Integrating Static and Dynamic Models to Predict Cannabidiol‐Mediated Metabolic Drug–Drug Interactions

Abstract

Cannabidiol (CBD) inhibits multiple cytochrome P450 (CYP450) enzymes in vitro via reversible and time‐dependent mechanisms, raising concerns for metabolic drug–drug interactions (DDIs). Clinical DDI data for CBD are limited, and the translation of in vitro derived inhibition parameters into in vivo DDI predictions remains uncertain. We implemented a stepwise, model‐informed framework to evaluate CBD‐mediated metabolic DDIs using (i) basic, (ii) mechanistic static (MSM), and (iii) physiologically based pharmacokinetic (PBPK) models. The basic model and MSM incorporated in vitro inhibition parameters for CBD and its primary active metabolite, 7‐hydroxycannabidiol (7‐OH‐CBD), to identify sensitive metabolic pathways. A PBPK model for CBD and 7‐OH‐CBD was developed and validated against clinical pharmacokinetic data, then used to simulate DDIs. Where necessary, inhibition parameters were refined using the PBPK model to capture clinical observations, and the optimized parameters were used to revisit the basic model and MSM, facilitating comparison across modeling approaches. The MSM predicted 3.6‐, 4.1‐, and 2.2‐fold increases in exposure of sensitive CYP1A2, CYP2C9, and CYP3A4 substrates, respectively, and a strong CYP2C19‐mediated interaction. In contrast, the PBPK model predicted lower DDI magnitudes consistent with clinical data, identifying CYP2C19 inhibition as the most clinically relevant liability, with a moderate interaction risk for sensitive substrates, while CYP1A2 inhibition increased sensitive substrate exposure twofold. The validated PBPK model was applied to evaluate exposure in special populations and DDI risk with commonly co‐prescribed medications. Overall, this framework reconciles in vitro and clinical DDI data and supports population‐specific assessment of CBD‐mediated DDIs across diverse clinical scenarios.

Study Highlights.

• WHAT IS THE CURRENT KNOWLEDGE ON THIS TOPIC?

Cannabidiol (CBD) is often used with concomitant medications across diverse populations. In vitro, CBD inhibits multiple CYP450 enzymes, but clinical drug–drug interaction (DDI) data remain limited.

• WHAT QUESTION DID THIS STUDY ADDRESS?

Can a stepwise, model‐informed framework predict CBD‐mediated DDIs and extend risk assessment to special populations and clinically relevant co‐medications?

• WHAT DOES THIS STUDY ADD TO OUR KNOWLEDGE?

Integrating basic, mechanistic static, and PBPK models showed that broad in vitro CYP450 inhibition and DDI risks flagged by static approaches do not uniformly translate into clinically meaningful interactions. The PBPK model identified CYP2C19 as the most clinically relevant DDI liability, resulting in a moderate DDI risk with sensitive substrates, while CYP1A2 inhibition increased sensitive substrate exposure twofold.

• HOW MIGHT THIS CHANGE CLINICAL PHARMACOLOGY OR TRANSLATIONAL SCIENCE?

This work supports a focused approach to CBD DDI risk assessment by leveraging PBPK modeling to prioritize clinically actionable risks, particularly CYP2C19‐mediated interactions involving sensitive substrates, prodrugs, active metabolites, and CYP1A2 substrates with narrow therapeutic indices, while extending risk assessment to populations underrepresented in clinical studies.

In the United States, evolving regulations have expanded access to medical marijuana and cannabinoid‐containing products, reflecting broader societal and clinical interest in their therapeutic potential. Cannabidiol (CBD) is one of the most abundant and clinically relevant phytocannabinoids. ¹ Unlike delta‐9‐tetrahydrocannabinol (Δ9‐THC), CBD has a low affinity for the cannabinoid receptor 1 (CB1) and does not produce the characteristic euphorigenic and psychotomimetic effects associated with Δ9‐THC, which has contributed to its increasing use in medical and consumer settings. ² In 2018, the US Food and Drug Administration (FDA) approved Epidiolex® (CBD), a purified oral solution, for the treatment of rare and severe forms of epilepsy, including Lennox–Gastaut syndrome, Dravet syndrome, and tuberous sclerosis complex‐associated seizures. ³ Beyond these labeled indications, CBD‐containing products are used for a wide range of off‐label conditions, including chronic pain, post‐traumatic stress disorder, anxiety, depression, and sleep disorders. ⁴

This expanding use raises important drug–drug interaction (DDI) concerns. CBD and its primary active metabolite, 7‐hydroxycannabidiol (7‐OH‐CBD), inhibit multiple cytochrome P450 (CYP450) enzymes in vitro at concentrations that may be clinically relevant. ² , ⁵ , ⁶ In addition, case reports have raised concerns about potential CBD‐mediated DDIs with commonly used medications, including escitalopram, warfarin, clopidogrel, and tacrolimus. ⁷ These risks are particularly relevant for children with severe epilepsy, older adults, and medically complex patients who often receive multiple concomitant medications. In addition, CBD‐containing products are widely available outside prescription settings, increasing the likelihood of unsupervised coadministration with drugs that have narrow therapeutic indices, active metabolites, or CYP450‐dependent clearance. Despite this, systematic clinical DDI evaluations remain limited, partly due to the availability of only one FDA‐approved CBD product, as well as marijuana's regulatory status, which has historically created barriers to controlled clinical studies. Moreover, CBD affects multiple metabolic pathways, and fully characterizing its DDI potential through dedicated clinical studies across all relevant substrates and populations would be impractical.

