Rare Diseases Are Common: How a Harvard Gene Hunter Uses o1-pro on the Cases Sequencing Already Missed

Nathan Labenz introduces Dr. Catherine Brownstein, MPH, PhD, an assistant professor at Boston Children's Hospital and Harvard Medical School whose job is to find genetic causes for previously unexplained rare and orphan diseases. She had just appeared on OpenAI's o1-pro announcement as a ChatGPT Pro grantee. Labenz's frame is that even obsessive AI users have only a few thousand hours with large language models, most of them on older systems, so a specialist who shows up at the current frontier as a beginner is not behind. She may be first.

Asked for a lay introduction to her work, Brownstein does not start with a gene. Rare diseases, she says, are quite common. There are more people with them in the United States than there are natural blondes. Autism is common; autism due to a de novo variant in `KCNJ8` is not. The official bar is also larger than intuition suggests: fewer than 200,000 people in the United States, or fewer than one in 2,000. A city of 200,000 does not feel like a small town. Epilepsy, by comparison, hits about one percent of the population. She calls herself a gene hunter. The work is diagnosing the undiagnosed.

The typical path to her desk is a diagnostic odyssey. A family sees a local doctor, then specialist after specialist, often for years, and finally reaches Boston Children's, an internationally known tertiary hospital. Savvy parents sometimes shortcut the maze by finding the person who publishes on that phenotype and calling until they get an appointment. Many others arrive after a long bounce. The Manton Center for Orphan Disease Research, which she works with, is a philanthropically funded virtual center. Patients can self-refer. Negative genetic cases from across the hospital flow in, with records, prior testing, a multidisciplinary review, reanalysis, and sometimes a new technology such as RNA-seq or long-read sequencing.

When she started in 2011 she was hired as a project manager to help clinicians get patients sequenced. An exome — one percent of the genome, the coding region — cost about $3,800. Clinicians had freezers of DNA waiting for the technology. Selection was rigorous: you bet the condition was genetic, you sequenced a trio, and you often found a de novo stop codon or a huge deletion in a gene already hypothesized for a related condition. Diagnosis rates hit about 80 percent. It was, she says, shooting fish in a barrel. Whole-genome sequencing in 2007 had cost upwards of a million dollars. Labenz notes that the drop toward a couple of hundred dollars is close to a 10,000x, one of the few cost curves that rivals large language models.

The same exome now prices at $160. For-profit labs and working geneticists absorb the slam dunks. What reaches Manton is a negative genome: no obvious variant in a known gene. Hit rates fell to about 10 percent because the remaining cases have already been reviewed by good geneticists. Industry-wide, only 25 to 33 percent of genetic tests come back positive. Brownstein thinks AI can still shorten the odyssey earlier in the chain — flag the right specialist, the right test — so families are jerked around less, not through anyone's fault, but by how the system is built. The cases that remain are solved as cohorts: four families, ten families, the same gene or gene family, a hypothesis you can actually test.

In 2014, Zak Kohane — a previous guest on the show — ran an international contest on three families with seemingly Mendelian disorders. Twenty-three teams worldwide received the data. Two of the three families walked away with diagnoses. The result that stuck: diverse teams beat rooms of bioinformaticians. Research assistants, genetic counselors, clinicians, and clinical geneticists together solved more. Brownstein treats that as a theme for LLMs as well. None of this exists in a vacuum. They still see slam dunks. One family had three generations of a rare bone disorder; the patriarch was in his 90s. He had already had surgeries he did not need. Being able to explain the cause in ten seconds, rather than three minutes of symptoms, meant a great deal.

A case on her desk now usually arrives with a thick record and prior genetic testing, increasingly on a thumb drive. She runs more than one genomic pipeline because each has a personality: some are comprehensive and noisy, some are intuitive black boxes that hide why a variant was dropped. She is a PhD, not an MD, and prefers cases anonymous so she is not a walking HIPAA violation, though she does meet families. Phenotypes get coded into ontologies — she likes HPO best — where "no tears" is alacrima and "lazy eye" is strabismus. Melissa Haendel's Monarch Initiative work on mapping those terms onto animal phenotypes, plus layperson HPO, is a field unto itself. Bad translation makes the downstream ranking worse. Inputs range from raw FASTQ off the machine to BAM reads to a still-huge VCF of variants. Pipelines then order variants by HPO, gene, and predicted pathogenicity, and the better ones can promote a gene that binds a known disease gene even if it is not on the list yet.

