Insomnia, proteomic characteristics, and mortality risk in depression: a UK Biobank study Authors: Xuesong Liu, Hongli Xu, Xiaoqian Zhu, Jiabin Tu, Kaiyang Lin, Yansong Guo Published: 2026 ABSTRACT AbstractBackgroundAlthough insomnia is associated with increased mortality in patients with depression, the underlying biological pathways remain unclear. As functional gene products, proteins are well suited to bridge behavioural risk factors and clinical outcomes.ObjectiveTo explore the associations between insomnia, proteomic characteristics, and mortality risk in depressed individuals.MethodsWe included 4,498 patients with depression from the UK Biobank. Insomnia was self-reported; all-cause mortality was the primary outcome. Cox models, LASSO regression, single- and multi-mediator joint mediation analyses, subgroup analyses, competing-risk analyses, and incremental predictive value assessments were performed.Results”Usually insomnia” was an independent predictor of increased mortality (HR = 1.38, 95% CI: 1.05–1.81) and was associated with a 2.3-year shorter life expectancy at age 45. Twenty insomnia-related proteins were identified via LASSO regression and used to construct a proteomic signature, which was significantly associated with mortality (T3 vs. T1: HR = 2.54, 95% CI: 1.95–3.32). Multi-mediator joint analysis revealed that these 20 proteins collectively accounted for 37.2% of the total effect of insomnia on mortality, among which TNFRSF6B, CSTB, and LGALS4 retained statistically significant independent mediating contributions. The proteomic signature demonstrated significant incremental predictive value beyond conventional clinical risk factors.ConclusionInsomnia is associated with increased mortality risk in depressed patients. Multi-mediator analysis identified three proteins (TNFRSF6B, CSTB, and LGALS4) as independent mediators of this association. The insomnia-related proteomic signature may serve as a biomarker for mortality risk stratification.Clinical trial numberNot applicable.Supplementary informationThe online version contains supplementary material available at 10.1186/s12888-026-08270-z.Keywords: Depression, Insomnia, All-cause mortality, Proteomic signature, UK Biobank Introduction According to data released by the World Health Organization (WHO) in 2021, approximately 280 million people are affected by depression worldwide, accounting for 4.4% of the global population [1]. Depression severely impairs the well-being and quality of life of those affected while imposing substantial societal burdens. Depression and insomnia are highly comorbid, with approximately 80–90% of patients with depression exhibiting symptoms of insomnia [2, 3]. Although insomnia has been associated with increased mortality among individuals with depression, the specific biological pathways involved remain to be elucidated [4]. As functional end products of gene expression, proteins are crucial for elucidating underlying disease pathophysiology and serve as ideal biomarkers that bridge behavioural risk factors and clinical outcomes [5, 6]. The effects of insomnia on proteomic signatures may be mediated by various biological pathways. Specifically, individuals with insomnia exhibit hyperactivation of the hypothalamic–pituitary–adrenal (HPA) axis and the sympathetic nervous system [7, 8], which results in elevated stress hormone secretion, an increased metabolic rate and impaired cardiac autonomic function [9–11]. This state of hyperactivation is associated with elevated risks of hypertension and diabetes and ultimately increased mortality [12, 13]. Additionally, insomnia is associated with amplified inflammatory responses through upregulation of key proinflammatory pathways. In healthy older adults, LPS stimulation was found to increase monocytic IL-6 and TNF-α production and enhance the activation of immune-related transcription factors under poor sleep quality [14]. A network analysis study using UK Biobank data further revealed that insomnia severity modulates anxiety–depression–CRP interrelationships [15]. Nevertheless, the proteomic landscape underlying the interaction between insomnia and mortality in depressed populations remains largely uncharacterized. Accordingly, this study aimed to investigate the association between insomnia and mortality risk in individuals with depression, and to explore the potential underlying mechanisms. Methods Study population The United Kingdom Biobank (UKB) is an ongoing prospective cohort of more than 500,000 adults recruited throughout England, Scotland, and Wales. Baseline data on demographics, health history, and lifestyle were gathered via touchscreen questionnaires, nurse-led interviews and links to electronic health records; anthropometry and biospecimens were obtained with standardized protocols (full details published elsewhere [16]). UKB has approval from the North West Multi-centre Research Ethics Committee, the National Information Governance Board for England and Wales, and the Community Health Index Advisory Group in Scotland. All the participants provided written informed consent. Proteomics assays Plasma collected at enrolment was stored at − 80 °C or in liquid nitrogen. A subset of these samples was sent to Olink Analysis Services for high-throughput proteomic profiling using proximity extension assays (PEAs). Protein expression levels were reported as normalized protein expression (NPX) units, which were on a log2 scale and had been pre-processed by the Olink laboratory to minimize batch effects and technical variability. Briefly, each assay included internal controls for incubation, extension, and detection, and the data were normalized using a plate-specific correction factor and bridge samples for harmonisation across plates. A total of 2,923 proteins were measured. Quality control was performed at both the sample and protein levels: samples failing to meet Olink’s internal QC criteria (deviation from median NPX, insufficient number of detectable proteins) were excluded; proteins with >20% missing values were removed. After these exclusions, 2,911 proteins remained for analysis. Details of the rigorous QC and external validation methods have been previously reported [17]. Among the 4,568 depressed patients for whom data from Olink were available, 51 with cancer, 11 without sleep data, and 8 who were lost to follow-up were excluded. Finally, 4,498 participants remained in the analysis. The definition of depression The depression cases were identified by combining multiple UK Biobank data sources, including primary care records, hospital inpatient episode data, and