Catalyst Twin · technical report · demonstration model, version 2 · 2026-10-07

caffeine.twin: a model of caffeine and its metabolites built only from quoted published values, tested against 21 studies it was not built from

Catalyst · catalystproject.ai

Not peer reviewed. caffeine.twin is the first module of Catalyst Twin, a model of the human body in development; this report describes that one module. Nothing here is a prediction for any person or medical advice.

Abstract

Background. Pharmacokinetic models are usually fitted to a handful of studies and judged against the literature they came from. We asked whether a model assembled only from published values, each traced to a verbatim quote, predicts studies it never saw.

Methods. Caffeine is one compartment with first-order absorption; CYP1A2 forms paraxanthine, theobromine and theophylline, each a compartment with its own measured clearance; brain A1-receptor occupancy follows from the half-maximal plasma concentration measured in people by PET. Every value comes from 7 papers. A validation protocol was committed before any prediction (commit 46b2c662). The test set combined 29 values from the graph’s study records with 18 values transcribed from 343 screened abstracts by two independent readers, kept only where both agreed and the quote occurred verbatim. Each study arm was simulated as reported, with 1,000 virtual subjects.

Results. 42 of 47 predictions (89%) fell within twofold of the reported value, 34 within 1.5-fold; average fold error 0.98, absolute average fold error 1.38, across 21 papers. Of the 18 values new to this test, 17 were within twofold. The one paraxanthine observation was predicted at 0.95 times the reported value. Peak concentration was under-predicted (fold error 0.74) and clearance under-predicted (fold error 0.70), consistent with low source values for clearance. The exported SBML reproduced the closed-form solution in libRoadRunner to a relative difference of 1.1 × 10⁻⁹.

Conclusions. A model built only from quoted literature values met the conventional twofold criterion for most of the summaries tested against independent studies. Its systematic biases point at specific inputs, absorption and clearance, which a next version can change and then test on data it has not seen.

1. Introduction

Caffeine is the most widely consumed psychoactive substance, and its kinetics are among the best documented in pharmacology: absorption is rapid and nearly complete^[6], clearance is almost entirely hepatic CYP1A2^[3], and its three dimethylxanthine metabolites have been characterised individually^{[5, 4]}. That makes it a fair first test of a simple question. If a model is assembled only from what papers report, with every number traceable to the sentence it came from, how well does it predict studies it was never shown?

Two features distinguish this work from a conventional model. First, provenance: every parameter carries the verbatim quote it was read from, most of them rows of a curated graph’s study records, and every block of the model links to them. Second, the test was fixed before it was run: the protocol, including the primary endpoint, was committed to version control before any prediction was made on the data it governs, and every observation is reported, inside the criterion or not.

2. Methods

2.1 Structure

Figure 1. Model structure. Solid arrows carry mass; the dotted arrow carries no mass (receptor occupancy is read from plasma concentration).

For a dose D at time zero, with amounts in mg:

dG/dt = −k_a G, G(0) = F·D

dA/dt = k_a G − (CL/V) A

dM_i/dt = f_i (CL/V) A − k_i M_i, i ∈ {paraxanthine, theobromine, theophylline}

C = A/V, occupancy = C / (C + IC₅₀)

The system is linear and solved in closed form: caffeine as a Bateman function and each metabolite as a three-exponential chain, summed over doses. A habitual day is the same doses laid down for the seven preceding days. Clearance scales with body mass and with CYP1A2 activity a through its CYP1A2 share f_m: CL = CL₀(1 − f_m + f_m a s), where s is the induction implied by the half-lives of smokers and non-smokers^[2] at unchanged volume. Metabolite formation scales with the same activity. The reference subject of Figure 2 is a 70 kg woman, non-smoker, who takes 100 mg at 07:30 and 15:00; her caffeine half-life is 6.8 h.

2.2 Parameters

Table 1 lists every parameter with its source sentence. Values were transcribed, never fitted. Most are rows of the study records the Catalyst graph keeps for each paper it reads (marked), whose anchors are checked mechanically against the paper’s text.

