Lower daily rhythm strength in hair cells is linked to irregular sleep in schizophrenia

Sixty-six participants were enrolled and assigned to one of three groups: (1) individuals with schizophrenia and obesity (SCH-OB), (2) individuals with obesity but no psychiatric diagnosis and no psychotropic medication (OB), and (3) normal weight individuals without psychiatric diagnosis and without psychotropic medication use (LEAN). The OB group served as an obesity-only comparator, allowing obesity-associated differences (OB vs LEAN) to be distinguished from schizophrenia-with-obesity differences (SCH-OB vs OB and SCH-OB vs LEAN). All groups were matched based on age (±5 years) and sex. SCH-OB and OB were additionally group-wise matched on BMI (±3 kg/m²) and OSA status. The study was preregistered on ClinicalTrials.gov (Identifier: NCT05413486↗), which originally included a fourth, unmatched comparison group of individuals with bipolar disorder and obesity. However, due to recruitment constraints, the final study design focused on the three matched groups described above.

Procedures for hair follicle sampling, RNA extraction, and cDNA synthesis were conducted as described in our published protocol [21]. Briefly, total RNA was extracted from plucked beard or scalp hair follicles collected at four time points (0, 6, 12 h after waking, and just before bed) across two consecutive days using a RNeasy Micro kit (QIAGEN, Hilden, Germany) according to the manufacturer’s instructions. The RNA concentration was quantified using a NanoDrop spectrophotometer (Thermo Fisher Scientific, Waltham, MA) after which RNA was reverse transcribed to cDNA using a high-capacity cDNA reverse transcription kit (Applied Biosystems, Foster City, CA) according to the manufacturer’s protocol. The qPCR was performed on a QuantStudio 5 real-time PCR system (Applied Biosystems, Foster City, CA) using pre-designed, commercially available TaqMan Custom Arrays against six core circadian clock genes PER2 (Hs01007553_m1), PER3 (Hs00213466_m1), NR1D1 (Hs00253876_m1), NR1D2 (Hs00233309_m1), BMAL1 (Hs00154147_m1), and RORA (Hs00536545_m1) (all from Applied Biosystems, Foster City, CA). Gene expression was normalized to the geometric mean of two reference genes, RPLP0 and TBP, selected based on expression stability across time points and study groups. All samples were run in triplicates, and the relative expression was calculated using the 2^–ΔCt method [22].

Saliva samples were handled as described, and in accordance with our established procedure for home-sampling [21, 23]. Briefly, samples were collected at waking, 6 and 12 h after waking, every 2 h thereafter, 1 h before bed, and just before bed across two consecutive days. Immediately after collection, samples were refrigerated by participants and subsequently centrifuged within 72 h of sampling time. They were then stored at –80 °C until analysis. Cortisol levels were quantified using a chemiluminescence immunoassay (IDS-iSYS, Immunodiagnostic Systems, Tyne & Wear, United Kingdom; IS-4900), while melatonin concentrations were measured via ELISA (IBL International, Hamburg, Germany; RE54041). Internal quality control measures were in place and the inter-assay coefficient of variation (CV) for the melatonin assay was calculated at two levels to be 24% (2.3 ng/L) and 28% (11.8 ng/L) with an average intra-assay CV of 18% (at the 2.3 ng/L level). For the cortisol assay, the intra-assay CV was maximum 15% at three different levels (4.7, 12, and 48 nmol/L). When sample volume was insufficient for both hormonal analyses, cortisol measurement was prioritized. Salivary pH was assessed using a standard pH electrode. Additional statistics and information regarding cortisol and melatonin testing are available in Supplementary Figs. S7-9.

Sleep quality was assessed using the Pittsburgh Sleep Quality Index (PSQI) [24], and chronotype the Morningness–Eveningness Questionnaire (MEQ) [25]. Anthropometric measurements were collected at the first clinical visit as previously described [21].

Sleep variables were derived from wrist-worn accelerometer raw-data using R [26] and the open-source R package GGIR [27], developed by Van Hees et al. [28, 29]. For this study, accelerometry was used to calculate the Sleep Regularity Index (SRI), sleep duration, sleep onset, and sleep offset [30]. All data were processed using default GGIR parameters and summary variables were averaged across available valid days.

