Deep Dive
The diagnostic odyssey: why rare disease takes years
Stav Rones grew up in Boston playing multiple sports but noticed age-appropriate performance gaps in his teens. Physical symptoms—pain, movement limitations—persisted for years without explanation. The medical system failed to connect these early: it took 3-4 years just to recognize a neurological component, then another 5+ years of genetic testing and consultation to land on a diagnosis. This isn't unusual. Alan Beggs explained the math: the human genome has 3 billion bases encoding roughly 20,000 genes, but between 8,000 and 9,000 genes are known to cause disease when mutated. Each gene can break in many ways. The sheer combinatorial space means doctors historically tested one gene at a time. Modern sequencing solved that—now clinicians can read all 3 billion bases at once—but created a new problem: drowning in data. What used to be a few testable hypotheses became hundreds of thousands of genetic variants to evaluate. The human brain can't prioritize that load, so cases stay unsolved for years. Katherine Brownstein's workflow typified the grind: filter for common variants, leaving 5,000 to 10,000 candidates; cross-reference each against disease databases; read literature on every gene; check for de novo mutations or inherited pairs; evaluate deletions and duplications. Hours of work per variant, only to conclude 'no, that's not it.' Rare disease families call this the diagnostic odyssey—and Katherine herself admitted that every diagnosis she reaches haunts her because of the cases she can't solve.
Why AI works: turning haystacks into ranked lists
Yash Shringerpur and the team realized the bottleneck wasn't computation—it was human prioritization. Instead of asking the AI to replace Katherine, they asked it to do the literature work and surface the top candidates. They built a prompt that combined three inputs: the patient's genetic data (filtered to 5,000-10,000 variants), phenotypic codes (numerical ontological terms describing symptoms), and the LLM's training on all known gene-disease associations. The model's job: intersect those three datasets and nominate 2-6 variants worth human attention. This only works because the model can ingest all 8,000+ disease genes in its weights—something no single human expert knows. They tested on solved cases first, discovering the model made systematic mistakes: ignoring sequencing depth, misreading complex inheritance patterns, missing nuances of genetic damage. Each error mode was addressed in the prompt. Iteration lifted accuracy from initial passes to 80-90% on known cases. Once confidence was high, they moved to real unsolved cases. The 376-case study surfaced 18 new diagnoses—some straightforward (a known pathogenic variant Katherine had simply missed), others novel (the S1PR1 case where the model threaded together decades of scattered literature). Alan Beggs noted the model doesn't replace the diagnostician; it tractably reduces an impossible search space to a human-reviewable list. This is the critical shift: AI as attention director, not decision maker.
New diagnoses and the S1PR1 discovery
Katherine described what it felt like watching the model work. Sometimes the output was anticlimactic—a known pathogenic variant that simply hadn't been linked to this patient's phenotype yet. But the S1PR1 case showed the model's real power. A child presented with vitiligo, transposition of the great vessels, and pulmonary hypertension—a rare constellation. The model nominated S1PR1, a gene Katherine had never heard of in this context. She was skeptical. Then she dug: the model had found a paper from 1995 arguing S1PR1 should be considered in vitiligo etiology, plus supporting literature on its role in vascular development. Katherine ran a quick search and discovered the researcher who cloned S1PR1 in 1988 works in the building next door at Boston Children's. She emailed him, he said yes, and they're now collaborating. This is the leverage: not that the model diagnoses, but that it can read 30 pages of PubMed results where a human stops at page 4. Katherine also mentioned a moment of unintended insight: when testing on 20 solved cases, the model got 19 right. But the one it 'missed'—the team had actually been wrong. The case had two genetic causes, but they were fixated on one. The model surfaced both, suggesting it's better at untangling complex phenotypes than human pattern-matching when anchored to a single hypothesis. These wins are why Katherine said 'how can you argue with diagnosing kids'—the model is solving real cases.
Making diagnosis faster and accessible to everyone
Right now, families have to come to Boston Children's Hospital and the Manton Center to use the workflow. That's the immediate bottleneck. Katherine and the team are building a tool to democratize access—one that lets clinicians upload data securely and run the model without flying to Massachusetts. The grant from the OpenAI Foundation is funding this. Alan pointed out that whole genome sequencing is already cheap and fast: under $1,000, costing less than an MRI that insurance routinely covers. The real innovation is making reanalysis routine. As new disease genes are discovered and papers published, old unsolved cases can be automatically re-run. An agentic model could flag new findings and route them back to the researcher. This transforms the equation: instead of one analyst spending weeks on a case once, the case gets continuously re-evaluated for free as knowledge grows. Unsolved cases from 10-15 years ago now have a second life. Of the Manton Center's 5,000 enrolled families, roughly 2,000 still lack a diagnosis. Alan and Katherine estimate the model can rapidly point to answers for a good number of them. The vision is broader still: Katherine imagines a world where an undiagnosed case runs through the system, matches with a diagnosis, and the output helps a clinician select the right treatment or clinical trial. Families can self-refer to the Manton Center—they don't need to be gatekept by a medical team—so reducing time-to-diagnosis matters immediately.
What comes next: from speed to synthesis
Yash articulated the future. Today, the model speeds up existing workflows. Tomorrow, it could autonomously flag new findings as papers drop. But the frontier is synthesis: proposing novel hypotheses about genes and diseases the team doesn't already know. The S1PR1 case is a glimpse. Alan warned he's not worried about losing his job because the work still needs human interpretation. Genetic diagnosis is only the start—understanding implications, sequencing next steps, communicating with families—these all require judgment. But AI makes the process much faster and more efficient. Katherine summed it up: 'We're just getting started.' She envisions a world where undiagnosed cases are rapidly matched with diagnoses and routed to the right treatment or trial. The foundation has shifted. Whole genome sequencing is becoming routine in neonatal ICUs. Some colleagues already achieved diagnosis within 24 hours. Costs will drop further. The bottleneck is no longer reading the genome—it's reading the literature and synthesizing it into actionable leads. O3 and future models handle that. What remains is human expertise: validating, explaining, deciding next steps. And unlike the old diagnostic odyssey, this cycle will repeat. Every new paper, every fresh bit of knowledge automatically makes old cases relevant again.