**Table 1.** Parameters, with the quote each was read from.
Block	Parameter	Value	Source
Gut	Mean absorption time, women	16 min	No significant differences in mean MAT (14 versus 16 min) or Tmax (63 versus 72 min) were found for men versus women ^[1]study record
Gut	Spread between people	51%	Interoccasion variability was smaller compared to interindividual variability in clearance, volume and absorption rate (27% vs. 33%, 10% vs. 15% and 23% vs. 51% respectively) ^[2]study record
Plasma caffeine	Volume, women	591 mL/kg	Vd ss/F (733 versus 591 mL/kg, p < 0.001) ^[1]study record
Plasma caffeine	Spread between people	15%	Interoccasion variability was smaller compared to interindividual variability in clearance, volume and absorption rate (27% vs. 33%, 10% vs. 15% and 23% vs. 51% respectively) ^[2]study record
CYP1A2 · liver	Clearance, women	1.0 mL/min/kg	CL/F (1.6 versus 1.0 mL/min/kg, p = 0.015) ^[1]study record
CYP1A2 · liver	Share through CYP1A2	95%	The partial or systemic caffeine clearance measured in plasma is considered to be the gold standard for CYP1A2 phenotyping since 95% of the systemic clearance of caffeine is estimated to be due to hepatic CYP1A2 ^[3]study record
CYP1A2 · liver	Smokers against non-smokers	×1.43 clearance	The extrapolated elimination half-lives of caffeine in the non-smokers and the smokers were 4.3 +/- 1.5 and 3.0 +/- 0.7 h respectively ^[2]study record
CYP1A2 · liver	Spread between people	33%	Interoccasion variability was smaller compared to interindividual variability in clearance, volume and absorption rate (27% vs. 33%, 10% vs. 15% and 23% vs. 51% respectively) ^[2]study record
Other routes	Share of clearance	5.9%	The mean (+/- s.d.) fractional conversion of CA to PX, TB and TP was 79.6 +/- 21.0%, 10.8 +/- 2.4% and 3.7 +/- 1.3%, respectively ^[4]study record
Paraxanthine	Share of caffeine converted	79.6%	The mean (+/- s.d.) fractional conversion of CA to PX, TB and TP was 79.6 +/- 21.0%, 10.8 +/- 2.4% and 3.7 +/- 1.3%, respectively ^[4]study record
Paraxanthine	Its own clearance	2.2 mL/min/kg	The total plasma clearances of CA and PX were similar in value (2.07 and 2.20 ml min-1 kg-1, respectively) as were those for TP and TB (0.93 and 1.20 ml min-1 kg-1, respectively) ^[5]study record
Paraxanthine	Its half-life	3.1 h	The half-lives of TP and TB (6.2 and 7.2 h, respectively) were significantly longer than those of CA and PX (4.1 and 3.1 h, respectively) ^[5]study record
Theobromine	Share of caffeine converted	10.8%	The mean (+/- s.d.) fractional conversion of CA to PX, TB and TP was 79.6 +/- 21.0%, 10.8 +/- 2.4% and 3.7 +/- 1.3%, respectively ^[4]study record
Theobromine	Its own clearance	1.2 mL/min/kg	The total plasma clearances of CA and PX were similar in value (2.07 and 2.20 ml min-1 kg-1, respectively) as were those for TP and TB (0.93 and 1.20 ml min-1 kg-1, respectively) ^[5]study record
Theobromine	Its half-life	7.2 h	The half-lives of TP and TB (6.2 and 7.2 h, respectively) were significantly longer than those of CA and PX (4.1 and 3.1 h, respectively) ^[5]study record
Theophylline	Share of caffeine converted	3.7%	The mean (+/- s.d.) fractional conversion of CA to PX, TB and TP was 79.6 +/- 21.0%, 10.8 +/- 2.4% and 3.7 +/- 1.3%, respectively ^[4]study record
Theophylline	Its own clearance	0.93 mL/min/kg	The total plasma clearances of CA and PX were similar in value (2.07 and 2.20 ml min-1 kg-1, respectively) as were those for TP and TB (0.93 and 1.20 ml min-1 kg-1, respectively) ^[5]study record
Theophylline	Its half-life	6.2 h	The half-lives of TP and TB (6.2 and 7.2 h, respectively) were significantly longer than those of CA and PX (4.1 and 3.1 h, respectively) ^[5]study record
A1 receptors · brain	Half-maximal plasma caffeine	67 μM	Half-maximal displacement was achieved at a plasma caffeine concentration of 67 μM, which corresponds to 450 mg in a 70-kg subject or approximately 4.5 cups of coffee. ^[7]
Coffee	Bioavailability	≈100%	The absolute bioavailability of caffeine is very high and reaches near 100% ^[6]study record
Gut	Mean absorption time, men	14 min	No significant differences in mean MAT (14 versus 16 min) or Tmax (63 versus 72 min) were found for men versus women ^[1]study record
Plasma caffeine	Volume, men	733 mL/kg	Vd ss/F (733 versus 591 mL/kg, p < 0.001) ^[1]study record
CYP1A2 · liver	Clearance, men	1.6 mL/min/kg	CL/F (1.6 versus 1.0 mL/min/kg, p = 0.015) ^[1]study record

Between-person variability is sampled log-uniformly one coefficient of variation either side of each typical value: absorption rate 51%, clearance 33%, volume 15%^[2]. Sex sets volume, clearance and absorption time^[1].

