Clozapine (CLO) is an atypical antipsychotic drug with proven efficacy in treatment-resistant schizophrenia (Kane et al., 1988; Siskind et al., 2016). Numerous studies have shown that clinical response to CLO is closely related to its blood levels, which better reflects the patient’s actual exposure to the drug than the daily dose (Wagner et al., 2021; Tralongo et al., 2023; Schoretsanitis et al., 2020). Despite clear recommendations for the use of therapeutic drug monitoring (TDM) to optimize CLO treatment (Tralongo et al., 2023; Schoretsanitis et al., 2020), in practice, therapeutic decisions are often based solely on dosage and clinical assessment (Schoretsanitis et al., 2020).

It is generally accepted that there is a linear relationship between the administered dose of CLO and its concentration in the blood of a given patient (Jiménez-Fernández et al., 2024; de Leon et al., 2020; Flanagan et al., 2023a). This assumption reflects the first-order kinetics that characterize most psychotropic agents, enabling predictable titration and the clinical use of concentration-to-dose ratios to guide individualized dosing. When linearity holds, stable metabolic conditions ensure proportional increases in drug exposure with dose escalation (Hiemke et al., 2018). However, our clinical observations indicate that the actual relationship between CLO dose and concentration is often much more complex. In everyday practice, we regularly observe cases of “unexpected” CLO concentrations that do not fit a simple linear model. In some patients, a significant increase in drug dose does not result in a proportional increase in concentration; on the contrary, a plateau or even a decrease in drug levels is observed in some individuals. In other cases, even small changes in dose lead to unpredictable concentration values, both too low and dangerously high. “Unexpected” or non-linear CLO dose-concentration relationship has been suggested by other authors (Flanagan et al., 2023a; Alfaro et al., 2001; Couchman et al., 2010; Chang et al., 1997). “Unexpectedly” low clozapine levels (200–250 ng/mL) were detected in two male, medication-compliant inpatients receiving CLO doses as high as 800–1,000 mg/day for at least 2 months (Alfaro et al., 2001). In contrast, “unanticipated”, high CLO concentrations were observed in 1.2% of samples from a TDM service with CLO levels >1,000 ng/mL at prescribed doses up to 150 mg/day, with no clues for overcompliance (Couchman et al., 2010). Such deviations from linearity carry important clinical implications. Subproportional increases in exposure may contribute to under-treatment, whereas supraproportional increases can produce disproportionate rises in blood levels and elevate the risk of concentration-dependent toxicity (de Leon et al., 2020; Couchman et al., 2010).

Based on literature data and our own observations, we assumed that the problem of difficult-to-interpret CLO concentrations exists in real-life clinical practice and cannot be explained by well-established variables associated with CLO concentrations. It is likely to occur more frequently at higher doses of the drug, but it cannot be ruled out that this phenomenon also occurs at lower doses (de Leon et al., 2020; Flanagan et al., 2023a; Hiemke et al., 2018).

The aim of the present study was two-fold. First, we reviewed our in-hospital database for representative cases of “unanticipated” changes in CLO concentrations during dose adjustment that did not conform to the assumption of within-subject linearity. Second, we conducted an exploratory analysis using the Multivariate Adaptive Regression Splines (MARS) (Friedman, 1991) to evaluate the assumption of linearity in the single-time-point, between-subject CLO dose-concentration association and to identify potential breakpoints [for similar biomedical applications of the MARS, see (Rockwood et al., 2017; Park and Kim, 2018; Lu et al., 2021; Yuan et al., 2025)]. This analysis was designed to reflect real-life conditions in which therapeutic decisions are made without access to prior TDM results for individual patients.

In the present study, the representative cases of individual patients with difficult-to-interpret CLO levels were identified in order to validate previous case reports and to illustrate within-subject deviations from linearity (Alfaro et al., 2001; Couchman et al., 2010). The MARS models were also used for the identification of non-linear CLO dose-concentration relationships in a cross-sectional, between-subject setting mimicking real-life challenges of CLO dose adjustments without TDM data. To the best of our knowledge, this is the first study employing the MARS to analyse the CLO dose-concentration relationship and to identify non-linear associations in real-life clinical practice.