self-reported medical conditions confirmed by verbal interviews. Specifically, participants were included if they had at least one primary care record indicating depression or at least one hospital admission coded with a primary or secondary ICD-10 diagnosis of F32 (depressive episode) or F33 (recurrent depressive disorder). The definition of insomnia In the UKB database, participants’ sleep status was assessed via questionnaire surveys. Participants self-reported their insomnia status in the questionnaires. The responses were categorized as never, sometimes, or usually. The definition of outcome The primary outcome of the study was all-cause mortality. Mortality data were obtained from the NHS Information Centre (England & Wales) and the NHS Central Register (Scotland). The follow-up period concluded in December 2022. The definition of covariates Covariates were chosen through a literature review and clinical judgement. Age, sex, ethnicity, smoking status and alcohol use were self-reported. Body mass index (BMI) was derived from baseline height and weight. Seated systolic blood pressures (SBP) were recorded with automated digital sphygmomanometers. Plasma collected at enrolment was assayed for the levels of glycated haemoglobin (HbA1c), creatinine (Cr), triglycerides (TG), and total cholesterol (TC) by UKB technicians. Socioeconomic status (SES) was constructed from pretax household income, occupation, and education and grouped into low, medium, and high levels [18]. Antidepressant and anti-insomnia medication use was self-reported at baseline. For depression severity-stratified subgroup analyses, antidepressant use was further categorized into three groups: none (no antidepressant use), standard treatment (first- or second-line monotherapy, including SSRIs and SNRIs) and severe proxy (tricyclic antidepressant use or polytherapy with two or more agents). We calculated the diet quality score on the basis of the intake of fruits, vegetables, processed meat, red meat, fish, whole grains, and refined grains. One point was assigned for each food item meeting the following criteria: average intake of fruits ≥ 3 servings/day, vegetables ≥ 3 servings/day, fish ≥ 2 servings/week, processed meats ≤ 1 serving/week, red meats ≤ 1.5 servings/week, whole grains ≥ 3 servings/week, and refined grains ≤ 1.5 servings/week. Therefore, the diet quality score ranged from 0 to 7, with a higher score indicating better diet quality [19]. Physical activity was calculated as the sum of the metabolic equivalent of task (MET) minutes of all activities per week. Chronotype (morning vs. evening preference) and sleep duration (hours per night) were self-reported at baseline via touchscreen questionnaire. For subgroup analyses, sleep duration was dichotomized at the median (7 hours) into short sleep (<7 h) and adequate sleep (≥7 h). Statistical analysis Missing values in covariates with less than 20% missingness were imputed using random forest-based multiple imputation implemented in the MICE package. The imputation model included all the covariates listed in Table 1 and all the proteins. Five imputed datasets were generated with 10 iterations each. The imputation process was repeated to ensure convergence. The results across imputed datasets were combined using Rubin’s rules for the final estimates. Variables with more than 20% missing data were excluded. Normally distributed continuous data are presented as means (standard deviations), skewed continuous data as medians (interquartile ranges), and categorical data as counts (percentages). Table 1Comparison of the baseline characteristics of the study populationInsomniaCharacteristicOverallNeverSometimesUsuallyP valueN449867819381882Age (median [IQR])56.00 [49.00, 62.00]56.00 [47.00, 62.00]56.00 [49.00, 62.00]56.00 [50.00, 62.00]0.142Sex (%) Female2,915 (64.8)404 (59.6)1,272 (65.6)1,239 (65.8)0.008 Male1,583 (35.2)274 (40.4)666 (34.4)643 (34.2)Ethnic (%)0.458 White4,091 (91.0)621 (91.6)1,770 (91.3)1,700 (90.3) Other407 (9.0)57 (8.4)168 (8.7)182 (9.7)Smoke (%)<0.001 Never2,125 (47.2)341 (50.3)933 (48.1)851 (45.2) Previous1,633 (36.3)239 (35.3)746 (38.5)648 (34.4) Current740 (16.5)98 (14.5)259 (13.4)383 (20.4)Drink (%)<0.001 Never222 (4.9)33 (4.9)88 (4.5)101 (5.4) Previous335 (7.4)40 (5.9)114 (5.9)181 (9.6) Current3,941 (87.6)605 (89.2)1,736 (89.6)1,600 (85.0)SES (%)0.011 Low SES1,779 (39.6)265 (39.1)752 (38.8)762 (40.5) Medium SES1,798 (40.0)286 (42.2)813 (42.0)699 (37.1) High SES921 (20.5)127 (18.7)373 (19.2)421 (22.4)Diet quality score (median [IQR])3.38 [2.00, 4.00]4.00 [3.00, 5.00]3.00 [2.00, 4.00]3.00 [2.00, 4.00]0.011Physical activity MET (median [IQR])2,387.76 [894.25, 2507.62]2,378.00 [1007.38, 2809.12]2,387.76 [924.00, 2534.50]2,387.76 [819.25, 2387.76]0.211Antidepressants (%)1,988 (44.2)266 (39.2)823 (42.5)899 (47.8)Sleeping pills (%)175 (3.9)4 (0.6)45 (2.3)126 (6.7)Biochemical (median [IQR]) HbA1c, mmol/mol35.80 [33.20, 38.00]35.25 [32.70, 37.60]35.60 [33.10, 37.80]36.10 [33.40, 38.58]<0.001 Cr, μmol/L69.30 [60.40, 78.47]70.20 [61.50, 79.18]69.20 [60.80, 78.40]68.90 [59.90, 78.20]0.118 TG, mmol/L1.65 [1.13, 2.26]1.50 [1.05, 2.08]1.64 [1.11, 2.24]1.70 [1.19, 2.35]<0.001 TC, mmol/L5.71 [4.94, 6.42]5.68 [4.84, 6.36]5.71 [4.97, 6.40]5.71 [4.95, 6.47]0.128 SBP, mmHg135.00 [123.00, 147.00]135.00 [122.00, 145.00]135.00 [123.00, 147.00]135.91 [123.00, 148.00]0.159 BMI, kg/m227.52 [24.65, 31.00]26.78 [24.29, 30.10]27.30 [24.69, 30.75]27.99 [24.80, 31.78]<0.001All-cause mortality (%)538 (12.0)65 (9.6)192 (9.9)281 (14.9)<0.001Abbreviations: N, Number; IQR: Interquartile range; SES: Socioeconomic Status; MET: Metabolic equivalent of task; HbA1c: Glycated haemoglobin; Cr: Creatinine; TG: Triglycerides; TC: Total cholesterol; SBP: Systolic blood pressure; BMI: Body mass index Comparison of the baseline characteristics of the study population Abbreviations: N, Number; IQR: Interquartile range; SES: Socioeconomic Status; MET: Metabolic equivalent of task; HbA1c: Glycated haemoglobin; Cr: Creatinine; TG: Triglycerides; TC: Total cholesterol; SBP: Systolic blood pressure; BMI: Body mass index Kaplan‒Meier curves were used to assess all-cause mortality among groups stratified by insomnia frequency. In this study, a Cox proportional hazards model was used to investigate the associations between insomnia and all-cause mortality. The analysis was divided into the following three models: Model 1 was an unadjusted model; Model 2 was adjusted for age, sex, and ethnicity; and Model 3 was further adjusted for variables such as smoking status, alcohol consumption, HbA1c, SBP, Cr, TG, TC, BMI, SES, diet quality, physical