Internal Research Brief

Pharmacokinetic Research Brief

Individualized Caffeine Metabolism Modeling for CafOS

Prepared for clinical review — March 2026 • v1.0 • CafOS Research

Section 01

The Problem

The standard advice — "stop drinking coffee by 2 PM" — is a population-level heuristic that ignores 3–12× individual variation in caffeine metabolism. A fast metabolizer clears caffeine in 3 hours; a slow metabolizer on oral contraceptives may require 12+ hours. One-size-fits-all timing advice is pharmacologically indefensible.

Caffeine (1,3,7-trimethylxanthine) is eliminated almost entirely by hepatic metabolism. The CYP1A2 enzyme handles ~95–97% of primary biotransformation via N-demethylation. Less than 2% of ingested caffeine is excreted unchanged in urine — kidney function is irrelevant to caffeine clearance.

FIG. 1 — Primary caffeine metabolism pathway via hepatic CYP1A2

Section 02

The Pharmacokinetic Model

CafOS uses a one-compartment oral absorption model with first-order kinetics, replacing the commonly used instant-absorption simplification.

C(t) = (F × D × k_a) / (V_d × (k_a − k_e)) × (e^−k_e·t − e^−k_a·t)

Parameter	Symbol	Population Mean	Range	Unit
Bioavailability	F	0.99	0.95 – 1.00	—
Dose	D	varies	28 – 400	mg
Absorption rate	k_a	2.0	0.5 – 4.5	h⁻¹
Volume of distribution	V_d	0.6	0.4 – 0.8	L/kg
Elimination rate	k_e	0.139	0.058 – 0.231	h⁻¹
Half-life	t_½	5.0	3.0 – 12.0	hours

FIG. 2 — 95mg caffeine dose: full PK model (solid) vs simplified instant-absorption model (dashed)

The full model captures the absorption phase (T_max ~45 min) that the simplified model misses. This matters for calculating peak plasma concentration and the timing window before caffeine levels begin to decline.

Section 03

Individual Variation Factors

CYP1A2 activity — and therefore caffeine half-life — varies by up to 8× between individuals due to genetics, hormones, drugs, and lifestyle. The following modifiers shift the population-mean t_½ of 5.0 hours:

Factor	Modifier on t_½	Effective t_½	Mechanism
CYP1A2 fast metabolizer	×0.70	3.5h	1A/1A homozygous
CYP1A2 normal	1.00	5.0h	Population mean
CYP1A2 slow metabolizer	×1.80	9.0h	1F or 1C carriers
Smoker (10+ cigs/day)	×0.65	3.3h	CYP1A2 induction via AhR
Combined oral contraceptive	×1.80	9.0h	Estrogen inhibits CYP1A2
Pregnancy (3rd trimester)	×3.00	15.0h	Hormonal CYP1A2 suppression
Age 60+	×1.20	6.0h	Reduced hepatic capacity
Fluvoxamine co-admin	×5.00+	25h+	Potent CYP1A2 inhibitor

FIG. 3 — Same 200mg dose: fast metabolizer (teal, t_½=3h) vs slow metabolizer (amber, t_½=9h)

Key insight: A slow metabolizer drinking Cold Brew (200mg) at 2 PM will still have 30mg of active caffeine at midnight — enough to delay sleep onset by 20+ minutes. The same dose in a fast metabolizer clears to <5mg by 8 PM.

Section 04

CYP1A2 Triage Questionnaire

Since >99% of users will never have CYP1A2 genotyping results, CafOS infers metabolizer status through a 6-question onboarding triage. Each question is designed to be factual and observable — not a guess about sleep causation.

Design principle: We do NOT ask "How does coffee affect your sleep?" — that question is circular. If users knew the answer, they wouldn't need the app. Every question below asks something the user definitely knows about themselves.

Q1: Smoking Status — strongest environmental modifier

"Do you smoke cigarettes or use nicotine regularly?"

Response	Modifier (M₁)	Mechanism
No / quit >3 months ago	1.00	Baseline
Quit within last 3 months	0.85	CYP1A2 de-induction in progress
Yes, <10 cigarettes/day	0.75	Moderate AhR-mediated induction
Yes, 10+/day	0.65	Strong CYP1A2 induction via PAHs

Polycyclic aromatic hydrocarbons in cigarette smoke activate the aryl hydrocarbon receptor (AhR), inducing CYP1A2 transcription. This is the most replicated finding in caffeine pharmacokinetics — heavy smokers clear caffeine 30–50% faster.