2.3 Implementation and verification

The model runs in the reader’s browser. Three checks confirm that the code computes the equations above. A unit test integrates the same system by explicit Euler steps and matches the closed form for caffeine and metabolites. Mass balance holds to 10⁻⁶ mg across a repeated day. And the model exported as SBML Level 3 Version 2 passes libSBML 5.21 with no errors, and libRoadRunner 2.10 (CVODE) reproduces the closed-form caffeine, paraxanthine, theobromine and occupancy for two subjects at eight times to a relative difference of 1.1 × 10⁻⁹. A control perturbation of 1% is detected by the same comparison. Anyone can repeat this: python scripts/twin-caffeine-sbml-check.py https://catalystproject.ai/twin/caffeine downloads the model and the page’s values and prints the difference.

2.4 Validation protocol

The protocol was committed as 46b2c662 with the model frozen at SHA-256 98af0315d6dbff4c9c40f801d7b16f99e8420854bd54114ac08501f02691d987. It fixes the data sources, the extraction rule, the scope, the simulation and the metrics, and states that no observation may be removed after prediction. Two amendments are recorded in it: a fault in parsing a record’s sample size, which changed no primary result, and the spellings of units the converter already handled, made before the first run.

Data. (i) Every value already held in the graph’s study records for papers not used to build the model (29 values, 11 papers). (ii) Every abstract in the graph’s caffeine library (2,667 papers) passing a fixed text screen (343 abstracts). Two readers transcribed each abstract independently into a fixed schema without seeing each other’s output or any prediction. A value was kept only where both reported the same arm, analyte, summary, value and unit, and only where the quote occurred verbatim in the stored abstract with the value inside it.

Scope. Adults given a single oral dose of caffeine as solution, beverage, capsule or tablet, fasted or not stated, measured in plasma or serum; the healthy control or placebo arm of an interaction or disease study was in scope. Excluded: infants and children, pregnancy, disease groups, oral contraceptive users, co-administered CYP1A2 inhibitors or inducers, repeated dosing, gum, inhaled, intravenous or sustained-release caffeine, fed arms, saliva, and groups named as elderly or obese.

Simulation and metrics. Each arm was simulated with its dose, sex (half and half when not stated), stated mean body mass (else 70 kg) and smoking status, as a single dose, with 1,000 virtual subjects. The primary endpoint was the fold error of the median virtual subject against the reported value, with success defined as within twofold, the conventional criterion for physiologically based predictions. We also report the proportions within 1.25- and 1.5-fold, the average fold error (AFE, a measure of bias) and the absolute average fold error (AAFE, a measure of precision). Two secondary tests: whether the reported value lay within the 5th–95th percentile of the virtual subjects, and whether it lay within the 90% interval of 500 virtual trials of the study’s own size.

2.5 Sensitivity

Normalised local sensitivity (elasticity) of five outputs of the reference day to eight inputs, by central difference at ±1%. An elasticity of −1 means that a 1% rise in the input lowers the output by 1%.

3. Results

3.1 The reference day

Figure 2. The reference subject’s day, repeated for a week: plasma caffeine and its three metabolites (left axis), and the share of brain A1 receptors caffeine occupies (right axis). Shaded: 23:00 to 06:00. A computation, not an observation.

3.2 The test set

Figure 3. How the test set was assembled. Every value not carried forward is listed with its reason in the JSON download.

The readers transcribed 153 and 149 values; 141 agreed, and every agreed quote was verbatim (141). 20 values reported by one reader only were set aside. Scope excluded 123, most often a disease population (28), route not oral, or not agreed (18), another drug, food or supplement given with caffeine (18), readers disagree on the population (16). The final set held 47 observations from 21 papers (11 through study records, 10 through the library), in 25 simulated arms. Several library papers studied a disease or an interaction; only their healthy control arms entered.

3.3 Accuracy

Figure 4. Predicted (median virtual subject) against observed, log axes. Solid line: identity. Dashed lines and band: twofold. Filled markers: study records; open: library abstracts. Colour: summary type.

