Machine learning-guided deconvolution of plasma protein levels

Machine learning-guided deconvolution of plasma protein levels | medRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-P4HH5NV'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search Machine learning-guided deconvolution of plasma protein levels View ORCID Profile Maik Pietzner , Carl Beuchel , Kamil Demircan , Julian Hoffmann Anton , Wenhuan Zeng , Werner Römisch-Margl , Summaira Yasmeen , Burulça Uluvar , View ORCID Profile Martijn Zoodsma , Mine Koprulu , View ORCID Profile Gabi Kastenmüller , Julia Carrasco-Zanini , Claudia Langenberg doi: https://doi.org/10.1101/2025.01.09.25320257 Maik Pietzner 1 Computational Medicine, Berlin Institute of Health at Charité – Universitätsmedizin Berlin , Berlin, Germany 2 Precision Healthcare University Research Institute, Queen Mary University of London , London, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Maik Pietzner For correspondence: maik.pietzner{at}bih-charite.de claudia.langenberg{at}qmul.ac.uk Carl Beuchel 1 Computational Medicine, Berlin Institute of Health at Charité – Universitätsmedizin Berlin , Berlin, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Kamil Demircan 1 Computational Medicine, Berlin Institute of Health at Charité – Universitätsmedizin Berlin , Berlin, Germany 2 Precision Healthcare University Research Institute, Queen Mary University of London , London, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site Julian Hoffmann Anton 2 Precision Healthcare University Research Institute, Queen Mary University of London , London, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site Wenhuan Zeng 1 Computational Medicine, Berlin Institute of Health at Charité – Universitätsmedizin Berlin , Berlin, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Werner Römisch-Margl 3 Institute of Computational Biology, Helmholtz Zentrum München – German Research Center for Environmental Health , Neuherberg, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Summaira Yasmeen 1 Computational Medicine, Berlin Institute of Health at Charité – Universitätsmedizin Berlin , Berlin, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Burulça Uluvar 1 Computational Medicine, Berlin Institute of Health at Charité – Universitätsmedizin Berlin , Berlin, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site Martijn Zoodsma 1 Computational Medicine, Berlin Institute of Health at Charité – Universitätsmedizin Berlin , Berlin, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Martijn Zoodsma Mine Koprulu 2 Precision Healthcare University Research Institute, Queen Mary University of London , London, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site Gabi Kastenmüller 3 Institute of Computational Biology, Helmholtz Zentrum München – German Research Center for Environmental Health , Neuherberg, Germany Find this author on Google Scholar Find this author on PubMed Search for this author on this site ORCID record for Gabi Kastenmüller Julia Carrasco-Zanini 2 Precision Healthcare University Research Institute, Queen Mary University of London , London, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site Claudia Langenberg 1 Computational Medicine, Berlin Institute of Health at Charité – Universitätsmedizin Berlin , Berlin, Germany 2 Precision Healthcare University Research Institute, Queen Mary University of London , London, UK Find this author on Google Scholar Find this author on PubMed Search for this author on this site For correspondence: maik.pietzner{at}bih-charite.de claudia.langenberg{at}qmul.ac.uk Abstract Full Text Info/History Metrics Supplementary material Preview PDF ABSTRACT Proteomic techniques now measure thousands of proteins circulating in blood at population scale, driving a surge in biomarker studies and biological clocks. However, their potential impact, generalisability, and biological relevance is hard to assess without understanding the origins and role of the thousands of proteins implicated in these studies. Here, we provide a data-driven identification of factors explaining variation in plasma levels of ∼3,000 proteins among 43,240 participants of the UK Biobank that explain their links to ageing and diseases, and help guide protein biomarker and drug target discovery. We use machine learning to systematically identify a median of 20 factors (range: 1-37) out of >1,800 participant and sample charateristics that jointly explained an average of 19.4% (max. 100.0%) of the variance in plasma levels across protein targets. Proteins segregated into distinct clusters according to their explanatory factors, with modifiable characteristics explaining more variance compared to genetic variation (median: 10.0% vs 3.9%). We identify proteins for which the factors explaining varying levels in blood differed by sex (n=1374 proteins) or across ancestries (n=74 proteins). We