Machine Learning Analyses Reveal Circadian Features Predictive of Risk for Sleep Disturbance

Abstract

Introduction: Sleep disturbances often co-occur with mood disorders, with poor sleep quality affecting over a quarter of the global population. Recent advances in sleep and circadian biology suggest poor sleep quality is linked to disruptions in circadian rhythms, including significant associations between sleep features and circadian clock gene variants.

Methods: Here, we employ machine learning techniques, combined with statistical approaches, in a deeply phenotyped population to explore associations between clock genotypes, circadian phenotypes (diurnal preference and circadian phase), and risk for sleep disturbance symptoms.

Results: As found in previous studies, evening chronotypes report high levels of sleep disturbance symptoms. Using molecular chronotyping by measuring circadian phase, we extend these findings and show that individuals with a mismatch between circadian phase and diurnal preference report higher levels of sleep disturbance. We also report novel synergistic interactions in genotype combinations of Period 3, Clock and Cryptochrome variants (PER3B (rs17031614)/ CRY1 (rs228716) and CLOCK3111 (rs1801260)/ CRY2 (rs10838524)) that yield strong associations with sleep disturbance, particularly in males.

Conclusion: Our results indicate that both direct and indirect mechanisms may impact sleep quality; sex-specific clock genotype combinations predictive of sleep disturbance may represent direct effects of clock gene function on downstream pathways involved in sleep physiology. In addition, the mediation of clock gene effects on sleep disturbance indicates circadian influences on the quality of sleep. Unraveling the complex molecular mechanisms at the intersection of circadian and sleep physiology is vital for understanding how genetic and behavioral factors influencing circadian phenotypes impact sleep quality. Such studies provide potential targets for further study and inform efforts to improve non-invasive therapeutics for sleep disorders.

Keywords: chronotype; circadian clock; circadian misalignment; machine learning; sleep disturbance; sleep quality.

Figure 1

Dual processes controlling the sleep-wake cycle and potential mechanisms by which circadian mutations may influence sleep disturbance and related outcomes. Process S describes the sleep homeostat; the body’s drive for sleep builds during the day as our time awake increases and energy is expended. Sleep drive gets stronger the longer we are awake. After a full night of sleep, the drive reaches its lowest point when we have no more need for sleep. Process C describes the circadian rhythm—a self-sustained oscillator that influences the timing of the sleep-wake cycle. These two processes of sleep regulation are continuously interacting but are guided by independent mechanisms.

Figure 2

Heat Map of Sleep Disturbance Scores and Diurnal Preference (MEQ). Higher MEQ scores indicative of morning preferences are strongly associated with low sleep disturbance scores and lower MEQ scores indicative of evening preference are associated with higher sleep disturbance scores. The heatmaps were made using Fisher’s Exact Test at varying cutoff levels for both dependent and independent variables. Fisher’s Exact Test shows significant association between high sleep disturbance scores (25+) and evening type individuals (MEQ<41). The p-values obtained from the analysis were log-transformed (base 10).

Figure 3

Circadian misalignment predictive of sleep disturbance. Approximately 25% of our sample size reported sleep disturbance with average sleep disturbance score of 20.68 (SD=5.91, SE=0.195). Advanced-morning types have significantly lower sleep disturbance scores than the average (p=0.007). The odds of sleep disturbance tended to be approximately 4.7 times higher if an individual was delayed vs advanced in their chronotype (OR=4.667 (0.5538–39.3242), p=0.157). There was a significant difference in sleep disturbance scores between ET and MT groups for which we had molecular chronotypes (F(1,63)=4.63, p=0.035). Average PROMIS sleep disturbance scores for the population sample is denoted by the red dashed line.

Figure 4

MEQ acts as a mediator between genotypic and clinical characteristics and human sleep disturbance. The network is constructed using the ARACNE method in conjunction with the mi-empirical method and bootstrapping. All links with bootstrap support greater than 50% are shown. MEQ mediates all associations with sleep disturbance.

Figure 5

Heat map of prediction accuracy for feature selection and classifier methods. Our analyses yielded up to 19.9% higher prediction accuracy than baseline (50%) on a balanced data set using top features.

Figure 6

Genotype combination of CLOCK3111_TC and CRY2_AG predictive of sleep disturbance in humans. (A) Sleep disturbance scores (±SD), measured using the self-reported Patient Reported Outcomes Measurement Information System (PROMIS™) Sleep Disturbance instrument, for males and females with the CLOCK3111_TC and CRY2_AG genotype. Average sleep disturbance scores were higher for both males and females with CLOCK3111_TC and CRY2_AG (males: 23.6 ±1.4; females 24.5±0.88) than for individuals with other genotypes (males: 20.0 ±0.59; females 21.2±0.35; two-way ANOVA: genotype F(1463)=15.94, p<0.001; gender F(1463)=1.30, p=0.254; genotype by gender F(1463)=0.018, p<0.89)). (B) Females with the CLOCK3111_TC and CRY2_AG genotypes have significantly higher sleep disturbance scores than females with other genotypes (p=0.002) and males with other genotypes (p<0.001).

Figure 7

Genotype combination of PER3B_AG and CRY1_CG show significant interaction effect. (A) Sleep disturbance scores (±SD), measured using the self-reported Patient Reported Outcomes Measurement Information System (PROMIS™) Sleep Disturbance instrument, for males and females with a combination PER3B_AG and CRY1_CG genotype. Sleep disturbance scores reveal a significant gender-by-genotype interaction (genotype F(1430)=2.123, p=0.146; gender F(1430)=3.03, p=0.083; genotype by gender F(1430)=5.05, p=0.025). (B) The highest average sleep disturbance score was reported in males with the PER3B_AG and CRY1_CG genotypes (27.2±2.72), with a single male of genotype PER3B_AA/CRY1_CG recording the highest individual sleep disturbance score (34.2).

Figure 8

Association rules networks for sleep disturbance. (A) In females, CLOCK311-TC and CRY2-AG co-occurred most frequently with no mediation via MEQ. PER3B-GG and age had the highest average lift (>1.45) in the analysis. (B)) In males, no robust co-occurrence across genotypic and clinical features was found, although average lift values were higher for males than females.