Study populations
The association analysis of metabolite levels with depression was performed in 13,596 participants separately recruited in five different cohort studies. The following cohort studies were included: the Rotterdam Study (RS), the Study of Health in Pomerania (SHIP-TREND), the Cooperative Health Research in the Region of Augsburg (KORA) study, the European Prospective Investigation into Cancer (EPIC)-Norfolk Study, and the Netherlands Epidemiology of Obesity (NEO) study. Detailed information on these cohorts is provided in the Supplementary Materials. All participants provided written informed consent, studies were approved by their local ethics committees and conformed to the principles of the declaration of Helsinki. Patients or the public were not involved in the design, or conduct, or reporting, or dissemination plans of our research
Association of depression with the dietary sources of the depression-associated metabolites was performed in the UK Biobank study. UK Biobank is a prospective cohort study including ~ 500,000 participants aged 40-69 years at baseline recruited between 2006 and 2010. The aim of the study is to investigate the effects of genetic and environmental factors on the risk of common multifactorial diseases. Participants have provided a detailed information on lifestyle, medical history and nutritional habits; basic variables such weight, height, blood pressure etc. were measured; and blood and urine samples were taken. Detailed information about the cohort is provided in the Supplementary Materials.
To ascertain the effects of various depression treatments including cognitive behavioural therapy (CBT) and antidepressants SSRI (escitalopram) and SNRI (duloxetine) on the depression-associated metabolites we performed a lookup in the PReDICT study. The design of PReDICT study has been published previously.(28) Details on the study and the metabolomics assessments are provided in the Supplementary Materials.
To select instruments/proxies for metabolites for Mendelian Randomization we used the results of the genome-wide association study (GWAS) performed using the NIHR BioResource (NBR). NIHR BioResource (NBR) – Rare Disease Study is a multi-center whole-exome and whole-genome sequencing study including up to 13,600 patients (http://bioresource.nihr.ac.uk/rare-diseases/rare-diseases/). The NBR–Rare Diseases study was approved by the East of England Cambridge South national research ethics committee (REC) under reference number: 13/EE/0325. The inclusion and exclusion criteria, as well as other steps of quality control, adjustment and transformations followed the same analytical steps as described before.(29)
Depression assessment
In the RS, depressive symptoms were assessed with the 20-item version of the Centre for Epidemiologic Studies Depression (CES-D) scale, a self-report measure of depressive symptoms experienced during the prior week.(30) The total score ranges from 0 – 60, where a higher score indicates more depressive symptoms. In the SHIP-trend and KORA cohorts, depressive symptoms were assessed with the Patient Health Questionnaire 9 (PHQ-9),(31) where each of the nine DSM-IV criteria for depression are scored from 0 – 3. The total score ranges from 0 – 27 where higher score indicates a greater depression severity. In KORA a brief interview version of PHQ-9 called Patient Health Questionnaire Depression (PHQ-D) module was used to measure depression.(31, 32) In the EPIC-Norfolk study depression was assessed using the following question: “Has the doctor ever told you that you have any of the following: depression requiring treatment?” with answers “yes” or “no”. In the NEO cohort, depressive symptoms were assessed using the Inventory Depressive Symptomatology Self Report questionnaire (IDS-SR30),(33) which assesses specific depressive symptoms (via a 4-level response system) during the last week and their severity. The total score ranges from 0 to 84, with higher scores indicating higher severity. Thus, in all cohorts except EPIC-Norfolk, depression in participants was measured on a quantitative scale and used as such in the analysis.
In the UKB study, we used the derived lifetime probable major depressive disorder measure as described in Smith et al. 2013.(34) We further defined current depressive symptoms by summing the responses to four questions related to mood in the past two weeks. These include, (1) Over the past two weeks, how often have you felt down, depressed or hopeless?, (2) Over the past two weeks, how often have you had little interest or pleasure in doing things?, (3) Over the past two weeks, how often have you felt tense, fidgety or restless? and (4) Over the past two weeks, how often have you felt tired or had little energy? Answers could be given on a four-point scale ranging from 0-3 (0 = not at all, 1 = several days, 2 = more than half of the days and 3 = nearly every day). The total score ranged from 0-12 where higher score indicating more severe depression.
In the PReDICT study, participants were treatment-naive adults defined as having never previously received a minimally adequate course of treatment with an antidepressant medication or evidence-based psychotherapy for a mood disorder, aged 18 to 65 years with moderate-to-severe, non-psychotic MDD depression as assessed by the Structured Clinical Interview for DSM-IV(35) and a psychiatrist’s evaluation, and if they scored ≥18 on the HRSD17. Eligible patients were randomized equally to one of three 12-week treatment arms: (1) cognitive behavior therapy (CBT, 16 sessions); (2) duloxetine (30–60 mg/d); or (3) escitalopram (10–20 mg/d).
Metabolomics measurements
In all studies, the metabolome was quantified using the Metabolon platform (Metabolon Inc., Durham, USA). Different versions of the platform have been used and details on the platforms are included in the Supplementary Materials. In all studies, metabolites with ≥ 40% missing values were removed and for the remaining metabolites missing metabolite values were replaced with half of the detection limit for that particular metabolite.(36) Subsequently, a natural logarithm transformation was applied to all metabolites and metabolites were scaled to standard deviation units.
In the PReDICT study, metabolites were quantified using targeted metabolomics platforms including ultra-performance liquid chromatography triple quadrupole mass spectrometry (UPLC-TQMS) (Waters XEVO TQ-S, Milford, USA) and gas chromatography time-of-flight mass spectrometry (GC-TOFMS) (Leco Corporation, St Joseph, USA). Metabolites with >20% missing values were excluded. Then, metabolites were log-transformed, imputed and scaled to mean zero and variance 1. Details are provided in the Supplementary Materials.
