Effect of Cumulative Exposure on the Efficacy of Paroxetine: A Population Pharmacokinetic‐Pharmacodynamic and Machine Learning Analyses

ABSTRACT

Selective serotonin reuptake inhibitors (SSRIs) are widely used in depression treatment. However, the relationship between treatment efficacy and plasma concentrations remains unclear. We assessed whether the anti‐depressive response can be predicted based on the pharmacokinetic (PK) data of paroxetine, a frequently used SSRI. During treatment, we measured the plasma paroxetine concentrations in 179 paroxetine‐treated patients with major depressive disorder. Of these patients, 50 patients had received a pre‐treatment personality assessment using the Temperament and Character Inventory at baseline, and their depression severity was assessed using the Montgomery‐Asberg Depression Rating Scale (MADRS) at baseline and 1, 2, 4, and 6 weeks after treatment initiation. We conducted population PK modeling followed by a population PK‐pharmacodynamic (popPK/PD) model to analyze the enhancement in depression severity until 6 weeks of paroxetine treatment using nonlinear mixed‐effects modeling. Additionally, we developed machine learning models to predict the likelihood of remission after 6 weeks. The contribution of each feature to the prediction was explained using SHapley Additive exPlanations (SHAP) values. The area under the plasma paroxetine concentration‐time curve during the first week (AUC_0–1week) and MADRS score after 1 week of treatment (MADRS_W1) were incorporated into the popPK/PD model. The SHAP values indicated that the AUC_0–1week and MADRS_W1 were the significant predictors of remission. Our results indicate that therapeutic responsiveness to paroxetine can be anticipated from its cumulative exposure, highlighting the clinical relevance of assessing SSRI blood concentrations.

Summary.

• What is the current knowledge on the topic?

  • ○
    The relationship between SSRI blood concentrations and clinical efficacy remains unclear, and it is unclear whether treatment response can be adequately predicted based on these concentrations.
• What question did this study address?

  • ○
    We aimed to develop models that predict responses to paroxetine while considering the effects of individual factors such as patient personality, which may be associated with paroxetine PK and PD.
• What does this study add to our knowledge?

  • ○
    This study demonstrated that the anticipated improvement rate in depression severity with paroxetine treatment was higher in patients with higher total exposure up to the first week of treatment, whereas the onset time of treatment efficacy was earlier in patients with lower depression severity at week 1.
• How might this change drug discovery, development, and/or therapeutics?

  • ○
    This study presents novel findings in clinical pharmacology and emphasizes the potential of personalized antidepressant therapy based on measuring blood drug concentrations, including setting the dose and predicting non‐responders. Our modeling approach can be applied to other antidepressants, which may contribute to the selection of treatment agents.

1. Introduction

Major depressive disorder is common and often recurrent, affecting over 200 million people globally. It severely limits psychosocial functioning and reduces the quality of life [1]. Depression is closely associated with increasing suicide rates that have become a social issue [1, 2]. Treatment with antidepressants—specifically selective serotonin reuptake inhibitors (SSRIs)—is an essential option for patients with moderate or severe depression. However, patients often experience delayed benefits of treatment that contribute to lower treatment retention rates [3]. SSRIs require at least 4 weeks to develop therapeutic effects, and over 30% of patients do not respond to the treatment [4]. Therefore, predicting patient responses to antidepressants and the timing of their responses is crucial to enhance their long‐term prognosis.

Therapeutic drug monitoring (TDM) may represent a useful tool to identify patients who will benefit from an increased or decreased antidepressant dosage [5, 6]. However, the relationship between blood concentration and clinical efficacy has only been demonstrated for older antidepressants, such as tricyclic antidepressants, with limited data available for SSRIs [7]. No linear association between blood concentration and clinical efficacy was observed for SSRIs [8, 9]. The latest consensus guidelines for TDM outline a therapeutic reference range for antidepressants, but do not recommend routine TDM for SSRIs, except for citalopram [6].