**Table 2.** Accuracy by summary. AFE below 1 means the model predicts low.
Summary	n	≤1.25×	≤1.5×	≤2×	AFE	AAFE	5th–95th	Trials
Peak concentration	13	6	10	12	0.74	1.35	2	1
Time to peak	9	3	5	8	1.43	1.47	4	2
Total exposure, AUC	6	5	6	6	1.02	1.18	5	2
Exposure to a stated time	2	0	0	1	0.52	1.91	0	0
Half-life	10	3	8	9	1.37	1.37	8	0
Clearance	6	3	4	5	0.70	1.44	4	3
Paraxanthine, peak concentration	1	1	1	1	0.95	1.05	1	1
Study records	29	11	19	25	1.05	1.41	14	3
Library abstracts	18	10	15	17	0.88	1.32	10	6
All	47	21	34	42	0.98	1.38	24	9

42 of 47 predictions (89%) met the primary criterion. Overall bias was small (AFE 0.98), but it averages opposing errors. By summary, total exposure was the most accurate (AAFE 1.18); peak concentration was predicted low in 13 of 13 studies (AFE 0.74) and its timing late (Tmax AFE 1.43); clearance was predicted low (AFE 0.70) and half-life correspondingly long (AFE 1.37). The secondary test with virtual trials of each study’s size, which asks the model to reproduce a study’s mean rather than to contain its subjects, was met by 9 of 47: the biases above are small relative to the spread between people but large relative to the precision of a study mean.

Figure 5. Fold error (predicted over observed) of every prediction, by summary. Tick: the summary’s average fold error. Band: twofold.

3.4 Sensitivity

Figure 6. Elasticity of each output (columns) to each input (rows) for the reference day. Orange: the output rises with the input; blue: it falls.

Clearance dominates every caffeine output: a day’s exposure falls exactly in proportion to it (elasticity -1.00, as the algebra requires at steady state), and the concentration at lights out more than in proportion. Absorption rate moves the peak but not the day’s exposure. That is why the clearance bias of §3.3 matters more for an evening than the absorption bias does.

4. Discussion

A model with 23 transcribed values and no fitted parameter predicted 89% of summaries from independent studies within twofold, including the 18 values transcribed specifically for this test (17 of 18). The errors are not random. Peaks come earlier and higher in people than in the model, which takes its absorption time and volume from a single study of coffee and energy drinks^[1], while many test studies gave caffeine as a solution, capsule or tablet. Clearance runs low most plausibly because the same study’s clearance in women, 1.0 mL/min/kg, sits below every clearance reported in the test set (lowest 1.27 mL/min/kg). Both are changes to one input each, and the sensitivity analysis says which outputs they would move.

Limitations. Library values come from abstracts, where reporting is brief and arms are described loosely; the protocol’s requirement that both readers agree on the population excluded many values whose context was ambiguous. Several observations are means compared with a distribution of individuals; the virtual-trial test addresses this and is stricter. Paraxanthine was tested once. A1 occupancy rests on one PET study after intravenous caffeine^[7] and is not tested against any observation here; it is reported as a computation. The model has no food effect, no body-composition term, no saliva compartment and no CYP1A2 genotype. The paraxanthine contribution to receptor blockade is known and not quantified.

Next. A version 3 that changes absorption or clearance will be a new model under a new protocol, tested on study records the graph acquires after it is fixed. The test set grows as the graph reads more papers; the report recomputes from it.

5. Data and code

The model, the protocol, the test data and this report’s code are public: source, protocol, SBML, every comparison, data with audit trail. The model runs on /twin, where the validation can be recomputed in the browser.

