Citrini Therapeutics

Citrini Research is an investment newsletter, and this write-up explains why one went into a wet lab. The team says it was looking for a neglected sector with AI upside, settled on life sciences, and concluded that surveying experts would not resolve whether AI drug discovery was real, because the answers ranged from never to five years. So they ran the experiment themselves: five people, three with no scientific background, with a budget of $200,000, trying to get from an idea to a package a contract lab could synthesize and test, keeping the model as autonomous as possible in between. Their own stated hypothesis was that if that group produced a molecule that did something in a test tube, something had changed.

The target choice is where the exercise got its shape. They wanted a problem where molecular design was the bottleneck, the biology was well enough established for a lab to tell them whether they had made progress, and the test would be quick and cheap, and they landed on binders against scorpion toxins. The design is a minibinder — a small protein built to latch onto a target protein — and the article notes that AI design is further along for proteins than for small molecules like aspirin, which is why a protein was the higher-probability attempt. The practical appeal is that existing scorpion antivenom is made by immunizing horses with harvested venom and purifying their antibodies, a laborious supply chain; the article cites a figure of $39 million per gallon for scorpion venom and US prices in the thousands of dollars per dose. Being an acute treatment, given once or a few times, also carries different development hurdles than a chronically administered protein. The team also set a deliberate moonshot — one molecule binding all five toxins — after the model warned them it was very likely to fail.

What the campaign looked like matters for reading the result. Because of biosecurity restrictions on the relevant model, the prompting was done in person at Anthropic's offices on its machines using an internal research model the article says is more advanced than the released ones; the team describes being blocked by the bio-classifier on their first attempt and being sent a questionnaire about security measures they did not have. The instructions became a roughly 35-page prompt specifying targets, design constraints, evaluation tools and candidate allocation. The article's account is that the model searched the literature, coordinated the specialist design and prediction software that performed the underlying molecular computation, managed compute, and assembled the candidate package — including emailing a vendor contact a numbered list of manufacturing-feasibility questions about which construct formats could actually be built, pausing that subset of designs until the reply arrived hours later and then continuing accordingly. The initial 180 designs came from five human messages and the campaign ended 64 hours after it began; follow-up prompts grew the pool to the 211 sent for synthesis, and the whole design process used a little over 10,000 H100 GPU-hours, with compute billed just under $50,000 and about $150,000 going to the two contract research organizations for synthesis, expression, purification, binding measurements and a developability assay.

The lab results are narrower than the campaign's scale suggests, and the article is careful about the distinction. Thirty-five design-toxin pairs registered as hits, of which 17 passed the team's criteria, spanning four of the five toxins — designed without any laboratory feedback loop to refine them. Binding was measured by surface plasmon resonance, which immobilizes the binders, washes the toxin over them and reads the optical response to infer affinity. The strongest leads were two designs against the toxin Cn2, binding in the low-nanomolar range, and a known antibody against that toxin blocked them in every scorable competition test, which the article presents as encouraging evidence that they attach at an overlapping or nearby site. The moonshot failed as predicted: of 91 designs aimed at all five toxins, one qualified, and only against a single toxin. Crucially, no experiment has yet tested whether any binder neutralizes the toxin — the article says the next step would be to confirm the toxin preparations disrupt the electrical signaling they are supposed to affect, then use a patch-clamp assay, which measures current through cells' ion channels, to see whether a binder prevents or reverses that disruption. Binding, in other words, is necessary but not sufficient; it says nothing about blocking activity, how long the protein survives in a body, or safety.

That gap is the team's own conclusion, and it is more interesting than either enthusiasm or dismissal. Their formulation is that AI in biology will progress in a way that is both alarmingly fast and not fast enough: the design work that would have taken a long time compressed into days, and then they spent two months learning laboratory turnaround times, because DNA still has to be synthesized, proteins expressed and purified, and controls established. What the campaign demonstrates is a design-and-coordination capability — people without the relevant expertise directing specialist tools they could not operate, and getting partially validated novel proteins out the other end — not a treatment. Anthropic published its own technical post, report and assay-level data release on the same campaign, which the corpus covers as Claude Scorpion Toxin Binder Design; the investment conclusions this article was written to support sit behind its paywall and are not part of what is described here.