Author response: Epigenetic scores for the circulating proteome as tools for disease prediction Article Swipe

MAKE IMPACT

Danni A. Gadd , Robert F. Hillary , Daniel L. McCartney , Shaza B. Zaghlool , Anna J. Stevenson , Yipeng Cheng , Chloe Fawns‐Ritchie , Clifford Nangle , Archie Campbell , Robin Flaig , Sarah E. Harris , Rosie M. Walker , Liu Shi , Elliot M. Tucker‐Drob , Christian Gieger , Annette Peters , Mélanie Waldenberger , Johannes Graumann , Allan F. McRae , Ian J. Deary , David J. Porteous , Caroline Hayward , Peter M. Visscher , Simon R. Cox , Kathryn L. Evans , Andrew M. McIntosh , Karsten Suhre , Riccardo E. Marioni · YOU? · · 2021 · Open Access · · DOI: https://doi.org/10.7554/elife.71802.sa2

Article Figures and data Abstract Editor's evaluation eLife digest Introduction Results Discussion Materials and methods Data availability References Decision letter Author response Article and author information Metrics Abstract Protein biomarkers have been identified across many age-related morbidities. However, characterising epigenetic influences could further inform disease predictions. Here, we leverage epigenome-wide data to study links between the DNA methylation (DNAm) signatures of the circulating proteome and incident diseases. Using data from four cohorts, we trained and tested epigenetic scores (EpiScores) for 953 plasma proteins, identifying 109 scores that explained between 1% and 58% of the variance in protein levels after adjusting for known protein quantitative trait loci (pQTL) genetic effects. By projecting these EpiScores into an independent sample (Generation Scotland; n = 9537) and relating them to incident morbidities over a follow-up of 14 years, we uncovered 137 EpiScore-disease associations. These associations were largely independent of immune cell proportions, common lifestyle and health factors, and biological aging. Notably, we found that our diabetes-associated EpiScores highlighted previous top biomarker associations from proteome-wide assessments of diabetes. These EpiScores for protein levels can therefore be a valuable resource for disease prediction and risk stratification. Editor's evaluation This is an important study that demonstrates the potential utility of the circulating proteome for disease prediction and risk stratification. https://doi.org/10.7554/eLife.71802.sa0 Decision letter eLife's review process eLife digest Although our genetic code does not change throughout our lives, our genes can be turned on and off as a result of epigenetics. Epigenetics can track how the environment and even certain behaviors add or remove small chemical markers to the DNA that makes up the genome. The type and location of these markers may affect whether genes are active or silent, this is, whether the protein coded for by that gene is being produced or not. One common epigenetic marker is known as DNA methylation. DNA methylation has been linked to the levels of a range of proteins in our cells and the risk people have of developing chronic diseases. Blood samples can be used to determine the epigenetic markers a person has on their genome and to study the abundance of many proteins. Gadd, Hillary, McCartney, Zaghlool et al. studied the relationships between DNA methylation and the abundance of 953 different proteins in blood samples from individuals in the German KORA cohort and the Scottish Lothian Birth Cohort 1936. They then used machine learning to analyze the relationship between epigenetic markers found in people's blood and the abundance of proteins, obtaining epigenetic scores or 'EpiScores' for each protein. They found 109 proteins for which DNA methylation patterns explained between at least 1% and up to 58% of the variation in protein levels. Integrating the 'EpiScores' with 14 years of medical records for more than 9000 individuals from the Generation Scotland study revealed 137 connections between EpiScores for proteins and a future diagnosis of common adverse health outcomes. These included diabetes, stroke, depression, Alzheimer's dementia, various cancers, and inflammatory conditions such as rheumatoid arthritis and inflammatory bowel disease. Age-related chronic diseases are a growing issue worldwide and place pressure on healthcare systems. They also severely reduce quality of life for individuals over many years. This work shows how epigenetic scores based on protein levels in the blood could predict a person's risk of several of these diseases. In the case of type 2 diabetes, the EpiScore results replicated previous research linking protein levels in the blood to future diagnosis of diabetes. Protein