- Six conditions in scope: endometriosis, PMDD, PCOS, adenomyosis, vulvodynia, and menopause
- Six evidence pipelines running across three evidence arms
- 111 signals, each scored against a published five-dimension rubric
- An independent biological-plausibility cross-reference from Every Cure's MATRIX model, shown beside our own grades rather than blended into them
- An external-validation grade on every drug and condition pair, with the highest grade reserved for signals backed by named clinical-guideline strength and certainty
- Every drug and condition resolved to canonical biomedical registries (ChEMBL, MONDO, EFO), with ambiguous cases held for human review
- A knowledge-graph cross-check, computed over Open Targets, that shows beside each signal whether the graph supports it or stays silent
- A sex-specific pharmacokinetics layer, seeded from FDA labels and the curated sex-PK literature, shown beside the relevant signals
- A cyclical-phase layer that records where a treatment's effect depends on the menstrual-cycle phase, seeded for the strongest-evidence PMDD cases and shown beside the relevant signals
- Run the two-rater validation study and publish the agreement score
- Add disproportionality statistics to the adverse-event arm so a real safety signal separates from reporting noise
- Flag where two or more arms support the same drug and condition pair
- Feed the knowledge graph into scoring at prompt time, beyond the beside-signal disclosure already live
- Ground every summary sentence against its source, extending the citation validation already running
- Publish an open, citable data export
- Extend to further under-researched women's health conditions
- Add complementary pipelines once the existing ones are solid
- Deepen coverage of the conditions already in scope
- Extend the substrate to hold sex-specific pharmacokinetics and cyclical hormonal phase as first-class variables, so a relationship carries the body and the cycle phase it holds in rather than a single averaged value
- Add an independent validation layer of open knowledge graphs and models, such as DRKG, PrimeKG, and TxGNN, shown beside each signal as an outside cross-reference and kept separate from the core architecture
- Once the validation work above is in place, partner with formal women's health advocacy organizations, such as IAPMD, PCOS Challenge, and Endometriosis UK, to bring structured patient-reported signal beyond Reddit into the evidence base
What Whel is, and who it is for
Whel is a research instrument: a structured, scored, searchable evidence base for the drug-repurposing signals that already exist across under-researched women's health conditions. We build the thing the field currently lacks, which is a single place where that evidence is gathered, graded, and made citable.
The signal is scattered today across published literature, clinical-trial registries, adverse-event databases, genetic-target platforms, and patient communities. We bring it together, grade each piece against a published rubric, and attach its source so it can be checked. The result is closer to a small evidence lab than a website, and it serves the research community in several different ways.
Researchers
Scored, sourced repurposing hypotheses worth taking into a formal study.
Clinician-researchers
The state of the evidence for a condition: every signal and source, at a glance.
Graduate students
An open, fundable research question in a field with unclaimed ground.
Journalists & advocates
Traceable, scored evidence to ground reporting and advocacy in more than anecdote.
Institutions & funders
Where the evidence is thinnest, and where new attention would go furthest.
Why these six conditions
We chose these six conditions deliberately, against three explicit criteria, and those same criteria determine how the database grows. They are a starting point in a much larger field.
Shared biology
The six conditions converge on the same handful of systems: estrogen signaling, chronic inflammation, metabolic regulation, and pain processing. That biological overlap is what makes cross-condition reasoning valid, because a signal in one condition can be informative about the others.
Documented neglect
Each condition carries a measurable evidence gap: long diagnostic delays, few treatments that address the underlying disease rather than the symptoms, and thin research funding. Whel is most useful exactly where the published literature is thinnest.
A focus on women
Whel exists to address the structural under-study of women's hormonal and reproductive health, a field that, until the NIH Revitalization Act of 1993, did not even require women in clinical research. That focus is deliberate and permanent, and every condition we add will be a women's health condition.
A thread runs through all three criteria: these six conditions are rarely fatal but routinely life-debilitating, lived with and managed over years rather than cured. That is exactly the kind of problem the cure-focused blockbuster model overlooks, and where repurposing the drugs that already help matters most. More on building for management →
The six conditions, mapped
Each condition is tagged with the biological systems it shares with the others, the overlap that makes cross-condition reasoning valid.
| Condition | Shared pathways | The gap |
|---|---|---|
| Endometriosis | EstrogenInflammationPain | Affects up to 10% of women of reproductive age, with a 7 to 10 year average diagnostic delay and no disease-modifying drug. |
| PMDD | EstrogenMoodPain | Clinically severe and cyclical, still treated mainly with imprecisely prescribed SSRIs. |
| PCOS | MetabolicEstrogenInflammation | One of the most common endocrine disorders in women of reproductive age, and chronically under-represented in research. |
| Adenomyosis | EstrogenInflammationPain | Long under-recognized, and historically confirmable only after hysterectomy. |
| Vulvodynia | PainInflammation | A chronic pain condition, and among the least-studied of the six. |
| Menopause | EstrogenMetabolic | A transition every woman who lives long enough experiences, and widely acknowledged to be poorly managed. |
The database is built to grow beyond its first six conditions.