Paroxetine, a frequently used SSRI, is the most potent serotonin reuptake blocker clinically available, and it exhibits lower selectivity for the serotonin reuptake site than either fluvoxamine or sertraline [10]. Paroxetine is one of the most commonly prescribed medications, with a broad spectrum of applications, such as in major depressive disorder, panic disorder, obsessive‐compulsive disorder, social anxiety disorder, generalized anxiety disorder, and post‐traumatic stress disorder. We previously performed a population pharmacokinetic (PK) analysis of paroxetine in Japanese patients with depression. CYP2D6*10/*10 genotype, age, sex, and body weight are associated with large interpatient variability in the plasma paroxetine concentrations [11]. However, large individual variability exists in the PK and pharmacodynamics (PD) of SSRIs, and current blood paroxetine concentrations are not adequately predictive of treatment response [6, 12, 13].

The personality traits of patients with depression may be associated with both antidepressant efficacy and genetic factors. Specific temperament traits in the Temperament and Character Inventory (TCI) are associated with the activity of neurotransmitter systems [14, 15]. Numerous studies have reported associations between TCI characteristics and therapeutic response or clinical course in patients with depression [16, 17]. Our group reported that responders to paroxetine 1 week after treatment initiation exhibited low harm avoidance, high self‐directedness scores, and a significant negative association between paroxetine plasma concentration and enhancement rate in depressive symptoms [17]. This indicates that patients with specific personality traits may benefit from a lower dose of paroxetine. However, to establish a personality trait‐based antidepressant dosing design, more detailed PK‐PD relationships that incorporate personality traits should be assessed.

Recently, various approaches have been reported to predict treatment responsiveness to antidepressants [18, 19], with machine learning (ML) emerging as a promising approach [20, 21]. Recently, the ML approach using PK data was used to identify patients at risk of adverse effects from escitalopram and sertraline [22]. However, there are limited data that assess a PK‐PD relationship to predict antidepressant responsiveness, and no consensus has confirmed that blood concentrations of SSRIs can predict efficacy [7].

In this study, we assessed whether blood concentration can predict the treatment responsiveness of paroxetine—a frequently used SSRI—using population PK‐PD analyses and ML techniques while carefully considering other patient factors, such as personality traits.

2. Methods

2.1. Patients and Study Protocol

This retrospective study included Japanese patients diagnosed with a major depressive disorder based on the Diagnostic and Statistical Manual of Mental Disorders, Fourth Edition, at Hirosaki University Hospital (Hirosaki, Japan) or Dokkyo Medical University School of Medicine (Shimotsuga Tochigi, Japan). In this study, we included patients from a previous study [11] who met the following criteria—have not received any antidepressants before starting paroxetine treatment; received a stable paroxetine dose for ≥ 9 days (based on its half‐life [−21 h] in humans); not taking any drugs that may alter the clearance of paroxetine; normal renal and hepatic functions; and the availability of detailed medical data. An initial dose of 10–20 mg/day of paroxetine (Paxil; GlaxoSmithKline, Tokyo, Japan) was administered, followed by weekly increases of 10 mg/day (once daily) to the maximum tolerated dose. During the paroxetine treatment, blood samples (10 mL) were collected. The plasma paroxetine concentrations were measured using high‐performance liquid chromatography, as previously reported [23]. Any suspected non‐adherence was determined by assessing whether the plasma concentrations were below the limit of quantification and whether the medications were taken based on pill counts, and patients suspected of being non‐adherent were excluded from the study. This study was approved by the Ethics Committee of the Hirosaki University Graduate School of Medicine and the Clinical Research Review Committee of Dokkyo Medical University Hospital. Detailed methodological procedures are described in the Data S1.

