Investment firm Citrini Research reports 49% accuracy for a computational assessment of protein binding to a toxin in an AI protein design project, says collecting laboratory data to improve it took about 30 times as long as generating molecules, and prompts debate over whether computation can solve biology

Investment firm Citrini Research reports 49% accuracy for a computational assessment of protein binding to a toxin in an AI protein design project, says collecting laboratory data to improve it took about 30 times as long as generating molecules, and prompts debate over whether computation can solve biology

On October 25, Citrini published a post on X that received about 340 thousand views. The firm divides researchers into an “AI camp” and a “biology camp” and cites two figures from a project with Anthropic.

The “AI camp” believes in scaling laws (models improve with more computation and data) and the “bitter lesson” described in AI researcher Richard Sutton’s 2019 essay: general methods that use computation ultimately prevail. This camp expects powerful AI to deliver medical discoveries. In a talk on AI and discovery, Sutton said that hypotheses also need to be tested and selected. According to Citrini, biologists find the “bitter lesson” rather absurd and expect reliable computational modeling of drug toxicity in humans to be at least 5–10 years away. Citrini adds that even most committed AI supporters firmly reject the idea that “superintelligence” will discover treatments from existing knowledge without substantially more biological data.

In the project with Anthropic, the model generated 318 484 protein variants designed to bind scorpion toxins in 64 hours. Laboratory testing of 211 proteins took about two months, according to Citrini’s article; 17 bound their intended toxin. Citrini reports that the computational assessment of binding was 49% accurate. The figure of 49% also appears in the Anthropic report, where it refers to the proportion of pairwise comparisons: among 60 proteins, each designed to bind a single toxin, the assessment ranked a binding protein above a nonbinding protein in 49% of pairs. Random ranking would yield 50%. According to Citrini, improving the assessment requires laboratory data, and collecting those data took about 30 times as long as generating the molecules. These were test-tube experiments; studies in humans would have taken longer.

A user posting as Mira responds that exponentially more computation would allow biology to be simulated accurately without experiments: “Experiments are a way to save computation, a holdover from when computation was much more expensive.” Mira also argues that “AlphaFold worked,” referring to the program that predicts protein structure. Aging researcher Morgan Levine asks where the data for such a simulation would come from: the genomes of only an estimated 1% of eukaryotic species have been sequenced, and “infinite computation will not help simulate a system that has never even been observed or measured.”

Tanishq Abraham, head of the medical AI company Sophont, argues that solving biology depends on data. With nearly infinite computation, it could probably be solved, but a simulation requires the system’s initial state, and we cannot yet describe that state in full. Abraham considers Mira’s example unconvincing because AlphaFold succeeded only thanks to the Protein Data Bank, a database of protein structures determined experimentally over decades.

Martin Borch Jensen, founder of Norn Group, which funds longevity research, frames the debate as a calculation: $50 million for data from long-term human studies is equivalent to about one day of a leading AI laboratory’s computing budget, while the time needed to collect those data cannot be compressed. “If pure simulation does not need these data, treatments will be delayed by one day.” If the data help the models, treatments will arrive months or years sooner; if the data are essential, treatments will arrive years sooner. Norn Group is raising $15 million in grants for such studies. It says that unless the studies begin now, future AI will lack those data.