Moderate predictive performance was obtained for the MARS models with the R ² value of 0.25 for CLO and 0.23 for NCLO. The analyses included variables with well-documented relevance for CLO metabolism, such as age, sex, CLO dose, smoking status, and the impact of concomitant medications other than CLO (β-blockers or other psychotropic drugs) (Flanagan et al., 2023c; Albitar et al., 2020; Singh et al., 2015; de Leon et al., 2025). As might be expected, the strongest predictors in both models were CLO dose and smoking status. Despite incorporating the well-established set of clinical variables (Flanagan et al., 2023c; Albitar et al., 2020; Singh et al., 2015; de Leon et al., 2025), a significant proportion of the variance in CLO concentrations remained unexplained. This finding indicates that even the well-established variables may not predict CLO concentrations with reasonable accuracy in some real-life TDM datasets.

The main finding of the MARS modeling was the identification of a non-linear relationship between CLO dose and serum concentration, along with the detection of a provisional hinge point. The hinge point was estimated at approximately 400 mg/day, suggesting that within the moderate-to-high dose range, increases in CLO dose are associated with a roughly linear rise in serum concentration (∼1.5 ng/mL per 1 mg of drug). Below this hinge point, the relationship weakens, and CLO levels become less predictable on the basis of CLO dose.

It is accepted that the relationship between the dose of CLO and its blood concentration is mostly linear, which forms the basis for estimating the individual optimal dose (de Leon et al., 2020). This means that in the absence of changes in factors affecting metabolism (i.e., without the presence of inducers or inhibitors), the concentration-to-dose ratio (C/D ratio) remains constant for a given patient (de Leon et al., 2022). However, the literature data are not consistent as to the extent to which such linearity indeed occurs (de Leon et al., 2025; Cheng et al., 1988; Choc et al., 1990; Schoretsanitis et al., 2021). In one of the early pharmacokinetic studies, Guitton et al. suggested that CLO exhibits linear kinetics in the concentration range of 145–1,411 ng/mL (Guitton et al., 1998). Other researchers have suggested that the relationship between CLO dose and concentration can be considered stable, provided that two criteria are met - no changes in factors that induce or inhibit metabolism and a CLO concentration above 150 ng/mL (Jiménez-Fernández et al., 2024). International guidelines also assume a linear relationship between CLO dose and concentration within the therapeutic range of CLO doses and in the absence of metabolism inducers or inhibitors (de Leon et al., 2022). It is worth noting that the range of CLO therapeutic doses may vary depending on ethnic origin or metabolic rate (de Leon et al., 2022; EMEA, 2002; Schoretsanitis et al., 2024). A recent experts’ opinion indicates that linear kinetics most often apply to the CLO concentration range of 150–1,000 ng/mL, if there is no influence of inducers or inhibitors (de Leon et al., 2025). Our results suggest the loss of linearity at lower doses with a critical value of approximately 400 mg/day. In clinical practice, the pharmacokinetics of the drug at lower doses may become less predictable and even a small modification of the dose may lead to a disproportionate change in drug concentration. Our finding may be related to the postulated non-linear increase in the risk of side effects (Couchman et al., 2010; Kikuchi et al., 2025), especially in the early phase of CLO dose adjustment. Previous studies have also shown non-linear pharmacokinetics at high CLO concentrations (>1,000 ng/mL) at which CLO metabolism is likely to approach saturation (de Leon et al., 2025). We did not observe a loss of linearity at high doses, which may be due to the smaller number of samples with high CLO concentrations (>1,000 ng/mL) in our database.

The loss of linear pharmacokinetics at lower CLO doses could be explained by several mutually non-exclusive factors, including complex metabolic pathways of the drug. Clozapine is metabolised in the liver, mainly by demethylation to NCLO and oxidation to N-oxide clozapine (Schoretsanitis et al., 2019). Cytochromes P450 isoenzymes, including CYP1A2, CYP3A4, CYP2C9, CYP2C19, and CYP2D6, play a key role in CLO demethylation (de Leon et al., 2022; Bertilsson et al., 1994; Fang et al., 1998). However, literature data are inconsistent regarding the percentage contribution of individual CYP enzymes to CLO metabolism (Thorn et al., 2018; Olesen and Linnet, 2001). The contribution of CYP isoforms varies depending on the concentration of the substrate. At low CLO concentrations, CYP1A2 is estimated to account for 70% of CLO metabolism, CYP2C19 and CYP3A4 for 20%–25%, with a small contribution from other CYP isoenzymes. However, at high drug concentrations, the contribution of cytochromes to CLO metabolism shifts in favour of CYP3A4 and CYP2C19 (Olesen and Linnet, 2001). The predominance of highly inducible CYP1A2 at lower CLO concentrations may underlie the greater variability and lack of linear association between CLO dose and concentration observed for lower doses. For example, it is possible that even minor day-to-day fluctuations in the number of cigarettes smoked and/or changes in the depth of inhalation of smoke could increase this variability by altering the degree of enzyme induction (Tralongo et al., 2023; Flanagan et al., 2023a).

