For Rare Disease, OpenAI's o1 Reasons Where Human Memory Fails

Boston Children’s Hospital geneticist Catherine Brownstein just let a new kind of AI loose on one of medicine’s most frustrating puzzles: matching a patient’s baffling symptoms to a single broken gene out of 20,000 possibilities. Her early verdict? The new model catches things its predecessors missed, and it does so by showing its work in a way that looks almost… methodical.

Brownstein’s lab deals with patients who have often spent years without a diagnosis, their data sitting in databases that human experts have already combed through. She tested OpenAI’s o1 model on these deeply complicated cases — the ones where the low-hanging fruit had been picked clean by earlier tools. The model identified gene candidates that had previously been overlooked, not by scraping the web for new research, but by connecting existing dots in a more logical, step-by-step fashion. Where a standard large language model might spout a probable-sounding guess based on statistical likelihood, o1 appeared to reason through biochemical pathways and inheritance patterns.

The core of the advance lies in the “thinking before responding” architecture. For a field dealing with 20,000 genes, each with endless mutations, the limitation has always been human cognitive bandwidth. Brownstein admitted it plainly: “It’s impossible to be an expert in every single gene… you can’t keep everything straight. But AI can.” The o1 model acts less like a search engine and more like a tireless resident working through a differential diagnosis, which is a significant shift from the previous generation of tools that were easily tripped up by the sheer noise in genomic data.

This isn’t a replacement for the clinical lab yet. The model still requires expert validation, and the stakes for a wrong answer are catastrophic. But the ability to surface a fresh candidate gene from a stack of negative tests is a concrete utility, not hype. It suggests the future of AI in medicine won’t just be about summarizing patient notes, but about reasoning through the raw, fundamental data of biology itself — finding the signal that the exhausted human eye simply can’t retain.