2.2. Clinical Data

Demographics and clinical data were retrospectively collected from the patient's medical records. Depression severity was assessed using the Montgomery‐Asberg Depression Rating Scale (MADRS) [24] at baseline and 1, 2, 4, and 6 weeks after starting paroxetine treatment. Physicians assessed patients through interviews, rating each of the 10 items on a scale of 0–6 points. “Remission” was defined as a MADRS score < 10 [25]. The patient's personality was assessed using the TCI [14] before starting paroxetine treatment. The TCI consisted of 240 questions, each in a “yes” or “no” format. Each patient's personality was scored on seven items, including four temperaments (novelty seeking, harm avoidance, reward dependence, and persistence) and three character dimensions (self‐directedness, cooperativeness, and self‐transcendence).

2.3. Population Model Development

Population‐based modeling was conducted using a nonlinear mixed‐effects model approach in NONMEM (version 7.5.1; ICON plc, Dublin, Ireland) with the first‐order conditional estimation with interaction method. For population PK (popPK) modeling, one‐ and two‐compartment models were assessed, with and without the absorption process and saturation of metabolism, respectively. The one‐compartment model—that does not account for the absorption processes and saturation of metabolism—best fits the data. To accurately estimate the PK parameters, the volume of distribution (V_d/F) was fixed at 1010 L, obtained by multiplying the mean body weight (kg) of this study population with the previously reported V_d/F value of 17.2 L/kg [26]. In the subsequent population PK‐PD (popPK/PD) analysis, the enhancement rate in depression severity with paroxetine treatment was predicted. This analysis included patients who had MADRS scores at baseline (MADRS_W0) and 1 (MADRS_W1), 2, 4, and 6 weeks after starting paroxetine treatment. The enhancement rate in depression severity was calculated as the percentage reduction in the MADRS score from the baseline. The E_max model visually best described the relationship between the treatment duration and enhancement rate, which is represented by the following equation:

$$\mathrm{EFF}=\frac{{\mathrm{E}}_{\max }\times \text{Time}}{{\mathrm{ET}}_{50}+\text{Time}}$$

where EFF represents the enhancement rate, E_max represents the maximum enhancement rate, Time represents the treatment duration, and ET₅₀ represents the treatment duration corresponding to half of E_max.

The Text S1 presents the procedure of covariate analysis and the method for assessing model appropriateness. The NONMEM model code is provided in the Texts S2 and S3.

2.4. ML Approach

We aimed to develop ML models to predict whether patients can be in remission at week 6 of paroxetine treatment. The patients were assigned to the remission and non‐remission groups based on MADRS scores at week 6. Random forest–recursive feature elimination (RF‐RFE) was performed for feature selection, and followed by predicted remission with Logistic Regression (LR), Decision Tree (DT), K‐Nearest Neighbor (KNN), Support Vector Machine (SVM), Random Forest (RF), and Multi‐Layer Perceptron (MLP) models. Additionally, we assessed the significance of the features using SHapley Additive exPlanations (SHAP) [27]. The detailed model development procedure and Python code for developing ML models are described in Texts S1 and S4, respectively.

3. Results

3.1. Patient Demographics

Of the 327 patients willing to participate in this study, 179 met the inclusion criteria and were included in the popPK analyses. Among them, 50 patients with available MADRS scores were included in the popPK/PD and ML analyses. Tables 1 and 2 summarize the characteristics of the patients involved in the popPK and popPK/PD analyses.

TABLE 1.

Summary of the characteristics of 179 patients (PK population).

Characteristics	Number (%)/mean ± SD (range)
Sex, male	73 (40.8%)
Age (years)	41.5 ± 14.2 (16–80)
Body weight (kg)	58.6 ± 10.5 (40–85)
Paroxetine dose (mg/day)	22.3 ± 11.1 (10–40)
Number of samples	1.8 ± 0.99 (1–5)
Plasma paroxetine concentration (ng/mL)	51.9 ± 49.0 (2.6–286.5)
CYP2D6 genotype
PMs	0 (0%)
IMs	121 (67.6%)
EMs	58 (32.4%)

Note: We designated CYP2D6*1 and *2 as functional alleles, *10 and *41 as reduced functional alleles, and *5 as a non‐functional allele. PMs are defined as patients with two non‐functional alleles. IMs are defined as patients with two reduced functional alleles, one functional and one non‐functional allele, or one reduced functional and one non‐functional allele. EMs are defined as patients with two functional alleles or one functional and one reduced functional allele.