activity, antidepressants, and sleeping pills. Subgroup analyses were performed to assess the consistency of the associations of insomnia and the insomnia-related proteomic signature with all-cause mortality across strata defined by ethnicity, antidepressant treatment status, chronotype, and sleep duration. Interaction terms were included in the Cox models to test for effect modification. To identify proteins associated with insomnia, we first performed multivariable linear regression for insomnia status with each protein, adjusting for all the covariates included in Model 3. To control for multiple testing, we applied the Bonferroni correction, considering proteins with a p value < 0.05/2911 as significantly associated with insomnia. This initial step provided a set of proteins that were robustly linked to insomnia after accounting for potential confounders. Given that many of these proteins were correlated and might collectively capture overlapping biological information, we subsequently applied least absolute shrinkage and selection operator (LASSO) regression to further refine the selection. LASSO is particularly suitable for high-dimensional data with multicollinearity, as it imposes a penalty on the absolute values of regression coefficients, shrinking some coefficients to zero and thus performing automatic variable selection. This approach helps to identify a parsimonious set of proteins that jointly predict insomnia status, reducing the risk of overfitting. The penalty parameter λ was chosen via 10-fold cross-validation using the one-standard-error rule to balance model fit and complexity (Supplementary Fig. S1). The proteins retained by LASSO were subsequently used to construct an insomnia-related proteomic signature. Specifically, we fitted a LASSO regression model with insomnia status as the outcome and the selected proteins as predictors; the linear predictor (the weighted sum of protein levels with LASSO-derived coefficients) was computed for each participant and served as the continuous proteomic signature. Participants were then divided into tertiles (T1, T2, and T3) on the basis of this signature for subsequent survival analyses. To evaluate the predictive performance of the insomnia-related proteomic signature for all-cause mortality, internal validation was performed using bootstrap resampling with 1,000 iterations. The C-index was calculated for each bootstrap sample, and the mean C-index with its 95% confidence interval was derived from the bootstrap distribution. Time-dependent receiver operating characteristic (ROC) analysis was further conducted at 5-, 10-, and 15-year horizons to assess the discriminative capacity of the proteomic signature in predicting all-cause mortality, with the area under the curve (AUC) and 95% confidence intervals calculated using the timeROC and pROC packages in R. To assess whether the proteomic signature provides incremental predictive value beyond established clinical risk factors, we compared two nested Cox models: a Clinical model adjusted for all covariates included in Model 3, and a Clinical + Proteomic model that additionally incorporated the insomnia-related proteomic signature. The two models were compared with respect to the C-index and time-dependent AUC at 5, 10, and 15 years. The added prognostic value contributed by the proteomic signature was further quantified using the integrated discrimination improvement (IDI) and the continuous net reclassification improvement (cNRI) at the same three time points, with 95% confidence intervals and p values estimated from 1,000 bootstrap resamples via the survIDINRI package. In addition, ROC analyses were performed for individual common clinical indicators and for the proteomic signature, with the corresponding AUC reported. A parametric survival model was used to estimate years of life lost (YLL). With age as the time scale, the area under the survival curve up to 100 years of age was calculated using the integral principle, and YLL was defined as the difference in areas under the survival curves between comparison groups. To select the optimal parametric distribution, four candidate distributions were compared (Weibull, exponential, log-logistic and Gompertz) using the Akaike information criterion (AIC). The Gompertz distribution yielded the lowest AIC (Supplementary Table S11) and was therefore selected for all subsequent parametric survival analyses. To assess model fit, predicted survival curves derived from the Gompertz model were visually compared against nonparametric Kaplan‒Meier estimates across insomnia frequency groups, demonstrating adequate concordance in the direction and magnitude of survival differences (Supplementary Figs. S2 and S3). The detailed methodology is consistent with that outlined in previous studies [20, 21]. To account for the competing risks of cause-specific mortality, subdistribution hazard models were fitted using the Fine‒Gray method. CVD death, cancer death, and other-cause death were treated as separate competing events. The associations of insomnia frequency and the proteomic risk score with each cause-specific mortality outcome were estimated as subdistribution hazard ratios (sHR) with 95% confidence intervals, adjusting for the same covariates as in the primary Cox regression analyses. Mediation analysis was conducted to evaluate whether insomnia-related proteins mediated the association between insomnia and all-cause mortality in patients with depression. We adopted a counterfactual framework using the mediation package in R. The average causal mediation effect (ACME), average direct effect (ADE), and proportion mediated were estimated using nonparametric bootstrapping with 1,000 bootstrap resamples to obtain 95% confidence intervals and p values. The proportion mediated was calculated as the ACME divided by the total effect (the sum of ACME and ADE). Each protein was analysed individually while adjusting for covariates. Proteins with a statistically significant ACME (p < 0.05) were considered mediators. All p values were corrected using the Bonferroni method. For the 15 proteins identified as significant mediators in the single-mediator analyses, KEGG enrichment analyses were performed using the online software Hiplot (https://hiplot.com.cn/) to examine pathways and biological processes related to the target genes. Given that the LASSO-selected proteins are highly correlated and that single-mediator analyses may overestimate each protein’s contribution due to shared mediating effects across correlated mediators, a multi-mediator joint analysis was performed under