Ok now AI is writing genomes!

Adam Spencer
10 hours ago
5 min read

An AI model was handed the opening fragment of a virus genome and asked to write the rest. It produced thousands of designs. Sixteen came alive in a dish, and a mixture of them conquered a seriously bad-ass bacterium.

(First published on my Substack where you can get #NerdNews, marvellous maths and general geekery.)

(“The Evo Revo-lution": created by the author via midjourney) — (“*The Evo Revo-lution"*: created by the author via midjourney)

Venter capital.

Earlier this year one of the giants of genetics died.

Craig Venter gained international geek celebrity in 2007 when he led a team that published the first ever entire genome of an individual human. It was his. But hey, if you’re going to go to the trouble of 5 years work and a lazy $40 million to immortalise a chunk of humanity, why not make it your own.

In 2010, another first: his team designed a modified bacterial genome, built it from bottles of chemicals, and transplanted it into a cell of a related species. The implanted genome took over the running of the cell.

Fast forward to 2026. You can read an entire human genome in a few hours, setting you back, say, $500. CRISPR technology regularly edits genetic material altering the running of cells.

And in a beautiful handover, months after Craig shuffled off to the great petri dish in the sky, a genetic GPT wrote an entire virus genome, pretty much from scratch.

“The ability to compose viral genomes using generative AI now exists; the governance to safely steer it does not.” — Thomas Inglesby and Moritz Hanke, Johns Hopkins Center for Health Security, writing in Science.

Phage against the machine.

Brian Hie and his team at Stanford, along with the Arc Institute, have made this breakthrough using a pair of genome language models. In the same way ChatGPT is trained on millions of sentences and phrases to then take a sentence and complete or respond to it, Evo 1 and Evo 2 have been trained on millions of samples of DNA. Feed them a genetic starter, they’ll complete the story.

The subject here was a virus called ΦX174. (That’s phi for the non-Greek scholars among us). It is a bacteriophage, a virus that infects bacteria. Not just any bacteriophage but one of the OGs. Plucked from a Parisian sewer in 1935 no less, in 1977 it ‘Craig Ventered’ Craig Venter, becoming the first DNA genome ever sequenced (props to Fred Sanger).

Clocking in at just 5,386 bases and bearing 11 genes, it is a masterpiece of genetic efficiency. The genes even overlap they are so tightly bunched into ΦX174.

Hie and his crew trained Evo on 14,466 chunks of DNA from viruses related to ΦX174. They then fed in an opening fragment of ΦX174’s genome. Evo wrote thousands of candidate genomes. The researchers applied their genius to build 285 of them and test them against E. coli.

Think of it as training a GPT on the music of 1,000 boy bands. Everything from Backstreet, to 5 Seconds of Summer, with a sprinkle of K-Pop thrown in. You give the GPT the opening lyric and chords from a One Direction song and ding, it writes thousands of singles. A few hundred are taken into the studio and sixteen, it turns out, are absolute bangers.

So the phages packed a punch?

You bet they did.

Of the 285 that were reproducible in the lab, sixteen were viable for the task at hand. Fired at a harmless lab strain of E. coli, they infected it, reproduced inside it and smashed it to pieces.

And the geeks got excited. Samuel King, first author on the paper, described clear patches appearing where the bacteria had been wiped out. Hie says the room broke into applause.

Now for the real test.

Cocktail hour.

Fighting bacteria with phages is standard treatment in some countries and it provides an alternative to antibiotics alone. Unfortunately, these bacteria are crafty little devils and can develop resistance to single phages, just like the antibiotic resistant superbugs you’ve probably heard about. The usual response is a cocktail of phages, but to overcome resistance, it helps to have phages that are different in their make-up.

The Stanford team pooled their designs and tested the mixture against a nastier strain of E. coli that had already become resistant to natural ΦX174. The AI-designed cocktail overcame that resistance quickly. A comparable mixture of naturally sourced ΦX174-like phages did not.

“If the bacteria gain resistance to a single phage, it's game over for the medication. But if you have multiple genetically distinct phages in a mixture, it would be harder for the bacteria to develop resistance to the entire cocktail.” — Brian Hie, Stanford University, quoted in the Stanford Report.

The fantastic phage family.

Every AI-designed phage differed from every known natural phage.

Some competed strongly against the natural version in laboratory tests.

One, Evo-Φ2147, was so different to its nearest natural relative you could consider it a new species of phage.

A particular beast is Evo-Φ36. You heard it here first, this phage will go far!

Φ36 contains a key protein normally found in a distant relative, phage G4. Researchers had tried that swap by hand, but the phage they constructed did not function.

Evo, writing the whole genome at once rather than editing one gene, produced a version that did. The researchers think other changes elsewhere in the genome made room for it, though they cannot yet say which ones.

Screen test.

Um … AI designing viruses that can kill things. Ethics anyone? How do we balance the potential for powerful, lifesaving new drugs against the possibility of AI-designed genomes bringing misery?

Hie and his team were thorough. Evo 1 and 2 were fine-tuned only on ΦX174-like phages. They combatted only laboratory E. coli. They were designed only to combat those bacteria.

Inglesby and Hanke rightly mention governance. Order genetic material from a synthesis company and many suppliers will screen the sequence against databases of sequences of concern, but that screening varies between providers, is not mandatory in the United States, and gets less reliable the more novel the design, because it works by looking for resemblance to what is already known.

If AI wrote an entirely new viral genome, the relevant constituent parts might not feature in any safety screening.

“Whole-genome generative design is now an engineering problem rather than an open question, and engineering problems, I think, will eventually get solved on a schedule — rather than waiting and hoping for a research breakthrough.” — Amy Webb, author of The Genesis Machine.

Brian Hie points out it is far easier, and potentially far more devastating, to simply alter an existing pathogen. Malfeasants turning up the infectiousness of, say, bird flu is a more realistic scenario than inventing an entirely new bio-baddie.

He is probably right … for now.