Heritability and Genetic Correlations for Sleep Apnea, Insomnia, and Hypersomnia in a Large Clinical Biobank

Abstract

Rationale:

Comorbid insomnia and sleep apnea (COMISA) is reported to have worse outcomes than either condition alone. The local genetic correlations of these disorders is unknown.

Objectives:

To identify local genomic regions with heritability for clinically diagnosed sleep apnea (SA) and insomnia; and identify local genetic correlations between these disorders and/or hypersomnia.

Methods:

50,217 patients of European ancestry were examined. Global and local heritability and genetic correlations for independent regions were calculated adjusting for obesity and other covariates.

Results:

SA and insomnia were significantly globally heritable and had 118 and 168 genetic regions with local heritability p-values < 0.05 respectively. One region had a significant genetic correlation for SA and hypersomnia (p-value = 9.85 × 10⁻⁴).

Conclusions:

Clinically diagnosed SA and insomnia have minimal shared genetic architecture, supporting genetically distinct COMISA components. However, additional correlated regions may be identified with additional sample size and methodological improvements.

Introduction

Sleep apnea (SA) is a common disease characterized by repetitive upper airway obstructions, leading to intermittent hypoxemia and increased risk of a range of disorders [1]. SA may present both with insomnia symptoms (comorbid insomnia and sleep apnea termed “COMISA” [2]) or hypersomnia symptoms. Patients with comorbid SA presentations may conceivably have distinct comorbidity profiles and/or SA endophenotypes when compared with patients presenting with SA without insomnia or hypersomnia. SA as measured through objective recordings is heritable, with a family-based estimate of ≈ 0.35, over half of which is explained by non-obesity pathways [3]. Multiple genetic associations with SA have been identified, e.g. [4]. Complex trait analyses may require >250,000 participants to fully discover grouped pathway effects [5]. Electronic health record studies may boost study power, but issues including data completeness, billing diagnoses, and the ability of ICD codes to fully describe complex phenotypes must be addressed [6]. We have demonstrated a SA phenotyping algorithm based on natural language processing (NLP) with improved positive predictive value in chart reviews [7] that may provide utility in large-scale genetic analyses.

Insomnia is also a common disorder with a significant genetic basis [8]. Recent epidemiological studies have identified increased mortality risks for individuals with COMISA compared to individuals with OSA [9-11]. Insomnia hazard ratio point estimates were sometimes higher compared to OSA, underscoring a need to further characterize the potential relationship between insomnia and SA. Global genetic correlation analyses are one means of understanding heritable and shared genetic architecture between two traits. Whether these clinically overlapping phenotypes map to distinct genetic regions is unknown. Genome-wide genetic relationships between pairs of traits may be nuanced, and local interactions may differentially contribute to risk [12]. Local genetic correlation analyses can be used to identify and prioritize specific regions with positive or negative genetic heritability correlations between two traits that may be obscured when performing global genetic correlation analyses.

In this study, we performed global and local genetic heritability analyses of SA and insomnia considering obesity and other covariates in a large clinical biobank. We consider potential local genetic correlations between the two disorders. Since cluster analyses indicate increased risk for patients with SA who report sleepiness [13,14], we also evaluated genetic correlations between SA and hypersomnia.

Methods

This study was approved by the Mass General Brigham Institutional Review Board (2021P013546). Studied participants from the MGB biobank [15] were aged ≥ 18 with European ancestry with a minimum data floor of three clinical notes, diagnoses, and encounters and available age, body mass index (BMI), and sex at birth information.

We performed NLP on clinic notes, extracted relevant SA ICD diagnoses (seven ICD9 and four ICD10 codes, Table S1, and phenotyped SA as per our published algorithm [7], excluding patients with SA ICD codes who were classified as controls by the algorithm. We used grouped ICD codes and used a semi-supervised NLP algorithm to classify patients with insomnia (eight ICD9 and ten ICD10 codes) and hypersomnia (eleven ICD9 and twelve ICD10 codes, Table S1) [16,17]. Insomnia and hypersomnia NLP terms matched the text of relevant ICD diagnoses.