establish a knowledge graph that integrates our findings with genetic studies and drug characteristics to guide identification of potential markers of drug target engagement. We demonstrate the value of our resource 1) by identifying disease-specific biomarkers, like matrix metalloproteinase 12 for abdominal aortic aneurysm, and 2) by developing a framework for phenotype enrichment of protein signatures from independent studies to identify underlying sources of variation. All results are explorable via an interactive web portal ( https://omicscience.org/apps/prot_foundation ) and can be readily integrated into ongoing studies using an associated R package ( https://github.com/comp-med/r-prodente ). INTRODUCTION High-throughput plasma proteomics is now fuelling a new wave of biomarker studies at unprecedented scale 1 , but is about to repeat the failures of decades of previous studies with very few, if any, of thousands of tested candidates ever improving clinical practice 2 . The ever-expanding content of proteomic assays, now surpassing half of the protein coding genome, further exceeds our ability to understand the origins and relevance of the many proteins that are detectable but have no established role in blood ( Fig. 1a ) 3 , 4 . Download figure Open in new tab Figure 1 A coordinate system of plasma protein variation. a Secretome assignment 4 for 2919 protein targets analysed in the present study. b Number and type of UK Biobank participant characteristics considered in the study. c Stacked bar charts displaying the achieved variance explained for each protein target. Proteins are ordered by explained variance across all factors. Colours indicate domains of explaining factors. d Uniform Manifold Approximation and Projection (UMAP) mapping of the variance explained matrix across 2853 protein targets for which we identified at least one feature explaining the variance in plasma levels. Each protein has been assigned a cluster based on k-means clustering and is coloured accordingly. e Violin plots showing the explained variance across protein targets according to categories of selected characteristics. The centre of each violin plot is a boxplot giving median (white dot) and interquartile range. Data information: In E, data are presented as boxplots, indicating median, interquartile range (IQR), and whiskers for >1.5 times IQR Proteome-wide genome-wide association studies 5 – 11 provided evidence that changes in plasma protein abundance can reflect altered production in tissues and identified at least one protein quantitative trait locus (pQTL) for most protein targets. Some pQTLs explained jointly as much as 70% of the variance in plasma levels 5 , 8 and they have been widely advocated as instruments for causal inference or even to impute the plasma proteome based on comparatively cheap genotyping 12 – 14 . However, on average pQTLs explain relatively little of protein variation in plasma, partly because static germline genetic variation does not capture dynamic adaptions that indicate early disease states or worsening of conditions. Atlas-like efforts are now emerging that statistically associate changes in plasma protein levels with hundreds of diseases, ageing clocks, or health characteristics 7 , 15 – 17 , but that do rarely translate into biological knowledge or understanding of the factors that underly disease associations or predictive models. Here, we present a framework for the integration of multimodal data to systematically identify factors, including characterisitics of human health and disease but also technical measures, explaining variation in plasma protein levels building on our previous work 18 . We demonstrate that a relatively small number of factors (median: 20; range: 1-37) of all >1,800 tested explain a considerable part of protein variation (>25% for most protein targets) in plasma. Protein targets thereby segregated into distinct clusters explained by indicators of human health but also pre-analytical variation, such as accidental activation of platelets, and identify proteins that are best explained by different characteristics characteristics across the sexes and ancestral strata. We demonstrate how the collective knowledge generated here integrated with human genetic studies can guide the identification of disease-specific biomarkers and drug response markers from observational studies. We create a resource that is publicly available to the community to explore results ( https://omicscience.org/apps/prot_foundation/ ) and a statistical framework to implement phenotype enrichment analyses in external studies ( https://github.com/comp-med/r-prodente ). RESULTS We identified 411 diverse, modifiable and non-modifiable, participant and technical characteristics (out of a total of 1879; Fig. 1b and Tab. EV1 ) to explain variation in plasma levels of one or more of 2,853 protein targets (97.7%; Tab. EV2 ) among 30,268UK Biobank participants (54.1% female; 5.2% Non-European) using regularised linear regression models with stability selection 19 . A median of 20 characteristics (range: 1 – 37) was selected across these protein targets and cumulatively explained on average 19.8% (range: 0.0005%-100.0%) of variation in plasma protein levels in an independent validation set (n=12,972; Fig. 1c and Tab. EV3 ). The highest amount of variance was explained for proteins actively secreted into blood (median of 25.5%) and those demonstrated to be reliably detected using the assay technology (median of 20.2% for 1,990 proteins with ≤5% of values below the limit of detection). The ability to reliably detect mRNA levels of the protein coding gene in one or more tissues was further associated with a higher amount of explained variance. ( Fig. EV1a-h ). We sought replication of our results in our previous work in a different cohort 18 . Despite difference in the applied proteomic technology and available participant characteristics, the total amount of explained variance was significantly correlated (r=0.28; p-value≤6.7×10 −37 ) across 1,968 protein targets measured on both platforms ( Tab. EV4 ). The correlation coefficient improved to 0.46 (p-value≤2.8×10 −37 ) among 699 protein targets that have been reported to correlate well across proteomic technologies 7 , demonstrating good generasability of our results. Improvements in explained variance by the present study were most strongly associated with the inclusion of disease status (+14.4% per 1%-increase in explained variance by disease status; p-value≤4.8×10 −17 ), genetically inferred ancestry (+8.5%; 1.3×10 −31 ), or considering parameters of general health (+18.4%; p-value≤9.8×10 −7 ). Results that showcased the importance to expand the phenotypic space compared to our pioneering work. A coordinate system for plasma protein variation Projecting the entire matrix of proteins times explaining factors into a lower-dimensional space established a coordinate system along which proteins segregated into distinct clusters according to shared major influences but also tissue and cell-type origin ( Fig. 1d ). The largest cluster showed evidence for enrichment by multiple explanatory factors, whereas remaining clusters distinguished by at most a few or even single characteristics explaining most of the variance in plasma levels of proteins within the cluster ( Tab. EV5 and Fig. EV2 ). For example, one cluster contained 617 protein targets that were strongly enriched for participant and sample characteristics indicating effects of technical variation and contamination by blood cell activation. For example, recruitment centre (beta: 0.18; p-value<4.9×10 −−08 ; mean explained variance (MEV): 1.82%) or plateletcrit (PCT) (beta: 1.73; p-value<2.8×10 −264 ; MEV: 7.2%) were most commonly selected for those proteins ( Fig. EV2k-l ). Our findings are in line with previous studies 20 – 22 and are likely a result of platelet activation, and subsequent release of contained proteins into plasma 21 . Collectively, these results provide evidence, that protein signatures in blood can be partly deconvoluted into distinct origins. Modifiable participant characteristics thereby outweighted non-modifiable ones, such as age, genetic sex, ancestry, or common protein quantitative trait loci, on average (p<7.5×10 −47 ; two-sided Wilcoxon rank sum test; Fig. 1d ). Organ and cell-type contributions Proteins circulating in blood have diverse origins and we therefore integrated gene expression data 3 , 23 to understand whether the selection of participant characteristics can be explained by effects in certain tissues or cell-types. More than 80% (n=2,420) of all detected protein targets showed evidence that the corresponding mRNA is preferentially expressed in one or at most a few tissues or cell types enabling enrichment analysis in up to 13 tissues and 28 cell-types ( Tab. EV2 ). This identified 94 characteristics significantly more frequently selected for proteins with tissue or cell-type specific mRNA expression ( Fig. 2 ; Fig. EV3 and Tab. EV6-7 ). Many of these enrichments were likely driven by tissue or cell damage. For example, proteins with enhanced mRNA expression in the liver, specifically hepatocytes, were explained by medications with adverse hepatic effects such as oral contraceptives 24 (e.g., conjugated oestrogens; odds ratio: 5.62; p-value<3.3×10 −18 ) or carbamazepine (odds ratio: 8.98; p-value<1.3×10 −13 ). Other medication associations more likely represented effects on target organs/cells, such as proton pump inhibitors and stomach (e.g., esomeprazole; odds ratio: 237.10; p-value<3.9×10 −15 ) or immune suppressants and immune cells (e.g., azathioprine; odds ratio: 13.01; p-value<2.3×10 −16 ). Download figure Open in new tab Figure 2 Plasma proteins link cell-types to indicators of health and