectively the combination of the model, filters, and laboratory testing identifies viable sequences compared with random mutations and serial phage selection.
Across the 16 viable variants, mean sequence identity with ΦX174 was about 97 percent. In the phylogenetic analysis, they fell within the range of natural phages. The preprint authors classify this combination as optimization: the procedure identified viable variants substantially more effectively than the baseline methods while remaining within a closely related sequence family.
Genomic similarity and viral function measure different properties. In the viable Evo-Φ36 variant, gene J encoded a protein that packages DNA. This gene came from the distantly related G4 phage and functioned within the genetic context of ΦX174. In an earlier experiment, directly replacing the ΦX174 J gene with the G4 version produced a nonviable phage. This example shows why both genomic similarity and functional configuration need to be assessed.
The authors propose comparing future genomic systems using two measurable characteristics: the sequence distance between their outputs and natural variants, and the effectiveness of the path from the model through filtering to laboratory testing.