AI model designs novel bacteriophages that kill E. coli in lab tests

Stanford chemical engineer Brian Hie and graduate student Samuel King used a generative AI tool called Evo 2 to design nearly 300 new bacteriophages based on the natural virus ΦX174. The model wrote entire genomes in a single pass from a short DNA snippet, producing thousands of candidate sequences.

The researchers synthesized and tested nearly 300 AI-designed phages against E. coli, narrowing the candidates to 16 that showed exceptional effectiveness. Several of these AI-designed phages displayed higher fitness than the natural ΦX174 virus.

Because bacteria can evolve immunity to a single phage, the team assembled the 16 genetically distinct phages into a cocktail. In proof-of-concept experiments, this mixture rapidly overcame E. coli strains that were immune to the natural ΦX174 virus.

The ΦX174 genome is relatively short at under 6,000 base pairs, making it a tractable test case for whole-genome design. King developed a computational framework to evaluate thousands of AI proposals, selecting candidates for synthesis based on predicted viability and genetic diversity.

DNA synthesis costs remain high, so the computational screening step was essential to focus resources on the most promising candidates. The framework evaluated genomes for traits derived from ΦX174 and related phages before synthesis and experimental validation.

Hie has released Evo 2 as open-source software, arguing that broad access accelerates beneficial research. He acknowledges potential misuse but contends that existing natural pathogens pose a greater immediate risk and that AI tools can improve defenses against both natural and engineered biological threats.

The researchers are now working to extend Evo 2 to longer and more complex DNA sequences, including small bacterial genomes. Open questions include how to achieve greater genetic novelty and more precise control over the functional outcomes of designed sequences.

The study, published in Science, provides a proof of concept for generative design of entire viral genomes. The researchers envision applying similar approaches to phages targeting other drug-resistant pathogens such as MRSA, tuberculosis, and Pseudomonas aeruginosa.