We searched the general purpose pharmacokinetics database PK-DB with the search query https://pk-db.com/api/v1/filter/?concise=false&download=true&interventions__substance=caffeine on 2021–10–06 and PKPDAI (Hernandez et al., 2021) with the search query https://app.pkpdai.com/?term=caffeine on 2021–10–06. PKPDAI is a dedicated search engine for pharmacokinetics data based on PubMed. This search resulted in 745 studies from which 599 could be accessed as full text.

Studies were included or excluded based on the following eligibility criteria. Reviews or publications describing computational models were excluded. For a study to be eligible caffeine must be applied, it must report pharmacokinetics data for caffeine or its metabolites, data must be measured in humans, and individuals and groups must consisted of adults (age > = 18 years), and data must be in vivo data. The application route of caffeine must be either oral, intravenous, or intramuscular. All application forms of caffeine (e.g., tablet, capsule, solution) are eligible. No restrictions were imposed on the dosing amount of caffeine or coadministrations. Relevant outcome measures are concentration-time profiles in plasma, serum, blood, and saliva of caffeine and metabolites and corresponding pharmacokinetic parameters. Studies containing pharmacokinetic parameters of caffeine and metabolites (i.e., clearance, maximum concentration, time of maximum concentration, half-life, AUC) and metabolic ratios of caffeine and its metabolites were included. In total 286 studies met our eligibility criteria, 313 studies were excluded for various reasons as described in Figure 1. Of the 286 studies, 145 were excluded due to low priority, e.g., if only urinary data was reported. Data was extracted and curated from the remaining 141 studies. During the curation process outliers from four studies (Stille et al., 1987; Harder et al., 1988, 1989; Balogh et al., 1992) were identified and excluded from all subsequent analyses. All four studies probably originate from the same clinical investigation.

Pharmacokinetics data was curated manually as part of the pharmacokinetics database PK-DB (https://pk-db.com/) (Grzegorzewski et al., 2020) using established workflows. The pharmacokinetics data was stored in combination with relevant metadata on groups, individuals, interventions, and outputs. PK-DB provided support in the curation process with strong validation rules, unit normalization, and automatic calculation of pharmacokinetic parameters from time-courses. As part of the curation process and the presented analyses, pharmacokinetic parameters and other commonly reported measurements were integrated from multiple studies. The meta-analyses and data integration allowed to identify and correct/remove outlier data which were mostly due to either curation errors or incorrect reporting. Pharmacokinetic parameters calculated from time-courses are included in the analyses. For more details see (Grzegorzewski et al., 2020).

In Figures 3A,B, Figures 5A–D, and Figure 4 the data is displayed in a similar manner. For collectively reported subjects, the group size and standard error is displayed as the marker size and error-bar, respectively. In the legend, (I), (G), and (TI) stand for individual participant data, the number of groups, and the total number of subjects, respectively. Data points are depicted as circles if reported equivalently in the source, as squares if calculated from concentration-time profiles and as triangles if inferred from corresponding pharmacokinetic data and body weights of the subjects. Typically, dosing is reported in mass units, AUC in mass per volume units, clearance in volume per time units, and half-life in time units. Occasionally, dosage, AUC, and clearance are reported in body weight units. In case of reported subject weight, the data is harmonized to similar units by multiplying with the reported weights.

In Figure 3, the depicted subjects were healthy. Male subjects were assumed to not take any oral contraceptives. Substances with negligible caffeine-drug interactions were determined by an effect size analysis in Section 3.5. All other co-administrations and investigations containing multiple caffeine dosages were excluded.

In subplot Figure 5A, included subjects were healthy. The area under the caffeine concentration curves measured at least up to 12 h after a single application of caffeine and AUC extrapolated to infinity were included. Multiple subsequent caffeine dosages were excluded, other administrations and co-administrations included. In subplot Figure 5C, healthy and non-healthy subjects were included. Single applications of caffeine or caffeine administrated as a cocktail with negligible drug-drug interactions were included. All other co-administrations and investigations containing multiple caffeine dosages were excluded.