Because the selection rule is explicit and repeatable, the set expands as our capacity grows. The conditions below meet the same criteria and are under consideration for future versions.
Where the framework goes next
Because the selection rule is explicit, extending it is straightforward. Any women's health condition that shares biology with the existing six, carries a documented research gap, and has enough of an evidence base to surface signals is a candidate. The conditions below illustrate where the framework points, and the final list remains a research and editorial decision. The scope itself stays fixed within women's hormonal and reproductive health.
Interstitial cystitis / bladder pain syndrome
A chronic pelvic pain condition that frequently co-occurs with endometriosis, predominantly affects women, and carries long diagnostic delays.
PainInflammationUterine fibroids
Extremely common and estrogen-driven, yet undertreated relative to prevalence, and sharing hormonal biology directly with adenomyosis and endometriosis.
EstrogenInflammationPrimary ovarian insufficiency
Hormonal and metabolic, extending the existing menopause arm to women who reach that transition far earlier than expected.
EstrogenMetabolicPerinatal mood conditions
A hormonally driven transition with severe consequences and a thin, only recently growing treatment literature.
EstrogenMoodLipedema
A metabolic and inflammatory condition that affects women almost exclusively and is routinely misdiagnosed, and among the most neglected in the field.
MetabolicInflammationIllustrative only; these conditions are not yet in the database.
The engine, and the sources it is built on
This is Whel's technical architecture: the original sources each condition and signal is built from, and the engine work that makes the evidence behind each signal more rigorous. The most valuable near-term work is to strengthen that engine before widening the data it draws on, because a new condition is worth little if the reasoning underneath it is not as solid as it can be. The priorities come directly from our own methods document and from the project's first independent review. The independent layers Whel is checked against, rather than built from, are kept separate in the validation layer below.
The sources Whel is built on
| Source | Role | Status |
|---|---|---|
| PubMed | Published literature | ● Live |
| ClinicalTrials.gov | Trial registry; also the trial-stage read in the regulatory & development-status panel | ● Live |
| FDA openFDA | Adverse-event data | ● Live |
| DailyMed | FDA drug labels; the on-label / off-label approved-indication read in the regulatory & development-status panel | ● Live |
| FDA Orange Book | Approved Drug Products with Therapeutic Equivalence Evaluations; the generic-availability and patent-supply read in the regulatory & development-status panel | ● Live |
| Open Targets | Genetic-target and pathway data | ● Live |
| Reddit communities | Patient-reported signal | ● Live |
| Patient-advocacy organizations | Structured patient-reported signal beyond Reddit, through planned partnerships with formal women's health advocacy groups, taken on once the validation work is in place. | ● Planned |
| EudraVigilance | European adverse-event data | ● Under review |
| SIDER | Drug side-effect reference | ● Under review |
| DrugBank | Drug-target and indication data | ● Planned |
Method upgrades in progress
| Source | Role | Status |
|---|---|---|
| Two-rater validation study | A stratified sample of signals re-scored blind by two independent raters, with agreement reported as a measured score. | ● Planned |
| Disproportionality statistics | Separating a real adverse-event signal from background reporting noise, using data we already hold. | ● Planned |
| Ontology-grounded entity resolution | Every extracted drug and condition resolved against canonical registries (ChEMBL, RxNorm, MONDO, EFO), with anything that fails to resolve held for human review. Applied across the corpus; the per-pipeline audit numbers are the remaining piece to surface. | ● Live |
| Knowledge-graph grounding | A domain-restricted graph of drug, target, and condition relationships built over Open Targets, surfacing a 'graph supports' or 'graph silent' layer beside each signal in the gated view. Feeding the graph into scoring at prompt time, and a property-graph version, are follow-ons; the property-graph would move to BioCypher or Neo4j for richer graph tooling, with Apache AGE as a lighter Postgres-native fallback. | ● Live |
| Phase-aware relationships | Holding cyclical hormonal state as a first-class variable, so a drug and condition relationship can carry the menstrual-cycle phase in which it holds rather than being averaged into a single static edge. Seeded for the strongest-evidence PMDD cases (luteal-phase SSRI dosing; drospirenone cycle suppression) from ACOG guidance, FDA labels, and a placebo-controlled RCT, and shown beside the relevant signals. Validation basis is the DRSP and the ISPMD consensus. Broader population is ongoing. | ● Live |
| Sex-stratified pharmacokinetics | Per-compound pharmacokinetic structure held by sex, so documented differences in metabolism and clearance inform scoring rather than being assumed uniform across bodies. Seeded for an initial set of compounds from FDA labels and the curated sex-PK literature, each carrying its source, and shown beside the relevant signals. Broader population is ongoing. | ● Live |
| Regulatory & development status | Where each candidate sits in the US regulatory landscape, read from three public FDA / NLM sources and shown beside the score: whether the target condition is an FDA-approved (on-label) use or off-label (DailyMed labels, counting only NDA/ANDA/BLA approvals); whether the molecule is a generic or a single-source brand still under patent (FDA Orange Book, single-ingredient products only); and how far it has been studied as a therapy for the condition (ClinicalTrials.gov, mechanistic and post-marketing studies excluded). Descriptive landscape context only, not a viability assessment; live across all six conditions. | ● Live |