All laboratory tests and accelerometer-based assessments were conducted by personnel blinded to group membership.

To determine whether physiological variables and gene expression exhibited 24-hour rhythmicity, we applied linear mixed-effects cosinor models. For each outcome, we fitted two models: a nested model including only a random effect for subject (ID), and a full cosinor model with fixed-effect sine and cosine terms and a random intercept for subject ID. Likelihood ratio tests (LRTs) were used to compare the models. A significant LRT p-value indicated that the inclusion of rhythmic components improved model fit, consistent with the presence of a 24-hour sinusoidal rhythm and supporting the appropriateness of cosinor-based regression [31]. Changes in Akaike’s Information Criterion (ΔAIC = AIC_{(null model)}- AIC_{(full model)}) were used as a secondary indicator to quantify improvements in model fit [32]. To test for group differences in clock gene amplitude (SCH-OB, OB, LEAN), we applied mixed-effects cosinor models using the cosinor.lm function from the cosinoRmixedeffects R package [33]. These models included sine and cosine terms as fixed effects, subject ID as random effects and group as fixed effect. Post-hoc contrasts were used to estimate differences in amplitude, MESOR (midline estimating statistic of rhythm) and time of peak expression (acrophase) between groups. Pairwise comparisons were Tukey-adjusted to control the family-wise error rate within each outcome/parameter. Significance was assessed at α = 0.05. 95% CIs for model-derived parameters were obtained via parametric bootstrapping (n = 1000). At the individual level, we applied standard cosinor linear regression models with sine and cosine terms, and used F-tests to assess whether the rhythmic terms jointly contributed to model fit [31]. To examine whether sleep metrics or psychotropic medication dose (expressed as chlorpromazine equivalents (CPZeq) [34]) were associated with modulation of clock gene amplitude, we constructed a separate set of models in which the cosinor framework was extended to include interaction terms between PSQI, SRI, CPZeq and the harmonic predictors (sine and cosine terms). These models tested whether PSQI, SRI, or CPZeq modulated the amplitude of clock gene expression. For each gene, model-based estimates of amplitude were computed across the observed range of PSQI or SRI values. 95% CI’s were obtained via parametric bootstrapping (n = 1000), and p-values for the slope of the amplitude~predictor association were calculated from the bootstrap distribution and corrected for multiple comparisons [35]. Marginal and conditional R² values were computed using the MuMIn R package to summarize the proportion of variance explained by fixed effects alone and by the full model (including random effects), respectively. To deal with missing data, mixed-effects models incorporate all observed data, enhancing the robustness of the estimates [33]. All statistics were conducted in R version 4.3.1.

Participants in the SCH-OB group showed lower amplitude of PER3 compared to both OB and LEAN groups (Fig. 2A, rightmost panel). OB had higher PER2 amplitude than SCH-OB (Fig. 2B, rightmost panel), while LEAN showed a higher amplitude of BMAL1 compared to OB (Fig. 2E, rightmost panels). No significant group differences were observed in MESOR or time of peak expression for any of the tested clock genes.

Oral pH and temperature showed 24-hour rhythmicity. pH MESOR was highest in LEAN, intermediate in OB, and lowest in SCH-OB, with all between-group differences significant (Fig. 3C). Oral temperature peaked 8–12 h after sleep offset, earlier in LEAN than SCH-OB, with no significant group differences in amplitude or MESOR (Fig. 3D).

Participants in the SCH-OB group reported significantly worse sleep quality than both the OB and LEAN groups, as measured by the Pittsburgh Sleep Quality Index (PSQI).

One SCH-OB participant was diagnosed with OSA post-enrollment, resulting in a minor deviation from the intended OSA matching; however, groups did not differ significantly in OSA severity based on continuous Apnea–Hypopnea Index (AHI) (Fig. 4J).

