ai gen

Estradiol pharmacokinetics — fact-check + recalibrated model

A multi-species kinetic model of oral, transdermal, sublingual, IM, sub-Q and COC estrogen administration, with E2 / E1 / E1S / EE / SHBG tracked as state variables. Built on top of (and corrects) an earlier AI chat-bot analysis of the same topic.

This site is entirely AI-generated and has been primarily AI-fact-checked (via two independent agents, Claude and GPT/codex). It has not been personally verified. Use it for thinking and intuition — not for medical decisions. Numerical predictions are good to roughly a factor of 2; many parameters are uncertain or inter-individual.

Quick links

NEW Three-document reference (Wikipedia-style)

synthesis.html tight reference with inline C1–C5 confidence tags on every claim
evidence.html primary-data figures + 64 direct source quotes
reference.html glossary, abbreviations, parameter definitions, 13 chemical structures
synthesis-narrative.html the long narrative version (preserved, less rigorously source-tagged)

The synthesis cross-links to evidence and reference for verification. Every load-bearing claim has a confidence tier and a primary-source link. Best entry point for new readers: synthesis.html.

NEW Interactive viewer

Open interactive time-animation viewer →

Pick a scenario, scrub through time, watch concentrations and metabolic fluxes update live. ECG-like display, model output annotated with ratio-vs-literature-anchor in real time.

NEW Sankey flux diagrams

One per scenario. Open the v4 figures folder and click any sankey_*.html.

Steady-state mass flow (µg/day) from sources through plasma species to metabolic sinks. Hover any flow to see the exact flux value.

FACT-CHECK What the old AI got wrong

Headline fact-check summary →

The previous AI's notes had ~8 substantial errors (SHBG Kd, ESTHER OR, PMP E1 production rate, sublingual Tmax, …). Two independent agents reconciled the corrections; they agreed on every major call.

v4 model writeup

v4 extensions (EE + SHBG + new routes) →
v3 architecture & parameters →
Diff from old AI v3 →

Research notes

Claude fact-check (5000 words)
GPT/codex fact-check
Published PK/PBPK models
EE parameters

Source code

v1: simplest 3-species
v2: dynamics + sublingual
v3: fact-checked v3
v4: EE + new routes

v4 steady-state validation

11 of 12 scenarios within 2× of literature anchors. The one failure (sublingual E1S 2.18×) is against a literature anchor that is itself genuinely uncertain.

Validation grid: model vs anchors for E2, E1, E1S, EE, SHBG
Scenario E2 model / anchor E1 m/a E1S m/a EE m/a SHBG m/a (nM)
cycling_follicular72/50 ×1.4458/50 ×1.16819/960 ×0.8550/50 ×1.00
oral_1mg48/35 ×1.37218/250 ×0.874096/2560 ×1.60105/80 ×1.31
oral_2mg94/70 ×1.34414/500 ×0.838013/5000 ×1.60129/110 ×1.17
oral_4mg186/130 ×1.43807/900 ×0.9015825/9000 ×1.76150/140 ×1.07
transdermal_50ug36/50 ×0.7229/50 ×0.58409/600 ×0.6850/50 ×1.00
transdermal_100ug70/100 ×0.7037/70 ×0.53644/900 ×0.7250/55 ×0.91
sublingual_1mg_BID139/100 ×1.39333/200 ×1.666537/3000 ×2.18106/80 ×1.32
IM_E2V_5mg_q5d393/200 ×1.97113/200 ×0.562841/1500 ×1.8951/60 ×0.85
IM_EC_5mg_q14d79/100 ×0.7940/70 ×0.57758/900 ×0.8450/55 ×0.91
SC_EUn_25mg_monthly193/200 ×0.9766/90 ×0.741519/1200 ×1.2750/60 ×0.84
COC_EE30_PO8.2/10 ×0.8217/30 ×0.58178/250 ×0.7134/40 ×0.84179/170 ×1.05
pregnancy_term20672/20000 ×1.038133/7000 ×1.1653658/100000 ×0.54211/350 ×0.60

Key model insights

Route effect on hepatic load

EE vs Oral E2 — same plasma E2 but very different SHBG induction

The right panel above is the punchline of the COC story: a 30 µg/day EE dose gives plasma estrogen (EE) similar to what oral 2 mg E2 gives, but SHBG rises ~50% more than with oral E2. Per molar at the hepatic level, EE is ~100× more potent than E2 at SHBG induction, because the 17α-ethinyl group blocks the 17β-HSD2 inactivation step that "uses up" oral E2 on first pass.

Oral dose-response

Oral E2 dose response: 0.5, 1, 2, 4 mg

Model predictions vs Kuhl 2005 / Lindberg 2005 anchors. Linear in E2/E1/E1S across the 1–4 mg range; SHBG response is sub-linear (Hill saturation).

What the model does not capture

What changed from the old AI's notes

The previous chat-bot analysis (saved in old-ai-analysis/conversation.md and the HTML artifacts) had the right qualitative picture but ~8 substantial errors that propagated through its v3 kinetic model. Two independent fact-checking agents (Claude and GPT/codex via the codex CLI in a tmux session) caught the same major issues. Full list →

Most important:

Files

Generated by Claude Opus 4.7 in a /workspace/trans/estradiol session, 2026-05-14. AI-generated, AI-fact-checked, not personally verified.