Recently, the International Council for Harmonization (ICH) issued the M12 Drug Interaction Studies guidance outlining a framework for investigating CYP450 enzyme‐ and transporter‐mediated DDI risk, emphasizing the use of model‐informed approaches to translate in vitro findings into in vivo DDI predictions. ⁸ Models ranging from conservative basic screening tools to mechanistic static model (MSM) and dynamic physiologically based pharmacokinetic (PBPK) models are applied in a stepwise manner to identify sensitive metabolic pathways and inform the need for clinical DDI studies, thereby reducing unnecessary trials. This framework is well‐suited to compounds such as CBD, where extensive clinical DDI evaluation is not feasible, and model‐informed strategies offer a valuable means of integrating in vitro and clinical data to characterize interaction risk.

Although PBPK models for CBD have been previously developed, most have focused on characterizing CBD pharmacokinetics, ⁹ , ¹⁰ , ¹¹ with fewer studies evaluating its metabolic DDI potential. ¹² , ¹³ Existing DDI models have generally relied on in vitro‐derived inhibition parameters, which may overpredict clinical DDI magnitude and fail to reconcile discrepancies between in vitro signals and observed clinical outcomes. ¹⁴ Consequently, uncertainty remains regarding which CBD‐mediated CYP450 inhibition signals are clinically meaningful in vivo. In this study, we addressed this gap using a learn–confirm–apply modeling strategy. First, we applied basic and mechanistic static models to screen and prioritize potential CYP450‐mediated DDI liabilities. Next, we developed and validated a CBD and 7‐OH‐CBD PBPK model and used clinical DDI data to refine inhibition parameters through reverse translation when direct in vitro–in vivo extrapolation (IVIVE) was inconsistent with observed outcomes. Finally, we applied the refined PBPK model to evaluate CBD‐mediated DDI risk with clinically relevant concomitant medications and in special populations that are often underrepresented in clinical DDI studies.

METHODS

The magnitude of CYP450‐mediated metabolic DDIs precipitated by CBD and 7‐OH‐CBD across different metabolic pathways was evaluated using a stepwise, model‐informed approach consisting of: (I) basic model, (II) MSM, and (III) dynamic PBPK model (Figure 1 ).

Figure 1.

Stepwise modeling workflow for CBD‐mediated DDI assessment. 7‐OH‐CBD, 7‐hydroxycannabidiol; AUCR, area under the plasma concentration–time curve ratio; CAF, caffeine; CBD, cannabidiol; CLB, clobazam; DDI, drug–drug interaction; IV, intravenous; MD, multiple dose; MDZ, midazolam; N‐CLB, N‐desmethylclobazam; PBPK, physiologically based pharmacokinetic; R ₁, reversible inhibition in the liver; R _1,gut, reversible inhibition in the gut; R ₂, time‐dependent inhibition; SD, single dose; STP, stiripentol.

Assessment of CBD‐mediated metabolic DDIs using the basic model and MSM

To parameterize the basic model, we conducted a comprehensive literature review to collect reversible inhibition (K_i) and time‐dependent inhibition (TDI) (K_I and k_inact) parameters for CBD and 7‐OH‐CBD from different in vitro systems. We selected the inhibition parameters reported by Bansal et al. ⁵ because their experimental conditions accounted for the extensive nonspecific binding of CBD and 7‐OH‐CBD to assay materials. Moreover, they measured the unbound fraction in the incubation (f_u,inc), enabling the correction of K_i and K_I values for binding (Table S1 ). Because published studies have reported conflicting evidence regarding CBD‐mediated TDI under different experimental conditions, ⁵ , ⁶ models with and without the reported TDI parameters were evaluated. We calculated R₁ and R_1,gut to assess reversible inhibition in the liver and gut, respectively, while R₂ was calculated to account for TDI, using the following equations:

$$R_1=\frac{{\mathrm{AUC}}_i}{\mathrm{AUC}}=1+\frac{{\left[I\right]}_{\max ,u}}{K_{i,u}}$$

$$R_{1,\mathrm{gut}}=1+\frac{{\left[I\right]}_{\mathrm{gut}}}{K_{i,u}}$$

$$R_2=\frac{{\mathrm{AUC}}_i}{\mathrm{AUC}}=\frac{1}{\frac{k_{\mathrm{deg}}}{k_{\mathit{deg}}+\frac{{\left[I\right]}_{\max ,u}*k_{\mathrm{\text{inact}}}}{{\left[I\right]}_{\max ,u}+K_{I,u}}}}$$

For the inhibitor concentration ([I]), we used the maximum unbound systemic concentration (C _max,u) observed following the 750 mg twice daily regimen used in clinical DDI studies. ¹⁵ Because CBD has poor aqueous solubility, we used its maximum solubility in gastrointestinal fluids as a surrogate for its intestinal luminal concentration ([I]_gut). ¹⁶ Following the FDA guidance, we considered an R ₁ ≥ 1.02, an R _1,gut ≥ 11, or an R ₂ ≥ 1.25 as indicators of potential in vivo inhibition of the respective metabolic pathway by CBD and 7‐OH‐CBD. ¹⁷ Enzymes that met any of these thresholds were further evaluated using the MSM.

