Women’s medicine is borrowed medicine.

The reason women’s medicine runs on borrowed drugs is that, for most of the modern era of drug development, women were written out of the part where drugs get designed.

In 1977 the U.S. Food and Drug Administration issued guidance recommending that women of childbearing potential be excluded from early-phase drug trials, a precaution born of the thalidomide era that hardened into a default. For roughly the next two decades, the basic safety and dosing of new drugs was established largely in men, and the conclusions were then prescribed to everyone. The United States did not require the inclusion of women in federally funded clinical research until the NIH Revitalization Act of 1993. The same year, the FDA reversed its 1977 stance, which means the modern era of actually studying drugs in women is barely three decades old.

This was never only a fairness problem; it was a data problem with measurable consequences. When the Government Accountability Office reviewed the prescription drugs withdrawn from the U.S. market between 1997 and 2000, it found that eight of the ten posed greater health risks to women than to men. The mechanism is mundane and well documented: women often metabolize drugs differently. In 2013 the FDA cut the recommended dose of the sleep drug zolpidem in half for women after data showed they clear it from the body more slowly, leaving morning blood levels high enough to impair driving, two decades after the drug first reached the market.

Carry that forward and the picture for women’s hormonal and reproductive conditions is worse still, because those conditions were not merely under-dosed; they were under-developed. Few drugs were ever designed from the ground up for endometriosis, PCOS, PMDD, adenomyosis, or perimenopause. So the question for women’s health is rarely “which new molecule cures this?” It is “which drug that already exists, designed for something else, happens to help, and can we prove it?”

Women’s health is already the largest unstructured drug-repurposing experiment in medicine. Almost everything that works for these conditions was borrowed.

The drugs on the homepage are not Whel’s discoveries; they are the public, decades-deep record of what borrowing looks like in practice. Each began life for another organ, another disease, sometimes another sex, and arrived at women’s health late, off-label, or by accident. Read as a set, they tell a single story: women’s medicine has run on repurposing for decades, and we are the first to treat that as something worth building on.

Starting here also tracks where drug development itself is heading. Developing a brand-new drug typically costs well over a billion dollars, takes ten to seventeen years, and fails roughly nine times out of ten. A repurposed drug starts from an established human safety record, so it reaches patients faster, at a fraction of the cost, and is approved at far higher rates. Regulators built a lane for exactly this: the FDA’s 505(b)(2) pathway lets a sponsor lean on existing safety data for an approved ingredient and begin clinical work for a new indication much closer to the finish line. Each Whel candidate now carries a regulatory & development-status panel that maps where it already sits in exactly that landscape.

The market reflects the same logic. Drug repurposing was worth roughly 36 billion dollars in 2025 and is forecast to keep climbing through the next decade. Almost all of that effort, though, points back at the conditions the wider industry already prioritizes. The borrowing that built women’s health happened in the open, over decades, in off-label prescribing, in trial registries, in adverse-event databases, and in the communities where patients log what helps. The signal is sitting there. What has been missing is a system that reads it for female biology and proves it rigorously enough to act on. That is where we start.

We chose premenstrual dysphoric disorder as the first condition to build in depth because it is where the method is hardest to fake and easiest to prove.

PMDD is severe, common, and recent. It affects an estimated 3 to 8 percent of menstruating women and was only added as a formal diagnosis to the DSM-5 in 2013, after decades in the manual’s appendix. That lateness is the point: a condition recognized this recently has a thin, scattered evidence base, which is exactly the terrain where a system that reads evidence carefully earns its keep.

It is also the cleanest possible test of the substrate, because PMDD is cyclical by definition. Symptoms track the luteal phase and lift with menstruation, which means the right question is never just “does this drug help?” but “does it help at the right point in the cycle, at the right dose, for the right person?” A platform that cannot represent cyclical hormonal state cannot reason about PMDD at all. Building it here forces the substrate to handle the thing male-default graphs ignore.

And the signal is dense. PMDD has large, articulate patient communities that record, in fine detail and in real time, what they take and when it works, the off-label reality the formal literature has not yet caught up to. If the method works anywhere, it works here first, and what we learn building PMDD transfers to the conditions next in line.

Whel and the other AI-native drug-repurposing companies run on similar machinery. What sets us apart is the medical model that machinery is built to serve.

Most drug discovery is built to chase cures, the single molecule that can eliminate a disease, because for cancer and rare genetic disorders that cure is the blockbuster, and it is where the AI drug-discovery companies point their engines. Endometriosis, PCOS, PMDD, adenomyosis, perimenopause, and vulvodynia rarely work that way. They are lived with and managed over years, rarely fatal but routinely life-debilitating, reshaping careers, relationships, and decades of daily function without ever showing up on a mortality table. Managing a chronic condition needs a different evidence base than curing an acute one. The fuller case for why discovery is built around cures, and what that has cost women, is laid out in the manifesto.