Q2: Hormonal Contraceptives / HRT (women only)

"Are you currently taking hormonal birth control or hormone replacement therapy?"

Response	Modifier (M₂)	Mechanism
No	1.00	Baseline
Combined pill/patch/ring (estrogen + progestin)	1.80	Ethinylestradiol inhibits CYP1A2
Progestin-only pill/IUD/implant	1.10	Minimal CYP1A2 effect
Hormone replacement therapy	1.50	Conjugated estrogens, moderate inhibition
Not sure which type	1.40	Conservative estimate

Q3: Pregnancy (women only)

"Are you currently pregnant?"

Response	Modifier (M₃)	Effective t_½
No	1.00	5.0h
1st trimester (weeks 1–13)	1.20	6.0h
2nd trimester (weeks 14–27)	2.00	10.0h
3rd trimester (weeks 28+)	3.00	15.0h

Progressive CYP1A2 suppression from rising estrogen and progesterone. Triggers automatic 200mg/day cap per ACOG guidelines.

Q4: Perceived Effect Duration — key phenotypic proxy

"After drinking a cup of coffee, how long do you typically feel the alertness/energy boost?"

Response	Modifier (M₄)	Implied phenotype
Less than 2 hours	0.70	Fast metabolizer
About 2–3 hours	0.85	Fast-normal
About 3–5 hours	1.00	Population average
About 5–7 hours	1.25	Slow-normal
More than 7 hours / jittery for a long time	1.50	Slow metabolizer

Why this works and "how does it affect your sleep?" doesn't: This question asks about waking alertness duration — something people actually observe during the day. "I stopped feeling my morning coffee by lunch" is an observation. "Coffee at 3PM affects my sleep" is a guess about causation across many confounding variables (stress, screens, kids, alcohol). The perceived pharmacodynamic duration correlates with time from C_max to sub-threshold plasma concentration without requiring users to understand sleep physiology.

Q5: Daily Caffeine Intake — calibrates Q4

"How many caffeinated drinks do you have on a typical day?"

Response	Modifier (M₅)	Role
0–1 per day	1.05	Low tolerance — Q4 answers highly credible
2–3 per day	1.00	Normal
4–5 per day	0.95	Minor CYP1A2 autoinduction
6+ per day	0.90	Heavy use — tolerance may mask Q4 perception

Q6: Age

Response	Modifier (M₆)
18–29	0.95
30–44	1.00
45–59	1.10
60+	1.20

Scoring Algorithm

t_½ = 5.0 × M₁ × M₂ × M₃ × M₄ × M₅ × M₆
Clamped to [2.5, 15.0] hours.

Cross-Validation Logic

Tolerance masking correction: If Q5 = "6+ cups/day" AND Q4 = "less than 2 hours" → shift M₄ from 0.70 to 0.85. Heavy consumers may not perceive the effect due to adenosine receptor tolerance, not because they metabolize faster. Conversely, if Q5 = "0–1 cups" AND Q4 = ">7 hours" → use full M₄ = 1.50 (highly credible slow metabolizer signal).

Worked Examples

Example A: Average Male Professional, 35

Non-smoker (×1.00) • N/A hormones • Not pregnant • Effect: 3–5h (×1.00) • 2–3/day (×1.00) • Age 35 (×1.00)
t_½ = 5.0 × 1.00 × 1.00 × 1.00 × 1.00 × 1.00 × 1.00

= 5.0 hours

Example B: Woman on Combined OC, Feels Coffee 5–7h

Non-smoker (×1.00) • Combined OC (×1.80) • Not pregnant (×1.00) • Effect: 5–7h (×1.25) • 1/day (×1.05) • Age 28 (×0.95)
t_½ = 5.0 × 1.00 × 1.80 × 1.00 × 1.25 × 1.05 × 0.95

= 11.2 hours → morning coffee only

Example C: Heavy Smoker, Coffee Wears Off Fast

10+/day smoker (×0.65) • N/A • N/A • Effect: <2h (×0.85 adjusted — 6+ cups masks perception) • 6+/day (×0.90) • Age 42 (×1.00)
t_½ = 5.0 × 0.65 × 0.85 × 0.90 × 1.00