To further contextualize these individual-level detection rates and the accuracy of phase estimation at the individual level, we performed simulation-based phase-precision analyses calibrated to the observed PER3 amplitude and within-person residual variability. Overall, results were consistent with phase precision being primarily driven by the empirical amplitude-to-noise relationship, with only modest gains from extending the number of sampling days at 4 samples/day (Supplementary Figs.,). S10 S11

We investigated whether clock gene amplitude varied as a function of SRI, subjective sleep quality (PSQI), and antipsychotic medication dose, calculated as chlorpromazine equivalent units (CPZeq) [34] (Supplementary Table S3).

A significant positive association was observed between SRI and PER3 amplitude (p = 0.006), which remained significant after correction for multiple comparisons (adjusted p = 0.030). A nominal association was observed between PSQI and BMAL1 amplitude (p = 0.032), though this did not survive correction for multiple testing (Supplementary Fig. S12). No associations were found between CPZeq and the amplitude of any clock gene.

Clock gene expression profiles derived from hair follicles have mainly been used to determine phase shifts in relation to sleep disorders [38], shift work [19, 39] and chronotype [40, 41]. The PER3 gene has emerged as the most utilized and reliable marker of circadian phase [42]. Supporting this, Watanabe et al. [43] identified PER3 as the clock gene exhibiting rhythmicity most consistently across individuals, and notably found no significant PER3 phase difference when comparing its expression in leukocytes and hair follicle cells. This cross-tissue concordance supports the validity of using hair follicle–derived PER3 rhythms as a potential marker of systemic circadian timing [43].

While PER3 is frequently studied for its phase properties, its amplitude, reflecting the strength of its rhythmic expression, has received comparatively little attention. One exception is Takahashi et al., who found that the amounts of physical activity and peak oxygen uptake in older adults were positively correlated with PER3 amplitude, suggesting that physical activity may help sustain, or even improve, peripheral clock rhythms [41]. In our study, PER3 amplitude was significantly reduced in individuals with schizophrenia and showed a positive association with the SRI, indicating that irregular rest-activity patterns may contribute to blunted rhythmic output. This interpretation is consistent with findings from Wulff et al. [44], who observed severe circadian and sleep-wake disruption in individuals with schizophrenia, including highly variable sleep timing and desynchronized melatonin rhythms in a substantial subset of individuals. Notably, Wulff et al. restricted their sample to unemployed individuals, indicating that marked irregularity in sleep timing can persist even in the absence of employment-related scheduling constraints. In a supplementary sensitivity analysis, adding employment status as a covariate did not improve model fit (LRT p = 0.54). Taken together, these results point to PER3 amplitude as a peripheral clock readout of sleep–wake regularity, with particular relevance in psychiatric populations where sleep–wake schedules are often more irregular [44]. In contrast, subjective sleep quality did not show any association with PER3 amplitude, suggesting that objective measures of sleep–wake regularity may be more closely aligned with molecular rhythm parameters than perceived sleep quality. This may indicate that PER3 amplitude relates more to sleep regulation and behavioral timing stability than to subjective sleep disturbance, consistent with PER3 polymorphism studies linking PER3 variation to homeostatic sleep markers without clear differences in melatonin/cortisol rhythms or peripheral PER3 phase [45].

While the OB group showed higher PER2 amplitude than the SCH-OB group and lower BMAL1 amplitude than the LEAN group, these isolated differences did not mirror the broader PER3 amplitude attenuation observed in the SCH-OB group. This distinction suggests that schizophrenia is associated with additional alterations in clock gene rhythm parameters beyond the effects of metabolic burden alone. In line with this, model-based estimates suggested earlier melatonin peak timing in the OB group than in SCH-OB (relative to sleep offset), while actigraphy showed delayed sleep offset in the SCH-OB group compared with both control groups, pointing to a broader desynchronization of behavioral and hormonal rhythms in schizophrenia. In mice, quetiapine or ziprasidone treatment have been shown to increase PER2 and BMAL1 expression [46, 47]. When controlling for the use of antipsychotic medication, we found no indication that psychotropic medication use in the SCH-OB group attenuated PER3 amplitude. This indicates that the blunted rhythmicity could be ascribed to the disease itself or to life-style factors rather than to the use of antipsychotic medication. Supporting this interpretation, Sun et al. (2016) reported persistently reduced PER3 amplitude in leukocytes of patients with schizophrenia, which remained unchanged following 8 weeks of clozapine treatment despite marked symptomatic improvement [14].