EpiScores could therefore allow researchers to identify people with the highest risk of disease, making it possible to intervene early and prevent these people from developing chronic conditions as they age. Introduction Chronic morbidities place longstanding burdens on our health as we age. Stratifying an individual's risk prior to symptom presentation is therefore critical (NHS England, 2016). Though complex morbidities are partially driven by genetic factors (Fuchsberger et al., 2016; Yao et al., 2018), epigenetic modifications have also been associated with disease (Lord and Cruchaga, 2014). DNA methylation (DNAm) encodes information on the epigenetic landscape of an individual and blood-based DNAm signatures have been found to predict all-cause mortality and disease onset, providing strong evidence to suggest that methylation is an important measure of disease risk (Hillary et al., 2020a; Lu et al., 2019; Zhang et al., 2017). DNAm can regulate gene transcription (Lea et al., 2018), and epigenetic differences can be reflected in the variability of the proteome (Hillary et al., 2019; Hillary et al., 2020b; Zaghlool et al., 2020). Low-grade inflammation, which is thought to exacerbate many age-related morbidities, is particularly well captured through DNAm studies of plasma protein levels (Zaghlool et al., 2020). As proteins are the primary effectors of disease, connecting the epigenome, proteome, and time to disease onset may help to resolve predictive biological signatures. Epigenetic predictors have utilised DNAm from the blood to estimate a person's 'biological age' (Lu et al., 2019), measure their exposure to lifestyle and environmental factors (McCartney et al., 2018c; McCartney et al., 2018a; Peters et al., 2021), and predict circulating levels of inflammatory proteins (Stevenson et al., 2020; Stevenson et al., 2021). A leading epigenetic predictor of biological aging, the GrimAge epigenetic clock incorporates methylation scores for seven proteins along with smoking and chronological age, and is associated with numerous incident disease outcomes independently of smoking (Hillary et al., 2020a; Lu et al., 2019). This suggests there is predictive value gained in utilising DNAm scores relevant to protein levels as intermediaries for predictions. Methylation scores also point towards the pathways that may act on health beyond the protein biomarker that they are trained on. A portfolio of methylation scores for proteins across the circulating proteome could therefore aid in the prediction of disease and offer a different, but additive signal beyond methylation or protein data alone. Generation of an extensive range of epigenetic scores for protein levels has not been attempted to date. The capability of specific protein scores to predict a range of morbidities has also not been tested. However, DNAm scores for interleukin-6 (IL-6) and C-reactive protein (CRP) have been found to associate with a range of phenotypes independently of measured protein levels, show more stable longitudinal trajectories than repeated protein measurements, and, in some cases, outperform blood-based proteomic associations with brain morphology (Stevenson et al., 2021; Conole et al., 2021). This is likely due to DNAm representing the accumulation of more sustained effects over a longer period of time than protein measurements, which have often been shown to be variable in their levels when measured at multiple time points (Koenig et al., 2003; Liu et al., 2015; Moldoveanu et al., 2000; Shah et al., 2014). DNAm scores for proteins could therefore be used to alert clinicians to individuals with high-risk biological signatures, many years prior to disease onset. Here, we report a comprehensive association study of blood-based DNAm with proteomics and disease (Figure 1). We trained epigenetic scores – referred to as EpiScores – for 953 plasma proteins (with sample size ranging from 706 to 944 individuals) and validated them using two independent cohorts with 778 and 162 participants. We regressed out known genetic pQTL effects from the protein levels prior to generating the EpiScores to preclude the signatures being driven by common SNP data that are invariant across the lifespan. We then examined whether the most robust predictors (n = 109 EpiScores) associated with the incidence of 12 major morbidities (Table 1), over a follow-up period of up to 14 years in the Generation Scotland cohort (n = 9537). We also tested for associations between EpiScore levels and COVID-19 disease outcomes. We regressed out the effects of age on protein levels prior to training and testing; age was also