For the MSM, we simulated the coadministration of CBD and 7‐OH‐CBD with sensitive substrates of major CYP450 enzymes: midazolam (CYP3A4), omeprazole (CYP2C19), theophylline (CYP1A2), ticlopidine (CYP2B6), dextromethorphan (CYP2D6), repaglinide (CYP2C8), and diclofenac (CYP2C9). The area under the curve ratio (AUCR) of sensitive substrates in the presence and absence of CBD and 7‐OH‐CBD was calculated using the following equation:

$$\text{AUCR}=\left(\frac{1}{\left[A_H\times B_H\right]\times f_m+\left(1-f_m\right)}\right)\times \left(\frac{1}{\left[A_G\times B_G\right]\times \left(1-F_G\right)+F_G}\right)$$

where A represents reversible inhibition, B represents TDI, H denotes the liver, and G denotes the gut. F _G is the fraction of the substrate drug escaping gut‐wall metabolism, and f _m is the fraction of the substrate drug metabolized by the inhibited enzyme. The MSM incorporates organ‐specific inhibitor concentrations at the enzyme site in the liver and gut. Because direct measurement is not clinically feasible, we estimated the in vivo unbound maximum hepatic inlet concentration of the inhibitor ([I] _h ) using the equation described by Ito et al., ¹⁸ and approximated the luminal concentration ([I]_gut) using CBD's maximum solubility in gastrointestinal fluids. ¹⁶ The parameters used in the MSM are summarized in Tables S1 and S2 . Based on the predicted AUCR, we classified the resulting DDIs according to the M12 guidance as weak (1.25 ≤ AUCR < 2), moderate (2 ≤ AUCR < 5), or strong (AUCR ≥ 5). ⁸

Development and validation of the CBD and 7‐OH‐CBD PBPK model

Using the Simcyp® Simulator v22 (Certara, Sheffield, UK), we developed the intravenous (IV) PBPK model for CBD in healthy volunteers to characterize its disposition following a 20 mg IV dose. ¹⁹ Depletion studies in human liver microsomes (HLMs) indicated that UDP‐glucuronosyltransferases (UGTs) and CYP450s contribute to approximately 80% and 20% of CBD metabolism, respectively. ¹² Using the reported f _m values ¹² and the plasma clearance after IV administration, ¹⁹ we derived enzyme‐specific intrinsic clearance (CL_int) values using the well‐stirred liver model in a middle‐out approach, whereby mechanistic in vitro metabolism data were integrated with clinically observed plasma clearance. Detailed calculations are provided in the supplementary material, and the final enzyme‐specific CL_int values are summarized in Table S3 . Furthermore, we predicted tissue‐to‐plasma partition coefficients (Kp) using the Poulin and Theil method, which was selected given the high lipophilicity of CBD (logP = 6.33) and reported limitations of the Rodgers and Rowland method for highly lipophilic compounds. ²⁰ Subsequently, we expanded the model to describe the pharmacokinetics of CBD and 7‐OH‐CBD following single and multiple oral dose administrations. We used the first‐order absorption model and estimated the fraction absorbed (f _a ) to capture the less‐than‐dose‐proportional increase in CBD exposure under fasted conditions and the higher exposure following multiple dosing and food administration (Table S4 ). For the fasted single‐dose condition, model development incorporated data across the 750–6,000 mg dose range ²¹ , ²² with external validation performed at 1500 mg ²¹ ; independent validation datasets were not available for the remaining dose levels. For the fasted multiple‐dose condition, we used the 1,500 mg dataset for model development ²¹ and the 750 mg dataset for validation. ²¹ Under fed conditions, we stratified doses into lower‐dose (200 mg) and higher‐dose (750–1,500 mg) groups, each characterized by a single f _a (Table S4 ). For the higher‐dose group, we used the 750 mg single‐dose dataset for model development ²² and independent 750 mg multiple‐dose ¹⁵ and 1,500 mg single‐dose ²¹ datasets for validation. For the lower‐dose group, we used the 200 mg single‐dose dataset ²³ for model development and an independent 200 mg single‐dose dataset for validation. ²⁴ Tables S5 and S6 summarize the clinical studies used for model development and validation, respectively.

For the 7‐OH‐CBD metabolite model, which is formed primarily by CYP2C9 and CYP2C19, ²⁵ we parametrized clearance using whole‐organ metabolic clearance. We scaled the in vitro CYP2C19 CL_int value for 7‐OH‐CBD ²⁵ by a factor of 0.3 to recapitulate the observed clinical data. Furthermore, using our PBPK model and Simcyp's virtual population libraries, we further validated the model by simulating two published studies: (i) a single 200 mg dose administration in adults with mild, moderate, and severe hepatic impairment, ²⁴ and (ii) multiple weight‐based doses (5–20 mg/kg/day) in pediatric patients with Dravet syndrome aged 4–10 years. ²⁶ In these populations, we did not scale enzyme‐specific CL_int values in our model because Simcyp back‐calculates them based on the enzyme abundance incorporated in the population tab. Model performance was deemed acceptable when predicted C _max and AUC values were within twofold of the observed summary‐level clinical data (defined as a predicted‐to‐observed ratio of 0.5–2.0). This twofold acceptance criterion was considered appropriate given the high pharmacokinetic variability of CBD exposure. ²⁷ Finally, using the validated PBPK models, we explored CBD exposure in special populations. We simulated the administration of 750 mg CBD twice daily with food for 30 days in geriatrics, obese adults, CYP2C19 poor metabolizers (PMs), and patients with hepatic impairment, for whom we also evaluated the reduced doses recommended in the CBD label (375 mg and 150 mg in moderate and severe hepatic impairment, respectively). ³