In subplots Figures 6A–C, no data was excluded. In subplot Figure 6D and Figure 4, included subjects were healthy, non-smoking, non-pregnant, and non-oral contraceptive consumers. Interventions with caffeine administrated as a cocktail with negligible caffeine-drug interactions were included. All other co-administrations and investigations containing multiple caffeine dosages were excluded.

Within this work, the first comprehensive open pharmacokinetics data set on caffeine was established. The data set integrates data from 141 publications (Figure 2 and Table 1), with most of the publications corresponding to a distinct clinical trial. Studies were identified and included or excluded following the systematic approach described in the PRISMA flow diagram (Figure 1A). The focus of data curation was on pharmacokinetics data of caffeine, caffeine metabolites, and caffeine metabolic ratios in human adults. Importantly, the data set is enriched with meta-data on 1) the characteristics of studied patient cohorts and subjects (e.g., age, body weight, smoking status, health status, fasting); 2) the applied interventions (e.g., dosing, substance, route of application); 3) measured pharmacokinetic time-courses; and 4) pharmacokinetic parameters (e.g., clearance, half-life, area under the curve). In summary, data from 500 groups and 4,714 individuals is reported under 387 interventions resulting in 24 ,571 pharmacokinetic outputs and 846 time-courses. The data set is available via the pharmacokinetics database PK-DB (https://pk-db.com/) with a detailed description of the data structure provided in (Grzegorzewski et al., 2020). We demonstrate the value of the data set by its application to multiple research questions relevant for metabolic phenotyping and liver function testing (Figure 1B): 1) the effect of smoking and oral contraceptive use on caffeine elimination (Section 3.3); 2) the effect of the type of assay (Section 3.4); 3) the interaction of caffeine with other drugs (Section 3.5), 4) alteration of caffeine pharmacokinetics in disease (Section 3.6); and 5) the applicability of caffeine as a salivary test substance by comparison of plasma and saliva data (Section 3.7). In the following, we summarize the quality of reporting and provide example applications of the data set.

The study design as well as quality and details of reporting of results were very heterogeneous between studies. Major differences exist in the study design, number of study participants, and number of reported time-courses. Many studies report some individual participants data (84/141) but only a minority of the studies report individual participant data for all study participants (6/84) (Axelrod and Reichenthal, 1953; Brazier et al., 1980; Rietveld et al., 1984; McQuilkin et al., 1995; Perera et al., 2010; Millard et al., 2018). Many studies report only aggregated data on group level (57/141). In most studies, the application of a single dose of caffeine was studied (129/141). In the case of multiple interventions (49/141), mostly one additional substance was co-administrated (33/141). The main categories of studies were either 1) case-control studies which compare caffeine pharmacokinetics in two groups (e.g., healthy vs Disease) (64/141), 2) crossover studies on caffeine-drug interactions (comparing caffeine alone vs caffeine and additional substance) (33/141) 3) studies on metabolic phenotypes (including drug cocktails) (42/141); or 4) methodological studies (e.g., establishing mass spectrometry protocol for quantification or new site of sampling).

Intervention protocols, i.e., the applied substances, form, dose, and timing of application was typically reported in good detail. In crossover studies, the difference between treatments was generally reported in good detail. For the dosing with caffeine, the amount (136/141), route (e.g., oral, intravenous) (140/141), form (tablet, capsule, solution) (136/141), and the substance (141/141) are typically reported. Co-administrations of medication and other substances are often mentioned qualitatively (28/49), skipping either the amount, route, form, or exact timing of application.