a counterfactual framework, with all LASSO-selected proteins entered simultaneously as parallel mediators. The total effect, joint indirect effect, and the overall proportion of the insomnia–mortality association jointly mediated by these proteins were estimated, with 95% confidence intervals derived from 1,000 bootstrap resamples. To assess the extent of overlap among individual mediating effects, the indirect effect of each protein was additionally decomposed using the product-of-coefficients approach, and the sum of single-mediator contributions was compared with the joint estimate. To further determine whether the mediating proteins act independently or display synergistic or antagonistic interactions, pairwise interaction analyses were conducted among proteins that retained statistically significant independent contributions in the multi-mediator joint model. To mitigate potential reverse causality arising from undiagnosed subclinical conditions at baseline, a sensitivity analysis was performed by excluding participants who died within the first 2 years of follow-up. All statistical analyses were performed using R version 4.3.1 (R Foundation for Statistical Computing, Vienna, Austria). A two-sided p value < 0.05 was considered statistically significant. Results Participant characteristics A total of 4,498 participants were included in the study, with 678 (15.1%), 1,938 (43.1%), and 1,882 (41.8%) categorized as having never, sometimes, and usually insomnia, respectively. The median age (interquartile range, IQR) of the overall population was 56.00 (49.00, 62.00) years, and no significant difference was observed across the three insomnia groups (p = 0.142). In terms of demographic characteristics, a significant difference was observed in sex distribution (p = 0.008), with a lower proportion of females in the never insomnia group (59.6%) than in the sometimes (65.6%) and usually (65.8%) insomnia groups, whereas ethnicity did not differ significantly between the groups (p > 0.05). SES also differed significantly among the groups (p = 0.011), with a higher proportion of high SES in the usually insomnia group (22.4%) than in the never (18.7%) and sometimes (19.2%) insomnia groups. Lifestyle factors, including smoking and drinking status, significantly differed across the groups (both p < 0.001); current smokers were most prevalent in the usually insomnia group (20.4%), followed by the never (14.5%) and sometimes (13.4%) groups, whereas the proportion of current drinkers was lower in the usually insomnia group (85.0%) than in the never (89.2%) and sometimes (89.6%) groups. The median diet quality score (IQR) was significantly higher in the never insomnia group (4.00 [3.00, 5.00]) than in the sometimes and usually insomnia groups (both 3.00 [2.00, 4.00]; p = 0.011). Physical activity (MET) did not differ significantly among the groups (p > 0.05). Clinically, the use of antidepressants and sleeping pills varied with insomnia frequency. The prevalence of antidepressant use was 39.2% in the never insomnia group, 42.5% in the sometimes insomnia group, and 47.8% in the usually insomnia group; the prevalence of sleeping pill use was notably lower in the never insomnia group (0.6%) than in the sometimes (2.3%) and usually (6.7%) insomnia groups. For the biochemical and anthropometric indicators, significant differences were found in the HbA1c level, TG level, and BMI (all p < 0.001). The overall all-cause mortality rate was 12.0%, which was significantly different across groups (p < 0.001); the rates were 9.6% in the never insomnia group, 9.9% in the sometimes insomnia group, and 14.9% in the usually insomnia group. The detailed baseline characteristics of the participants are presented in Table 1. Associations between insomnia and mortality As presented in Table 2, the cumulative incidence of mortality increased with increasing insomnia frequency (log-rank p < 0.001; Fig. 1A). In the unadjusted model (Model 1), with the “never” group as a reference, the association between “sometimes” insomnia and mortality was not significant (HR: 1.02; 95% CI: 0.77–1.36), whereas the hazard ratio for mortality was 1.60 (95% CI: 1.22–2.09) for “usually” insomnia. After adjustment for age, sex and ethnicity in Model 2, “usually” insomnia was significantly associated with an increased risk of mortality (HR: 1.64; 95% CI: 1.25–2.15). After adjusting for all confounding variables (Model 3), “usually” insomnia remained an independent predictor of mortality (HR: 1.38; 95% CI: 1.05–1.81; P for trend = 0.001; Table 2). Furthermore, compared with patients who never experienced insomnia, patients with frequent insomnia had a shorter life expectancy. Specifically, at 45 years of age, patients with frequent insomnia had a shortened life expectancy by 2.3 years (Fig. 1B). Table 2Associations between insomnia, insomnia-related proteomic signature and all-cause mortalityModel 1Model 2Model 3HR (95% CI)P valueHR (95% CI)P valueHR (95% CI)P valueInsomnia statusNeverReferenceReferenceReferenceSometimes1.02(0.77,1.36)0.8691.03(0.78,1.37)0.8170.99(0.75,1.32)0.959Usually1.60(1.22,2.09)<0.0011.64(1.25,2.15)<0.0011.38(1.05,1.81)0.023P for trend<0.001<0.0010.001Proteomic characteristicsT1ReferenceReferenceReferenceT21.84(1.41,2.42)<0.0011.48(1.13,1.94)0.0051.36(1.03,1.79)<0.001T34.12(3.23,5.26)<0.0013.28(2.57,4.19)<0.0012.54(1.95,3.32)<0.001P for trend<0.001<0.001<0.001Note: T1, T2, T3: Tertiles of the proteomic signature (T1 = lowest tertile, T2 = middle tertile, T3 = highest tertile)Abbreviations: SES: Socioeconomic status; MET: Metabolic equivalent of task; HbA1c: Glycated haemoglobin; Cr: Creatinine; TG: Triglycerides; TC: Total cholesterol; SBP: Systolic blood pressure; BMI: Body mass index. Model 1 was unadjusted. Model 2 was adjusted for age, sex and ethnicity. Model 3 was adjusted for age, sex, ethnicity, smoking status, alcohol consumption, HbA1c, SBP, Cr, TG, TC, BMI, SES, diet quality, physical activity MET, antidepressants, and sleeping pills Associations between insomnia, insomnia-related proteomic signature and all-cause mortality Note: T1, T2, T3: Tertiles of the proteomic signature (T1 = lowest tertile, T2 = middle tertile, T3 = highest tertile) Abbreviations: SES: Socioeconomic status; MET: Metabolic equivalent of task; HbA1c: Glycated haemoglobin; Cr: Creatinine; TG: Triglycerides; TC: Total cholesterol; SBP: Systolic blood pressure; BMI: Body mass index. Model 1 was unadjusted. Model 2 was adjusted for age, sex and ethnicity. Model 3 was adjusted for age, sex, ethnicity, smoking status, alcohol consumption, HbA1c, SBP, Cr, TG, TC, BMI, SES, diet