Assessment of CBD‐mediated metabolic DDIs using PBPK modeling

We incorporated in vitro‐derived reversible inhibition and TDI parameters for CBD and 7‐OH‐CBD across CYP450 enzymes into the PBPK model. ⁵ We also included the reversible inhibition parameters for CBD against multiple UGT enzymes to capture potential autoinhibition effects on CBD exposure. ²⁸ We simulated two DDI studies in which CBD was coadministered with midazolam (a sensitive CYP3A4 substrate) ²⁹ and caffeine (a sensitive CYP1A2 substrate) ³⁰ using the validated PBPK models available in the Simcyp compound library. In addition, we simulated a DDI study that evaluated the inhibitory potential of CBD on drugs commonly prescribed for childhood epilepsy. ¹⁵ Specifically, we assessed DDIs with clobazam (CLB), a CYP2C19 and CYP3A4 substrate, its active metabolite N‐desmethylclobazam (N‐CLB), a sensitive CYP2C19 substrate, and stiripentol (STP), which is predominantly metabolized by CYP1A2, CYP2C19, and CYP3A4. These assessments were performed using our in‐house developed STP, CLB, and N‐CLB PBPK models. ³¹ We prospectively assigned the clinical DDI datasets either to model training or model validation, as shown in Table 2 . The midazolam, caffeine, CLB, and N‐CLB DDI datasets constituted the training set, in which model performance was assessed by comparing predicted and observed AUC and C _max ratios, and, when necessary, CBD inhibition parameters were refined through reverse translation when the original in vitro parameters failed to recapitulate the clinical DDIs. The final inhibition parameters were then fixed, and the STP datasets were used for model validation. Predictions were considered acceptable when the predicted‐to‐observed AUC and C _max ratios fell within the bioequivalence limits (0.80–1.25), and the predicted AUC and C_max ratios ideally within the 90% confidence interval of the observed clinical values. The final optimized inhibition parameters were subsequently incorporated into the basic model and MSM to enable a consistent comparison of DDI risk across modeling approaches.

Table 2.

PBPK model–predicted and observed ¹⁵ , ²⁹ , ³⁰ AUC and C _max ratios for CBD‐mediated metabolic DDIs

Substrate	Role in PBPK	Parameter	GMR predicted	GMR Observed (90% CI)	Pred/Obs
Midazolam	Training	AUC0–inf ratio	1.04	0.92 (0.78, 1.09)	1.13
		C max ratio	1.01	0.80 (0.67, 0.96)	1.26
Caffeine	Training	AUC0–inf ratio	2.02	1.95 (1.62, 2.35)	1.04
		C max ratio	1.07	1.15 (1.04, 1.26)	0.93
Clobazam	Training	AUCτ ratio	1.20	1.21 (1.05, 1.39)	1.00
		C max ratio	1.16	1.2 (1.05, 1.38)	0.97
N‐CLB	Training	AUCτ ratio	2.84	3.38 (2.62, 4.36)	0.84
		C max ratio	2.84	3.39 (2.61, 4.39)	0.84
Stiripentol	Validation	AUCτ ratio	1.45	1.55 (1.42, 1.69)	0.93
		C max ratio	1.22	1.28 (1.08, 1.52)	0.95

AUC_0–inf ratio, area under the plasma concentration–time curve from time 0 extrapolated to infinity; AUC_τ ratio, area under the plasma concentration–time curve over the dosing interval; C _max, maximum plasma concentration; GMR, geometric mean ratio; 90% CI, 90% confidence interval; N‐CLB, N‐desmethylclobazam; Pred/Obs, predicted/observed.

DDIs classified as moderate or strong were further evaluated in special populations, and the predicted AUC ratios were compared with those in healthy volunteers to assess whether the DDI risk differed across populations. Furthermore, we explored the inhibitory potential of CBD on several drugs commonly used by patients taking CBD, for which clinical case reports have raised concerns about possible DDIs. To assess these interactions, we simulated 750 mg of CBD administered twice daily for 21 days with a single oral dose of each victim drug given on day 21. The victim drugs evaluated were clopidogrel (75 mg), citalopram (40 mg), escitalopram (10 mg), omeprazole (20 mg), lansoprazole (30 mg), and esomeprazole (20 mg), using either previously published PBPK models ³² , ³³ , ³⁴ or compound files from the Simcyp compound library. For warfarin, we used a multiple‐dose design, in which CBD was administered for 30 days, and warfarin (10 mg) was initiated on day 14 and continued until day 30 to account for its long half‐life and steady‐state kinetics. We conducted all simulations under fed conditions, except for proton pump inhibitors (PPIs). We simulated 10 trials, each including 10 subjects aged 20–50 years, with a 1:1 male‐to‐female ratio.