References

White JR Jr, Padowski JM, Zhong Y, et al.. Pharmacokinetic analysis and comparison of caffeine administered rapidly or slowly in coffee chilled or hot versus chilled energy drink in healthy young adults. Clin Toxicol (Phila). 2016;54(4):308-12. PMID 27100333 · doi:10.3109/15563650.2016.1146740 · built the model
Seng KY, Fun CY, Law YL, et al.. Population pharmacokinetics of caffeine in healthy male adults using mixed-effects models. J Clin Pharm Ther. 2009;34(1):103-14. PMID 19125908 · doi:10.1111/j.1365-2710.2008.00976.x · built the model
Grzegorzewski J, Bartsch F, Köller A, et al.. Pharmacokinetics of Caffeine: A Systematic Analysis of Reported Data for Application in Metabolic Phenotyping and Liver Function Testing. Front Pharmacol. 2021;12:752826. PMID 35280254 · doi:10.3389/fphar.2021.752826 · built the model
Lelo A, Miners JO, Robson RA, et al.. Quantitative assessment of caffeine partial clearances in man. Br J Clin Pharmacol. 1986;22(2):183-6. PMID 3756066 · doi:10.1111/j.1365-2125.1986.tb05247.x · built the model
Lelo A, Birkett DJ, Robson RA, et al.. Comparative pharmacokinetics of caffeine and its primary demethylated metabolites paraxanthine, theobromine and theophylline in man. Br J Clin Pharmacol. 1986;22(2):177-82. PMID 3756065 · doi:10.1111/j.1365-2125.1986.tb05246.x · built the model
Guest NS, VanDusseldorp TA, Nelson MT, et al.. International society of sports nutrition position stand: caffeine and exercise performance. J Int Soc Sports Nutr. 2021;18(1):1. PMID 33388079 · doi:10.1186/s12970-020-00383-4 · built the model
Elmenhorst D, Meyer PT, Matusch A, et al.. Caffeine occupancy of human cerebral A1 adenosine receptors: in vivo quantification with 18F-CPFPX and PET. J Nucl Med. 2012;53(11):1723-9. PMID 22966134 · doi:10.2967/jnumed.112.105114 · built the model
Culm-Merdek KE, von Moltke LL, Harmatz JS, et al.. Fluvoxamine impairs single-dose caffeine clearance without altering caffeine pharmacodynamics. Br J Clin Pharmacol. 2005;60(5):486-93. PMID 16236038 · doi:10.1111/j.1365-2125.2005.02467.x · tested it
Desmond PV, Patwardhan RV, Johnson RF, et al.. Impaired elimination of caffeine in cirrhosis. Dig Dis Sci. 1980;25(3):193-7. PMID 7371463 · doi:10.1007/BF01308138 · tested it
Callahan MM, Robertson RS, Arnaud MJ, et al.. Human metabolism of [1-methyl-14C]- and [2-14C]caffeine after oral administration. Drug Metab Dispos. 1982;10(4):417-23. PMID 6126344 · tested it
Blanchard J, Sawers SJ. The absolute bioavailability of caffeine in man. Eur J Clin Pharmacol. 1983;24(1):93-8. PMID 6832208 · doi:10.1007/BF00613933 · tested it
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):425-8. PMID 4029248 · doi:10.1007/BF00544361 · tested it
Abernethy DR, Todd EL, Schwartz JB. Caffeine disposition in obesity. Br J Clin Pharmacol. 1985;20(1):61-6. PMID 4027137 · doi:10.1111/j.1365-2125.1985.tb02799.x · tested it
Wang T, Kleber G, Stellaard F, et al.. Caffeine elimination: a test of liver function. Klin Wochenschr. 1985;63(21):1124-8. PMID 4079279 · doi:10.1007/BF02291094 · tested it
Kamimori GH, Somani SM, Knowlton RG, et al.. The effects of obesity and exercise on the pharmacokinetics of caffeine in lean and obese volunteers. Eur J Clin Pharmacol. 1987;31(5):595-600. PMID 3830245 · doi:10.1007/BF00606637 · tested it
Scott NR, Stambuk D, Chakraborty J, et al.. Caffeine clearance and biotransformation in patients with chronic liver disease. Clin Sci (Lond). 1988;74(4):377-84. PMID 3356110 · doi:10.1042/cs0740377 · tested it