quality, physical activity MET, antidepressants, and sleeping pills Kaplan‒Meier curves and life expectancy. (A) Kaplan‒Meier curves of all-cause mortality in patients with different frequencies of insomnia. (B) the reduction in life expectancy among patients with different frequencies of insomnia. (C) Kaplan‒Meier curves of all-cause mortality in patients with different proteomic characteristics related to insomnia. (D) the reduction in life expectancy among patients with different proteomic characteristics related to insomnia. T1, T2, T3: tertiles of the proteomic signature (T1 = lowest tertile, T2 = middle tertile, T3 = highest tertile) Proteomic signatures of insomnia After adjusting for all confounding factors, insomnia was significantly associated with 43 proteins (Supplementary Table S1). After further selection using LASSO regression, 20 of these proteins remained significantly associated with insomnia (Supplementary Table S2). The relationships between these 20 proteins and the risk of death in patients with depression are depicted in Fig. 2A and Supplementary Table S3. As shown in Table 2, the proteomic signature of insomnia derived from these 20 proteins was also significantly associated with the risk of death. After adjustment for all covariates, the highest risk tertile (T3) exhibited a more than twofold increased risk of death (HR: 2.54, 95% CI: 1.95–3.32; P for trend < 0.001; Table 2). Participants in the group with higher insomnia-related proteomic signatures had higher all-cause mortality rates and significantly shorter life expectancies (Fig. 1C and D). Bootstrap internal validation yielded a mean C-index of 0.68 (95% CI: 0.65–0.70). ROC analysis further demonstrated an AUC of 0.69 (95% CI: 0.66–0.71) for mortality prediction using the continuous proteomic risk score, which was consistent with the C-index estimate (Fig. 3). ROC analyses demonstrated that the insomnia-related proteomic signature (AUC = 0.69) achieved comparable discriminative performance to age (AUC = 0.69) and outperformed all other individual clinical indicators, including BMI, HbA1c, SBP, TG, TC, and creatinine (Fig. 4). To further evaluate the incremental predictive value of the proteomic signature beyond established clinical risk factors, we compared a Clinical model (adjusted for all covariates in Model 3) with a Clinical + Proteomic model. Adding the proteomic signature significantly improved the overall C-index from 0.75 (95% CI: 0.73–0.77) to 0.76 (95% CI: 0.74–0.78, p < 0.001). More notably, the time-dependent AUC consistently increased across all evaluated time horizons: from 0.76 (95% CI: 0.71–0.81) to 0.79 (95% CI: 0.74–0.84) at 5 years (p = 0.010), from 0.76 (95% CI: 0.73–0.79) to 0.78 (95% CI: 0.75–0.81) at 10 years (p = 0.010), and from 0.77 (95% CI: 0.74–0.81) to 0.79 (95% CI: 0.76–0.83) at 15 years (p = 0.008; Supplementary Table S4). The IDI and cNRI were both statistically significant at all three time horizons (IDI: 0.02, 0.02, and 0.03 at 5, 10, and 15 years, respectively, all p ≤ 0.05; cNRI: 0.34, 0.21, and 0.22 at 5, 10, and 15 years, respectively, all p < 0.05; Supplementary Table S5), indicating that the proteomic signature provides meaningful incremental prognostic information beyond conventional clinical indicators. Associations between insomnia-related proteins and all-cause mortality and KEGG enrichment analysis. (A) association between insomnia-related proteins and all-cause mortality. (B) mediating effect of insomnia-related proteins between insomnia and all-cause mortality. (C) KEGG enrichment analysis of mediator proteins Bootstrap internal validation and ROC analysis for predicting all-cause mortality using the insomnia-related proteomic signature. (A) bootstrap internal validation analysis for predicting all-cause mortality using the insomnia-related proteomic signature. (B) ROC analysis for predicting all-cause mortality using the insomnia-related proteomic signature ROC curves comparing the discriminative performance of the insomnia-related proteomic signature and individual clinical indicators for predicting all-cause mortality Subgroup analyses Subgroup analyses were performed across ethnicity, antidepressant treatment status, chronotype, and sleep duration to assess the consistency of the associations of insomnia and the insomnia-related proteomic signature with all-cause mortality (Supplementary Table S6). The associations were broadly consistent across ethnicity, antidepressant treatment, and chronotype strata (all P for interaction > 0.05). A significant interaction was observed with sleep duration for both insomnia (P for interaction = 0.019) and the proteomic signature (P for interaction = 0.046), with stronger associations among participants with a sleep duration of less than 7 hours. The predictive performance of the insomnia-related proteomic signature did not differ significantly between White and non-White individuals (AUC = 0.68, 95% CI: 0.65–0.71 in White; AUC = 0.74, 95% CI: 0.67–0.81 in non-White; DeLong p = 0.086, Supplementary Fig. S4). Cause-specific mortality analysis using a competing risks framework Among participants with frequent insomnia (usually), a significant association was observed with CVD mortality (sHR = 2.14; 95% CI: 1.12–4.07; p = 0.021), whereas no significant associations were detected with cancer mortality (sHR = 1.20; 95% CI: 0.75–1.94; p = 0.451) or other-cause mortality (sHR = 1.17; 95% CI: 0.79–1.75; p = 0.428). The proteomic risk score demonstrated associations across all three competing endpoints: CVD death (T3 vs. T1: sHR = 6.70; 95% CI: 3.36–13.35; p < 0.001), cancer death (sHR = 2.83; 95% CI: 1.78–4.49; p < 0.001), and other-cause death (sHR = 2.50; 95% CI: 1.71–3.66; p < 0.001). The detailed results are presented in Table 3. Table 3Cause-specific mortality analysis using a competing risks frameworkStatusOutcomesHR95% CIP valueSometimesCVD death1.580.82–3.050.178UsuallyCVD death2.141.12–4.070.021Proteomic characteristics T2CVD death2.691.31–5.520.007Proteomic characteristics T3CVD death6.703.36–13.35<0.001SometimesCancer death0.890.54–1.460.650UsuallyCancer death1.200.75–1.940.451Proteomic characteristics T2Cancer death1.520.94–2.440.086Proteomic characteristics T3Cancer death2.831.78–4.49<0.001SometimesOther death0.870.58–1.310.492UsuallyOther death1.170.79–1.750.428Proteomic characteristics T2Other death1.450.98–2.130.062Proteomic characteristics T3Other death2.501.71–3.66<0.001Adjusted for age, sex, ethnicity, smoking status, alcohol consumption, HbA1c, SBP, Cr, TG, TC, BMI, SES, diet quality, physical activity MET, antidepressants, and