RESULTS

Assessment of CBD‐mediated metabolic DDIs using the basic model and MSM

Using the basic model, CBD, and to a lesser extent, 7‐OH‐CBD, had the potential to precipitate DDIs across all tested CYP450 enzymes through reversible inhibition and TDI. Further evaluation of DDI risk using the MSM revealed that CBD can precipitate moderate inhibition risk for CYP1A2, CYP2C9, and CYP3A4 and strong inhibition risk for CYP2C19. In contrast, 7‐OH‐CBD was predicted to pose a moderate inhibition risk for CYP2C19. The results of the basic model and MSM are presented in Table 1 .

Table 1.

CBD‐ and 7‐OH‐CBD‐mediated DDI risk assessment using the basic model and mechanistic static model (MSM)

Enzyme	Compound	Basic model			MSM
		R 1	R 1,gut	R 2	AUCR
CYP1A2	CBD	15.47			3.58
	7‐OH‐CBD	1.00			1.00
CYP2B6	CBD	1.69			1.47
	7‐OH‐CBD	1.03			1.04
CYP2C19	CBD	1.43	500.38	75.63	5.51
	7‐OH‐CBD	1.01		56.09	3.10
CYP2C8	CBD	1.25			1.14
	7‐OH‐CBD	1.02			1.02
CYP2C9	CBD	2.73			4.10
	7‐OH‐CBD	1.04			1.66
CYP2D6	CBD	1.29			1.26
	7‐OH‐CBD	1.01			1.01
CYP3A	CBD	1.39	445.44		2.22
	7‐OH‐CBD	1.01			1.71

Cutoff values for the basic model were defined per the FDA guidance as R ₁ ≥ 1.02, R _1,gut ≥ 11, or R ₂ ≥ 1.25. ¹⁷ AUCR was classified per M12 criteria as weak (1.25 ≤ AUCR < 2), moderate (2 ≤ AUCR < 5), or strong (AUCR ≥ 5) DDI risk. ⁸ All calculations were based on the final optimized inhibition parameters. R ₂ values for CYP1A2 and CYP3A4 were excluded following model refinement; for the remaining blank entries, R ₂ was not calculated because no TDI was detected in vitro. Values in bold indicate ratios exceeding the predefined DDI cutoffs or corresponding to moderate or strong DDI risk.

Development and validation of the CBD and 7‐OH‐CBD PBPK model

The final drug‐specific parameters for CBD and 7‐OH‐CBD used to develop the PBPK model are presented in Table S3 . The developed PBPK model successfully recapitulated the plasma concentration–time profile of CBD following IV administration, with predicted AUC in agreement with the observed value (ratio = 1.01; Figure S1 , Table S5 ). For oral administration, f_a was reduced with increasing single doses in the fasted state, from 0.45 at 750 mg to 0.19 at 6000 mg, to reflect the less‐than‐dose‐proportional increase in exposure. In contrast, f_a was increased following multiple dosing and under the fed state to reflect the higher exposure seen under these conditions (Table S4 ). The PBPK model successfully captured the systemic exposure of CBD and 7‐OH‐CBD following single and multiple oral doses in both fed and fasted conditions (Figure S2 ), as well as in patients with hepatic impairment (Figure S3 ) and pediatric patients with Dravet syndrome (Figure S4 ). Predicted C _max and AUC values of CBD and 7‐OH‐CBD fell consistently within the twofold range of the observed values (Figure 2 ). However, the model tended to overpredict CBD C _max under fasted conditions (predicted‐to‐observed ratios of 1.25–1.6), potentially reflecting the use of a fixed absorption rate constant (k _a  = 0.82 h⁻¹) across prandial states, which may not fully capture food‐related differences in CBD absorption kinetics.

Figure 2.

Predicted and observed ¹⁵ , ²¹ , ²² , ²³ , ²⁴ , ²⁶ exposure parameters of CBD and 7‐OH‐CBD following oral administration across various dosing regimens and populations. *Fasted arm from the food effect study. MD, multiple doses; SD, single dose.

Figure 3 shows the predicted CBD exposure across special populations. As a high‐extraction drug with extensive hepatic metabolism, CBD exposure was sensitive to changes in hepatic clearance. Accordingly, CBD AUC increased progressively with the severity of hepatic impairment, with predicted exposures approximately 2.1‐, 2.9‐, and 3.8‐fold greater in individuals with mild, moderate, and severe hepatic impairment, respectively, than in healthy adults (Table S7 ). The reduced 375 mg dose partially normalized exposure in moderate hepatic impairment, resulting in 1.4‐fold higher exposure than in healthy adults. Notably, the 150 mg dose recommended in severe hepatic impairment resulted in approximately 30% lower systemic exposure than that predicted in healthy adults receiving 750 mg twice daily. These simulations support dose reductions in moderate and severe hepatic impairment and suggest that dose reduction may also be warranted in mild hepatic impairment, depending on the target exposure and clinical context. CBD exposure was also approximately 1.5‐fold greater in geriatric individuals than in healthy adults, likely reflecting age‐related changes in body composition and metabolic capacity. In contrast, obesity and CYP2C19 poor metabolizer status resulted in CBD exposures approximately 0.8‐ and 1.0‐fold those of healthy adults, respectively, suggesting that dose adjustment is not necessary in these populations.