Abbreviations: CYP2D6, cytochrome P450 2D6; EMs, extensive metabolizers; IMs, intermediate metabolizers; PK, pharmacokinetic; PMs, poor metabolizers.

TABLE 2.

Characteristics of 50 patients included in the popPK/PD analysis.

Characteristics	Number (%)/mean ± SD (range)
Sex, male	19 (38.0%)
Age (years)	46.6 ± 13.7 (20–70)
Body weight (kg)	56.3 ± 10.4 (40–85)
Daylight hours (h)	120.6 ± 43.4 (54.3–173.2)
Paroxetine dose (mg/day)	32.2 ± 11.0 (10–40)
MADRS score
Baseline	39.6 ± 9.7 (12–53)
Week 1	28.6 ± 11.3 (4–56)
Week 2	21.1 ± 12.0 (0–49)
Week 4	16.8 ± 12.3 (0–44)
Week 6	13.7 ± 12.9 (0–46)
TCI scale
Novelty seeking	16.9 ± 4.4 (5–28)
Harm avoidance	27.9 ± 4.1 (18–35)
Reward dependence	13.7 ± 3.3 (7–20)
Persistence	3.6 ± 1.7 (1–8)
Self‐directedness	20.3 ± 6.3 (7–39)
Cooperativeness	26.3 ± 4.2 (18–36)
Self‐transcendence	10.7 ± 5.3 (3–24)

Abbreviations: MADRS, Montgomery‐Asberg Depression Rating Scale; popPK/PD, population pharmacokinetic‐pharmacodynamic; TCI, temperament and character inventory.

3.2. PopPK Modeling

A total of 329 steady‐state concentrations from 179 patients were used for the popPK analysis. The inter‐individual variability in clearance (CL) and residual variability were best represented by the exponential and proportional error models, respectively. Table S1 summarizes the effects of covariates on the objective function. Age, sex, and body weight significantly reduced the objective function value of CL. Owing to the correlation between sex and body weight—with the mean body weight greater in males—CL was described only by age and body weight. In contrast, none of the factors had a statistically significant effect on V_d/F (Table S1). The final popPK model is as follows:

$$\mathrm{CL}\left(\mathrm{L}/\mathrm{h}\right)=23.6\times {\left(\frac{\mathrm{Age}}{41.5}\right)}^{-0.536}\times {\left(\frac{\mathrm{BW}}{58.6}\right)}^{0.563}$$

$${\mathrm{V}}_{\mathrm{d}}/\mathrm{F}\left(\mathrm{L}\right)=1010$$

where BW represents the total body weight (kg).

The goodness‐of‐fit (GOF) plots demonstrated correlations between predicted and observed concentrations (Figure S1). The conditional weighted residuals were primarily distributed within ±2 and exhibited a homogeneous distribution around zero across time. All 1000 bootstrap runs converged successfully. All estimated parameters were within the 95% confidence intervals (CIs) obtained from the bootstrap analysis, confirming the model validity (Table S2). The shrinkage values for the inter‐individual variability of CL and residual variability were below 20%–30%, which are generally acceptable [28]. Additionally, we confirmed that the trend in paroxetine concentration was adequately described through simulations using visual predictive checks (VPC) (Figure S2).