The quantification protocol, i.e., type of assay (e.g., high-performance liquid chromatography, gas-liquid chromatography) (140/141), the site of sampling (e.g., plasma, serum, saliva, urine) (141/141), and the time points when samples were taken (128/141), were mostly reported in good detail. However, in some studies the protocol is not mentioned explicitly but only via additional references, which complicates the curation. For a single study no information on the quantification method could be accessed (Blanchard and Sawers, 1983c). The reporting of type of assay is very heterogeneous with many studies not providing sufficient details (e.g., HPLC is often mentioned as the separation technique but the subsequent detection or quantification method is not stated).

The information on subject characteristics was less often reported in sufficient detail with large differences in the quality and quantity of reporting between studies. Any information on sex (131/141), weight (71/141), and age (89/141) was relatively often reported on group or individual level. However, age and weight were rarely reported on an individual level and often not even for all groups. Other anthropometric factors such as height (15/141), body mass index (BMI) (15/141), and ethnicity (25/141) are rarely reported. The genotype of CYP1A2 (gCYP1A2) is rarely measured or reported (5/141), even though there is evidence that genetic variation can play an important role in caffeine metabolism. Further, the nomenclature is not standardized. It is worth noting that low-cost genotyping methods were not available for early studies included in the data set. The phenotype (pCYP1A2, pXO, pNAT2) of enzymes involved in the metabolism of caffeine, i.e., CYP1A2, xanthine oxidase (XO), or N-acetyltransferase type 2 (NAT2) were investigated occasionally (38/141). The information on other factors influencing the pharmacokinetics of caffeine is reported very heterogeneously. The strong influence of smoking (105/141) and the use of oral contraceptives (42/141) on the enzyme activity of CYP1A2 and thereby on the apparent clearance of caffeine is covered relatively well in many publications. Health status and patient diseases are often covered (134/141). However, often categorized in broad and general disease classes, with more specific disease classification lacking. Markers related to cardiovascular health (e.g., blood pressure, cholesterol level, bilirubin level) are basically not reported in the context of caffeine pharmacokinetics. In case of cirrhosis, further information of severity is reported sparsely. Important information on the abstinence of caffeine or methylxanthines and consumption of caffeine or other caffeine-containing beverages is often missing.

Individual-level reporting is essential for subsequent pharmacokinetic modeling, as biological mechanisms responsible for the pharmacokinetics strongly correlate with these factors and large inter-individual variability exists in caffeine pharmacokinetics. Despite the importance of individual subject data, information on individuals is rarely provided.

Most studies report pharmacokinetics outputs on caffeine (126/141). Data on the main product paraxanthine (46/141) and secondary metabolites theobromine (20/141) and theophylline (19/141) are reported sometimes. Additional metabolites such as 137U, 17U, 13U, 37U, 1X, 1U, 3X, 3U, 7X, 7U, AFMU, AAMU, ADMU, A1U, and A3U are seldom reported, mostly as urinary measurements and not directly but as part of a metabolic ratio.

In a first analysis, we were interested in the effect of smoking and oral contraceptive use on the pharmacokinetics of caffeine in healthy subjects (see Figure 3). Both have repeatedly been reported as key exogenous factors affecting caffeine elimination. A main question was how reproducible the effect is and if by integrating data from multiple studies a more consistent picture of the effects can be gained. For the analysis the data set was stratified into smokers, oral contraceptive users and a control group (neither smoking nor using oral contraceptives). Smoking results in increased caffeine clearance (Figures 3A,C) and decreased half-life of caffeine elimination (Figures 3B,D) whereas oral contraceptive use has the opposite effect over a wide dose range of caffeine. An important result from our analysis is that a consistent and reproducible effect can be found over more than 50 years of pharmacokinetic research. With exception of a few outlier studies probably from a single clinical trial (see methods) all data was highly consistent. This provides a strong argument for the applied methods and protocols.