Age was set to the first diagnosis date for a given case or the last encounter date for a control. BMI was set as the average of the two measurements from the two closest dates to the assigned age. Current and former smoking status was assigned based on a questionnaire administered by the MGB biobank supplemented by tobacco use disorder diagnoses (Table S1).

Biobank participants were genotyped using Illumina MEGA and GSA arrays and imputed to the NHLBI Trans-Omics for Precision Medicine (TOPMed) panel using the TOPMed Imputation Server [18]. Imputed variants with a minimum r² of 0.95 and present in high-coverage whole-genome sequencing of 1000 Genomes European ancestry participants at 0.5% minor allele frequency were retained. All regions are listed using Build 38 coordinates.

Global heritability analyses were performed using PCGC and LDAK [19,20], adjusting for age, sex at birth, BMI, and 10 population principal components [21], both with and without smoking status to evaluate potential heritability improvements in this clinically diagnosed sample. LAVA was used to calculate local genetic heritability analyses [22]. Bivariate genetic correlation analyses were conducted for independent genomic regions. Genetic correlations considering SA versus insomnia and hypersomnia were performed for loci with univariate heritability p-values < 0.05.

Results

Sample characteristics

We examined 50,225 adult patients with European ancestry. The median (interquartile range) of age was 61.2 (24.1) and BMI was 27.3 (7.8). 54.2% of this sample were women, with 17.6% classified as current smokers and 20.2% classified as former smokers. The prevalence rates of NLP-defined clinically diagnosed SA, insomnia, hypersomnia, and COMISA were 8.3%, 7.4%, 1.4%, and 1.3% respectively. Patients with both insomnia and SA diagnoses on average had later insomnia diagnoses (men median 1.1 years later, IQR 5.2 years; women 0.2 years later, IQR 5.5 years), although a large portion of the patients had earlier insomnia diagnoses (Figure S1).

Global heritability estimates

Global heritability estimates are listed in Table S2. SA was significantly heritable (h²g [SE] 0.168 [0.067], p = 2.79 × 10⁻⁵), as was insomnia (h²g [SE] 0.077 [0.031], p = 1.24 × 10⁻⁴). Hypersomnia, which was much rarer, had a non-significant heritability estimate of 0.126 (SE 0.102; p = 0.08). Adjustment for smoking did not appreciably change the heritability estimates.

Local heritability estimates and genetic correlations

Local heritability analyses can highlight specific regions of biological interest, even when genome-wide effects are largely neutral. We calculated local genetic heritability estimates for SA and insomnia in 2,818 autosomal regions. Associated regions are shown in Table S3. Lead results among the 118 SA-associated regions and 168 insomnia-associated regions are summarized in Table 1. The most heritable region was on chromosome 5 (81,174,678 – 82,414,318 SA h²g p-value = 1.86 × 10⁻¹¹).

Table 1.

Lead sleep apnea and insomnia local genetic heritability regions.