sleeping pills Cause-specific mortality analysis using a competing risks framework Adjusted for age, sex, ethnicity, smoking status, alcohol consumption, HbA1c, SBP, Cr, TG, TC, BMI, SES, diet quality, physical activity MET, antidepressants, and sleeping pills Mediation analysis of proteins Among the 20 proteins selected by LASSO regression, 15 were identified as significant mediators of the association between insomnia and mortality. The key mediators included TNFRSF6B, CSTB, LGALS4, and CCL22, which had median proportion estimates of 14.44%, 14.26%, 12.08%, and 10.58%, respectively (Fig. 2B and Supplementary Table S7). KEGG pathway enrichment analysis performed on these 15 mediators revealed significant enrichment in several biological processes, including viral protein interactions with cytokines and cytokine receptors, cytokine−cytokine receptor interactions, chemokine signalling pathways, phenylalanine metabolism and histidine metabolism (Fig. 2C). To address the substantial correlations among the LASSO-selected proteins and the potential overestimation of mediating contributions in single-mediator analyses, a multi-mediator joint analysis was further performed (Supplementary Table S8). The joint indirect effect mediated by all 20 LASSO-selected proteins was statistically significant, accounting for 37.2% of the total effect of insomnia on mortality, while the direct effect contributed the remaining 62.8%. Among the 20 proteins, only TNFRSF6B, CSTB, and LGALS4 retained statistically significant independent mediating contributions after accounting for between-protein correlations (all p < 0.05; Supplementary Table S9). Pairwise interaction analyses among these three proteins revealed a significant synergistic interaction between TNFRSF6B and LGALS4 (HR for interaction = 1.175, 95% CI: 1.03–1.34, p = 0.019), whereas the remaining pairs (TNFRSF6B–CSTB and CSTB–LGALS4) showed no significant interaction (Supplementary Table S10). Sensitivity analyses Sensitivity analyses excluding participants who died within the first 2 years of follow-up yielded results consistent with the primary findings: the associations of insomnia and the insomnia-related proteomic signature with all-cause mortality were materially unchanged (Supplementary Tables S12–S14). The discriminative performance of the proteomic signature, both alone and incremental to clinical risk factors, was likewise preserved (Supplementary Tables S15–S16; Supplementary Figs. S5–S6). Discussion Our findings demonstrate that insomnia is an independent risk factor for mortality in depression. To explore potential biological mediators, we performed a large-scale proteomic analysis. Single-mediator analyses identified 15 proteins as statistical mediators of the insomnia–mortality association, and these proteins were enriched in biological pathways including cytokine–cytokine receptor interactions and chemokine signalling. Multi-mediator joint analysis further confirmed that TNFRSF6B, CSTB, and LGALS4 retained independent mediating contributions after accounting for inter-protein correlations. The insomnia-related proteomic signature demonstrated significant incremental predictive value for mortality beyond conventional clinical risk factors. Several studies have investigated the relationship between insomnia and depression. Li et al. reported an association between insomnia and depression risk in a meta-analysis involving thirty-four cohort studies [22]. Subsequent larger-scale studies conducted by Chen et al. and network analyses by Zhang et al. confirmed this relationship [23, 24]. A bidirectional relationship hypothesis posits that insomnia constitutes a core manifestation of depression and represents an independent risk factor for disease exacerbation [25, 26]. Longitudinal epidemiological investigations have clarified the link between comorbid insomnia and depression [27]. However, the molecular mechanisms underlying this association remain unexplored. Proteomics has emerged as a powerful tool for understanding disease mechanisms at the molecular level. Proteomic analyses have been instrumental in elucidating the underlying molecular alterations in the pathogenesis of depression [28], offering valuable insights into disease development and potential interventions [29]. Therefore, our study utilized large-scale longitudinal proteomic data to investigate the potential mechanisms linking insomnia to mortality in patients with depression. In this study, “sometimes insomnia” was not associated with the risk of death, but “usually insomnia” was significantly correlated with the risk of death. This gradient is consistent with prior evidence [30] and reflects a threshold effect, whereby only frequent insomnia confers clinically meaningful mortality risk or may be attributable to limited statistical power to detect smaller effect sizes at intermediate exposure levels [31]. Our study revealed a set of insomnia‑related proteins associated with mortality in depression, among which TNFRSF6B, CSTB, and LGALS4 were identified as the principal independent mediators in the multi‑mediator joint analysis. TNFRSF6B is a soluble decoy receptor that neutralizes FasL, LIGHT, and TL1A, and is conventionally regarded as an anti‑inflammatory mediator [32]. However, elevated circulating TNFRSF6B levels have been consistently observed in states of chronic systemic inflammation, cardiovascular disease, and depression, where they may reflect a compensatory but ultimately insufficient counter‑regulatory response to sustained inflammatory activation [33–35]. Thus, elevated TNFRSF6B in our cohort likely serves as a biomarker of underlying inflammatory burden rather than a direct pro‑inflammatory driver. CSTB, a cysteine protease inhibitor involved in neuronal homeostasis and apoptosis regulation, has been associated with neurodegenerative, cardiovascular, and malignant diseases in prior observational studies [36–40]. LGALS4, a galectin family member predominantly expressed at synapses, has been linked to brain ageing in proteomic studies; given the observed correlation between accelerated brain ageing and systemic biological ageing, this protein may represent a molecular marker of age‑related multi‑organ decline [41–43]. The multi‑mediator joint analysis revealed that the 20 LASSO‑selected proteins collectively accounted for 37.2% of the total effect of insomnia on mortality, a proportion substantially lower than the sum of individual mediating effects from single‑mediator analyses. This discrepancy confirms that the single‑mediator estimates are inflated by the strong inter‑protein correlations