Figure 3.

CBD exposure (AUC) across healthy volunteers and special populations, summarized using trial‐level median AUC values (10 trials per group). Boxes represent the 25th–75th percentiles with the median line; whiskers indicate the minimum and maximum values. PMs, poor metabolizers.

Assessment of CBD‐mediated metabolic DDIs using PBPK modeling

Using the reported TDI parameters for CYP1A2 and CYP3A4 in HLMs, ⁵ , ¹⁶ the initial PBPK model predicted AUC ratios of 5.52 for midazolam and 7.23 for caffeine, corresponding to strong interactions that were not observed clinically. ²⁹ , ³⁰ Moreover, the model overpredicted the increase in CLB exposure (2.4‐fold compared with the clinically observed 1.21‐fold increase) and underpredicted the increase in N‐CLB exposure (1.3‐fold compared with the observed 3.4‐fold increase), ¹⁵ likely due to overprediction of CYP3A4 inhibition and consequent suppression of CYP3A4‐mediated conversion of CLB to N‐CLB. We therefore evaluated an alternative model excluding the CYP3A4 and CYP1A2 TDI components. In addition, the reversible CYP1A2 inhibition parameter was refined using a sensitivity analysis‐guided approach, reducing the total K_i from the in vitro value of 1.83 μM ⁵ to 0.08 μM to capture the clinically observed CBD–caffeine interaction. The impact of these refinements on predicted AUC ratios across the evaluated CYP450 substrates is shown in Figure S5 . The final PBPK model successfully captured the clinically observed DDI magnitude (Figures S6 and S7 ), predicting AUC ratios of 1.04 for midazolam, 2.02 for caffeine, 1.20 for CLB, 2.84 for N‐CLB, and 1.45 for STP (Table 2 ). Collectively, all predicted‐to‐observed AUC and C _max ratios fell within bioequivalence limits (0.80–1.25) and all the AUC and C_max ratios, except for midazolam, within the 90% confidence intervals of the observed clinical data. The predicted midazolam C _max ratio (1.01) was slightly above the observed upper 90% confidence limit of 0.96 (Table 2 ). Overall, model predictions were consistent with clinical observations, identifying CYP2C19‐mediated inhibition as the most clinically significant interaction. This interaction resulted in an approximately threefold increase in N‐CLB exposure and was classified as a moderate DDI risk according to the M12 guidance (2 ≤ AUCR < 5). Subsequently, we evaluated whether this DDI classification changed when CBD was coadministered with CLB across different special populations. Figure 4 shows the predicted change in N‐CLB exposure (AUC) in special populations relative to healthy volunteers. Although absolute CBD and N‐CLB exposures varied across populations, the magnitude of the interaction (AUCR) was comparable to or lower than that observed in healthy volunteers. For instance, the DDI magnitude was attenuated in patients with hepatic impairment, likely due to reduced CYP2C19 activity. Similarly, in obese individuals, lower CYP2C19 expression and an increased volume of distribution were associated with a reduced DDI magnitude. In CYP2C19 PMs, no interaction was predicted, consistent with the absence of functional enzyme activity. Overall, these simulations suggest that CBD‐mediated DDI risk on N‐CLB is unlikely to be exacerbated in these populations despite higher systemic exposure.

Figure 4.

Observed ¹⁵ and PBPK‐predicted AUC geometric mean ratios (AUCR) with 90% confidence intervals for CBD‐mediated DDIs on N‐CLB across special populations. The shaded yellow region denotes moderate drug interaction risk.

Figure 5 illustrates the predicted DDIs following coadministration of CBD with commonly used drugs. We predicted weak CBD‐mediated DDI risk for escitalopram and warfarin. In addition, CBD reduced the exposure of clopidogrel's active metabolite, clopidogrel thiol H4, by approximately 30% (AUC ratio = 0.70; Table S8 ), consistent with CYP2C19 inhibition affecting metabolic activation. Predicted interactions with PPIs varied according to their f _m by CYP2C19. Omeprazole, a highly sensitive CYP2C19 substrate, showed an approximately fourfold increase in exposure, whereas lansoprazole, a moderately sensitive substrate, showed a 3.2‐fold increase. Esomeprazole, which is not classified as a sensitive CYP2C19 substrate by the FDA, exhibited a 2.3‐fold increase in exposure.

Figure 5.

Predicted exposure (AUC) of commonly co‐prescribed drugs in the presence vs. absence of CBD, expressed as geometric mean ratios (GMRs) with 90% confidence intervals. The dashed vertical line indicates no interaction (GMR = 1). The shaded yellow region denotes moderate drug interaction risk.