included as a covariate in disease prediction models. We controlled for common risk factors for disease and assessed the capacity of EpiScores to identify previously reported protein-disease associations. Figure 1 Download asset Open asset EpiScores for plasma proteins as tools for disease prediction study design. DNA methylation scores were trained on 953 circulating plasma protein levels in the KORA and LBC1936 cohorts. There were 109 EpiScores selected based on performance (r > 0.1, p < 0.05) in independent test sets. The selected EpiScores were projected into Generation Scotland, a cohort that has extensive data linkage to GP and hospital records. We tested whether levels of each EpiScore at baseline could predict the onset of 12 leading causes of morbidity, over a follow-up period of up to 14 years; 137 EpiScore-disease associations were identified, for 11 morbidities. We then assessed whether EpiScore associations reflected protein associations for diabetes, which is a trait that has been well characterised using SOMAscan protein measurements. Of the 29 SOMAscan-derived EpiScore-diabetes associations, 23 highlighted previously reported protein-diabetes associations. Table 1 Incident morbidities in the Generation Scotland cohort. Counts are provided for the number of cases and controls for each incident trait in the basic and fully adjusted Cox models run in the Generation Scotland cohort (n = 9537). Mean time-to-event is summarised in years for each phenotype. Alzheimer's dementia cases and controls were restricted to those older than 65 years. Breast cancer cases and controls were restricted to females. Basic modelFully adjusted modelMorbidityN casesN controlsYears to event(mean, SD)N casesN controlsYears to event(mean, SD)Rheumatoid arthritis6592816.1 (3.5)5477366.4 (3.3)Alzheimer's dementia6937648.3 (2.7)5231378.2 (2.7)Bowel cancer7793986.4 (3.2)6578176.5 (3.2)Depression10183063.9 (3.3)8069763.7 (3.3)Breast cancer12953556 (3.4)11044015.9 (3.4)Lung cancer20192655.2 (3.1)17277055.1 (3.1)Inflammatory bowel disease20390835 (3.5)16375674.9 (3.5)Stroke31790236.5 (3.4)24875466.4 (3.5)COPD34689396.2 (3.4)27374596.1 (3.4)Ischaemic heart disease39586465.8 (3.3)30972485.9 (3.3)Diabetes42887565.7 (3.4)32273315.7 (3.4)Pain (back/neck)149453415.2 (3.5)122144755.3 (3.5) COPD: chronic obstructive pulmonary disease. A video summarising this study and detailing how the 109 EpiScores can be calculated in cohorts with DNAm data is available at: https://youtu.be/xKDWg0Wzvrg. Our MethylDetectR shiny app (Hillary and Marioni, 2020) has CpG weights for the 109 EpiScores integrated such that it automates the process of score generation for any DNAm dataset and is available at: https://www.ed.ac.uk/centre-genomic-medicine/research-groups/marioni-group/methyldetectr. A video on how to use the MethylDetectR shiny app to generate EpiScores is available at: https://youtu.be/65Y2Rv-4tPU. Results Selecting the most robust EpiScores for protein levels To generate epigenetic scores for a comprehensive set of plasma proteins, we ran elastic net penalised regression models using protein measurements from the SOMAscan (aptamer-based) and Olink (antibody-based) platforms. We used two cohorts: the German population-based study KORA (n = 944, mean age 59 years [SD 7.8], with 793 SOMAscan proteins) and the Scottish Lothian Birth Cohort 1936 (LBC1936) study (between 706 and 875 individuals in the training cohort, with a total of 160 Olink neurology and inflammatory panel proteins). The mean age of the LBC1936 participants at sampling was 70 (SD 0.8) for inflammatory and 73 (SD 0.7) for neurology proteins. Full demographic information is available for all cohorts in Supplementary file 1A. Prior to running the elastic net models, we rank-based inverse normalised protein levels and adjusted for age, sex, cohort-specific variables and, where present, cis and trans pQTL effects identified from previous analyses (Hillary et al., 2019; Hillary et al., 2020b; Suhre et al., 2017) (Materials and methods). Of a possible 793 proteins in KORA, 84 EpiScores had Pearson r > 0.1 and p < 0.05 when tested in an independent subset of Generation Scotland (The Stratifying Resilience and Depression Longitudinally [STRADL] study, n = 778) (Supplementary file 1B). These EpiScores were selected for EpiScore-disease analyses. Of the 160 Olink proteins trained in LBC1936, there were 21 with r > 0.1 and p < 0.05 in independent test sets (STRADL, n = 778, Lothian Birth Cohort 1921: LBC1921, n = 162) (Supplementary file 1C). Independent test set data were not available for four Olink proteins. However, they were included