disease. Chord diagram of phenotype associated protein signature enrichment among protein coding genes with enhanced cell type expression (‘marker genes’). Each line represents a significant enrichment (Fisher’s test; p<2.9×10 −6 ) of proteins associated with a participant characteristic among protein coding genes with enhanced expression in a cell type. Enhanced expression estimates were derived from single-cell RNA sequencing data in the Human Protein atlas. Corresponding statistics can be found in Tab. EV7. Associations with marker genes of fibroblasts have been highlighted by darker colours. On a cell population level, we observed yet unreported links between participant characteristics and proteins with enhanced mRNA expression in fibroblasts that synthesize the extra cellular matrix (ECM; Fig. 2 ). This included drug usage like beclomethasone dipropionate (odds ratio: 14.24; p-value<4.0×10 −9 ), diseases like atrial fibrillation (odds ratio: 6.13; p-value<2.7×10 −9 ), but also participant’s age (odds ratio: 2.88; p-value<6.7×10 −10 ), body mass index (BMI; odds ratio: 3.49; p-value<7.7×10 −13 ) and genetically inferred ancestry (odds ratio: 2.71; p-value<7.2×10 −9 ). These results provide evidence that changes in plasma protein levels can to some extent be explained because of adverse, but also intentional, effects of drugs and diseases on specific tissues or cell types. Sex– and ancestral differential effects on the plasma proteome Genetically inferred ancestry (n=1,139) and participant’s sex (n=1,199) were among the most frequently selected non-modifiable participant characteristics. To better understand potentially differential or specific effects across ancestries (41,369 White Europeans, 849 British Africans, and 823 British Central South Asians) and the sexes (23,601 females, 20,055 males), we repeated the feature selection procedure separately within each of the groups. We identified significantly lower levels of explained variance in participants of British African (median: –5.07%; IQR: –12.64% – –0.03%; p<1.5×10 −91 ) and to a much lesser extent British Central South Asian ancestry (median: –0.03%; IQR: –4.36% – 3.90%; p<1.3×10 −2 ) compared to White Europeans. Results, that were, at least among participants of British African ancestry, not entirely explained by the lower sample size. We still observed a difference of ≥1.4% in explained variance in plasma levels for more than half of the protein targets following matching for the number of selected features across the White-European and British African cohorts (IQR: –6.58% – 3.15%, p-value<1.0×10 −66 ; Fig. 3a-b ) or when rerunning feature selection in age– and sex-matched subsets of Europeans with the same sample size (median: –2.62%; IQR: –9.34% – 0.00%; p-value<6.1×10 −29 ). Download figure Open in new tab Figure 3 Summary of ancestral-stratified analysis. a Comparison of the achieved explained variance in plasma proteins levels within White Europeans compared to British Africans participants. Proteins are coloured based on cluster membership and those with extreme differences are annotated. *The number of selected features has been matched with the smaller ancestry to account for differences in the power of discovery. b Same as a , but now comparing to participants of British Central South Asian ancestry. c-d Contrasting differences in allele frequencies for cis-pQTL scores with changes in explained variance between White Europeans and British African (c, AFR) or British Central South Asian (d, CSA). e Variance explained estimates for 34 cis-pQTLs with evidence for distinct ancestral-lead signals. Darker colours indicate the effect of the trans-ancestral signal, whereas shades indicate the explained variance by the ancestry-specific lead signal. We note that the effect of ancestry-specific lead signals might be overestimated, since they have not been selected from an independent cohort. Data information: In (A-B), point estimates were derived as variance explained (r 2 ) from a multivariable linear regression model; In E, partial variance explained was computed for each genetic variant based on a multivariable linear regression model The single strongest contributor of differential estimates of explained variance in plasma protein levels across ancestries was the contribution of cis– and trans-pQTL scores ( Tab. EV8 ). Possible explanations include that pQTLs are present at different minor allele frequencies (MAF; Fig. 3c-d ) across ancestries or may have differential effects on protein levels. We identified 34 protein targets for which frequency-enriched ancestry-specific lead signals (non-overlapping haplotype blocks; Fig. 3d ) accounted for strong differences in explained variance. For example, the missense variant rs2071421 (p.Asn352Ser) conferring arylsulfatase A pseudo deficiency was most common in participants of British African ancestry (MAF