Using AI to create novel viruses

Scientists from Stanford University and the Broad Institute reported in the journal Science that they used an artificial intelligence model called Evo, a DNA language model trained on millions of phage genomes, to generate functional bacteriophages de novo. The researchers selected 302 AI-generated phage genomes for testing, chemically synthesised the DNA of 285, and 16 of those produced bacteriophages. The work marks the first time functional phage genomes not previously found in nature have been designed from scratch rather than edited from a natural template.

In 1915, a British bacteriologist named Frederick Twort noticed something strange while studying the bacterium Staphylococcus aureus. On his agar plates, he saw clear patches where bacterial growth seemed to have disappeared, as if the bacteria had died. Twort suggested that the phenomenon might be caused by a virus that attacked bacteria..Two years later, French-Canadian microbiologist Félix d’Hérelle noticed something similar while studying patients with dysentery. When he cultured Shigella, the bacterium that causes dysentery, he found clear patches appearing in the bacterial growth. He investigated further and found that the agent responsible could pass through a filter that retained bacteria yet still kill Shigella. He also showed that this agent could multiply when transferred from one bacterial culture to another. D’Hérelle called this agent a “bacteriophage”, meaning “bacteria-eater”..Soon after the discoveries of Twort and d’Hérelle, scientists discovered several new bacteriophages, and as they studied and characterised them, they began to realise how abundant these organisms are..The advent of DNA sequencing in the early 1970s allowed scientists to read phage genome sequences an organism’s total DNA , confirming what they had long suspected: that phages are the most abundant biological entities on Earth. Estimates suggest that there may be around 1031 phage particles on our planet, more than the estimated number of stars in the universe..Nature generated this diversity with the help of its greatest ally — time. Nature does not invent a new phage from scratch. Instead, when existing phages reproduce, their DNA gradually changes through small errors called mutations, such as insertions of DNA pieces the phage acquired from its host or deletions of existing DNA fragments..Natural selection then acts on these phages with modified genomes, allowing some to survive and causing others to perish. This cycle of reproduction, DNA change, selection and replication happens repeatedly. Over millions of years, these changes can accumulate, resulting in a phage very different from its ancestor..For a long time, we could only mimic this process in the laboratory. We could take an existing phage, edit its genome, or let it evolve under controlled conditions and select variants with the properties we wanted..However, we could not design a completely new phage from scratch. Designing a new, functional phage genome de novo, that is, without using a natural phage genome as a starting point, remained out of reach..Now, however, scientists from Stanford University and the Broad Institute, both in the United States, have reported in a new publication in the journal Science that they have used artificial intelligence to generate functional bacteriophages de novo, meaning that they have created phage genomes that did not previously exist in nature..The researchers did this by using an artificial intelligence model called Evo, a “language model” for DNA. Evo was trained on vast quantities of genomic DNA sequences of several organisms, including several million phages. This allowed the model to learn patterns that tend to occur in functional phage DNA..The researchers then trained the AI on a particular family of small bacteriophages that are known to infect the gut bacterium Escherichia coli. The researchers then gave the AI the DNA sequence of a well-known phage called ΦX174 as a template and asked it to generate new DNA sequences that retained the features necessary for a functional phage while introducing genetic variation..Anthropic quietly sets up biology lab as it ramps AI drug program.The researchers then screened the resulting sequences computationally for characteristics of a typical phage, such as genome size, gene organisation, and host specificity..The researchers ultimately selected 302 such AI-generated phage genomes for experimental testing and chemically synthesised the DNA of 285 of them. 16 of these produced bacteriophages capable of infecting and surviving on their own..Tweaking the genome.However, at the DNA sequence level, these phages were not completely unrelated to anything found in nature. Their genomes retained between 93% and 98% sequence identity to one or another naturally occurring phage. However, at the genome level, even a few per cent difference represents a substantial change..A 5% difference at the DNA level is commonly used as a threshold for distinguishing phage species. Some of the AI-generated phages crossed that threshold. That is, they were sufficiently different from their closest known natural relatives to be considered a different phage species..The researchers then demonstrated the potential of the AI-generated phages by testing a cocktail of the 16 AI-generated phages against two strains of E. coli engineered to resist ΦX174. Neither ΦX174 nor a cocktail of 16 natural ΦX174-like phages could overcome the resistance and infect the bacterium..However, a cocktail of 16 AI-generated phages overcame this resistance and infected both strains, showing that AI engineering endowed the phages with genetic properties the original phage lacked..The authors suggest that this approach could eventually be used to design different phage cocktails with improved ability to overcome bacterial resistance, making phage-based therapies more effective. The authors also imply that their work has broader applications in biotechnology and other antimicrobial applications..While the paper represents a significant advance in our ability to design biological systems, it also reminds us that nature got there first. A phage is classified as “new” when just 5% of its DNA differs from its closest natural relative. We use the same threshold to call the phages we generated by AI de novo. T.hat also means nature solved the other 95% for us. To copy nature, we are assembling the building blocks of life, and in the process, realising that nature built the blocks in the first place.. The writer is an assistant professor of Molecular Biology at a Chennai-based research centre Disclaimer: The views expressed above are the author's own. They do not necessarily reflect the views of DH.