Entrepreneur Egan Peltan proposes using AI to let one leader manage several drug development projects
Entrepreneur Egan Peltan proposes using AI to let one leader manage several drug development projects
In an essay dated 30 October, Peltan distinguishes between two reasons drug development is slow: lengthy testing and organizational delays. Jesse Brodkin, a participant in the discussion, proposes first investing in animal experiments that more accurately predict treatment outcomes in humans. Peltan suggests considering philanthropic funding for this work.
Peltan is building a startup to preserve physical function in older adults. He starts with a question: how can we tell whether a new molecule will help a person? Drug developers use animal models to study processes associated with human disease and test potential treatments. Peltan emphasizes that these experiments can be poor predictors of effects in humans. When testing a drug to prevent dementia, which affects memory and thinking, researchers need to follow participants for a long time to learn whether they develop the disease. Safety testing also takes time:
Giving a drug to twelve men for one month cannot establish its safety over a year of use.
Peltan builds on drug developer Cody Tranbarger's idea of running several projects with AI. In Peltan's view, funding and coordination among specialists, each holding a different part of the data, create additional delays. He sees a role for AI here: gathering information and coordinating work, freeing the leader to make scientific decisions. Contract laboratories carry out the experiments, and several projects can share overhead costs. If AI takes on more coordination, one leader could manage several projects in parallel, working on others while awaiting results from one.
In Peltan's proposed setup, specialized AI assistants read and update a shared log of data and decisions, while a separate assistant coordinates their work with the leader. Peltan describes his own experience: when summarizing long scientific and manufacturing documents in sections, AI can produce conclusions without citing sources. Those conclusions then need to be checked again. Peltan asks how assistants can exchange information while retaining the source of each claim. A person must be able to trace a conclusion back to the original document.
Brodkin calls for investment in better animal models used to select drug candidates:
Prioritize building laboratories over data centers in the desert.
In his reply, Peltan agrees with this assessment of animal models, expects little near-term impact from AI on this problem, and asks whether philanthropists could fund the development of more accurate models.
Peltan suggests that companies developing AI are entering drug development partly to gain experience with human-AI collaboration. In this field, results take a long time to verify, data are scattered among specialists, and decisions depend on one another. Manufacturing cars and building desalination plants also require coordinating contractors and tracing data back to their sources. He suggests that experience gained at a small drug development company could also be useful in these industries.