3.3. PopPK/PD Modeling

We performed popPK/PD modeling of the paroxetine treatment duration and enhancement rate in depression severity. The mean ± standard deviation (SD) of the enhancement rate at weeks 1, 2, 4, and 6 was 25.9 ± 30.3, 45.7 ± 27.4, 57.5 ± 28.3, and 64.2% ± 30.6%, respectively. The proportional error model best accounted for the inter‐individual errors of E_max and ET₅₀, whereas the residual error was best with an additional error model. Among the paroxetine concentration‐related indices examined, we observed that the area under the plasma paroxetine concentration‐time curve from treatment initiation to a specific time point (AUC_0–t) was significantly associated with the MADRS score at that point. Therefore, the paroxetine AUC corresponding to the time point of MADRS assessment was calculated for each patient, and its effect on the enhancement rate was assessed, along with other individual factors. Subsequently, the AUC_0–1week significantly affected E_max, and MADRS_W1 affected ET₅₀ (Table S3). The final popPK/PD model is as follows:

$$\mathrm{EFF}=\frac{{\mathrm{E}}_{\max }\times \text{Time}}{{\mathrm{ET}}_{50}+\text{Time}}$$

$${\mathrm{E}}_{\max }=90.9-{\left(\frac{{\mathrm{AUC}}_{0–1\text{week}}}{2610.43}\right)}^{-8.56}$$

$${\mathrm{ET}}_{50}=0.984+{\left(\frac{{\text{MADRS}}_{\mathrm{w}1}}{28.64}\right)}^{2.38}$$

Figure 1 illustrates the time course of the enhancement rate as explained by this model. The high AUC_0–1week was associated with an increased E_max (Figure 1A), and the low MADRS_W1 was associated with an early enhancement (Figure 1B).

FIGURE 1.

Relationship between the treatment duration and enhancement rate in depression severity based on the population pharmacokinetic‐pharmacodynamic model. (A) The relationships between the treatment duration and the enhancement rate in severity are represented as a solid curve for patients with a high value of AUC_0–1week, and are presented as a dotted curve for average patients. (B) The relationships between the treatment duration and the enhancement rate in severity are represented as a solid curve for patients with a low MADRS score at week 1, and are presented as a dotted curve for average patients. AUC, area under the plasma paroxetine concentration‐time curve; AUC_0–1week, AUC during the first 1 week of treatment; MADRS, Montgomery‐Asberg Depression Rating Scale; E_max, maximum enhancement rate in depression severity; ET₅₀, treatment duration corresponding to 50% of the E_max.

The GOF plots and VPC demonstrated that the final model accurately represents the patient data (Figures S3 and S4). Among 1000 bootstrap runs, 937 successfully converged and were included in the bootstrap analysis. All estimated parameters of the final model were within the 95% CIs, confirming the model validity (Table S4). The shrinkage values were generally reasonable, despite slightly higher than 30% for ET₅₀, suggesting that parameter estimates were relatively reliable. Figure 2 shows receiver operating characteristic (ROC) curves for the enhancement rate and presence or absence of remission at week 6. The popPK/PD model achieved the area under the ROC curve (ROC‐AUC) of 0.973 (95% CI: 0.931–1), accuracy of 0.880, sensitivity of 0.808, and specificity of 0.958 for the prediction of remission at week 6.

FIGURE 2.

ROC curve for the predicted enhancement rate and presence or absence of remission at week 6 of treatment. The solid line represents the ROC result curve of the final popPK/PD model. The dotted line represents the reference line. ROC, receiver operating characteristic; popPK/PD, population pharmacokinetic‐pharmacodynamic.

3.4. ML Approach

The remission and non‐remission groups included 26 and 24 patients, respectively. Compared to the remission group, the non‐remission group had a higher proportion of men, body weight, and MADRS_W1, and lower AUC_0–1week (p < 0.1) (Table S5). The RF‐RFE algorithm selected four significant predictors: AUC_0–1week, AUC_0–2weeks, MADRS_W1, and TCI score for reward dependence. SVM and RF exhibited superior performance with 0.761 ROC‐AUC and 0.720 accuracy, respectively (Table 3). According to the SHAP values based on SVM and RF, the most effective feature was MADRS _W1, followed by AUC_0–1week (Figure 3A,C). The SHAP values indicated that patients with a higher AUC_0–1week were more likely to achieve remission at week 6 (Figure 3B,D).

TABLE 3.

Predictive performance of ML models.