McDonagh JE, Nathan VV, Bonavia IC, et al.. Caffeine clearance by enzyme multiplied immunoassay technique: a simple, inexpensive, and useful indicator of liver function. Gut. 1991;32(6):681-4. PMID 2060878 · doi:10.1136/gut.32.6.681 · tested it
Mumford GK, Benowitz NL, Evans SM, et al.. Absorption rate of methylxanthines following capsules, cola and chocolate. Eur J Clin Pharmacol. 1996;51(3-4):319-25. PMID 9010706 · doi:10.1007/s002280050205 · tested it
Kamimori GH, Joubert A, Otterstetter R, et al.. The effect of the menstrual cycle on the pharmacokinetics of caffeine in normal, healthy eumenorrheic females. Eur J Clin Pharmacol. 1999;55(6):445-9. PMID 10492057 · doi:10.1007/s002280050654 · tested it
Akinyinka OO, Sowunmi A, Honeywell R, et al.. The effects of acute falciparum malaria on the disposition of caffeine and the comparison of saliva and plasma-derived pharmacokinetic parameters in adult Nigerians. Eur J Clin Pharmacol. 2000;56(2):159-65. PMID 10877011 · doi:10.1007/s002280050735 · tested it
Jodynis-Liebert J, Flieger J, Matuszewska A, et al.. Serum metabolite/caffeine ratios as a test for liver function. J Clin Pharmacol. 2004;44(4):338-47. PMID 15051740 · doi:10.1177/0091270004263468 · tested it
Skinner TL, Jenkins DG, Folling J, et al.. Influence of carbohydrate on serum caffeine concentrations following caffeine ingestion. J Sci Med Sport. 2013;16(4):343-7. PMID 22964452 · doi:10.1016/j.jsams.2012.08.004 · tested it
Laizure SC, Meibohm B, Nelson K, et al.. Comparison of caffeine disposition following administration by oral solution (energy drink) and inspired powder (AeroShot) in human subjects. Br J Clin Pharmacol. 2017;83(12):2687-2694. PMID 28758694 · doi:10.1111/bcp.13389 · tested it
Sadek P, Pan X, Shepherd P, et al.. A Randomized, Two-Way Crossover Study to Evaluate the Pharmacokinetics of Caffeine Delivered Using Caffeinated Chewing Gum Versus a Marketed Caffeinated Beverage in Healthy Adult Volunteers. J Caffeine Res. 2017;7(4):125-132. PMID 29230348 · doi:10.1089/jcr.2017.0025 · tested it
Morton K, Knight K, Kalman D, et al.. A Prospective Randomized, Double-Blind, Two-Period Crossover Pharmacokinetic Trial Comparing Green Coffee Bean Extract-A Botanically Sourced Caffeine-With a Synthetic USP Control. Clin Pharmacol Drug Dev. 2018;7(8):871-879. PMID 29659178 · doi:10.1002/cpdd.451 · tested it
Weiser T, Weigmann H. Effect of Caffeine on the Bioavailability and Pharmacokinetics of an Acetylsalicylic Acid-Paracetamol Combination: Results of a Phase I Study. Adv Ther. 2019;36(3):597-607. PMID 30758744 · doi:10.1007/s12325-019-0891-5 · tested it
Weiser T, Schepers C, Mück T, et al.. Pharmacokinetic Properties of Ibuprofen (IBU) From the Fixed-Dose Combination IBU/Caffeine (400/100 mg; FDC) in Comparison With 400 mg IBU as Acid or Lysinate Under Fasted and Fed Conditions-Data From 2 Single-Center, Single-Dose, Randomized Crossover Studies in Healthy Volunteers. Clin Pharmacol Drug Dev. 2019;8(6):742-753. PMID 30897305 · doi:10.1002/cpdd.672 · tested it
Thanawala S, Shah R, Abiraamasundari R, et al.. Comparative Bioavailability and Benefits on Mental Functions of Novel Extended-Release Caffeine Capsules against Immediate-Release Caffeine Capsules: An Open-Label, Randomized, Cross-over, Single-Dose Two-Way Crossover Study. J Diet Suppl. 2024;21(1):13-27. PMID 36602880 · doi:10.1080/19390211.2022.2160529 · tested it