Her first success was episodic ataxia: a patient who went stiff and locked in place. Sequencing found a variant in `KCNA1`, not the gene they had in mind, but related to it — and it rose straight to the top. When the known-gene pass is empty, she looks for special sauce: structural changes, translocations, deletions, epigenetic expression stuck on, or a burden of small hits via GWAS or SKAT. Most cases still go on the shelf. Reanalysis a year later matters because science moves. Often the number-one variant in the browser is now the answer, and you had stared at it. Alan Beggs and Monica Wojcik, Manton director and medical director, told her about three siblings who died of a myopathy twenty years ago. DNA ran out. An RNA-seq pilot on leftover sample found a `CFL2` variant. Surviving siblings, now planning families, could test partners and avoid passing on a fatal myopathy. The diagnosis did not change the past. It changed the future.

After a negative standard pass she goes looking for anomalies, then asks what the protein looks like. AlphaFold, STRING-DB, CADD, SIFT, and PolyPhen are in the toolbox, and she refuses to treat the scores as gospel. Some known gene-disease relationships would fail today's filters because a tiny perturbation in that gene is enough. Experience is a cold shower: some genes are mutated in almost everybody, and a veteran will say it is never that gene, "it's never lupus." Then you see a gene you have never met, conserved to zebrafish or *C. elegans*, mangled in AlphaFold, and you get excited. A lot of the time it still falls apart: the variant is common in an ethnicity that is hardly ever sequenced, and the patient happens to be from that group. That is why, she says, we need to sequence the whole world — to tell disease-causing change from background variation. In 2011 a finding like that went straight toward a paper. In 2024 the waterline is higher. You want other families, via Matchmaker Exchange or Beacon, or you join someone already writing up 19 cases.

Labenz asks whether the tiny n in rare-disease papers is a true scarcity of patients or a man-made scarcity of sharing. Brownstein will not fault a young investigator who sits on a beautiful case hoping to find a second and publish first-author instead of becoming a middle author in someone else's series. The problem is that it often never works out that way, and the patient does not benefit. Children's has a CRDC cohort system: de-identified genetic data, no names, no phenotypes. She can query a candidate gene across everyone sequenced in the hospital, email the treating physician, and join a study. She has found four other patients that way, reached a group in the Netherlands assembling a series, and ended up with an accurate, comprehensive description of the condition rather than a lonely case report.

If sharing moved from today's reality to something closer to ideal, she thinks the difference would be huge: more discoveries, more diagnoses. Cohorts sit in the back of the freezer unsequenced, not from apathy so much as from how hard it is to do the work and to get there mentally. Her advice to families who email her is to enroll in several programs and registries, not one. When a new finding lands, researchers go to the registry. That is how you keep a sample from waiting forever. Inertia shows up in defaults. One of her 2011 jobs was to help shift the biobank from opt-in to opt-out for leftover samples. Years later, she says, it is still opt-in in many places.

Research findings also have to be CLIA-confirmed on a new sample and returned through a counselor or physician before they count. Some doctors will not play: they do not see the value, or, in a case in her own family, the physician said he did not have email. Two floors down, Tim Yu made milasen, an n-of-1 drug for a girl with Batten disease, after an extremely motivated family broke through every barrier. Labenz's dark-matter point is that every inspirational story stands in for a hundred families who could not clear those hurdles. He reaches for nuclear power: thousands of weapons, still burning fossils, benefits stuck behind obstinance. Brownstein wants AI, among other things, to put more autonomy in families' hands so they are less dependent on infrastructure that does not work.

Labenz's own patient habit is to talk to Claude or ChatGPT before a visit, and if the model clams up, to say he is preparing for a conversation with his doctor. Brownstein's biggest research impact is more ordinary: summarizing articles and genes. Looking up a paper, hitting a paywall, logging into the Harvard library, skimming an abstract that is not what she wanted, and repeating that loop used to eat the day. Asking for a summary and either keeping the paper or moving on has, she says, changed her life and given her hours back. Boston Children's has put ChatGPT behind the hospital firewall, so she is not worried about data leaving. Citations still fail in ways she finds hilarious. Hallucinations are getting better. She hopes one bad headline will not be the only story people read.