Phenotype	Chromosome	Start	End	h2g p- value	Protein-coding Genes in Region
Sleep apnea	5	81,174,678	82,414,318	1.86 × 10−11	ACOT12, ATG10, CKMT2, RASGRF2, RPS23, SSBP2, ZCCHC9
Sleep apnea	3	168,549,584	169,880,814	2.67 × 10−6	ACTRT3, LRRC31, LRRC34, LRRIQ4, MECOM, MYNN
Sleep apnea	6	27,518,868	28,703,605	3.58 × 10−5	GPX5, GPX6, H1-5, H2AC13, H2AC14, H2AC15, H2AC16, H2AC17, H2BC13, H2BC14, H2BC15, H2BC17, H3C10, H3C11, H3C12, H4C11, H4C12, H4C13, NKAPL, OR2B2, OR2B6, OR2B8P, PGBD1, ZBED9, ZKSCAN3, ZKSCAN4, ZKSCAN8, ZKSCAN8P1, ZNF165, ZSCAN12, ZSCAN16, ZSCAN23, ZSCAN26, ZSCAN31, ZSCAN9
Sleep apnea	8	85,490,346	86,552,020	1.36 × 10−4	ATP6V0D2, CPNE3, PSKH2, RMDN1, SLC7A13, WWP1
Sleep apnea	6	11,217,931	11,775,181	2.36 × 10−4	ADTRP, NEDD9, TMEM170B
Sleep apnea	9	104,298,903	105,440,109	2.76 × 10−4	ABCA1, NIPSNAP3A, NIPSNAP3B, OR13C2, OR13C3, OR13C4, OR13C5, OR13C8, OR13C9, OR13D1, OR13F1, SLC44A1
Sleep apnea	8	9,979,082	10,620,638	2.79 × 10−4	MSRA, PRSS51, PRSS55, RP1L1
Sleep apnea	8	66,215,247	67,662,602	3.06 × 10−4	ADHFE1, ARFGEF1, C8orf44-SGK3, COPS5, CPA6, CSPP1, MCMDC2, MYBL1, PPP1R42, RRS1, SGK3, TCF24, VCPIP1, VXN
Sleep apnea	12	4,858,524	5,775,562	4.50 × 10−4	ANO2, KCNA1, KCNA5, NTF3
Sleep apnea	1	158,930,438	159,762,876	8.92 × 10−4	ACKR1, AIM2, APCS, CADM3, CRP, FCER1A, IFI16, OR10J1, OR10J3, OR10J4, OR10J5, PYDC5, PYHIN1
Insomnia	10	102,437,650	103,700,556	3.71 × 10−5	ACTR1A, ARL3, AS3MT, ATP5MK, BORCS7, BORCS7-ASMT, C10orf95, CALHM1, CALHM2, CALHM3, CNNM2, CYP17A1, INA, MFSD13A, NEURL1, NT5C2, PCGF6, PDCD11, RPEL1, SFXN2, SH3PXD2A, SUFU, TAF5, TRIM8, WBP1L
Insomnia	9	94,710,721	95,785,833	4.02 × 10−5	AOPEP, FANCC, PTCH1
Insomnia	2	1,377,422	1,991,457	4.73 × 10−5	MYT1L, PXDN, TPO
Insomnia	2	51,442,356	52,267,286	1.26 × 10−4
Insomnia	11	11,337,460	12,128,955	1.65 × 10−4	CSNK2A3, DKK3, GALNT18, MICAL2, USP47
Insomnia	4	148,808,693	149,733,447	1.88 × 10−4	IQCM
Insomnia	8	18,428,040	19,106,735	2.02 × 10−4	PSD3
Insomnia	14	65,832,798	66,559,780	2.76 × 10−4	CCDC196, GPHN
Insomnia	7	128,263,781	129,144,853	7.51 × 10−4	ATP6V1F, ATP6V1FNB, CALU, CCDC136, FLNC, GARIN1A, GARIN1B, HILPDA, IMPDH1, IRF5, KCP, METTL2B, OPN1SW, PRRT4, RBM28, TNPO3, TSPAN33
Insomnia	1	58,073,339	58,847,823	8.69 × 10−4	DAB1, JUN, MYSM1, OMA1, TACSTD2

All calculations adjusted for age, sex, body mass index, and 10 population principal components. Region borders were calculated by LAVA [22] using the Mass General Brigham data with Build 38 coordinates. Full results for nominally significant regions are provided in Table S3. h²g: single nucleotide variant-based heritability.

We calculated genetic correlations for regions with both univariate heritability p-values < 0.05 and performed equivalent SA-hypersomnia calculations. Six regions had overlapping insomnia and SA univariate heritability p-values < 0.05, and 9 regions had overlapping hypersomnia and SA univariate heritability p-values < 0.05 (Table S4). Within these 15 regions, one region had a significant genetic correlation for SA and hypersomnia (chr3:168,549,584 – 169,880,814; r² = 0.35; p-value = 9.85 × 10⁻⁴).