Appendix: every observation

**Table S1.** Every comparison, with the quote it was read from. Values in mg/L, h, mg·h/L, L/h or L.
Ref.	Arm	Summary	Reported	Predicted	Fold	Quote
[22]abstract	420 mg, healthy males, 6 mg/kg caffeine, fasted	peak concentration	8.54	7.27	0.85	Peak concentration was greater in the 9mgkg(-1) fasted trial than the corresponding fed condition (70±9μmolL(-1) and 56±6μmolL(-1), respectively) and both were greater than the 6mgkg(-1) conditions (44±8μmolL(-1) and 38±8μmolL(-1) for 6mgkg(-1) fasted and fed, respectively).
[22]abstract	630 mg, healthy males, 9 mg/kg caffeine, fasted	peak concentration	13.6	11.0	0.81	Peak concentration was greater in the 9mgkg(-1) fasted trial than the corresponding fed condition (70±9μmolL(-1) and 56±6μmolL(-1), respectively) and both were greater than the 6mgkg(-1) conditions (44±8μmolL(-1) and 38±8μmolL(-1) for 6mgkg(-1) fasted and fed, respectively).
[23]record	100 mg, energy drink	peak concentration	1.94	1.96	1.01	peak plasma concentration 1790 and 1939 ng ml-1
[24]record	51 mg, instant coffee	peak concentration	1.30	0.989	0.76	Cmax (ng/mL) \| 1180 (24.2) \| 1300 (28.6)
[25]record	60 mg, caffeine from green coffee	peak concentration	1.91	1.05	0.55	Cmax, μg/mL \| 15 \| 1.91 (0.876) \| 16 \| 2.09 (1.49)
[25]record	60 mg, synthetic caffeine	peak concentration	2.09	1.04	0.50	Cmax, μg/mL \| 15 \| 1.91 (0.876) \| 16 \| 2.09 (1.49)
[26]record	50 mg, with aspirin and paracetamol	peak concentration	1.06	0.862	0.81	Cmax (µg/ml) \| 1.06 \| 1.09 \| 25.9
[27]record	100 mg, with ibuprofen, fasted	peak concentration	2.53	1.95	0.77	Cmax (µg/mL) \| 31.4 (5.22) \| 31.8 (6.74) \| 44.9 (8.77) \| 2.53 (0.72)
[12]abstract	162 mg, non-smoking drug-free healthy control females, 162 mg caffeine	peak concentration	4.09	3.53	0.86	(3.99 vs 4.09 micrograms/ml)
[11]record	350 mg, oral solution, 5 mg/kg	peak concentration	10.0	6.09	0.61	In addition, the variation in the maximum plasma concentration (Cmax) was low, 10.0 +/- 1.0 micrograms/ml
[18]abstract	72 mg, cola, 72 mg caffeine	peak concentration	1.57	1.41	0.90	For cola, Cmax: 1.57 micrograms.ml-1
[18]abstract	72 mg, caffeine capsules 72 mg, treatment 1	peak concentration	1.93	1.41	0.73	capsule treatments 1 (Cmax: 1.93 micrograms.ml-1)
[18]abstract	72 mg, caffeine capsules 72 mg, treatment 2	peak concentration	2.05	1.40	0.68	and 2 (Cmax: 2.05 micrograms.ml-1)
[22]abstract	420 mg, healthy males, 6 mg/kg caffeine, fasted	time to peak	1.00	0.940	0.94	Peak caffeine concentration occurred 60min following ingestion for both the 6 and 9mgkg(-1) fasted (p<0.001) trials compared to 120 and 180min following ingestion for the 6 and 9mgkg(-1) fed trials
[22]abstract	630 mg, healthy males, 9 mg/kg caffeine, fasted	time to peak	1.00	0.940	0.94	Peak caffeine concentration occurred 60min following ingestion for both the 6 and 9mgkg(-1) fasted (p<0.001) trials compared to 120 and 180min following ingestion for the 6 and 9mgkg(-1) fed trials
[24]record	51 mg, instant coffee	time to peak	0.625	1.02	1.63	tmax a (h) \| 0.667 (0.250–1.50) \| 0.625 (0.333–1.50)
[25]record	60 mg, synthetic caffeine	time to peak	0.630	0.940	1.49	tmax, h a \| 15 \| 0.75 (0.50, 1.50) \| 16 \| 0.63 (0.50, 1.50)
[25]record	60 mg, caffeine from green coffee	time to peak	0.750	0.940	1.25	tmax, h a \| 15 \| 0.75 (0.50, 1.50) \| 16 \| 0.63 (0.50, 1.50)
[26]record	50 mg, with aspirin and paracetamol	time to peak	0.500	0.940	1.88	tmax (h) \| 0.50
[27]record	100 mg, with ibuprofen, fasted	time to peak	0.500	1.02	2.04	Median tmax (h) \| 1.88 \| 1.88 \| 0.50 \| 0.50
[28]record	200 mg, immediate-release tablet	time to peak	0.830	1.02	1.23	Mean tmax was 4.08 ± 2.13h for SR-Caffeine compared to 0.83 ± 0.39h for IR-Caffeine, (p < 0.0001)
[11]record	350 mg, oral solution, 5 mg/kg	time to peak	0.497	0.960	1.93	The oral absorption was very rapid, reaching a peak (Tp) plasma concentration after 29.8 +/- 8.1 min (mean +/- SEM)