Despite the large effect of smoking and oral contraceptive use on the pharmacokinetics of caffeine, the information is only reported for a subset of studies. Smoking as well as oral contraceptive use should be an exclusion criteria for subjects in studies of caffeine pharmacokinetics, due to the possible confounding effects. Importantly, in many groups smokers and non-smokers were mixed without reporting data for smokers and non-smokers separately. Without reporting of data on individuals or subgroups no stratification could be performed, which could strongly affect results if not balanced between groups. In summary, the integrative data analysis showed a consistent strong activating effect of smoking on caffeine elimination and an inhibiting effect of oral contraceptive use on caffeine elimination.

The reported type of assays were curated systematically (see Table 1 to study effects of the type of technique and assay on the reported pharmacokinetics data (see Figure 4, abbreviations explained in the legend of Table 1). Data affected by confounding factors (i.e., smoking, oral contraceptive use, caffeine-disease, and caffeine-disease interactions) was excluded. Based on our analysis no systematic difference between the reported type of assays could be detected. Immunoassays (RIA, EMIT, ELISA), mass spectrometry (LC MS, GC MS, LC MS/MS, HPLC MS/MS, HPLC-ESI-MS/MS, ES MS/MS), and chromatography with unspecified quantification method (HPLC, RP-HPLC, GLC, CGC, CC, GC) do not show any differences in half-life or clearance. Spectrophotometric methods (UV/Vis) are generally rare and were not applied in this data subset. Based on the analysis we conclude that no systematic correction was required to compare the data collected with the different types of assays.

An important question for metabolic phenotyping and liver function testing with caffeine is how the coadministration of other drugs and compounds affects caffeine clearance. Consequently, we studied in the second analysis the reported caffeine-drug interactions in the data set. The impact of drugs was quantified using the change in AUC between a coadministration and a respective control l o g ( A U C c o a d m i n i s t r a t i o n A U C c o n t r o l ) (see Figures 5A,C).

Only case-controlled studies, mostly cross-over trials with a washout phase were included in the analysis. Corresponding controls were not matched across different studies. Overall coadministration data with AUC difference was available for 33 substances in our data set. In accordance with FDA, EMA and PMDA guidelines (Sudsakorn et al., 2020) we classified substances as inhibitors or inducers of caffeine clearance based on changes in AUC: FDA–Clinical DDI guidance: strong, moderate and weak inhibitors increase the AUC ≥ 5 -fold, ≥ 2 to < 5 -fold, ≥ 1.25 to < 2 -fold, respectively; strong, moderate and weak inducers decrease the AUC by ≥ 80 percent, ≥ 50 to < 80 percent and ≥ 20 to < 50 percent, respectively.

Most substances do not affect the AUC of caffeine, with the exception of fluvoxamine, pipemidic acid and norfloxacin, which inhibit caffeine clearance. Tipranavir was the only substance showing a weak induction of caffeine clearance, but only in steady state dosing (not after a single dose) (Dumond et al., 2010)). Substances administered as a cocktail along side caffeine, aiming to phenotype several enzymes simultaneously did not affect the AUC of caffeine, substantiating the use of caffeine as part of a drug cocktail design (Turpault et al., 2009; Dumond et al., 2010; Oh et al., 2012; Doroshyenko et al., 2013; Kakuda et al., 2014; Tanaka et al., 2014; Armani et al., 2017; Edwards et al., 2017; Lammers et al., 2018). Our analysis shows that only a minority of studied drugs show an interaction with caffeine, confirming its value for phenotyping even under co-administration. As a side note, the protocols for studying caffeine-drug interactions were highly variable, e.g., the applied caffeine dose and the dose of the coadministrated substance varied between studies. Our results suggest that most medications can be safely consumed in combination with caffeine with exception of the antidepressant fluvoxamine, the antibacterial pipemidic acid and the antibiotic norfloxacin in which case caution is warranted.

An important question for using caffeine as a test substance for liver function testing and phenotyping, as well as for drugs metabolized via CYP1A2, is how disease affects the pharmacokinetics and elimination of caffeine. To study this question, we stratified caffeine clearance rates based on the reported disease of subjects and groups in the data set. To quantify the effect of disease the absolute clearance of caffeine (Figure 5B) and the logarithmic difference to a control group l o g ( A U C d i s e a s e A U C c o n t r o l ) (Figure 5D) were analyzed. The corresponding controls were not matched across different studies.