within shared inflammatory and immune pathways. After accounting for these correlations, only TNFRSF6B, CSTB, and LGALS4 retained statistically significant independent contributions, suggesting that these three proteins capture non‑redundant biological information linking insomnia to mortality. Pairwise interaction analyses further revealed a significant synergistic interaction between TNFRSF6B and LGALS4, suggesting that inflammatory dysregulation (indexed by TNFRSF6B) and accelerated biological ageing (indexed by LGALS4) may co-occur with elevated mortality risk, whereas the remaining protein pairs exhibited additive effects. One possible explanation is that chronic inflammation has been linked to accelerated biological ageing in prior studies, and the co-occurrence of both processes in the context of insomnia may be paralleled by elevated mortality risk. In summary, the circulating levels of these mediator proteins are associated with biological perturbations spanning central nervous system function and systemic physiological homeostasis, including inflammation, vascular dysfunction, and accelerated ageing. Notably, this proteomic signature may represent either a biological consequence of insomnia, a preexisting risk factor, or a combination of both. The single baseline measurement of proteomic markers precludes definitive causal inferences regarding the temporal directionality of these associations. Future longitudinal studies with repeated measurements are warranted to clarify temporal relationships. On the basis of KEGG enrichment analysis performed on the 15 proteins identified as significant mediators in the single-mediator analyses, these proteins primarily participate in viral protein interactions with cytokines and cytokine receptors, cytokine−cytokine receptor interactions, chemokine signalling pathways, phenylalanine metabolism and histidine metabolism. These findings suggest that insomnia may be associated with alterations in immune-inflammatory responses and neurotransmitter metabolism, potentially involving neuroinflammation, neurotransmitter imbalances, and an increased risk of systemic somatic diseases. Our results confirm the previously observed association between comorbid depression–insomnia and elevated inflammatory biomarkers, an association that persists after adjusting for covariates such as age, sex, and BMI [44]. Moreover, our findings are consistent with those of prior research indicating that the combination of depressive symptoms and increased inflammation significantly increases the risk of mortality [45]. Notably, the identified proteins were simultaneously enriched in both inflammatory and amino acid metabolic pathways, suggesting that these processes may be co-regulated and cooperatively involved in the biological alterations associated with insomnia. Dysregulated phenylalanine metabolism has been linked to neurotransmitter imbalance and a neurotoxic environment, which may impair brain function and worsen psychiatric outcomes [46]. Moreover, in addition to its well-established central role in sleep–wake regulation via histamine signalling [47], the dysregulation of histidine metabolism has significant peripheral implications. The systemic presence of metabolites such as imidazole propionate is implicated in the pathogenesis of type 2 diabetes, obesity, and the progression of atherosclerotic, renal, and heart failure conditions [48]. Collectively, these pathways suggest potential biological links between insomnia and adverse clinical outcomes, including the progression of depression and elevated mortality risk. Notably, the KEGG pathway enrichment findings presented herein are hypothesis-generating and await confirmation through experimental validation in cellular or animal models. Subgroup analyses revealed a significant interaction between insomnia and sleep duration, with the insomnia–mortality association being more pronounced among participants sleeping less than 7 hours per night. This finding is consistent with prior evidence that insomnia combined with short sleep duration represents a more biologically severe phenotype associated with greater mortality and cardiovascular risk [49, 50]. The combination of subjective sleep complaints and insufficient sleep duration may reflect a state of heightened physiological hyperarousal characterized by greater hypothalamic–pituitary–adrenal axis activation and sympathetic overdrive, thereby conferring a disproportionately elevated mortality risk. These results highlight the importance of jointly assessing insomnia symptoms and sleep duration in clinical risk evaluation for patients with depression. Fine–Gray subdistribution hazard analyses revealed that insomnia was associated with CVD mortality but not with cancer or other-cause mortality. These findings are consistent with the well-established mechanistic links between insomnia and cardiovascular dysregulation, including sympathetic nervous system hyperactivation, HPA axis dysregulation and endothelial dysfunction [7, 8], all of which converge on elevated cardiovascular risk. The absence of a significant association with cancer mortality may reflect the longer latency of carcinogenesis relative to the follow-up period or that insomnia per se is more weakly associated with cancer-specific mortality in the context of depression [51]. Notably, the subdistribution hazard ratios for insomnia were directionally elevated across all three cause-specific endpoints (sHR = 2.14, 1.20, and 1.17 for CVD, cancer, and other-cause mortality, respectively), suggesting a consistent direction of effect; the failure to reach statistical significance for cancer and other-cause mortality likely reflects the modest effect sizes of a behavioral exposure combined with limited event counts in these subcategories. In contrast, the insomnia-related proteomic risk score demonstrated statistically significant associations across all three competing endpoints, with the strongest association observed for CVD mortality (sHR = 6.70) followed by cancer (sHR = 2.83) and other-cause mortality (sHR = 2.50). This gradient parallels the pattern observed for insomnia itself. The proteomic signature, comprising biologically multifunctional proteins involved in inflammation, immune regulation, and cellular ageing, captures systemic perturbations that are influenced by but not exclusive to insomnia. These proteins participate in shared upstream processes that are recognized as common antecedents of diverse fatal diseases, which accounts for their broader predictive scope relative to the behavioural exposure alone. Taken together, these