Non-targeted metabolite detection and quantification was conducted by the metabolomics provider Metabolon, Inc. (Durham, USA) on fasting plasma samples of 10,654 participants from the UK Bioresource. The metabolomic dataset measured by Metabolon included 1069 compounds of known structural identity belonging to the following broad categories - amino-acids, peptides, carbohydrates, energy intermediates, lipids, nucleotides, cofactors and vitamins, and xenobiotics. Metabolites data were day-median normalized, and inverse normalized, as the metabolite concentrations were not normally distributed. Metabolic traits with more than 20% missing values were excluded leaving 722 metabolites of known chemical identity for analysis.
Genotyping
For the GWAS of metabolites, genotyping in the UK bioresource was carried out with a high-density array data (Affymetrix UK Biobank Axiom® Array). Genotypes were subsequently imputed using information from the Human Reference Consortium imputation panel (version r1.1, 2016).(37) Only individuals of full European ancestry (N=8,809) were included in the analyses in the discovery cohort.
Statistical analyses
Metabolites association analysis
All cohorts used linear regression analysis to test the association between the metabolite levels (dependent variable) and depression. Three different models were tested, where the first model (model 1) was adjusted for age and sex only, the second model (model 2) was additionally adjusted for antidepressant medication usage, and the third model was an extension of the second model (model 3) with additional adjustment for lipid-lowering medication (yes/no), antihypertensive medication (yes/no), antidiabetic medication (yes/no), BMI (kg/m2), and current smoking (yes/no). The summary statistics from all cohorts were combined in a sample size-weighted meta-analysis using METAL software.(38) Sample size weighted meta-analysis was used since the depression measurement scales were different among cohorts. Only metabolites that were present in two or more studies were included. To investigate the robustness of our findings, a sensitivity analysis was performed by including only cohorts that assessed metabolites with the most recent version of the Metabolon platform (HD4).
Association analysis of major depressive disorder with dietary sources of the metabolites in the UK Biobank
We used logistic regression analysis to test the association between major depressive disorder and dietary sources of metabolites (vitamin A supplements, retinol intake estimated from food, fresh fruits intake and vitamin K antagonists). Age, sex and principal components were used as covariates in the analysis. For the association of current depressive symptoms, we used linear regression analysis. We further tested the association of volume of white matter hyperintensities (WMH) with vitamin supplements to ascertain the impact of these supplements on brain pathology. Linear regression analysis was used with the volume of WMH as the dependent variable, vitamin supplements as the independent variable, and age, sex, BMI, head size and principal components as covariates. All analyses were performed in R.
Metabolite GWAS for Mendelian Randomization (MR) analysis
To test for association between metabolite levels and genotypes, we built linear regression models where the outcome was defined as the transformed level of each metabolite, predicted by the allele dosage at each polymorphic (MAF > 0.01) genotyped or imputed genetic variant. In addition, analyses were adjusted for age, sex and BMI. All analyses were conducted using the PLINK software (https://www.cog-genomics.org/plink/2.0/).
Mendelian Randomization (MR) analysis
To understand the relationship between the identified metabolites and major depression we performed bidirectional two-sample MR analysis. For major depression we used the independent genome-wide significant single nucleotide polymorphisms (SNPs) reported by Howard et al. 2019(39) as instrumental variables (IVs). Summary statistics for these IVs were extracted from Howard et al. The summary statistics for the metabolites were extracted from the GWAS performed in UK Bioresource. Of the identified metabolites in this study (model 3), GWAS results were available for six metabolites including 2-aminooctanoate, 10-undecenoate (11:1n1), 1-palmitoyl-2-palmitoleoyl-GPC (16:0/16:1), hippurate, mannitol/sorbitol and retinol (Supplementary Table 1). The IVs for these six metabolites and their summary statistics were extracted from the same GWAS. Because of scarcity in GWAS-grade significance for SNPs associated with these metabolites, we used independent SNPs that showed the strongest association with a p-value < 10-06 as instruments (Supplementary Table 2). The summary statistics for depression for these IVs were extracted from the publicly available dataset (2019 PGC UKB Depression Genome-wide; https://www.med.unc.edu/pgc/download-results/mdd/). For the analysis we used the ‘mr_allmethods’ option of the R (https://cran.r-project.org/) library “MendelianRandomization”(40) that reports the results from the median method (simple, weighted and penalized), Inverse variance weighted and Egger methods (penalized, robust and penalized & robust).
Effect of antidepressant therapy on metabolites in PReDICT study
To examine the strength and significance of metabolite concentration changes within each of the three treatment arms, i.e., (1) CBT (16 sessions); (2) duloxetine (30–60 mg/d); or (3) escitalopram (10–20 mg/d), linear mixed effect models (with random intercept) with metabolite levels (in log scale) as the dependent variable, were fitted while correcting for age, sex, BMI, and baseline HRSD17. Then, the R package “emmeans” was used to compute the least squared means of the contrasts of interest (week 12 vs. baseline) and their corresponding p-values.
To detect whether metabolites levels were associated with clinical outcomes, linear regression analyses were conducted corrected for age, sex and treatment arm. Dependent variables (Baseline HRSD17, Week 12 HRSD17, and 12 weeks change in HRSD17) were regressed on either of following independent variables: 1) baseline metabolite, 2) week 12 metabolite, 3) 2 weeks change in metabolites and 4) 12 weeks change in metabolites.
Linking metabolites to human and/or gut metabolism
To assess whether the identified metabolites are products of human metabolism, gut microbial metabolism, or both, we integrated our findings with those of the Virtual Metabolic Human (VMH) and Assembly of Gut Organisms through Reconstruction and Analysis (AGORA2) databases. Additional information is provided in the Supplementary Materials.