Algorithm	ROC‐AUC (95% CI)	Accuracy	Sensitivity	Specificity
LR	0.699 (0.552–0.845)	0.580	0.538	0.625
DT	0.662 (0.512–0.812)	0.540	0.549	0.542
KNN	0.592 (0.429–0.756)	0.560	0.577	0.542
SVM	0.761 (0.621–0.902)	0.680	0.923	0.417
RF	0.702 (0.553–0.851)	0.720	0.769	0.667
MLP	0.660 (0.503–0.818)	0.660	0.769	0.542

Abbreviations: CI, confidence interval; DT, decision tree; KNN, K‐nearest neighbor; LR, logistic regression; ML, machine learning; MLP, multi‐layer perceptron; RF, random forest; ROC‐AUC, area under the receiver operating characteristic curve; SVM, support vector machine.

FIGURE 3.

The interpretation of SVM and RF based on the SHAP values. (A) The significance ranking of the features based on the mean absolute SHAP values in SVM. (B) SHAP beeswarm summary plot for the effect of features on the output of SVM. (C) The significance ranking of the features based on the mean absolute SHAP values in RF. (D) SHAP beeswarm summary plot for the effect of features on the output of RF. SVM, support vector machine; RF, random forest; SHAP, SHapley Additive exPlanations; MADRS, Montgomery‐Asberg Depression Rating Scale; MADRS_W1, MADRS score at 1 week after starting paroxetine treatment; AUC, area under the plasma paroxetine concentration‐time curve; AUC_0–1week, AUC during the first 1 week of treatment; AUC_0–2weeks, AUC during the first 2 weeks of treatment; RD, reward dependence.

4. Discussion

This study demonstrated that cumulative exposure to paroxetine can predict treatment responsiveness through population‐based modeling and ML approaches. We observed that the time course of the enhancement rate in depression severity was significantly affected by AUC_0–1week, in conjunction with MADRS_W1.

In this study, we used two approaches—popPK/PD analysis and ML techniques—to predict treatment responsiveness to paroxetine. Recently, ML algorithms have become promising and innovative strategies for pharmacometrics modeling and have been applied variously in the context of pharmacometrics. For example, the ML approach would facilitate the selection of PK model structures [29] and covariate selection in pharmacometrics modeling [30]. In the present study, the validity of the findings obtained through the popPK/PD analysis that the therapeutic efficacy of paroxetine could be predicted based on AUC_0–1week and MADRS_W1 was confirmed using the ML approach. Among ML algorithms, SVM showed the highest AUC, 0.761, but it was still lower than the performance of the popPK/PD model. At present, we do not have a clear explanation for the potential reason why SVM has the highest predictive performance. The prediction accuracy of the SVM was high for sensitivity but low for specificity (Table 3), so the rate of false positives for predicting remission was high, which is thought to be one of the practical limitations in clinical settings. Furthermore, the possibility of model overfitting cannot be completely denied due to the small sample size, even though the model performance was carefully evaluated using the nested cross‐validation. Therefore, increasing the number of subjects and improving the prediction performance is necessary.