= 2.5 hours (clamped) → can drink until evening

Rejected Approaches

Approach	Why rejected
"How does coffee affect your sleep?"	Circular — users don't know, that's why they use the app
Automatic Bayesian learning from sleep data	Too noisy — stress, screens, kids, alcohol all affect sleep. Can't attribute causation.
Ethnicity	GDPR Article 9 risk. Q4 captures phenotype regardless of genotype source.
Diet (broccoli, chargrilled meat)	Effect too small (5–15%), confusing UX
Family history	Users often don't know. Complex genetics, too noisy.
Specific medications	Handled in Settings as a separate drug interaction check, not in onboarding.

Section 05

User-Driven Recalibration

The onboarding questionnaire provides an initial estimate. Rather than attempting automatic Bayesian inference from noisy sleep data (stress, screens, kids, and alcohol all confound sleep quality attribution), CafOS uses direct user feedback on its own output.

Why not automatic sleep learning? Sleep quality is affected by dozens of variables beyond caffeine — work stress, screen time, alcohol, room temperature, hormonal cycles, children. Automatically attributing every bad night to caffeine would systematically overestimate the half-life. The model would drift toward false-slow, pushing cutoff times unreasonably early.

Recalibration Prompt

After 2 weeks of use, CafOS asks a single question:

"Do you feel your cutoff time is: Too early | About right | Too late"

Response	Adjustment	Rationale
Too early	k_e × 1.10 (faster metabolism)	User metabolizes caffeine faster than estimated
About right	No change	Model is calibrated
Too late	k_e × 0.90 (slower metabolism)	User is still affected at cutoff time

This prompt is available once per month in Settings. It creates a clean feedback loop on the app's output rather than attempting to infer metabolism from noisy sleep physiology data. Each adjustment shifts the effective half-life by ~10%, allowing convergence over 2–3 months without overfitting.

The key insight: Let the human decide if the cutoff feels right. They know their own life context better than any algorithm. We optimize the app output, not the sleep data.

Section 06

Safety Guardrails

Parameter	Value	Basis
Sleep caffeine threshold	20 mg	Conservative; Drake et al. 2013
WHO daily maximum	400 mg/day	EFSA/WHO healthy adult limit
t_½ model clamp	[2.5, 15.0] hours	Physiological bounds
Pregnancy daily limit	200 mg/day	ACOG recommendation

Disclaimer: CafOS is a wellness tool, not a medical device. It does not diagnose, treat, or prevent any condition. Users on medications that interact with CYP1A2 (e.g., fluvoxamine, ciprofloxacin, theophylline) should consult their healthcare provider. Pregnant users are advised to follow ACOG guidelines independently.

References

Drake C, Roehrs T, Shambroom J, Roth T. Caffeine effects on sleep taken 0, 3, or 6 hours before going to bed. J Clin Sleep Med. 2013;9(11):1195-1200.
Sachse C, Brockmöller J, Bauer S, Roots I. Functional significance of a C→A polymorphism in intron 1 of the cytochrome P450 CYP1A2 gene. Br J Clin Pharmacol. 1999;47(4):445-449.
Cornelis MC, El-Sohemy A, Kabagambe EK, Campos H. Coffee, CYP1A2 genotype, and risk of myocardial infarction. JAMA. 2006;295(10):1135-1141.
Parsons WD, Neims AH. Effect of smoking on caffeine clearance. Clin Pharmacol Ther. 1978;24(1):40-45.
Abernethy DR, Todd EL. Impairment of caffeine clearance by chronic use of low-dose oestrogen-containing oral contraceptives. Eur J Clin Pharmacol. 1985;28(4):425-428.
Fredholm BB, Bättig K, Holmén J, Nehlig A, Zvartau EE. Actions of caffeine in the brain with special reference to factors that contribute to its widespread use. Pharmacol Rev. 1999;51(1):83-133.
Landolt HP. Sleep homeostasis: a role for adenosine in humans? Biochem Pharmacol. 2008;75(11):2070-2079.
Knutti R, Rothweiler H, Schlatter C. Effect of pregnancy on the pharmacokinetics of caffeine. Eur J Clin Pharmacol. 1981;21(2):121-126.