[19]record	300 mg, women, three cycle phases	total exposure, auc	83.0	74.6	0.90	83.0 mg 1(-1) h and 2.06 h(-1), respectively, during the luteal phase
[19]record	300 mg, women, three cycle phases	total exposure, auc	84.7	74.6	0.88	84.7 mg 1(-1) x h and 1.84 h(-1), respectively, during the follicular phase
[19]record	300 mg, women, three cycle phases	total exposure, auc	93.0	74.6	0.80	was 93.01 mg 1(-1) x h and the absorption rate constant (ka) was 2.88 h(-1) during the ovulatory phase
[24]record	51 mg, instant coffee	total exposure, auc	8.52	10.1	1.19	AUC0-∞ (h·ng/mL) \| 7440 (34.6) \| 8520 (35.2)
[26]record	50 mg, with aspirin and paracetamol	total exposure, auc	5.73	7.84	1.37	AUC0–∞ (µg h/ml) \| Test \| 5.73 \| 6.06 \| 35.0
[27]record	100 mg, with ibuprofen, fasted	total exposure, auc	18.7	19.9	1.06	AUC0–∞ (µg · h/mL) \| 138 (31.2) a \| 130 (26.4) \| 127 (26.4) \| 18.7 (13.1)
[25]record	60 mg, caffeine from green coffee	exposure to a stated time (0–4 h)	6.35	3.51	0.55	AUC0–4 h, μg·h/mL \| 15 \| 6.35 (3.34) \| 16 \| 6.99 (5.45)
[25]record	60 mg, synthetic caffeine	exposure to a stated time (0–4 h)	6.99	3.48	0.50	AUC0–4 h, μg·h/mL \| 15 \| 6.35 (3.34) \| 16 \| 6.99 (5.45)
[21]abstract	300 mg, healthy volunteers, 300 mg caffeine	half-life	4.30	6.31	1.47	values of clearance (0.035 vs. 0.094 L/h/kg), elimination coefficient (0.061 vs. 0.153 h(-1)), and half-life (11.4 vs. 4.3 h)
[8]record	250 mg, with placebo	half-life	4.90	6.26	1.28	prolonged its elimination half-life (4.9 vs. 56 h, P < 0.01; mean difference: 51 h, 95% CI: 26-76)
[24]record	51 mg, instant coffee	half-life	4.70	6.24	1.33	t1/2 b (h) \| 4.91 (1.44) \| 4.70 (1.26)
[26]record	50 mg, with aspirin and paracetamol	half-life	4.29	5.50	1.28	t1/2 (h) \| 4.29 \| 4.50 \| 31.7
[27]record	100 mg, with ibuprofen, fasted	half-life	5.64	6.23	1.10	Mean t1/2 (h) \| 2.54 (0.88) \| 2.13 (0.25) \| 2.26 (0.37) \| 5.64 (3.22)
[28]record	200 mg, immediate-release tablet	half-life	5.78	6.29	1.09	mean t½ was 7.07 ± 3.48h for SR-Caffeine compared to 5.78 ± 2.11h for IR-Caffeine (p = 0.04189)
[15]record	408 mg, lean men, at rest	half-life	2.59	5.41	2.09	longer serum half-life (t1/2 4.37 vs. 2.59 h)
[13]record	162 mg, normal body weight	half-life	5.40	6.35	1.18	elimination half-life (7.05 +/- 1.08-obese vs 5.40 +/- 0.40 h; NS)
[12]abstract	162 mg, non-smoking drug-free healthy control females, 162 mg caffeine	half-life	5.37	7.28	1.36	(mean 7.88 h vs 5.37 h in the controls)
[10]abstract	350 mg, adult male volunteers, 5 mg/kg radiolabeled caffeine	half-life	3.00	5.46	1.82	The half-life of caffeine in both serum and saliva was approximately 3 hr
[21]abstract	300 mg, healthy volunteers, 300 mg caffeine	clearance	6.58	5.00	0.76	values of clearance (0.035 vs. 0.094 L/h/kg), elimination coefficient (0.061 vs. 0.153 h(-1)), and half-life (11.4 vs. 4.3 h)
[17]abstract	300 mg, hospital and normal control subjects, 300 mg caffeine citrate	clearance	5.33	4.97	0.93	compared with control subjects (median 1.27 ml/min/kg, p less than 0.001)
[16]abstract	400 mg, healthy control subjects, 400 mg caffeine	clearance	5.46	5.01	0.92	the mean (+/- SD) serum caffeine clearance was 1.3 +/- 0.4 ml min-1 kg-1
[12]abstract	162 mg, non-smoking drug-free healthy control females, 162 mg caffeine	clearance	7.35	3.90	0.53	(1.05 vs 1.75 ml/min/kg, respectively)
[14]abstract	140 mg, healthy volunteers, 140 mg caffeine	clearance	13.0	5.04	0.39	(0.23 +/- 0.04 h-1; 3.1 +/- 0.9 ml/min per kg)
[9]abstract	250 mg, healthy controls, 250 mg caffeine	clearance	5.88	5.05	0.86	significantly higher (1.4 +/- 0.5 ml/min/kg) in controls as compared to cirrhotics (0.9 +/- 0.3 ml/min/kg)
[20]abstract	300 mg, healthy adult Nigerians, 300 mg caffeine	Paraxanthine peak concentration	1.40	1.33	0.95	in healthy controls (1.4 +/- 0.5 microg/ml)