Salivary melatonin and cortisol rhythms were largely preserved across groups, except for a modest reduction in melatonin amplitude in the SCH-OB group. This suggests that molecular markers of circadian rhythms may be more vulnerable or differently regulated than systemic hormonal outputs. Prior studies have reported highly inconsistent melatonin profiles in schizophrenia, including elevated, reduced, and unchanged levels compared to control individuals [48], highlighting the clinical and biological heterogeneity of the disorder and limiting the utility of melatonin as a uniform circadian marker. Similarly, while salivary cortisol holds value as a noninvasive indicator of hypothalamic-pituitary-adrenal (HPA) axis function [49], elevated levels are more often associated with vulnerability traits and may emerge prior to symptom onset rather than as a stable feature of schizophrenia itself [50]. The absence of inter-group cortisol alterations in the present study supports the view that HPA axis dysregulation, as assessed by salivary cortisol, is not an inherent or persistent feature of schizophrenia [51].

We included oral temperature and pH measures based on prior evidence suggesting that both exhibit 24-hour rhythmic patterns and may reflect underlying physiological processes, [52–54]. In addition to their biological relevance, these measures were chosen for their low cost, minimal participant burden, and ease of integration into field-based investigations. To support adherence to the manufacturer’s recommended passive saliva collection time, we selected a digital thermometer with a measurement duration that naturally occupied the same interval, helping to ensure sufficient saliva for analysis when these were measured concurrently [21].

Lower oral pH has previously been observed in individuals with mental disorders [55]. Our findings align with this but also suggest that obesity may exert a similar contributing effect. The consistently low pH observed across groups with elevated BMI highlights a potential metabolic influence on oral acid–base balance that remains underexplored [56]. This is particularly relevant given that poor oral health is more common among people with mental disorders and is also linked to increased risk of cardiovascular, respiratory, and metabolic diseases — primarily through shared vulnerabilities such as health-detrimental behaviors, chronic inflammation, and systemic immune activation [57]. If oral pH is used as a screening tool for assessing oral health, or for identifying health conditions that influence or are influenced by oral health, our and other studies indicate the importance of adjusting for circadian variation in oral pH [52–54].

Adherence to the sampling protocol was high across groups, and the RNA yield was consistently sufficient for qPCR analysis, even in participants with schizophrenia. This shows the method is feasible in vulnerable populations. While overall compliance was strong, participant feedback pointed to sampling-associated discomfort and practical challenges related to repeated hair follicle sampling, particularly over consecutive days. Other studies have utilized hair follicle self-sampling [39, 58–61], but none reported success rate or mean yield, nor compared assisted with self-sampling.

We and others have noted a higher RNA yield in male participants, likely due to stronger follicular adhesion in beard hairs [19, 62]. Bracci et al. found that pubic hair in female nurses yielded the highest RNA quality and quantity among the tested sites, though still requiring 15 follicles per sample [59]. At the same time, thicker, more strongly rooted hairs can increase discomfort, highlighting the need for researchers to balance RNA yield with participant burden when designing sampling protocols.

While group-level analyses revealed clear rhythmicity in PER3 expression, rhythm detection at the individual level was more variable. In our samples, one third (20/66) of the participants exhibited statistically significant PER3 rhythms based on cosinor fitting, with lower detection rates for other genes. Previous studies using similar protocols have produced varying results on individual rhythm detection. For example, Watanabe et al. detected PER3 rhythmicity in hair follicles of 6 of 8 participants, but rhythmic expression of PER2 and BMAL in only 2/8 and 3/8, respectively [43], using a composite 24/12-hour cosinor model in ChronoLab [63]. Hattammaru et al. reported significant PER3 rhythms in hair follicles in 64% of daytime workers and 0% of consecutive-night shift workers [39] using the acro.exe [64]. Haraguchi et al. [38] displayed clear rhythms in all participants (with no formal test) but excluded 12/27 participants based on criteria of yield <20 ng/μL or qPCR Ct-values > 33. These findings highlight both biological and methodological factors that constrain rhythm detection at the individual level. Technical limitations, including limited sampling density, variable sample quality, and gene-specific expression amplitudes, can obscure otherwise intact biological rhythms. Our supplementary simulation-based phase-precision analyses indicate that, under the present hair-follicle sampling design, individual-level phase estimation has limited precision and improves only modestly when extending the number of sampling days (Supplementary Fig. S10). One should accordingly be cautious when interpreting failure to detect rhythmicity as evidence of arrhythmicity or altered circadian function. Failure to meet statistical thresholds may reflect low amplitude, irregular timing, or technical noise, rather than true arrhythmicity.