None of the reported diseases increased the clearance rate of caffeine. Most of the diseases contained in this data set are diseases of the liver (e.g., alcoholic cirrhosis, primary biliary cholangitis) or are known to affect the liver (e.g., alcohol dependent). Cirrhotic liver disease had moderate to strong effects on the caffeine clearance with large variability in the reported data. Malaria and obesity had no effect on clearance with caffeine. An issue in the study of caffeine-disease interaction is that control group and disease group are different subjects (no cross-over design). In addition diseases were reported very heterogeneously (e.g., either only cirrhosis or with underlying cause such as alcoholic cirrhosis).

An important question for metabolic phenotyping and liver function tests with caffeine is how saliva measurements correlate with plasma or serum caffeine measurements, in the following referred to as blood-based measurements. A good correlation would allow simple non-invasive phenotyping using saliva samples. To study this question we analyzed 1) the relationship of blood-based concentrations of caffeine and paraxanthine with their respective saliva concentrations (Figures 6A,B); and 2) how the caffeine clearance measured in saliva correlate to blood-based measurements (Figure 6C).

Systematic errors due to different dosing protocols and different clinical investigation seem to be minimal as the data from multiple studies shows very consistent results. Linear regressions were performed to quantify the relation between saliva and blood-based caffeine and paraxanthine measurements (see Figures 6A–C). The resulting scaling factors of saliva to blood-based concentration of caffeine and paraxanthine are 0.79 ± 0.01 and 0.68 ± 0.02 ( x ̄ ± SD ), respectively. Pearson correlation coefficients between saliva and blood-based concentrations for caffeine and paraxanthine are 0.84 and 0.76, respectively.

When comparing saliva-based caffeine clearance against blood-based clearance (Figure 6C) an even stronger correlation of 0.88 with a scaling factor of 01.04 ± 0.02 ( x ̄ ± SD ) is observed. The integrated clearance data strongly indicates that clearance can either be calculated from saliva or blood-based measurements.

Paraxanthine/caffeine ratios are mainly used for metabolic phenotyping based on caffeine. Whereas most studies use 6 h to phenotype no clear consensus exists in the literature and metabolic ratios are reported for varying time points after caffeine application. Paraxanthine/caffeine ratios for caffeine administered either as a single dose or in a cocktail to healthy, non-smoking, non-pregnant, and non-oral contraceptive consuming subjects were investigated (Figure 6D). By applying this strict data filtering, the variability due to either smoking and oral contraceptive use (see Section 3.3), caffeine-drug interactions (see Section 3.5) or disease (see Section 3.6) could be removed from the metabolic ratios.

Early and late time-sampling are least suitable for phenotyping. At these time points concentrations are low, resulting in relatively high random errors and thus low single to noise ratio. In an early stage, the outcome of metabolic ratios are further influenced by the distribution phase of the substance and its absorption kinetics, both affected by form and route of administration. Main results are that the metabolic phenotyping with paraxanthine/caffeine ratios is strongly time dependent with increasing ratios with time; and that a clear caffeine-dose dependency exists in the phenotyping with smaller caffeine doses increasing the metabolic ratio. Our results show the importance of clear standardized protocols for metabolic phenotyping.

In summary, by integrating data from multiple studies we could show a very good correlation between saliva and plasma caffeine concentrations, paraxanthine concentrations and pharmacokinetic parameters calculated from saliva versus plasma concentrations. This pooled data provides a strong argument for caffeine phenotyping based on saliva samples. Furthermore, we analyzed the time dependency of paraxanthine/caffeine ratios often used for phenotyping of CYP1A2. Our time dependent correlation allows to correct caffeine/paraxanthine ratios depending on the time after application and caffeine dose.