| Actionability layer | A second axis beside the evidence score that weighs management endpoints, 505(b)(2) regulatory viability, and patient-community signal, so a livable-management candidate is ranked for what it is rather than against a cure-focused default. It builds on the descriptive regulatory & development-status layer above, which maps the landscape but stops short of judging whether a development path is viable. | ● Planned |
| Citation validation | Verifying every generated citation against its registry, so a fabricated or mis-attributed reference is caught before publication. | ● Live |
| Summary grounding | Checking every summary sentence against its source text and flagging any that drifts beyond what the source supports. The verifier is built; it is waiting on the per-source finding excerpts it reads from being populated, then a first run. A biomedical embedding model such as PubMedBERT-NLI is under evaluation to sharpen recall over the current general-purpose embeddings. | ● Planned |
| Per-claim synthesis and contradiction marking | Tagging individual claims where the model combined findings (a synthesis) or where the underlying sources disagree (a contradiction), and surfacing those markers beside each signal. The marking is built into the signal view; populating it across the corpus, so the markers appear wherever they apply, is the remaining work. | ● Planned |
| Cross-arm concordance flag | Marking where two or more arms support the same drug and condition pair. | ● Planned |
| Open data export | A citable CSV and JSON export under an open license, deposited with a DOI. | ● Planned |
Disproportionality statistics
Pharmacovigilance has a standard way of separating a real adverse-event signal from background reporting noise, the proportional reporting ratio and the reporting odds ratio. Adding that calculation to the adverse-event arm, using data we already hold, is the single most valuable near-term upgrade to the evidence.
The validation study
Our central claim is that we can grade evidence, and the methods document already designs the test: a stratified sample of signals, re-scored blind by two independent raters, with agreement reported as a measured score. Running it, and publishing the result whatever it turns out to be, converts every confidence tier from model output into a measured claim.
Open data
A CSV and JSON export under an open license, deposited with a DOI, makes Whel citable, and a tool that other researchers can cite enters the research record rather than staying a website they happen to read.
What Whel is checked against
Separate from the sources Whel is built on is the layer it is checked against: independent references shown beside each signal rather than blended into its grade. Some are live today. Others are open knowledge graphs and models that lead the biomedical drug-repurposing field, and are planned as outside cross-references. The fuller account of how each external layer is disclosed lives on the external references page.
| Source | Role | Status |
|---|---|---|
| Every Cure MATRIX | An independent treatment-probability cross-reference from Every Cure's graph-ML model, shown beside our grades rather than blended into them. | ● Live |
| Clinical-guideline curation | Strength and certainty drawn from named society guidelines, normalized into the highest external-validation grade where a named recommendation covers a pair. | ● Live |
| DRKG (Drug Repurposing Knowledge Graph) | An open, multi-source repurposing knowledge graph. Planned as an outside cross-reference shown beside a signal, not folded into Whel's own graph, since it carries the field's male-default coverage that Whel exists to correct. | ● Planned |
| PrimeKG (Precision Medicine Knowledge Graph) | An open precision-medicine graph spanning drugs, diseases, phenotypes, and pathways. Planned as a second independent cross-reference to widen the 'graph supports or graph silent' disclosure beyond a single source. | ● Planned |
| TxGNN (graph foundation model) | An open, zero-shot drug-repurposing model. Planned as a benchmark and hypothesis cross-reference whose predictions Whel would validate rather than trust outright, because the model inherits the same male-default training data. | ● Planned |
The planned graphs and models are kept outside Whel’s core architecture on purpose. Resources like DRKG, PrimeKG, and TxGNN are built on the same general biomedical record that under-covers women’s hormonal and reproductive health, so folding them into the engine would import the exact blind spot Whel exists to correct. They are most useful as an outside check: a place where the graph either agrees with a signal, stays silent, or disagrees, shown plainly beside Whel’s own grade. Where these layers are silent on a women’s health condition, that silence is itself a finding worth surfacing.
How they are read, once added, is asymmetric by design. Because these models draw on the same literature Whel does, their agreement is weak evidence rather than independent confirmation. The informative case is disagreement: a model confident where Whel’s own graph is silent is a flag that the model may be filling a gap, and Whel’s graph supporting a pair the models miss is a candidate worth a closer look. Agreement nudges confidence a little; disagreement starts an investigation.
How each external layer is disclosed →A living page
This roadmap is a dated document, and it will change. We release Whel as dated snapshots rather than a live feed, and the priorities here are shaped by feedback from researchers, clinicians, and the patient communities whose reported experience the database draws on, about which gaps matter most. If you work in one of these fields, that feedback is welcome.