This week I had the opportunity to give a talk at EURORDIS-Rare Diseases Europe about how AI can accelerate research and unlock new treatments for rare disease. I came to Barcelona to give a lecture, but I came away with a revised understanding of where medicine is today, and how rare disease research is transforming the future for everyone.
In a fundamental way, I see the personalized approach that rare patients demand - because there may only be a handful of them in the world - eventually become the standard of care for every disease.
But first, let me back up.
The reason I know so much about this is because my daughter Siena has an ultra-rare disease of which there are only a few thousand known cases world-wide. It is caused by a mutation in the CDKL5 gene, and while every patient with this disease gets the same label - CDKL5 Deficiency Disorder - the majority have different mutations, at different locations in the gene.
As far as we know, nobody else in the world has Siena’s exact genetic context.
Even children with the same exact mutation often show radically different clinical presentations. There are identical twins where one can walk and talk while the sibling is profoundly disabled.
In that sense, every child is unique.
Siena is the only person in the world with her disease.
In practice, that makes her financially ineligible for conventional medical research. A disease needs to have enough patients to justify the investment required to study it. There simply aren’t enough Sienas to make the conventional economics work. Common diseases have a larger number of clinical trials, published studies, treatment options, grant opportunities and commercial incentives for researchers and companies to explore approaches that may work for different subgroups. Your chances of finding an effective treatment if you have a common disease are therefore much higher - not because people with common diseases aren’t unique, of course they are - but because that heterogeneity can be hidden inside averages.
Common diseases give medicine the luxury of learning from populations. With enough data points, you can begin to identify groups of patients who are more or less likely to respond to a particular treatment. You can discover that a drug works particularly well in one subgroup, poorly in another, or causes a particular adverse effect in a third.
What we have today is an average-based medicine, not because everyone is treated the same, but because enough patients cluster around local averages to justify the investment. Our pharmaceutical system was largely built around the idea of the blockbuster: identify a disease affecting a large population, develop a treatment that works well enough across that population, and manufacture it at scale. It’s a functional financial model given the extreme costs involved in drug discovery, but it is not necessarily the model that best serves the individual patient.
In an ideal world everything would be tailor-made for you. Your doctor would know everything about you, your unique genetic and metabolic makeup, your epigenetics, historical diet, stress levels and environmental exposures. This unique fingerprint could then be matched against an enormous universe of potential interventions to identify the treatments most likely to work for you.
The reason we haven’t been able to deliver this is because it would require a team of specialists to read thousands of papers, integrate clinical and molecular information, formulate hypotheses, compare potential therapies and design experiments, over and over again for each patient. It would be economically absurd.
The average patient is a useful fiction, but for n=1 patients this concept breaks down.
Rare isn’t an exceptional category of human biology, it just doesn’t fit into a low-res system built around averages. This is one reason the treatment gap is so enormous: only around 5% of the roughly 10,000 known rare diseases have an FDA-approved treatment.
But something fundamental is changing. If increasingly capable AI systems can do much of the cognitive work for a fraction of that cost, the boundary moves. We can analyze individual mutations at the time of diagnosis, or predict the consequences of variants before a patient is even born. We can predict splicing behavior, protein structures and potential downstream effects. We can integrate transcriptomic and other molecular data. We can predict compounds that might interact with particular protein targets. A single promising candidate used to take months or years of dedicated teams to discover, now thousands can be predicted in seconds. That is a seismic shift in the economic dynamic.
The information bottleneck is starting to collapse. The wet lab, clinical trials, manufacturing all remain expensive, for now, but I predict we will see a boom of automated experimental validation companies soaking up some of the tsunami of in-silico biological predictions coming our way.
In rare disease research we frequently use the patients own biology to design treatments and predict their responses. These high-resolution datasets can further feed AI models, allowing us to identify patterns across patients, discover new response subgroups, and improve our ability to predict which treatments are most likely to work for a particular individual.
It’s a virtuous cycle:
patient → data → model → prediction → treatment → outcome → better model
What is now an exception, born of necessity for parents trying to save their children, may tomorrow become the norm. As AI makes it increasingly economically feasible to make treatment decisions at the level of the individual rather than the population average, patients will increasingly expect that level of care.
What begins as necessity in rare disease may become the standard of care for everyone.