SSRIs containing paroxetine are believed to inhibit the serotonin transporter at the presynaptic axon terminal of serotonergic neurons, thereby preventing serotonin reuptake, increasing serotonin concentration in the synaptic cleft, and producing antidepressant effects. SSRIs often take > 4 weeks to develop clinical efficacy owing to the desensitization of 5‐hydroxytryptamine 1A (5‐HT_1A) autoreceptors in the dendritic region of serotonergic nerves [31, 32]. The 5‐HT_1A receptor is one of the receptors that regulate the serotonergic nerve. Repeated SSRI treatment reduces the functional sensitivity of 5‐HT_1A autoreceptors. The delay in the therapeutic onset of SSRIs may represent the time required for autoreceptor desensitization that results in greater serotonin availability in the synapse [31]. Additionally, the time lag in the clinical response, which typically occurs after SSRI treatment initiation, is considered as the gradual accumulation of SSRIs in the brain over time and the relatively low affinity of SSRIs towards the BDNF receptor, TRKB [33]. Recent experiments using HEK cells have reported that fluoxetine, an SSRI, directly binds to the transmembrane region of the tyrosine kinase receptor 2 (TRKB), facilitates its transport to the cell surface, and activates the BDNF signaling pathway [34]. Additionally, chronic administration of antidepressants containing paroxetine may require the drug concentration in the brain to reach a threshold level in the micromolar range [35], and SSRIs gradually accumulate in the brain and reach a plateau after several weeks of treatment [36]. These insights suggest that the clinical response is only achieved when the drug reaches a brain concentration high enough to interact with a low‐affinity binding target, such as TRKB [34]. We speculate that AUC, which reflects the time course of blood concentration, may be more associated with the degree of accumulation in the brain than cross‐sectional information, such as trough concentrations and daily dose, and that AUC_0–1week may reflect the brain concentration high enough starting to interact with TRKB, although the exact mechanism remains unknown. Our findings regarding AUC_0–1week may be associated with the presumed mechanisms of action of SSRIs stated above and suggest that they are not specific to the paroxetine dose regimen but apply to SSRIs in general. Nevertheless, further studies are required to confirm the relationship between cumulative exposure to paroxetine for 1 week and the mechanisms of action.

The TCI scores did not affect the time course of the enhancement rate in depression severity in our popPK/PD model. In a previous study, early responders to paroxetine treatment showed lower harm avoidance scores and higher self‐directedness scores than nonresponders and late responders [17]. Indeed, MADRS_W1 scores were positively associated with harm avoidance scores and negatively associated with self‐directedness scores in the present study (data not shown). This led to the possibility that the impacts of self‐directedness and harm avoidance on treatment outcomes were not apparent. On the other hand, SVM indicated that reward dependence scores may be positively related to the likelihood of remission after 6 weeks (Figure 3B). Reward dependence quantifies individual differences in the extent to which a person is sociable, approval seeking, and warm versus aloof, detached, and cold [14]. Although it has been reported that low reward dependence scores may have a negative impact on the course and outcome of major depressive disorder [37], the underlying mechanisms are unclear. Therefore, further studies are needed to investigate the reason(s) why reward dependence was associated with the treatment outcome in patients treated with paroxetine.

Additionally, the results indicated that patients with lower MADRS_W1 may achieve an earlier enhancement in depression severity with paroxetine. A previous study on the timing of fluoxetine effect onset reported that 55.5% of patients who responded to fluoxetine at week 8 of treatment began to respond to antidepressants by week 2 [38]. In contrast, patients who did not respond by weeks 4–6 were less likely to react by week 8 [38]. Therefore, incorporating the early MADRS score into the popPK/PD model appears to be reasonable. We speculated that high enough concentrations of paroxetine lead to lower MADRS_W1, which may achieve an earlier enhancement in depression severity with paroxetine. However, further studies are required to verify the underlying pharmacological and pathophysiological mechanisms of the association between depression severity in the first week and the treatment response to paroxetine.

The popPK analysis showed that body weight may affect the CL of paroxetine. Since paroxetine is lipophilic and widely distributed throughout the body [10], it is likely to accumulate in the adipose tissue of obese patients, and the elimination of paroxetine may be delayed. Conversely, obesity may increase the activity of CYP2D6 [39], and may also be associated with increased paroxetine metabolism. Therefore, there are individual differences in the effect of obesity on the CL of paroxetine, and the small sample size of the present study potentially resulted in a high relative standard error and wide 95% CI.

There was no observed effect of CYP2D6 genotype on PK parameters in the present study. Our previously reported Michaelis–Menten elimination kinetics model demonstrated that the presence of one CYP2D6*5 allele or any CYP2D6*10 allele may have no significant effect on paroxetine PKs in the steady state [11]. In contrast, CYP2D6 genotypes were associated with blood paroxetine concentrations [23, 40]. Although we could not assess the effect of CYP2D6 null alleles because of their extremely low frequency in East Asians, it is possible that differences in the metabolic capacity of paroxetine were not observed in this study because numerous patients were treated at relatively low doses and few experienced metabolic saturation. However, these genotypes may affect the therapeutic efficacy of paroxetine, which requires further research.