based on their performance (r > 0.1 and p < 0.05) in a holdout sample of 150 individuals who were left out of the training set. We then retrained all selected predictors on the full training samples. A total of 109 EpiScores (84 SOMAscan-based and 25 Olink-based) were brought forward (r > 0.1 and p < 0.05) to EpiScore-disease analyses (Figure 2 and Supplementary file 1D). There were five EpiScores for proteins common to both Olink and SOMAscan panels, which had variable correlation strength (GZMA r = 0.71, MMP.1 r = 0.46, CXCL10 r = 0.35, NTRK3 r = 0.26, and CXCL11 r = 0.09). Predictor weights, positional information, and cis/trans status for CpG sites contributing to these EpiScores are available in Supplementary file 1E. The number of CpG features selected for EpiScores ranged from 1 (lyzozyme) to 395 (aminoacylase-1 [ACY-1]), with a mean of 96 (Supplementary file 1F). The most frequently selected CpG was the smoking-related site cg05575921 (mapping to the AHRR gene), which was included in 25 EpiScores. Counts for each CpG site are summarised in Supplementary file 1G. This table includes the set of protein EpiScores that each CpG contributes to, along with phenotypic annotations (traits) from the MRC-IEU EWAS catalog (MRC-IEU, 2021) for each CpG site having genome-wide significance (p < 3.6 × 10–8) (Saffari et al., 2018). GeneSet enrichment analysis of the original proteins used to train the 109 EpiScores highlighted pathways associated with immune response and cell remodelling, adhesion, and extracellular matrix function (Supplementary file 1H). Figure 2 with 2 supplements see all Download asset Open asset Test performance for the 109 selected protein EpiScores. Test set correlation coefficients for associations between protein EpiScores for (a) inflammatory Olink, (b) neurology Olink, and (c) SOMAmer protein panel EpiScores and measured protein levels are plotted. 95% confidence intervals are shown for each correlation. The 109 protein EpiScores shown had r > 0.1 and p < 0.05 in either one or both of the GS:STRADL (n = 778) and LBC1921 (n = 162) test sets, wherever protein data was available for comparison. Data shown corresponds to the results included in Supplementary file 1B-C. Correlation heatmaps between the 109 EpiScore measures (Figure 2—figure supplement 1) are provided, along with a summary of the most enriched functional pathways for the genes of the 109 proteins used to train EpiScores (Figure 2—figure supplement 2). EpiScore-disease associations in Generation Scotland The Generation Scotland dataset contains extensive electronic health data from GP and hospital records as well as DNAm data for 9537 individuals. This makes it uniquely positioned to test whether EpiScore signals can predict disease onset. We ran nested mixed effects Cox proportional hazards models (Figure 3) to determine whether the levels of each EpiScore at baseline associated with the incidence of 12 morbidities over a maximum of 14 years of follow-up. The correlation structures for the 109 EpiScore measures used for Cox modelling are presented in Figure 2—figure supplement 1. Figure 3 with 1 supplement see all Download asset Open asset Nested Cox proportional hazards assessment of protein EpiScore-disease prediction. Mixed effects Cox proportional hazards analyses in Generation Scotland (n = 9537) tested the relationships between each of the 109 selected EpiScores and the incidence of 12 leading causes of morbidity (Supplementary file 1I-J). The basic model was adjusted for age and sex and yielded 294 associations between EpiScores and disease diagnoses, with false discovery rate (FDR)-adjusted p < 0.05. In the fully adjusted model, which included common risk factors as additional covariates (smoking, deprivation, educational attainment, body mass index (BMI), and alcohol consumption), 137 of the basic model associations remained significant with p < 0.05. In a sensitivity analysis, the addition of estimated white blood cells (WBCs) to the fully adjusted models led to the attenuation of 34 of the 137 associations. In a further sensitivity analysis, 81 associations remained after adjustment for both immune cell proportions and GrimAge acceleration. The Cox proportional hazards assumption was assessed through the Schoenfeld residuals from the models. Two associations in the basic model adjusting for age and sex failed to satisfy the global assumption (across all covariates) and were excluded. There were 294 remaining EpiScore-disease associations with a false discovery rate (FDR)-adjusted p < 0.05 in the basic model. After further adjustment for common risk factor covariates (smoking, social deprivation status, educational attainment, body mass index [BMI], and alcohol consumption), 137 of the 294 EpiScore-disease associations from the basic model had p < 0.05 in the fully adjusted model (Supplementary file 1I-J). Eleven of the 137 fully adjusted associations failed the Cox proportional hazards assumption for the EpiScore variable (p < 0.05 for the association between the Schoenfeld residuals and time; Supplementary file 1K). When we restricted the time-to-event/censor period by each year of possible follow-up, there were minimal differences in the EpiScore-disease hazard ratios between follow-up periods that did not violate the assumption and those that did (Supplementary file 1L). The 137 associations were therefore retained as the primary results. The 137 associations found in the fully adjusted model comprised 78 unique EpiScores that were related to the incidence of 11 of the 12 morbidities studied. Diabetes and chronic obstructive pulmonary disease (COPD) had the greatest number of associations, with 33 and 41, respectively. Figure 4 presents the EpiScore-disease relationships for COPD and the remaining nine morbidities: stroke, lung cancer, ischaemic heart disease (IHD), inflammatory bowel disease (IBD), rheumatoid arthritis (RA), depression, bowel cancer, pain (back/neck), and Alzheimer's dementia. There were 13 EpiScores that associated with the onset of three or more morbidities. Figure 5 presents relationships for these 13 EpiScores in the fully adjusted Cox model results. Of note is the EpiScore for Complement 5 (C5), which associated with five outcomes: stroke, diabetes, IHD, RA, and COPD. Of the 29 SOMAscan-derived EpiScore associations with incident diabetes, 23 replicated previously reported SOMAscan protein associations (Elhadad et al., 2020; Gudmundsdottir et al., 2020; Ngo et al., 2021) with incident or prevalent diabetes in one or more cohorts (Figure 6 and Supplementary file 1M). Figure 4 Download asset Open asset Protein EpiScore associations with incident disease. EpiScore-disease associations for 10 of the 11 morbidities with associations where p < 0.05 in the fully adjusted mixed effects Cox proportional hazards models in Generation Scotland (n = 9537). Hazard ratios are presented with confidence intervals for 104 of the 137 EpiScore-incident disease associations reported. Models were adjusted for age, sex, and common risk factors (smoking, body mass index (BMI), alcohol consumption, deprivation, and educational attainment). IBD: inflammatory bowel disease. IHD: ischaemic heart disease. COPD: chronic obstructive pulmonary disease. For EpiScore-diabetes associations, see Figure 6. Data shown corresponds to the results included in Supplementary file 1J. Figure 5 Download asset Open asset Protein EpiScores that associated with the greatest number of morbidities. EpiScores with a minimum of three relationships with incident morbidities in the fully adjusted Cox models. The network includes 13 EpiScores as dark blue (SOMAscan) and grey (Olink) nodes, with disease outcomes in black. EpiScore-disease associations with hazard ratios < 1 are shown as blue hazard ratios > 1 are shown in COPD: chronic obstructive pulmonary disease. IHD: ischaemic heart disease. Data shown corresponds to the results included in Supplementary file 1J. Figure 6 Download asset Open asset of known protein-diabetes associations with protein EpiScores. EpiScore-incident diabetes associations in Generation Scotland (n = 9537). The 29 SOMAscan and four Olink associations shown with p < 0.05 in fully adjusted mixed effects Cox proportional hazards models. Of the 29 SOMAscan-derived 23 associations were with protein-diabetes associations in one or more of the studies that used SOMAscan protein levels. associations were Data shown corresponds to the results included in Supplementary and cell and GrimAge sensitivity analyses of the 78 EpiScores that were associated with incident disease < 0.05 in the proportional hazards with covariates relationships with both estimated white blood cell proportions and GrimAge (Figure supplement 1). These covariates were therefore to the Cox models (Figure There were associations that remained significant p < 0.05 in the basic model and p < 0.05 in the fully adjusted after adjustment for immune cell proportions, of which 81 remained significant when GrimAge scores were to this model (Supplementary file In a further sensitivity analysis, relationships between both estimated white blood cell proportions and GrimAge scores with incident diseases were assessed in the Cox model independently of EpiScores. Of the