The consequence is financial before it is clinical. Pharma recoups its development bills through patent-protected exclusivity, so the candidates worth a company’s attention are the ones it can own, and most of the drugs that already manage women’s conditions are cheap, off-patent generics. The repurposing literature has a name for them: financial orphans, clinically valuable and commercially unattractive, left unstudied because no one profits from confirming what they do. A platform scoring drug-disease pairs by their potential to cure, and by their potential to be owned, ranks a drug that manages endometriosis for millions of women as a near-miss. We rank it as the result.

All of this is why female biology deserves a knowledge graph entirely separate from the male-default graphs the rest of the field reasons over.

Every AI drug-discovery platform reasons over a knowledge graph, a structured map of how drugs, targets, pathways, and diseases connect. Those graphs were assembled from the same literature that under-studied women, so they inherit its priors: doses set in male tissue, mechanisms worked out without cyclical hormonal state, conditions that are thinly represented because they were thinly funded. Adding a “women’s health filter” on top of that substrate does not fix the substrate. The errors are underneath the filter.

So we are building the corrected version from the ground up, grounded in the same standard biomedical ontologies the field trusts (MONDO, EFO, RxNorm, ChEMBL) and then extended with the female-specific concepts no existing ontology captures adequately: sex-divergent pharmacokinetics, cyclical hormonal state, and the cross-condition mechanisms that only become visible once you stop treating the male body as the default.

This is meant as an additive layer that complements pharma rather than competing with it. The rest of the field is mapping the biology it was built to see; we are correcting and completing the half of it that was left out. A fuller account of the architecture, how each signal is graded, and the external resources we build on is on the technical architecture and external references pages.

A fair question from anyone who knows the field: could a general-purpose biomedical platform just query endometriosis in the graph it already has? On three counts, it would be looking in the wrong places.

The sources. The large biomedical AI platforms read the institutional record. Causaly, for one, ingests PubMed, MEDLINE, the trial registries, and patent filings, which is the right diet for most of biology. It is the wrong diet for conditions medicine left to off-label practice, because there the earliest and densest signal lives in patient communities no general platform reads.

The variables. A general knowledge graph holds a drug and a disease as a fixed relationship. Female pharmacology moves: drug metabolism and immune signaling shift across the menstrual cycle, and a condition like PMDD is defined by that timing. A platform that asks whether a drug affects a pathway, without asking how that changes across the cycle, passes over the candidates where the timing is the whole insight.

The candidates. General platforms serve pharma R&D teams whose budgets run on oncology, neuroscience, and immunology, and whose candidates have to be ownable. A cheap generic that manages a women’s condition is a financial orphan, clinically valuable and commercially uninteresting. Whel is built for a different audience, women’s health researchers, emerging biotech teams, advocacy organizations, and public funders, and our ranking follows their priorities.

Low-dose naltrexone for endometriosis sits at the intersection of all three. At low doses naltrexone appears to quiet the glial inflammation behind chronic pain, by a mechanism unrelated to the addiction treatment it was approved for, and women have documented its effects in endometriosis communities for years, often noting how the response tracks their cycle. The institutional evidence is thin and unresolved: the one randomized endometriosis trial was terminated with nine patients enrolled, and it is a generic no company can profit from confirming. A general platform ranks it near the bottom, or never reads the signal at all. We surface it, with its contradictions and uncertainty shown rather than smoothed away, because the signal and the mechanism are both real and no one else is reading them together.

The form of the bias. There is a deeper version of this, and it is the question a careful reader eventually asks: if women’s health was under-studied everywhere, is Whel not built on the same biased record as everyone else? It is. No biomedical source escaped that gap. What differs is the form the bias takes. In raw, granular sources, a single paper, or a typed drug-to-target relationship that carries its own provenance, the bias shows up as sparsity: fewer records, missing edges, gaps we can see, measure, flag, and fill with patient-community signal and our own sex-specific modeling. In a pre-trained predictive model, the same bias is collapsed into a confidence score, a hidden gap that emits an answer where there should be an honest silence.

That distinction is why we build where we do. We rely on the raw, traceable layer, and mostly on its mechanism: a drug binding a target, a gene linked to a disease, the part of biology least distorted by who was enrolled in a trial. That layer is not perfectly sex-neutral. The receptor binding is, but the signaling context it sits in is modulated by hormonal state, which is exactly the gap our cyclical-phase and sex-specific layer is built to fill. Every fact in that layer is grounded in a primary source, an FDA drug label or the curated sex-PK literature (Zucker and Prendergast 2020; Soldin and Mattison 2009), and cross-checked against it rather than asserted. Even the molecular reference data tilts male: roughly two-thirds of donors in GTEx, the standard tissue-expression atlas, are men, and sex shapes gene expression across nearly every tissue. The predictive graphs and models that have already digested the literature we keep beside the work as a cross-reference, never letting their averaged verdict become our ground truth. And where the evidence runs out, we show the gap rather than fill it: query Open Targets for vulvodynia or PMDD, two of our six conditions, and it returns nothing at all, a silence we surface rather than smooth over, because for these conditions a marked gap is worth more than a confident number resting on almost nothing.