What if we modeled the mechanism before prescribing the answer?

PMDD is the example that made this question impossible for me to ignore. We have research describing hormones, metabolites, receptors, brain networks, symptoms, and possible treatments. Most of that knowledge still lives in separate papers and static diagrams. I want to know what becomes possible when we turn it into a model that people can inspect, challenge, and test.

I do not mean a machine that tells someone what medication or supplement to take. I mean a virtual lab that makes every assumption visible. It could show which biological relationships are supported by human experiments, which come mostly from animal models, which are correlations, and which are still educated guesses. Then it could help researchers ask a better next question.

This matters far beyond PMDD. The same approach could help with conditions where the symptoms are real, the pathways are complex, and the evidence is fragmented across different kinds of studies. The goal is not to replace laboratory science or clinical trials. The goal is to make the path toward the next useful experiment more legible.

What PMDD can teach us

The simplest explanation of PMDD is often the wrong one. The evidence does not suggest that most people with PMDD simply have too much or too little progesterone. A leading model is that some brains respond differently to otherwise normal changes in ovarian hormones across the menstrual cycle. The American College of Obstetricians and Gynecologists describes the biology as unclear and probably multifactorial, with heightened sensitivity to normal estrogen and progesterone fluctuations and possible changes in serotonin and GABA signaling.¹

After ovulation, progesterone can be metabolized into allopregnanolone, often shortened to ALLO. ALLO is not GABA. It is a neuroactive steroid that changes the way GABA A receptors respond. GABA is the main inhibitory neurotransmitter in the brain, so changes in that system may affect stress sensitivity, anxiety, irritability, and emotional regulation.

The important word is may. Research on ALLO and GABA A receptor sensitivity is compelling, but it does not give us one proven receptor defect that explains every case. Some detailed ideas about receptor subunits and adaptation come from animal work, small human studies, or indirect measurements. A broad review in 2023 concluded that abnormal central nervous system responses to neuroactive steroid changes are likely important, while also stating that the findings do not yet form a complete biological explanation.^{2, 4}

Estradiol also matters. A 2025 hormone addback study expanded earlier work showing that symptoms can remit during ovarian suppression and recur when estradiol or progesterone is reintroduced in susceptible participants.³ That makes PMDD a powerful example of a dynamic system. The same hormone level can mean different things depending on timing, rate of change, receptor state, stress, genetics, and the rest of the network.

A useful treatment is not the same as a complete explanation

Selective serotonin reuptake inhibitors are among the best supported treatments for affective premenstrual symptoms. ACOG gives them a strong recommendation based on moderate quality evidence.¹ A 2024 Cochrane review of 34 randomized trials and 4,563 participants found that SSRIs probably reduce symptoms. Both continuous and luteal phase dosing appeared effective, although continuous dosing was probably more effective on average. The review also found common adverse effects and noted that 68 percent of the included studies were funded by pharmaceutical companies.⁵

That combination deserves a thoughtful response. We should take the evidence of benefit seriously. We should also take side effects, funding patterns, and unanswered mechanism questions seriously. It would be a mistake to dismiss a treatment simply because it does not map neatly onto one theory of the disease. It would also be a mistake to treat clinical usefulness as proof that we understand why it works.

The unusually fast SSRI response reported in PMDD is one clue. In a small study of 12 treated participants, symptom scores improved significantly by the second day of luteal phase fluoxetine.⁶ That is interesting because it differs from the slower response often associated with depression. It supports more research into rapid network and neurosteroid effects. It does not prove that a single ALLO pathway explains the response.

This is the gap I want the virtual lab to make visible. A therapy can have solid evidence for benefit while its full mechanism remains uncertain. A model should hold both facts at the same time.

Saffron is a good example of an evidence gap

Saffron is interesting because the signal is promising and the evidence is still thin. One randomized PMDD trial assigned 120 participants to saffron, fluoxetine, or placebo for the luteal phase of two cycles. Saffron improved the Daily Record of Severity of Problems score compared with placebo, while another depression scale did not show a significant difference among the three groups.⁷

A 2026 review found a positive pooled effect for saffron on broader premenstrual syndrome symptoms, but that evidence also included studies of PMS and dysmenorrhea rather than a large body of rigorously confirmed PMDD trials.⁸ Promising is not the same as established. A few studies are enough to justify a better question, not a treatment claim.

A responsible repurposing system could map saffron compounds to proposed molecular targets, compare those targets with the PMDD pathway graph, show where human evidence exists, and identify missing safety or interaction data. It could help prioritize a laboratory experiment or a properly designed trial. It should never convert a pathway match into personal medical advice.

What the virtual lab could actually do

The technical foundation already has a name: quantitative systems pharmacology. These models connect drugs, biological networks, disease processes, and measurable outcomes. Machine learning can help infer patterns from data, while mechanistic models encode what researchers believe causes what. The two approaches can strengthen each other when the difference between knowledge and data stays explicit.⁹

I would start with a narrow model, not a digital replica of the whole brain. Calling something a digital twin before it has personal data, continuous updates, and validation creates more confidence than the model deserves. A useful first version could be a literature backed causal map with a small set of equations or state transitions. Every connection would carry its source, evidence type, population, and uncertainty.

The FDA now has final guidance for model informed drug development and joint principles with the European Medicines Agency for AI used across the drug life cycle. Both emphasize a clear context of use, data provenance, multidisciplinary expertise, risk based validation, and documented limits.^{10, 11} Those are not details to add later. They are part of the product.

A hackathon built around uncertainty

I want the event to bring patients, clinicians, neuroscientists, systems biologists, pharmacologists, data scientists, and software builders into the same room. The focus would be gaps in health research and women's quality of life, with challenges selected together rather than centered on a single condition.

Teams could choose one of four challenges:

• Pathway mapping. Build a traceable map of a condition where every edge links to a source and carries an evidence label.
• Mechanism simulation. Build a small model that compares competing biological explanations and makes parameter uncertainty visible.
• Drug and supplement repurposing. Generate testable candidates from target overlap, then grade the supporting evidence, safety concerns, and missing data.
• Study design. Identify the next experiment that would be most informative and explain why the current data cannot answer it.

Judging should reward scientific honesty as much as technical ambition. A team that clearly shows why its model cannot support a conclusion may be doing better science than a team with a beautiful prediction and no credible validation.

The safest first event would use public, appropriately licensed, synthetic, or properly deidentified data. Any future work with personal health information would require a different level of security, governance, consent, and institutional oversight.

What this project must never pretend to be

Biology is not software. A pathway diagram is not the body. A language model can extract a claim incorrectly. A machine learning model can reproduce bias in its data. A mechanistic model can look persuasive while hiding an assumption that determines the result.

Digital twin research repeatedly runs into the same problems: incomplete biology, limited data, weak parameter estimates, uncertain translation across scale, and the need for real world validation.¹² The answer is not to avoid modeling. The answer is to build a system that shows where the model stops.

I do not want to build a machine that tells someone what to take. I want to build a place where researchers, engineers, clinicians, and patients can see exactly what is known, what is assumed, and which experiment would reduce uncertainty most.

That feels like a useful place to begin.