AI Designs a Virus That Infects Bacteria, but Only 5% Succeed
AI has, for the first time, designed a complete virus genome from scratch that can infect bacteria; the experimental success rate is only about 5%, far from the all-capable image portrayed in the media.
The video won't play here. Listen to the audio instead:
The argument · tap a timestamp to hear it
Vaccine switch as a natural control
The pitfall of the 'healthy vaccinee bias' is that vaccinated people tend to be more health-conscious than the unvaccinated, so comparing the two groups can mistake lifestyle differences for vaccine effects. This study exploits the 2017 US switch of the shingles vaccine from Zostavax to Shingrix: those vaccinated before the policy change got the old vaccine, those after got the new one, and both groups were actively vaccinated, theoretically eliminating the bias. However, it is still not a randomized trial; unmeasured factors such as socioeconomic status, smoking, and diet may still be unbalanced between the two groups, so it can only be considered a quasi-experiment.
— Vincent RacanielloWeak confounding can overturn vaccine benefit
How strong must unmeasured confounding be to overturn an observational conclusion? The E-value measures this. This vaccine study has an E-value of only 1.42, barely above 1, meaning a confounder that is not very strong and is associated with both 'which vaccine received' and 'cardiovascular outcome' could explain away the observed risk reduction. An E-value of 3.4 or even 5 would indicate a robust conclusion; 1.42 is far from robust. The authors themselves admit that randomized clinical trials are needed to confirm, and this observation should not be taken as evidence.
— Brianne BarkerAI creates first functional virus genome
The researchers used genomic language models EVO1 and EVO2 to write phage genomes: first training the models on massive public DNA sequences, then fine-tuning on PhiX174 to generate thousands of candidate sequences. When tested in the lab, only about 5% of the products could actually infect E. coli, marking the first time AI designed a complete genome from scratch that was proven functional. If the media summarizes this as 'AI directly creates viruses,' they miss a key point: the model only saw sequences, with no experimental annotations telling it which base does what, and the 5% hit rate shows it is still largely ignorant of how sequence determines function.
— Vincent RacanielloReplication origin is the model's most conserved memory
The prompt to the model only needed the first 4 to 9 nucleotides of the PhiX174 genome, and the accuracy of the generated results increased significantly. The authors found that this unremarkable sequence is part of the replication origin and is 100% conserved across all genomes generated by the model. It acts like a password saying 'I am a real virus': the model locks the replication origin as a core element that must be preserved, while allowing other positions to evolve freely. This also explains why such a short prompt is effective—the key is not prompt length but whether the prompt hits a functional Achilles' heel.
— Kathy SpindlerAI swaps a gene and adds 58 mutations itself
EVO36's design is not a full rewrite but a gene swap: replacing PhiX174's J gene with the shorter J protein from phage G4. In the past, when humans manually performed the same transplant, the product could not infect; but EVO36, given by AI, could infect normally because it accumulated an additional 58 mutations across the genome. These mutations appear to be compensatory mutations that allow the foreign J protein to re-adapt to the local environment. Humans might not think to change 58 positions at once, which is where generative models surpass single-point editing.
— Vincent RacanielloPhage cocktail targets drug-resistant bacteria
Natural PhiX174 hits a wall when encountering already resistant bacteria, failing to break through even after five passages; a mixture of 16 AI-synthesized phages infects the originally resistant strain within one or two passages. The resistance is not broken by a single phage but by several synthetic phages recombining during replication, accumulating new mutations, and finally breaking the barrier through collective cooperation. This result suggests that phage therapy should bet on cocktails of multiple synthetic phages rather than a single strain.
— Vincent RacanielloDon't call AI-made viruses an urgent threat
For the study generating PhiX174, a commentary used the word 'urgent' to describe its biosecurity implications. Vincent directly opposed this wording on the show: the authors have already proven through reproducible experiments that such research can be done safely, so the real action should not be to sound the alarm but to start building a safety framework. He also stressed that as long as the sequence is sufficient, reconstructing a virus is not science fiction, and precisely for that reason we need frameworks, not panic.
— Vincent RacanielloNot vaccinating is often just not knowing
When discussing the unvaccinated population, the article raises a frequently overlooked mechanism: some people do not get vaccinated not out of religious or ideological opposition, but simply because they do not know they should. Treating all vaccine hesitancy as a 'belief issue' misdirects policy—for these people, the key is not debate or persuasion but information outreach so 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 Shingrix group | 9% | 9:39 |
| E-value | 1.42 | 16:56 |
| EVO2 training data size | 9.3 trillion nucleotides | 40:15 |
| Effective rate of AI-generated sequences | about 5% (Evo 1: 5.3%, Evo 2: 6.9%) | 52:50 |
| Synthesized and functionally validated candidate genomes | 285 synthesized, 16 effective | 51:45 |
| Number of mutations added by EVO36 for G4J protein | 58 | 56:55 |
| Test score drop corresponding to using AI for homework | 25% | 1:23:27 |
Glossary
- E-value
- A measure of how strong unmeasured confounding must be to explain an observed association; the closer to 1, the more fragile the conclusion.
- healthy vaccinee bias
- Vaccinated people tend to be more health-conscious than the unvaccinated, which can overestimate vaccine effects in observational studies.
- genomic language model
- A large model trained on DNA sequences as text, learning sequence patterns to generate new genomes.
- AS01 adjuvant
- An immune enhancer in the shingles vaccine Shingrix, thought to trigger non-specific protective effects.
- compensation mutation
- An additional mutation that offsets the functional damage of another mutation, common in AI design.
- phage cocktail
- A therapy that mixes multiple phages to reduce the risk of bacterial resistance.
How to listen
Tech journalists and investors at the AI-biology intersection who are skeptical of 'AI-generated virus' reports, as well as evidence-based medicine practitioners who need to evaluate observational vaccine studies.
Readers not interested in vaccine confounding methods can jump to the AI phage section at 38 minutes.