A further challenge lies in the lack of consensus on the optimal method for detecting rhythmicity. A variety of statistical tools are available, each with different assumptions and sensitivities. Beyond classical cosinor analysis, other commonly used methods include spectral approaches such as the Lomb-Scargle periodogram [65] and non-parametric alternatives like the Jonckheere–Terpstra–Kendall test [66]. No gold standard exists, and detection performance can vary substantially depending on sample size, data resolution, waveform shape, and noise characteristics. Consequently, differences in reported rhythmicity across studies may reflect not only biological variation, but also methodological heterogeneity.

Several limitations should be acknowledged when interpreting our findings. First, while our sampling design was optimized for ecological validity, participants collected only four samples per day across two days, which may be insufficient for robust rhythm detection at the individual level, particularly for lower amplitude genes. Second, although the RNA yield was generally adequate, a subset of samples fell below our quantity threshold, and the quantity threshold utilized was lower than those in some other studies [38, 41]. Third, participants in the schizophrenia group were clinically stable and independently living, and findings may not generalize to individuals in acute phases or with more severe functional impairment. Fourth, although antipsychotic medication use was recorded, data on duration of exposure were not available, and our sample size did not permit stratified analysis by medication class or dose, or polypharmacy. Fifth, while we matched participants on key variables like age, sex, BMI, and sleep apnea status, residual confounding from unmeasured lifestyle or pharmacological factors cannot be ruled out. Moreover, the lack of a schizophrenia group without obesity limits inference about obesity-independent effects of schizophrenia. Sixth, contrary to our initial protocol [21], melatonin-based phase estimates were not derived from formal DLMO calculations; therefore, melatonin results should be interpreted cautiously and not as definitive measures of circadian phase. Finally, the cross-sectional design of this study limits causal inference, and future longitudinal or interventional studies are needed to determine whether sleep regularity or other behavioral factors can causally influence circadian gene expression.

Our findings raise the possibility that PER3 amplitude and sleep regularity may serve as peripheral markers of circadian vulnerability in schizophrenia. The association between the SRI and PER3 rhythmicity suggests a potential bio-behavioral feedback loop, in which irregular sleep weakens molecular rhythms, while dampened rhythms in turn impair behavioral stability. Given this, interventions that target sleep timing, such as light therapy, behavioral sleep strategies, or structured daily routines, may not only improve subjective and functional outcomes [67], but also strengthen underlying circadian biology. Moreover, aligning the timing of other behavioral cues, such as meal timing and physical activity, with the body’s natural rhythms could further support circadian entrainment and enhance the effectiveness of such interventions [38, 41]. Future longitudinal or interventional studies could incorporate hair follicle sampling to evaluate group-level changes in clock gene expression in response to such treatments, offering a minimally invasive and ecologically valid method for tracking biological rhythms in real-world settings. However, at the present sampling density, hair-follicle sampling appears unlikely to provide clinically reliable individual-level phase (acrophase) estimation.

Our findings demonstrate that repeated hair follicle self-sampling is a feasible and informative approach for assessing circadian gene expression. We identified a significantly attenuated amplitude of PER3 rhythmicity in individuals with schizophrenia and found that reduced PER3 amplitude was associated with lower sleep regularity, linking molecular rhythms to behavioral patterns. These results support the utility of PER3 amplitude as a potential marker of dampened molecular rhythmicity, while underscoring the need for future studies to refine molecular detection methods and evaluate responsiveness to behavioral or therapeutic interventions.