DISCUSSION

As medical cannabis and CBD‐containing products become more widely used, CBD is increasingly administered to patients with substantial polypharmacy burden, including children with refractory epilepsy, older adults, and individuals with chronic comorbidities requiring long‐term pharmacotherapy. In this context, the central clinical pharmacology question is not whether CBD inhibits CYP450 enzymes in vitro, but which of these inhibition signals translate into clinically meaningful DDIs in vivo. Our results indicate that, under the evaluated clinical dosing scenarios, CBD‐mediated DDI risk is more selective than suggested by broad in vitro CYP450 inhibition. Across the clinical DDI datasets and PBPK simulations evaluated here, CYP2C19 inhibition emerged as the most consistent clinically relevant liability. Interaction through CYP3A4 was not clinically meaningful despite notable in vitro inhibition signals, whereas the twofold increase in caffeine exposure indicates a borderline‐moderate but potentially clinically relevant CYP1A2 interaction. These observations are in line with prior reports showing that in vitro‐derived DDI parameters do not always translate directly to the in vivo setting and are routinely refined through a middle‐out approach using clinical data. ³⁵ , ³⁶ , ³⁷ , ³⁸ Therefore, for CBD, our findings support a focused risk‐assessment strategy prioritizing sensitive CYP2C19 substrates, CYP2C19‐dependent prodrugs and active metabolites, and CYP1A2 substrates with a narrow therapeutic index, rather than assuming equivalent clinical concern across all CYP450 pathways flagged by in vitro assays.

This work establishes a translational framework for CBD DDI risk assessment by integrating in vitro inhibition data, static modeling, PBPK simulations, and clinical DDI observations. Previously published PBPK models of CBD and 7‐OH‐CBD have primarily focused on CBD disposition, including enzyme contributions to clearance, ¹¹ and route‐ or absorption‐dependent pharmacokinetics. ⁹ , ¹⁰ A prior CBD perpetrator DDI model incorporated in vitro‐derived TDI parameters for CYP1A2, CYP3A4, and CYP2C19, ¹² but overpredicted the magnitude of several clinical DDIs. ¹⁴ For example, following multiple‐dose CBD administration, predicted AUC ratios for CYP1A2, CYP2C19, and CYP3A4 substrates were substantially higher than clinical observations with caffeine, N‐CLB, and midazolam. ¹⁵ , ²⁹ , ³⁰ These discrepancies highlight the limitations of directly applying in vitro inhibition parameters without clinical calibration. While this approach is widely used for CYP450 induction, where in vitro parameters are calibrated against a well‐characterized reference inducer such as rifampicin to control for in vitro data variability, ³⁹ no equivalent strategy is available for CYP450 inhibition. Therefore, the present study uses clinical DDI data with sensitive probe substrates to evaluate and, when necessary, refine in vitro‐derived inhibition parameters.

The HLM derived TDI parameters for CYP1A2 and CYP3A4 substantially overpredicted the clinically observed DDIs, and excluding these two components markedly improved agreement with the clinical data. This discrepancy may be partly explained by the ability of CBD to induce CYP1A2 and CYP3A4 in vitro, ⁴⁰ , ⁴¹ potentially offsetting TDI during repeated dosing. Consistent with this hypothesis, Balhara et al. showed that incorporating both CYP3A4 induction and inactivation into a CBD PBPK model resulted in no net change in hepatic intrinsic clearance, helping explain the absence of a clinically meaningful midazolam interaction after repeated CBD dosing. ⁴⁰ In addition, discrepancies between in vitro and in vivo TDI are biologically plausible because HLM systems may not fully recapitulate hepatocellular processes that mitigate enzyme inactivation in vivo, including conjugative metabolism and detoxification of reactive intermediates. ³⁷

The three DDI modeling approaches used in this research differed substantially in predictive performance. The basic model flagged potential inhibition across all tested CYP450 enzymes, consistent with its intended role as a conservative screening tool designed to minimize false negatives. ⁴² The MSM provided additional mechanistic context by incorporating organ‐specific inhibitor concentrations and substrate‐specific f _m values, but still overpredicted DDI risk for CYP1A2, CYP2C19, CYP2C9, and CYP3A4, leading to risk classifications that could prompt unnecessary follow‐up studies (Figure S8 ). In contrast, the PBPK model incorporated time‐varying inhibitor and substrate concentrations, repeated‐dose conditions, first‐pass exposure, and population‐specific physiology, which better reconciled in vitro inhibition signals with clinical observations. This progressive improvement in predictive capacity across the basic, MSM, and PBPK analyses is in agreement with prior systematic comparisons showing that PBPK models predict clinical DDI magnitude with higher accuracy than static approaches. ⁴² , ⁴³

We applied our validated PBPK model to clinically untested scenarios, including special populations and coadministration with commonly prescribed medications. Although CBD exposure increased in hepatic impairment and geriatric populations, the relative magnitude of the CBD–N‐CLB interaction was not proportionally increased. This likely reflects reduced CYP2C19 activity in these populations, which limits the additional effect of enzyme inhibition on N‐CLB clearance. Across concomitant medications, predicted DDI magnitude depended on the contribution of CYP2C19 to drug disposition and on whether the exposure change was likely to be clinically consequential. The model did not predict a clinically meaningful interaction with citalopram or escitalopram, consistent with clinical evidence showing only modest increases in steady‐state concentrations during CBD coadministration. ⁴⁴ Furthermore, despite the moderate CYP2C9 interaction risk predicted by the MSM, the PBPK model predicted a weak interaction with warfarin, with a 25% increase in exposure, consistent with the lack of a clinically meaningful interaction reported for tolbutamide, a CYP2C9 substrate, in the CBD prescribing information. ³ For clopidogrel, CBD reduced exposure to the active metabolite clopidogrel thiol H4 by approximately 30%. This prediction supports the CBD prescribing information, which notes that concomitant CBD administration may decrease therapeutic efficacy and recommends considering a dosage increase. ³ Finally, the predicted increases in exposure for omeprazole (4‐fold), lansoprazole (3.2‐fold), and esomeprazole (2.3‐fold) support a moderate CYP2C19‐mediated interaction risk, consistent with the CBD prescribing information, which identifies CBD as a moderate CYP2C19 inhibitor and recommends considering dose modification of concomitant CYP2C19 substrates. ³