Discussion

In this study, we used genetic analyses to begin to parse out the extent to which SA and its symptom-based presentations- COMISA and SA-hypersomnia represent genetically determined complex phenotypes or the co-occurrence of unique traits. We performed genome-wide local genetic heritability calculations considering clinically diagnosed sleep apnea and insomnia adjusted for obesity and other covariates in a large biobank and asked whether there may be a shared genetic architecture for comorbid insomnia and sleep apnea (COMISA). We also examined local genetic correlations between sleep apnea and hypersomnia, performed global heritability analyses, and examined the potential effects of adjustment for smoking. We identified regions with significant heritability, including regions with overlapping signals between SA and insomnia and SA and hypersomnia. We identified a single region that was significantly genetically correlated between SA and hypersomnia but did not identify significant local genetic correlations between SA and insomnia. These results indicate that COMISA is likely comprised of separate genetic contributions from SA and insomnia pathways, based on our current methods and sample size.

Accurate phenotyping using billing code information must address multiple challenges, including data completeness, biases arising from differential healthcare utilization, rule-out diagnoses, sparsity of individual ICD diagnoses and the ability of ICD codes to fully describe complex phenotypes [6]. We address these issues using healthcare utilization data floors and adjustments [7,17], grouped PheCode diagnoses, and NLP-informed phenotyping algorithms. Our SA phenotyping algorithm demonstrated improved positive predictive value compared to single-ICD case assignments using chart reviews [7] and here demonstrates elevated heritability relative to other biobank estimates [23], despite a smaller sample size, exclusion of some patients with uncertain case/control status, and adjustment for obesity and other covariates. Additional algorithm refinements that consider SA endotypes reflected in clinical note phrases and patterns of multimorbidity are possible and may further improve performance. Insomnia phenotyping would likely improve based on a similar strategy that separately considers the effects of early, middle, and late insomnia that may have heterogeneous genetic architecture and differential effects on comorbid disease risk.

The prevalence of insomnia, SA, and COMISA varies by study [9-11]. Insomnia, SA, and COMISA have reported prevalences of 6-36.5%, 23.4-53.5%, and 3-19.3% respectively. There are a number of design differences to consider when comparing the prevalences of these disorders, including but not limited to community versus hospital ascertainment, use of symptoms and/or AHI > 5 criteria, oversampling of self-reported snorers [24], and changing prevalences across years (e.g. [25]).

We identified multiple regions with local genetic heritability for SA and insomnia (Tables 1 and S3). Additional studies are required to identify if one or more protein coding genes or other mechanisms may be associated. Aggregating analyses into units and pathways is a pragmatic means of building power for analyses that will likely require hundreds of thousands of samples to saturate signals [5]. Local genetic correlations provide a means of identifying effects that may be obscured at a global level [22], improve power by reducing comparisons, and highlight regions of potential biological significance. Despite identifying regions where both SA and insomnia have local genetic heritability, we did not identify regions that have significant correlations between the two traits. LAVA has been used to identify local genetic correlations using smaller samples of trait pairs with strong biological plausibility [22], but we cannot rule out identifying shared genetic architecture between SA and insomnia in the future as we refine our methods and begin to meta-analyze results with other biobanks.

Strengths of this study include a large sample of clinically defined patients with SA and insomnia with NLP and healthcare utilization phenotyping refinements. We considered local genetic correlations between SA and insomnia for the first time. Limitations of this study include a focus on European-ancestry patients (our largest sample), and potential limits on study power. We will address these limitations as we refine our methods and begin to perform meta-analyses with data from other biobanks.

Conclusions

We have identified multiple regions of local genetic heritability for SA and insomnia. None of these regions have genetic correlations between SA and insomnia, indicating that the genetic architecture of COMISA as identified in medical records is likely comprised of separate SA and insomnia contributions.

Public Health Relevance Statement.

Insomnia and sleep apnea are both common, heritable diseases that increase the risk of a range of disorders. Some studies indicate that comorbid insomnia and sleep apnea (COMISA) increases mortality risks relative to the risks associated with either diseases alone. This could conceivably be due to shared genetic mechanisms that contribute to a distinct COMISA endotype. To our knowledge, this is the first study to examine local genetic architecture potentially shared between the two disorders. We identified novel genomic regions with clinically diagnosed sleep apnea and insomnia heritability but with minimal shared genetic architecture, supporting genetically distinct COMISA components.