Patrick Malone proposed evaluating AI drug discovery by how well it selects candidates for laboratory testing

Patrick Malone proposed evaluating AI drug discovery by how well it selects candidates for laboratory testing

On August 20, Patrick Malone described his model for AI drug discovery in a post. In this model, scientists direct many specialized programs that narrow thousands of protein or molecule candidates to a short list for laboratory testing.

The model can propose a very large number of plausible protein or molecule candidates, but a laboratory can synthesize and test only a small fraction of them. Malone describes this gap as follows:

“One hundred thousand plausible molecules are useful only if you can identify the twenty worth testing.”

His proposal builds on Claude’s recent miniprotein design campaign. The agent selected a protein region, design software, and candidate lists. Two laboratories then tested whether the candidates bound to the intended protein. Many decisions must be made between the initial model prompt and the experimental result.

Candidate selection is already a separate problem. In a blind evaluation of 511 antibodies, most AI participants performed worse than the baseline strategy on the second task, which required selecting candidates that bound to the target protein. A laboratory needs a list of candidates with the highest probability of producing the desired result in the next experiment.

In Malone’s scenario, the scientist assigns separate stages to specialized programs: literature review, modeling, design, result analysis, and experimental planning. He proposes converting knowledge that usually remains in a specialist’s head into rules, examples, tools, and checks that the programs can use.

Prioritizing candidates requires selecting the appropriate program for each question, allocating computational resources, establishing validation checks, preserving data provenance, and defining evaluation rules. Failures must then be investigated. This process turns a stream of candidates into a list that can be sent to the laboratory.

Malone predicts that a small group of scientists could direct hundreds of programs if they learn how to define tasks, verify the responses, and select candidates before testing. In his view, large flexible proteins and dynamic protein complexes that can adopt several conformations will remain difficult targets for this type of design and validation.