AI-Designed Phages: Machine Learning Creates Viral Genomes!

Is the age of generative biology here?

AI can design enzymes, DNA origami, and CRISPR systems. Can it create whole genomes? Well, the authors of today’s paper generated bacteriophage genomes that were actually viable!

You might have seen headlines of “AI creating viruses”, forgetting to mention that they’re simple bacterial viruses. Well, this is one of the coolest papers I’ve read recently, so let’s do it some justice!

All life runs on DNA.

And some things that aren’t quite life do too. Looking at you, viruses! With just 4 nucleotides, DNA encodes everything from molecular machines to whole organisms! So it’s no surprise that genomes are central to biological research, diagnostics, and synbio.

These days, we’re much better at understanding genomes. DNA sequencing lets us read them fast and (relatively) cheaply, and we’re getting better at writing them too! But there’s still a lot we don’t know, and this makes designing entire genomes incredibly difficult.

And why would you even want to do it?

Imagine creating bacterial strains optimized for bioproduction, cell lines that precisely reflect disease states, or plants engineered to resist pathogens! That could open new possibilities for biomedical research, biomanufacturing, and agriculture.

And AI could help us get there.

Genomes are complicated, and AI is perfect for learning these complex biological interactions. A couple of years ago, researchers at the Arc Institute published Evo, a genomic foundation model trained on over 300 billion nucleotides!

Then came Evo 1.5 and Evo 2, trained on trillions of DNA bases and expanded to genomes across all domains of life. These models have already been used to design genetic components, including CRISPR-Cas complexes, transposable elements, and toxin-antitoxin pairs.

But whole genomes are much harder!

Even a tiny bacteriophage still has to balance coding and non-coding elements, collaborative structural proteins, interactions with host proteins… Much more than a handful of genes.

So, can we go from designing genetic elements to full genomes?

That’s what today’s paper wants to do!

The team used Evo 1 and Evo 2 to generate complete bacteriophage genomes. Starting from the natural phage ΦX174, they evaluated thousands of AI-generated genomes, tested almost 300 of them in the lab, and found 16 viable ones!

And these aren’t just ΦX174 copies.

The generated phages are different from any natural ones, complete with packaging proteins from a distant relative! And the authors showed that the designed viruses can overcome bacterial resistance to phages, a core challenge in phage therapy.

Sounds incredible, right? Let’s see how they did it!

At this point, Evo is a family of genomic foundation models, with 3 versions.

The basic idea is to train on huge amounts of DNA and let the model learn the “rules” that make genomes work: local context, global structures, and evolutionary constraints.

Evo 1 and Evo 2 have already shown that this can be useful for designing individual genetic systems that work in real cells and for predicting the clinical effect of pathogenic gene variants.

But full genomes are much more complicated.

For this paper, the authors focused on bacteriophages, which have the simplest genome you can find. But still complex enough to be a real challenge!

The workflow has 6 steps:

This extensive pipeline shows that with these models, filtering is as important as training. After filtering, the authors kept around 15% of the generated sequences!

Okay, so the generated genomes look ΦX174-like. But do they actually work?

The team kept 302 candidate genomes after filtering. Of these, 285 were successfully synthesized and assembled. Most of the failures were due to complex DNA sequences: yeah, we still can’t assemble all the DNA we want!

Of the 285 tested:

So yeah: a genome model produced phage genomes, and some of them were actually viable in cells!

But how new are they, really?

These phages are in an interesting spot:

Honestly, it seems tricky! But viral taxonomy is hard.

An interesting example is the generated Evo-Φ36.

The authors found that its gene J, a protein that contributes to capsid structure, was replaced by a homolog from a distant virus. The new J protein is much shorter and lacks some domains, but Evo-Φ36 is still viable.

Cryo-EM suggests that capsid interactions are context-dependent, and that a distant packaging protein can be made compatible with the right genomic background. But without it, it doesn’t work: this new J protein is lethal in the base ΦX174!

The genome context is really important for deciding whether a protein swap works!

The authors asked a practical question: Can these generated phages help fight bacterial resistance?

Phage therapy uses bacteriophages to find and destroy bacterial infections. But bacteria evolved resistance quickly, which can make a phage useless. So, the authors tested if a cocktail of their generated viruses can help overcome resistance.

They evolved two ΦX174-resistant E. coli C strains. Then they tested:

Only the generated cocktail could consistently overcome resistance! After two passages, it evolved new mechanisms to attack the bacteria, and the growth inhibition was maintained for at least five passages.

Sequencing showed that the resistant phages were assembled mostly from segments of the generated phage pool, with only a few mutations in the capsid and spike protein. Makes sense, since they interact with the bacteria!

So, is this the beginning of AI-generated biology?

To a certain extent, yes. The team demonstrated the first generative design of complete bacteriophage genomes using genome language models. And it’s super cool! It’s so impressive.

On the other hand, this approach depended on a template genome and a known host. They didn’t really invent a phage from scratch. And the authors are clear that they picked ΦX174 exactly because it’s tractable and safe!

Talking about safety, this kind of work raises some biosecurity alarms.

One could imagine models like this used to create viruses that could infect humans. Evo can’t. The authors intentionally excluded eukaryotic viruses from the training data.

And there are also other bottlenecks between us and that future:

Still, something to keep in mind. But I think it’s incredibly cool that we can do this! And I can’t wait to see what people come up with. Until then, go and read the whole paper; it’s worth it!

If you made it this far, thank you! What do you think of genome language models? Do you think they have a place in biomedicine? Reply and let me know!