Within two or three years, expect a quiet shift in how pharma kitchens decide what to put on the menu for rare diseases: instead of a chef guessing which ingredient might work, they will hand a genetics-trained model a phenotype and ask, "What should we even be reaching for?" The dish in question today is RareGPS, and it arrives plated with unusual confidence.

The Problem on the Plate

Rare diseases are the tasting menu of medicine - thousands of tiny, exquisite, underserved courses, most of which no one has bothered to develop a proper recipe for. Common diseases get the genome-wide association studies, the big sample sizes, the well-stocked pantry. Rare diseases get scraps. When your patient population fits in a single lecture hall, the usual statistical machinery starts to taste thin and watery.

So the question the chefs at Mount Sinai asked is deceptively simple: given that a gene is genetically tied to a rare condition, can we predict whether a drug aimed at that gene is actually worth cooking? Genetic support for drug targets is not a new idea - we have known since 2015 that drugs with human genetic evidence behind them roughly double their odds of surviving the brutal march through clinical trials (Nelson et al., Nature Genetics, 2015). RareGPS just takes that insight and reduces it into something far more concentrated.

The Recipe

Here is where the technique gets interesting. Most prioritization tools treat a genetic association like a single binary seasoning: significant or not, in or out. RareGPS instead tastes the full distribution of associations across allele-frequency bins - an "allelic-series" model. In plain kitchen terms, it does not just ask whether a gene is linked to a disease; it asks how that link behaves as you move from common variants to the rare, the ultra-rare, and the barely-there. The shape of that gradient turns out to carry real flavor.

The framework folds in 11 sources of evidence - genetic, clinical, and experimental - into a single machine-learning stock. That is a lot of ingredients in one pot, and the obvious risk is a muddled, over-reduced sauce where nothing stands out. The pleasant surprise is that the layering holds. Across 161 phenotypes, RareGPS outperformed existing resources at predicting both real drug indications and how far a candidate climbs through clinical trials.

The Finish

The numbers on the back of the menu are the part worth lingering on. Targets ranked in the top 1% were 58 times more likely to advance from "never indicated" all the way to phase IV than the unremarkable middle 50%. Even starting from phase I, the top targets reached phase IV at 8 times the rate. For a field where most candidates die quietly somewhere between the appetizer and the entree, that is a striking separation of the good plates from the forgettable ones.

And the team did not just trust their own palate. They validated the predictions against the prescriptome of two million patients - essentially checking whether real doctors, off-label and on, had independently been reaching for the same ingredients. They cross-checked against AMELIE, an independent literature-mining tool, to make sure the model was not just confidently making things up like your uncle at Thanksgiving. Then, generously, they published predictions for 3,021,965 gene-phenotype pairs, which is roughly three million little suggestions for what to try next.

A Fair Critique

No dish is flawless. A model this rich inherits the biases of its pantry - genetic association data is still uneven across ancestries and conditions, and "off-label prescribing matches our prediction" is suggestive validation, not proof a drug works. Electronic health records are a famously noisy ingredient, full of coding quirks and missing courses. RareGPS sharpens the odds; it does not replace the trial. The authors, to their credit, plate it as a prioritization aid rather than a verdict.

Still, for an area of medicine that usually eats last, this is a confident, well-structured offering. The base is solid genetics, the seasoning is thoughtful, and the finish - those advancement odds - lingers in exactly the right way. Send it back to the kitchen for more validation, sure. But order it again.

Reference

Chen R, Duffy Á, Mort M, Cooper DN, Rocheleau G, Jordan DM, Do R. Genetically supported drug target prioritization for rare diseases. Genome Medicine (2026). DOI: 10.1186/s13073-026-01671-5. PMID: 42277932.

Supporting context: Nelson MR, et al. The support of human genetic evidence for approved drug indications. Nature Genetics (2015). DOI: 10.1038/ng.3314.

Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.