possible relationships between measures and the morbidities four were significant p < 0.05) in the basic model and remained significant with p < 0.05 in the fully adjusted model (Supplementary file A of cells was linked to risk of incident RA, diabetes, and pain The GrimAge score was associated with IHD, diabetes, and pain in the fully adjusted models (p < 0.05) (Supplementary file The of the GrimAge was to the EpiScore between EpiScores and COVID-19 Two previous studies proteomic measurements from the Generation Scotland cohort = as of analyses found that several proteins to our EpiScores were associated with COVID-19 outcomes et al., 2021; et al., 2020). These included proteins such as and all of which were associated with one or more incident diseases in this Two = and = of the Generation Scotland sample who COVID-19 were therefore used to test the that EpiScores associate with COVID-19 outcomes years after the blood for DNAm significant associations were identified that differences between the of or from COVID-19 that had p < 0.05 did not adjustment for multiple (Supplementary file Discussion Here, we report a comprehensive DNAm study of 953 circulating proteins. We 109 robust EpiScores for plasma protein levels that are independent of known pQTL effects. By projecting these EpiScores into a cohort with data we show that 78 EpiScores associate with the incidence of 11 leading causes of morbidity EpiScore-disease associations in but not associate with COVID-19 outcomes. we show that EpiScore-diabetes associations previously measured protein-diabetes The of EpiScore-disease associations are independent of common lifestyle and health factors, differences in immune cell and GrimAge acceleration. EpiScores therefore signatures for disease prediction and risk stratification. The between our EpiScore-diabetes associations and previously identified protein-diabetes relationships (Elhadad et al., 2020; Gudmundsdottir et al., 2020; Ngo et al., 2021) suggests that epigenetic scores identify biological In addition to the comprehensive of SOMAscan proteins with diabetes, several of the markers we identified for COPD and also previous associations with measured proteins et al., 2016; et al., 2021). The three studies used for the diabetes the protein of type 2 diabetes to and the top markers identified included sex and protein 2 (Elhadad et al., 2020; Gudmundsdottir et al., 2020; Ngo et al., 2021). Our EpiScores for these top markers were also associated with diabetes, in addition to EpiScores for several protein markers reported in these A growing body of evidence suggests that type 2 diabetes is by genetic and epigenetic and and proteins such as and are thought to a range of diabetes-associated and et al., that we identify through EpiScore associations, such as protein 2 have also been in type 2 diabetes onset through analysis et al., 2021). In the case of diabetes, EpiScores may therefore be used as risk many years prior to onset. be tested when data available for the remaining morbidities. test set r = and r = it is that such strong is between EpiScores for proteins that associated with diabetes and the with measured proteins. DNAm scores for and have previously been shown to in test sets (r to explained variance in protein but and often outperform the measured protein related to a range of phenotypes (Stevenson et al., 2020; Stevenson et al., 2021). to scores utilising DNAm for the prediction of our EpiScores the study of individual protein predictor signatures with outcomes. For levels of the EpiScore onset of Alzheimer's dementia, several years prior to by has been as a for Alzheimer's disease et al., et al., et al., 2020) and has been shown to protein and associate with accumulation of in models of Alzheimer's et al., et al., 2020). Our assessment of EpiScores a for future as predictors for may be using our EpiScore These be tested in incident disease when case data are Our results that the set of 109 EpiScores are likely to be enriched for and immune in addition to extracellular cell remodelling, and cell This previous work linking chronic and the (Zaghlool et al., 2020). also suggests that EpiScores could be in the prediction of morbidities that are characterised by inflammatory of this is the EpiScore for Complement 5 (C5), which was associated with the onset of five morbidities, the highest number for any EpiScore (Figure The EpiScore for is likely to the biological pathways in individuals with and could be utilised to alert clinicians to individuals at risk of levels of have been associated with and et al., 2019; et al., and and a range of are in et al.,