The broader set of interactions flagged in the CBD prescribing information reflects a precautionary approach that incorporates both in vitro and clinical evidence. The label recommends dose adjustment for substrates of CYP2B6, CYP2C8, and UGT1A9; however, several of these recommendations are based primarily on in vitro inhibition data rather than clinically demonstrated interactions and are therefore written with conditional language such as “may be observed” and “as clinically appropriate.” Clinically, no meaningful changes in exposure were observed with the UGT1A9 substrate mycophenolate mofetil or the CYP2B6 substrate bupropion, while no dedicated clinical DDI study has evaluated CYP2C8 substrates. ³ In contrast, the available clinical evidence, supported by our PBPK model predictions, indicates clinically relevant CYP1A2‐ and CYP2C19‐mediated interactions. CBD increased caffeine exposure approximately twofold (AUCR = 1.95), supporting the label recommendation for caution with CYP1A2 substrates for which small increases in exposure may lead to serious adverse reactions (e.g., theophylline). For CYP2C19, the model further contextualizes the well‐established CBD–CLB interaction observed during CBD development. While our simulations do not assess the relative antiseizure contributions of CBD and N‐CLB, they demonstrate that increased N‐CLB exposure is a predictable consequence of CBD‐mediated CYP2C19 inhibition, with minimal effect on parent CLB. This interaction should be considered when interpreting efficacy and safety outcomes during CBD and CLB coadministration.

We acknowledge some limitations in the current study. First, CBD absorption is complex and influenced by solubility‐limited dissolution, formulation, dose, prandial state, and meal composition. In the present model, absorption was represented using the first‐order absorption model with dose‐, food‐, and dosing‐regimen‐dependent adjustments to the f_a. While this approach adequately captured CBD systemic exposure, independent datasets were unavailable to validate several dose levels under fasted conditions. Nevertheless, the model was validated under fed conditions, consistent with the available clinical DDI studies, and was therefore used for prospective simulations. Moreover, the first‐order absorption model cannot mechanistically distinguish between low‐ and high‐fat meal conditions or explicitly resolve dissolution‐ and solubility‐limited absorption. A more mechanistic ADAM model may better describe these processes; however, the data needed for robust parameterization were limited. Second, the approximately twofold increase in CBD exposure following repeated dosing ²¹ was captured empirically by increasing f_a. Although the model accounts for UGT autoinhibition through the inclusion of reversible inhibition parameters against multiple UGT enzymes, ²⁸ a recent study reported at least 15‐fold greater inhibitory potency of CBD toward UGT1A9 and UGT2B7, ⁴⁵ which account for the majority of CBD metabolism. ¹² Thus, the empirical increase in f_a may compensate for underestimated UGT‐mediated autoinhibition in the current model. Third, our DDI assessment focused on CYP450‐mediated interactions and did not incorporate P‐gp inhibition. Although the CBD label indicates that CBD and 7‐OH‐CBD are not anticipated to interact with P‐gp, intestinal P‐gp inhibition cannot be excluded. ³ Clinically, CBD increased exposure to the dual CYP3A/P‐gp substrates everolimus and tacrolimus by approximately 2.5‐ and 3.1‐fold, respectively. ⁴⁶ , ⁴⁷ These interactions may also partly reflect CYP3A5 inhibition, which is not represented in our model and is more potently inhibited by CBD than CYP3A4 in vitro. ⁴⁸ Thus, our model may underestimate interactions involving CYP3A5 and P‐gp substrates. Finally, the PBPK model was developed and validated using clinical studies conducted at relatively high CBD doses compared with many nonprescription or recreational CBD exposures. ¹⁶ Although clinically meaningful DDIs are expected to be less likely at lower doses, extrapolation to low‐dose CBD products remains uncertain. This uncertainty reflects differences in product composition, formulation, purity, prandial conditions, and oral bioavailability. Therefore, predictions for low‐dose or nonprescription CBD products should be interpreted cautiously until additional pharmacokinetic and DDI data become available.

FUNDING

This work was funded by a grant from the State of Florida Consortium for Medical Marijuana Clinical Outcomes Research.

CONFLICT OF INTEREST

The authors declared no competing interests for this work.

AUTHOR CONTRIBUTIONS

S.A.A., B.E., B.C., and R.C. wrote the manuscript. S.A.A., B.E., and R.C. designed the research. S.A.A. and B.E. performed the research. S.A.A., B.E., B.C., and R.C. analyzed the data.

Supporting information

Table S1.

DATA AVAILABILITY STATEMENT

The Simcyp project file supporting this study is available from the corresponding author upon request.