Citrini Reports Nearly $150,000 Lab Bill for Claude Protein Project

Citrini Research’s laboratory bill for a Claude-assisted protein project came to just under $150,000, exceeding its $49,997 compute cost and putting a price on testing AI-generated designs.

The bill covered contract work across Adaptyv Bio and Twist Bioscience, including DNA synthesis, protein production and binding measurements. It gives a concrete cost for the physical work behind Citrini’s AI drug discovery investment thesis. The project does not establish that demand for laboratory services has increased.

Anthropic and Citrini used an internal research model operating within Claude Science to design small proteins aimed at scorpion neurotoxins. During a 64-hour campaign, it generated 318,484 designs and screened 553,052 design-toxin pairs. Researchers chose 180 designs for testing and later added 31, bringing the tested total to 211.

Seventeen proteins met the project’s binding standard, each against one toxin. None qualified against more than one. The designs targeted five neurotoxins from four scorpion species, but the laboratory tested only a small selection of the designs the model generated. The earlier binding results established an initial experimental result, not a treatment.

To qualify, a protein had to show binding in every replicate, reach an affinity of 1 micromolar or stronger and show no signal against an unrelated protein. The strongest designs, Cn2-B1 and Cn2-B2, bound the Cn2 toxin with affinities in the single-digit nanomolar range.

Those measurements used toxins produced in a laboratory. The proteins have not been tested for their ability to stop a toxin’s effects in cells or for efficacy in animals. Binding is an early step in assessing a possible therapy; the results do not establish that any design works as an antivenom.

The design campaign used about 10,000 GPU hours and 12 specialist tools. The roughly 35-page prompt developed for the project was followed by four additional prompts, and people supervised the model’s work. Producing candidate designs still left researchers with the separate tasks of making the proteins and measuring their behavior in the laboratory.

That distinction matters for the investment thesis: faster design can create candidates for testing, but this experiment alone cannot show whether laboratories will receive more work. Its reported costs describe one research project, while further testing would be needed to determine whether any of its proteins can block a toxin.