AI designs viruses that infect bacteria, but only 5% of the time
For the first time, AI designed a complete viral genome from scratch that can infect bacteria; experimental validation puts the success rate at only about 5%, nowhere near as omnipotent as the media suggests.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
A vaccine switch as a natural control
The trap in "healthy vaccinee bias" is that vaccinated people tend to be more attentive to their own health than unvaccinated people, so comparing the two groups mistakes lifestyle differences for vaccine effects. This study exploits the moment in 2017 when the US shingles vaccine switched wholesale from Zostavax to Shingrix: those vaccinated before the policy switch got the old vaccine, those after got the new one, and both groups were people who actively sought vaccination, so in theory the bias is erased. It is still not a randomised trial, however — socioeconomic status, smoking, diet and other unmeasured factors could still be unbalanced between the two groups, so it counts only as a quasi-experiment.
— Vincent RacanielloWeak confounding alone can overturn a vaccine benefit
How strong would unmeasured confounding have to be to overturn an observational conclusion? That is what the E-value measures. The E-value for this vaccine study is only 1.42, almost hugging 1, meaning a confounder that is not even very strong, one linked to both "which vaccine was received" and "cardiovascular outcomes", could explain away the observed risk reduction. An E-value needs to reach 3.4 or even 5 for a conclusion to count as solid; 1.42 is far from robust. The authors themselves concede that a randomised clinical trial is needed to confirm it, and that this observation should not be treated as evidence.
— Brianne BarkerAI produces a functional viral genome for the first time
The researchers took the genomic language models EVO1 and EVO2 and had them write phage genomes: first the models read massive public DNA sequence datasets, then they were fine-tuned on PhiX174 to generate thousands of candidate sequences. Back in the lab, only about 5% of the products could actually infect E. coli — the first time AI has designed a complete genome from scratch that was then proven functional. If the media summarises this as "AI directly made a virus", it misses a key point: the model has only ever seen sequences, with no experimental annotation telling it which base does what, and the 5% hit rate shows how ignorant it still is about how sequence determines function.
— Vincent RacanielloThe origin of replication is the model's most conserved memory
The prompt given to the model needed only the first 4 to 9 nucleotides of the PhiX174 genome for the accuracy of the output to rise markedly. The authors found that this unremarkable stretch is part of the origin of replication, and it was 100% conserved across every genome the model generated. It works like a password meaning "I am a real virus": the model treats the origin of replication as a core that must be locked down, and only dares to let other positions evolve. That also explains why such a short prompt works — what matters is not the prompt's length but whether it hits a functional pressure point.
— Kathy SpindlerSwap one gene and AI supplies 58 mutations itself
The design idea behind EVO36 was not to rewrite the whole thing but to swap genes: replace PhiX174's J gene with the shorter J protein from phage G4. When humans performed the same transplant by hand in the past, the product could not infect; but the AI's EVO36 infects normally, because it accumulated 58 additional mutations across the genome. These look like compensatory mutations that exist to re-adapt the foreign J protein to its local environment. Humans might not think to touch 58 positions at once, and this is exactly where generative models exceed the power of single-point editing.
— Vincent RacanielloMixed phages specifically target resistant bacteria
Natural PhiX174 runs into a wall against already-resistant bacteria, failing to break through even after five passages; a mix of 16 AI-synthesised phages infected the originally resistant strain within one or two generations. Resistance is not broken by any single phage but by several synthetic phages recombining with each other and accumulating new mutations during replication, finally breaking the barrier through collective action. The result suggests that phage therapy should not bet on a single strain but should deploy multiple synthetic phages as a cocktail.
— Vincent RacanielloDon't call AI-made viruses an urgent threat
For that study generating PhiX174, a contemporaneous commentary used the word "urgent" to describe its biosafety implications. Vincent pushed back on that wording directly on the show: the authors had already shown through reproducible experiments that this kind of research can be done safely, so what is needed is not alarm-ringing but the start of building safety frameworks. He also stressed that as long as you have enough sequence, reconstructing a virus is not science fiction — precisely why frameworks are needed rather than fearmongering.
— Vincent RacanielloNot getting vaccinated is often just not knowing
Discussing the unvaccinated population, the article raises an often overlooked mechanism: some people do not get vaccinated not out of religious or ideological opposition, but because they simply do not know they should. Treating vaccine hesitancy uniformly as a "matter of faith" misdirects policy — for this group the key is not debate or persuasion but information outreach, so that they first learn the vaccine exists.
In their own words · checked verbatim
It's only sequence. There is no biological knowledge associated with these sequences. But these are sequences from things that work. So the assumption is that you can learn from it.
Vincent Racaniello40:15
But it's not right. The genome sequence is everything. And if you have enough sequences, you could build a virus.
Vincent Racaniello1:09:46
I think it's also kind of cool in terms of helping us with basic discovery in that we might not have known that putting the J protein in in this way would have worked.
Vincent Racaniello1:10:46
It's not a religious thing. It's just they don't know that they should.
In his case, they said, well, your research is not of interest to the American public.
Figures
| Risk reduction for cardiovascular events in the Shingrix group | 9% | 9:39 |
| E-value | 1.42 | 16:56 |
| EVO2 training data scale | 9.3 trillion nucleotides | 40:15 |
| Success rate of AI-generated sequences | about 5% (5.3% for Evo 1, 6.9% for Evo 2) | 52:50 |
| Candidate genomes synthesised and functionally validated | 285 synthesised, 16 functional | 51:45 |
| Mutations EVO 36 added to the G4 J protein | 58 | 56:55 |
| Exam score decline associated with using AI to do homework | 25% | 1:23:27 |
Glossary
- E-value
- A measure of how strong unmeasured confounding would have to be to explain away an observed association; the closer to 1, the more fragile the conclusion.
- healthy vaccinee bias
- Vaccinated people tend to be more health-conscious than unvaccinated people, which makes observational studies overestimate vaccine effects.
- genomic language model
- A large model trained on DNA sequences as its corpus, learning sequence patterns in order to generate new genomes.
- AS01 adjuvant
- The immune enhancer in the shingles vaccine Shingrix, thought to be capable of triggering non-specific protective effects.
- compensation mutation
- An extra mutation that offsets the functional damage of another mutation, commonly seen in AI design.
- phage cocktail
- A therapy mixing multiple phages to reduce the risk of bacterial resistance.
How to listen
Tech journalists who are half-skeptical of "AI-generated virus" coverage and investors at the AI-biology intersection, plus evidence-based medicine practitioners who need to evaluate observational vaccine studies.
Readers uninterested in vaccine confounding methods can jump straight to the AI phage segment at 38 minutes.