Joe Betts-LaCroix proposes confidential FDA evaluation of AI trial predictions

Joe Betts-LaCroix proposes testing AI predictions against confidential FDA data

Joe Betts-LaCroix, CEO of Retro Biosciences, which develops therapies for aging, proposes testing how well artificial intelligence (AI) models predict clinical trial outcomes. He hopes that a sustained record of accurate predictions could eventually let regulators approve drugs without human trials.

That evaluation requires outcomes the model does not already know. If it has read a published trial report, a correct answer may reproduce a known outcome. That answer cannot establish how well the model predicts an unknown result.

Betts-LaCroix therefore proposes that the US Food and Drug Administration (FDA) conduct the evaluation. The agency holds confidential trial data and could compare model predictions with outcomes while keeping developers unaware of which trials it uses. After collecting thousands of predictions and waiting for the trials to finish, the FDA would give developers aggregate feedback on performance.

Moving from that evaluation to drug approval without trials remains Betts-LaCroix's aspiration. The report provides no evaluation results under the proposed scheme and specifies no sufficient accuracy threshold. His proposal also differs from the FDA's draft guidance, which places responsibility for establishing the credibility of a model used to support a drug application on the company. Betts-LaCroix wants the regulator itself to evaluate the models.

Retro itself is conducting a phase 1 trial of RTR242, a drug intended to restore autophagy, the process cells use to clear their own waste. According to Betts-LaCroix, results may arrive around 2027.

Open the related Eternal Search page

Sources
[1] youtube.com
[2] fda.gov
[3] pubmed.ncbi.nlm.nih.gov
[4] fda.gov
[5] t.me
[6] t.me