Type

peer-review

Language

en

Landing Page

https://doi.org/10.7554/elife.71802.sa2

OA Status

gold

Related Works

10

OpenAlex ID

https://openalex.org/W4214633900

Raw OpenAlex JSON

OpenAlex ID

https://openalex.org/W4214633900

Canonical identifier for this work in OpenAlex

DOI

https://doi.org/10.7554/elife.71802.sa2

Digital Object Identifier

Title

Author response: Epigenetic scores for the circulating proteome as tools for disease prediction

Work title

Type

peer-review

OpenAlex work type

Language

en

Primary language

Publication year

2021

Year of publication

Publication date

2021-11-25

Full publication date if available

Authors

Danni A. Gadd, Robert F. Hillary, Daniel L. McCartney, Shaza B. Zaghlool, Anna J. Stevenson, Yipeng Cheng, Chloe Fawns‐Ritchie, Clifford Nangle, Archie Campbell, Robin Flaig, Sarah E. Harris, Rosie M. Walker, Liu Shi, Elliot M. Tucker‐Drob, Christian Gieger, Annette Peters, Mélanie Waldenberger, Johannes Graumann, Allan F. McRae, Ian J. Deary, David J. Porteous, Caroline Hayward, Peter M. Visscher, Simon R. Cox, Kathryn L. Evans, Andrew M. McIntosh, Karsten Suhre, Riccardo E. Marioni

List of authors in order

Landing page

https://doi.org/10.7554/elife.71802.sa2

Publisher landing page

Open access

Yes

Whether a free full text is available

OA status

gold

Open access status per OpenAlex

OA URL

https://doi.org/10.7554/elife.71802.sa2

Direct OA link when available

Concepts

Epigenetics, Disease, Proteome, DNA methylation, Population stratification, Biomarker, Epigenome, Bioinformatics, Computational biology, Leverage (statistics), Medicine, Biology, Data science, Computer science, Genetics, Machine learning, Gene, Genotype, Internal medicine, Single-nucleotide polymorphism, Gene expression

Top concepts (fields/topics) attached by OpenAlex

Cited by

0

Total citation count in OpenAlex

Related works (count)

10

Other works algorithmically related by OpenAlex

Full payload