The MARS model also revealed a non-linear dose-concentration relationship for NCLO, although with a different pattern. The hinge point was located at slightly higher doses (approximately 500 mg/day) beyond which the NCLO concentration became less dependent on the CLO dose, possibly reflecting saturation of the CYP1A2 pathways responsible for the formation of the metabolite. NCLO has some pharmacological activity (Costa-Dookhan et al., 2020; Park et al., 2020; Jessurun et al., 2022) and may contribute to the occurrence of adverse reactions associated with CLO (Schoretsanitis et al., 2019; De Leon et al., 2003). For this reason, the non-linear relationship between CLO dose and NCLO concentration may make monitoring of treatment safety and predicting the risk of CLO side effects more difficult.

The present study has several limitations concerning both the case selection and the between-subject analysis. With regard to the case selection (the within-subject approach), although we made every effort to eliminate the most evident sources of laboratory error and the most plausible clinical explanations for the difficult-to-interpret cases of a non-linear clozapine dose–concentration relationship (Table 1), the selected cases may still reflect unrecognized sources of variability. These include potential issues related to blood sample handling, undiagnosed somatic conditions, missed doses of CLO or other medications, day-to-day fluctuations in the number of cigarettes smoked, variations in the depth of smoke inhalation, and - last but not least - various combinations of these and other unidentified biological or organizational factors. Because this study relied on a retrospective dataset, we were unable to simply repeat the measurements, which would be the most effective strategy for verifying “unexpected” results and therefore represents an important clue for future prospective research.

Regarding the between-subject analysis, our methodological choice of using MARS, carries both strengths and drawbacks. While MARS effectively captures non-linear patterns with high interpretability, its reliance on piecewise linear approximations may oversimplify more complex pharmacokinetic relationships. Moreover, model’s predictive performance was only moderate, suggesting that important sources of variability were missing. These may include not incorporated variables such as genetic factors (e.g., CYP polymorphisms), body composition, active inflammation, pharmacokinetic interactions (e.g., with caffeine or other dietary components), and subtle hepatic or renal dysfunction (Flanagan et al., 2023a; Diaz et al., 2018; Ammar et al., 2021; Alli-Balogun et al., 2023; Pardiñas et al., 2019; Ghoneim and Mansour, 2020; Hägg et al., 2000). Variables with potentially higher predictive power - such as genetic data - are, however, rarely available at the point of care. The moderate performance may also reflect the fact that some variables were coded only in binary form, which reduces granularity. For instance, the substantial difference between smoking one versus twenty cigarettes per day could not be represented. Likewise, we did not incorporate specific drug types (beyond two broad categories) or individual doses due to substantial heterogeneity. Finally, the lack of external validation of the model on non-training data prevents a robust assessment of generalizability. The retrospective design and single center setting further restrict external validity. Although our in-hospital TDM database reflects some aspects of real-world clinical practice, validation in larger, multicenter cohorts remains necessary. Importantly, our goal was not to build the most accurate or broadly generalizable model but to explore whether non-linear patterns emerge in the CLO dose-concentration relationship when examined alongside known, clinically accessible moderators in our sample. Accordingly, the presented model and hinge-point localization should be interpreted as preliminary indications of potential non-linearity rather than clinically applicable predictive tools.

In conclusion, the identified cases and the MARS analysis suggest that the non-linear relationship between CLO daily dose and its concentration can be a real-life clinical problem with CLO doses of approximately 400 mg/day as the provisional hinge point. Although preliminary, the present results warrant further investigations on non-linear aspects of CLO pharmacology, including “unexpected” CLO concentrations, toxicity, and lack of therapeutic activity.