An AI-designed virus has been given to a human patient, according to the company behind it.

The virus was designed to attack bacteria. Radical Numerics says a treatment containing phages designed with its Omnii-1 model was administered to someone with an antibiotic-resistant infection. The patient’s outcome has not been disclosed. The company’s report

That is already a story worth paying attention to. Put it beside recent announcements about cancer, brain-cell research and scorpion toxins, and I think there is a bigger one.

AI in biology is becoming one of the strongest reasons to be optimistic about AI.

I want more of the AI race to go here.

I am Alex. Welcome to ShortCu8 by Innov8.

Lets Dive Deep.

Phages are viruses that infect bacteria. Using them as a treatment is an existing field of medicine; the new development here is how AI helped design the candidates.

Finding a suitable phage can be difficult because one that attacks one bacterial strain might be useless against another. Radical Numerics used Omnii-1 to help design and select variants for this kind of problem. How the approach works

The company says a cocktail containing its designs eliminated the patient’s bacteria in a lab test. Doctors subsequently administered the investigational treatment through an authorised compassionate-use pathway, which allows access to experimental therapies in serious cases with no satisfactory alternatives.

We do not know whether it helped the patient. The company has kept the clinical outcome private, and this is not an approved treatment for general use.

Still, a biological design involving AI has reached a patient under medical oversight. I think that deserves attention on its own terms. We can be excited about the step without inventing a happy ending.

The cancer announcement starts with something hospitals already collect: a tissue slide.

Sophont, co-founded by Tanishq Mathew Abraham, introduced Holon-S1: a model that predicts where genes are active using images of tissue slides routinely prepared during cancer diagnosis. Those predictions could help researchers understand what makes a particular tumour different. Sophont’s technical report

Its most interesting example involves Sid Sijbrandij, whose cancer treatment involved extensive molecular investigations. Sophont went back to his archived slides and reported recovering several findings from those investigations, including a biological pattern associated with a target of a therapy he received.

The model did not choose his treatment. This was a look back at existing material, and Holon-S1 is currently for research use only. The retrospective case study

What interests me is the possibility of learning more from a sample that has already been collected. If the predictions hold up in further studies, researchers could investigate old slides for clues that previously required additional testing.

OpenAI, PerturbAI and Scripps Research approached another part of the problem: making sense of experimental results.

They used thousands of GPT-Rosalind agents to analyse existing data from millions of mouse brain cells and connect the results with scientific literature. The analysis took 32 minutes. The experiments that produced the data had already happened. OpenAI’s research account

One lead concerned how cholesterol moves inside neurons. Could problems with that movement contribute to degeneration? Could restoring it help?

Those are questions for further experiments. But giving scientists a specific, worthwhile question is useful work. There is a big difference between saying “find a treatment for brain disease” and identifying a mechanism a research team can investigate.

Anthropic’s scorpion project shows that next step.

Claude helped design small proteins intended to attach to dangerous scorpion toxins. Researchers tested 211 designs in a lab. Seventeen met their criteria for binding to the targets. Anthropic’s results

Attaching to a toxin does not establish that a protein neutralises it. The team explicitly says it has not produced an antivenom.

Even so, the result has gone beyond a plausible answer on a screen. People made the proteins and measured what they did.

Across these projects, that is the part I care about: AI helping researchers move from a clue to a candidate they can test.

These results give me a concrete reason to back AI research in biology. Scientists have more ways to investigate existing samples, connect scattered findings and design candidates for experiments.

The next headline is worth reading with one question in mind: what happened after the AI made its prediction?

In these stories, the answers range from a research hypothesis to laboratory measurements to an experimental treatment given to one patient. Those are different achievements. Each deserves credit for what it demonstrates.

For the phage patient, the answer I want next is whether the treatment helped. For Holon-S1, it is whether future studies show the predictions can reliably help guide care.

That is the progress I want to follow. Give me more evidence that someone’s treatment options could get better.

We read every reply - just reply to this email and let us know how we can improve !

#AD1

Your AI is only as good as what you feed it. Feed it a transcript with your product names, acronyms, and numbers spelled wrong, and you spend the afternoon hand-fixing the output.

Wispr Flow Notetaker uses your dictionary and calendar, so product names, acronyms, numbers, and uncommon names come out spelled right. It pops up when your meeting starts and captures Zoom, Google Meet, Teams, and Slack huddles with one click, no bot joining the call.

Then your meetings show up inside Claude or ChatGPT through the built-in connector. No copy paste. Try Notetaker free on Mac and Windows.

Try Notetaker Free

#AD2

Answering a question takes the right article. Doing the work (a refund, a plan change) takes permission. Running Agents in Customer Work, four weekly webinars. Register for All Four