competing-risk analyses support the interpretation that CVD represents the primary cause-specific pathway through which insomnia is most strongly associated with mortality, while the proteomic signature serves as a disease-agnostic biomarker reflecting cumulative biological risk in patients with depression and insomnia. In this study, the insomnia-related proteomic signature demonstrated significant incremental predictive value for all-cause mortality, with significant improvements in the C-index, time-dependent AUC, IDI, and cNRI. Notably, the discriminative performance of the proteomic signature was comparable to that of age and superior to all other individual clinical indicators, including BMI, HbA1c, SBP, TG, TC, and creatinine. These findings support its potential value as a biomarker for mortality risk stratification in patients with depression and insomnia. Limitations There were several limitations of our study that should be acknowledged. First, insomnia was assessed exclusively through self-report questionnaires using a three-category ordinal scale (never/sometimes/usually) without standardized frequency definitions. This reliance on subjective data renders the evaluation susceptible to misclassification bias. In particular, depressed individuals with greater symptom severity or cognitive negativity biases may systematically over-report insomnia symptoms, representing a form of differential misclassification that could inflate rather than attenuate effect estimates. Furthermore, this study did not distinguish insomnia subtypes (sleep-onset difficulty, sleep-maintenance difficulty, and early-morning awakening), which may have distinct biological profiles and differential associations with mortality outcomes. To partially address this limitation, we performed subgroup analyses stratified by chronotype and sleep duration as two complementary sleep-related biological dimensions; however, these cannot fully substitute for clinical insomnia subtyping. Refined subtyping of insomnia and objective sleep monitoring remain important directions for future large-scale clinical cohort studies. Second, the definition of depression was derived from ICD codes and medical records, without distinguishing episode severity, duration, or recurrence or excluding bipolar disorder depressive episodes. Since more severe depression may independently increase both insomnia frequency and mortality risk, unmeasured variation in depressive severity represents a potential source of residual confounding. Subgroup analyses using antidepressant treatment patterns as a proxy for severity showed no significant interaction, but future studies with validated severity scales are needed to further clarify this issue. Third, the single baseline assessment of both insomnia and proteomic profiles prevents the establishment of temporal ordering and limits causal inference. The mediation estimates reported herein should be interpreted as statistical decompositions under counterfactual assumptions rather than as direct evidence of biological causation, as the sequential ignorability assumption cannot be verified with single‑timepoint data. Fourth, despite adjusting for a comprehensive set of covariates, residual confounding cannot be fully ruled out. Finally, the generalizability of our findings is constrained by the UK Biobank cohort, which overwhelmingly comprises individuals of European ancestry. Although our ethnicity‑stratified analyses showed no significant interaction, the non‑White subgroup was relatively small, and differences in genetic background, lifestyle, and social determinants across populations may modify the observed associations. Thus, validation in more diverse and multiethnic cohorts is warranted. Conclusion Insomnia is associated with increased mortality risk in patients with depression. Multi-mediator analysis identified TNFRSF6B, CSTB, and LGALS4 as independent mediators of this association, and the insomnia-related proteomic signature may serve as a potential biomarker for risk stratification. Electronic supplementary material Below is the link to the electronic supplementary material. Supplementary Material 1: Supplemental Figure S1-6 Supplementary Material 2: Supplemental Table S1-16 Acknowledgements Not applicable. Author contributions YSG designed the research and is the guarantor of this work, with full access to all data and responsibility for data integrity and analysis accuracy. XSL conducted the analysis and wrote the first draft. HLX, XQZ, JBT and KYL revised the manuscript. All authors read and approved of the final manuscript. Funding This work was supported by the National Natural Science Foundation of China Youth Program (Grant number: 82300546), Yunnan Fundamental Research Projects (Grant number: 202401CF070013), National Natural Science Foundation of China General Program (Grant number: 82470335, 82171569), Natural Science Foundation of Fujian Province (Grant number: 2023J01176) and Youth Top Talent Project of Fujian Provincial Foal Eagle Program; Fujian Research and Training Grants for Young and Middle-aged Leaders in Health care. Data availability The data that support the findings of this study are available from the UK Biobank but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available. Declarations Human ethics and consent to participate declarations All UK Biobank participants provided written informed consent, and the North West Multi-centre Research Ethics Committee granted ethical approval (June 17, 2011 [11/NW/0382]; May 13, 2016 [16/NW/0274]). This research was performed in accordance with the principles of the Declaration of Helsinki and using the UK Biobank Resource (application No. 105,322). Further details about the UK Biobank can be found on the showcase website (https://biobank.ndph.ox.ac.uk/showcase/). Consent to participate All participants agreed to participate in the UKB survey. Consent for publication All the authors agreed to publish the manuscript. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Xuesong Liu, Hongli Xu and Xiaoqian Zhu contributed equally to this work. Contributor Information Kaiyang Lin, Email: lky7411@sina.com. Yansong Guo, Email: ysguo1234@126.com. References Associated Data Supplementary Materials Supplementary Material 1: Supplemental Figure S1-6 Supplementary Material 2: Supplemental Table S1-16 Data Availability Statement The data that support the findings of this study are available from the UK Biobank but restrictions apply to the availability of these data, which were used under license for the current study, and so are not publicly available.