We suggest that AUC_0–1week should be estimated using the popPK model before treatment initiation, and the dose should be increased in the first week while paying careful attention to side effects for patients who are expected to show insufficient treatment response. Furthermore, our findings may facilitate the determination of whether to continue paroxetine treatment or switch to different treatments after 1 week of treatment initiation. It would be difficult to lead a response‐driven titration regimen based on depression severity 1 week after treatment initiation with our findings. However, AUC_0–2weeks may also affect remission after 6 weeks (Figure 3B,D), and further investigation would be necessary to propose a more detailed dosing regimen that includes after the first week.

This study has certain limitations. First, we focused on paroxetine, and therefore, we should confirm whether the findings can be generalized to other SSRIs. We could not detect CYP2D6 allelic duplications, which may cause different metabolizer phenotypes [41]; therefore, further study is required to investigate the effect of CYP2D6 allelic duplications on clinical outcomes under paroxetine treatment. Additionally, because we did not assess any pharmacogenetic biomarkers of the therapeutic effect of paroxetine, it is necessary to assess the effects of other polymorphisms that may be associated with the PK and PD of SSRIs [42, 43]. Meanwhile, the absorption process was excluded and the Vd/F value of paroxetine was fixed in the present study, since sparse trough concentration data may lead to inaccurate estimation. We examined whether age, gender, and weight affect the Vd/F, but we found no significant effect of any of the factors (Table S1). Therefore, we should examine the covariates for each parameter in detail in the future, using information that can also consider the absorption process with a larger number of patients. However, we applied a Bayesian approach with prior information from literature‐based models and confirmed that our popPK model showed equivalent or even better performance based on the GOF plots and VPC (Figures S5 and S6). Finally, although this observational study aimed to accurately reflect the clinical conditions, it involved a relatively small cohort of Japanese participants and the generalizability of our findings would be limited. Many patient factors may have impacts on the PK and PD of paroxetine. Paroxetine undergoes significant first‐pass metabolism, and it is catalyzed largely, but not exclusively, by CYP2D6 [10, 44]. Large inter‐ethnic differences exist in the CYP2D6 allele frequencies [45, 46], which may be attributed to inter‐population variability in paroxetine metabolism and treatment responses. Meanwhile, there are significant differences in not only the PK but also the treatment response of paroxetine between younger and older patients with major depressive disorders [10, 47]. Moreover, we did not include any patients with impaired liver or kidney function, in whom the treatment responses to paroxetine may differ from those in our study subjects. Therefore, we should first verify that the study findings are not specific to the subjects of this study in retrospective studies for other patient groups with comparable patient backgrounds, and prospective validations in a large and diverse population are required to verify whether our findings can be generalized to other populations.

Despite these limitations, our results offer clinically significant insights into the potential of assessing the blood concentration of SSRIs to predict their therapeutic efficacy. This study demonstrated that treatment response to paroxetine can be predicted by assessing cumulative exposure along with depressive symptoms early in treatment. Although our findings require further assessment, they may help in proposing treatment strategies for patients with major depressive disorder.

Author Contributions

K.S., T.E., T.T., H.M., A.K.‐M., K.O., N.Y.‐F., and J.S. wrote the manuscript; K.S., T.E., T.T., A.K.‐M., K.O., N.Y.‐F., and J.S. designed the research; K.S., T.E., T.T., N.H., K.I., K.M., M.N., T.K., H.M., K.D.Y., A.K.‐M., K.O., N.Y.‐F., and J.S. performed the research; K.S., T.E., T.T., N.H., K.I., K.M., M.N., T.K., N.Y.‐F., and J.S. analyzed the data.

Conflicts of Interest

The authors declare no conflicts of interest.

Supporting information

Data S1.