She runs a core facility, so she has to learn new sequencing methods fast. A technique with one paper attached to it becomes a readable summary instead of an afternoon in the literature. Before meeting investigators she summarizes their work. There are close to 20,000 genes; a case that names a condition she does not know becomes a one-pager with the latest and who is working on it. Reviewer two's complaint that "there is a whole body of literature on this" becomes something she can put in context. All of it, she says, cuts the mundane part of the job so she can return to gene discovery — going through 20 candidates and narrowing to three to present in an hour and a half. True story. Rank by CADD and conservation, then a quick ChatGPT relevance pass.

The worked example is interstitial cystitis, severe undiagnosed bladder pain, sometimes so bad people cannot leave the house, with no slam-dunk gene. She has about 500 patients. With collaborators at Columbia she looks for genes and pathways with more variation than expected, then asks, pathway by pathway, whether the genes tie to bladder, bladder pain, or bladder cancer — often as a yes/no string, then she drills the yeses and still checks the nos. One small-molecule transport pathway had 60 percent of its genes tied to bladder cancer, meaning they are expressed in the bladder, with known perturbations that cause dysmorphology. She almost screamed when a gene linked to urothelial issues, a plausible mechanism; she had a meeting the next morning. She started in genetics after the genome was published and does not know how people did this before. There will be a whole generation of geneticists, she says, who will not know how the work was done without these tools.

She was skeptical at first that checking every claim would save any time. Accuracy and reasoning keep improving, and sometimes the model answers as a genetic counselor and then offers another way to look at it that surprises her. She is a convert, relatively early. The destination she wants is solved cases, a shorter odyssey, and more democratic access to genetics interpretation — if the system can convince people it is accurate enough of the time. She admits a double standard: a human can miss something and say sorry; an AI miss will not be forgiven. "I guess that's the way it should be." Labenz's self-driving prior is that the machine may need to be about ten times safer to be accepted in a patient-facing role. Tools for professionals can be used earlier, with the human still holding the bag.

In the nitty-gritty she is not using web search. She is mostly on o1, playing with GPT-4o, with no sophisticated taxonomy yet — comfort, then iteration. Four-o's reasoning impresses her. Web search still scares her a little because there is a lot of garbage online and medical answers have to be ones she can stand behind. Checking everything is still paramount. Thinking time has shortened by the day; prompts are getting better too. Labenz's failure mode is leaking a preconception so the model mirrors it back. Brownstein's is the opposite excess: she asks whether a gene relates to a phenotype and the model leaps three steps down an interaction graph. Intellectually impressive, not what she asked. She wants a paper that links the gene to the phenotype, or a yes that it is upregulated in cancer, not a mechanism she is not making. Is that Move 37 brilliance or a tangent? She does not know. Faith, an open mind, keep using it.

She is not dumping whole medical records into the model, even behind the firewall. She summarizes. Finding the limit — how much it can handle, how much it can replace her — is part of the OpenAI grant. Compute, tangents, and utility all bind. Not everyone who smokes gets cancer. Every question takes energy; you cannot throw every record and every genome at the model to see what falls out. The interesting future is predictive: mining the chart on steroids, and a system that says you might need a colonoscopy at 24, not 40. Labenz notes Sam Altman has said OpenAI still loses money on $200-a-month o1-pro. Relative to a medical team, that price looks cheap. The grant relationship is close: use cases, feedback, mail at 11 p.m. on a weekend. They were just getting back after the holiday. In a few months she hopes to have something more exciting to report.

Asked to picture a year out, Brownstein stays conservative. For patients, wrapping the free-form prompt into something guided and user-friendly is low-hanging fruit. For researchers, her institution is still surveying what people actually want the tools for. She would love to say they will be solving cases. She does not know if that will be true. She thinks the models will be more intertwined in day-to-day work. Labenz is more aggressive. He doubts she is on the o3 review team yet and cites FrontierMath: about 25 percent at a very high compute setting, and still about 10 percent at low compute, against a previous ceiling around 2 percent. He would take the side of the bet that some cases get thrown at an o3 — or an o4, given how short the gap from o1 to o3 was — and come back with meaningful conclusions within a year. They put a reunion on the calendar.

What a family can do without waiting for that reunion is already on the table. Talk to a model before the appointment, with the "I am preparing to see my doctor" prefix if it refuses. Stay neutral in the prompt. Demand a short answer when that is all you need. Use summaries as a literature filter. If you live with a rare disease, register the sample in more than one place, and push for CLIA confirmation when a research lab finds something. If you are a specialist in a narrow field, this is a reasonable moment to start. Brownstein is doing that in public, with a beginner's curiosity and a professional's refusal to trust a sentence she has not checked. Next year this time, she says, might be kind of different.