AFR =38.1%) but less frequent in participants of other ancestries (MAF EUR = 11.0%; MAF CSA = 14.8%) and accordingly explained considerably more variance (65.8% vs 5.9%) compared to the trans-ancestral lead signal rs873697 (MAF AFR =11.5%; MAF EUR = 4.4%; MAF CSA = 0.6%) in participants of British African ancestry. However, most ancestral differential effects, including 135 out 137 cis-pQTLs, were due to the same pQTL having different effect sizes across ancestries with no other nearby variant explaining those. The reason for different effects of the same genetic variant across ancestries remains to be established. We lastly identified few protein targets (n=22) that were much better explained (>4 s.d.) in one but not the other genetically inferred sex ( Fig. EV4a-c ). We, however, noted that selected participant characteristics differed for almost a third (n=895) of the protein targets (median Jaccard index = 0.56; Fig. EV4c ) indicating the need to consider sex-specific contributions to plasma protein levels ( Tab. EV9 ). Obvious examples included medications (n=1062 pairs, most frequently anticontraceptives) and diseases (n=30 pairs) given/occurring only in one sex, whereas abundant sex-differential effects were explained by associations with age, biomarkers, or body mass index ( Fig. EV4d and Tab. EV9 ). Notably, there were only few examples of sex-differential genetic effects, e.g., the cis-pQTL score for plasma oxytocin (OXT; female: 18.7%; male: 30.5%; p-value inter <2.8×10 −48 ; Fig. EV4e ). Protein biomarker discovery and pruning for incident diseases We next systematically explored how our results can guide the identification of plasma protein biomarkers in biobank scale studies. Among the 67,033 significant protein – disease associations (p<4.1×10 −8 ) observed here and reported in other studies 7 , 15 , we observed a more than 32-fold drop after regressing out characteristics explaining variation in plasma protein levels for 424 incident diseases ( Fig. 4a and Tab. EV10 ). Associations of more than two-thirds (1,333 out of 2,080) of protein targets were almost completely attenuated. Even among the 1,975 protein – disease associations with directionally concordant effects and persisting significance, >80% (n=1,691) showed considerable attenuation of effect sizes in Cox-models (≥20%). This suggests that preciser measurements of associated participant characterisitcs are likely to lead to further statistical attenuation ( Fig. 4b ). Notably, a >5-fold decrease in the number of significant protein – disease associations was achieved with as few as five selected characteristics (67,033 to 13,307), demonstrating the importance of understanding and considering protein determinants rather than solely statistical significance for distinguishing biologically relevant from false positive findings and type 1 errors. Download figure Open in new tab Figure 4. Protein – incident disease associations. a Summary of significant (Cox-proportional hazard model; p<4.1×10 −8 ) associations between plasma protein levels (x-axis indicated by the position of the protein-coding gene) and one or more of 424 diseases in UK Biobank using Cox-regression models. Each dot represents an association passing multiple testing, and black dots indicate those persisting after regressing out factors that explained plasma protein variation. The top panel illustrates the fraction of disease associations per protein that were still significant after accounting for associated protein characteristics, similar to right panel for diseases. Proteins or diseases with minimal effect attenuation were annotated. b Scatterplot comparing hazard ratios per 1 s.d. increase in protein levels from Cox-regression models adjusting for age and sex or additionally accounting for factors explaining variation in plasma protein levels. Only protein – disease associations passing statistical significance in at least the minimal model are shown and coloured according to significance and effect attenuation in the extended model. c-d Scatter plots opposing effect estimates for significant (see legend) protein – disease associations comparing those using measured plasma protein levels (x-axis; n=43,647) to those based on genetically imputed plasma protein levels in the entire unrelated White-European UKB cohort (y-axis; n=342,240). Data information: In (A-D), each dot reflects a significant association between plasma protein levels and disease onset based on Cox-proportional hazard models Robust protein – disease associations included established clinical screening markers such as prostate-specific antigen (referred to by Olink as KLK3) for prostate cancer (HR 3.11; p-value 5.3×10 −254 ), and markers of early tissue damage, such as the lung-specific surfactant protein D (SFTPD) for post-inflammatory pulmonary fibrosis 25 , or the eye-specific protein crystallin beta B2 (CRYBB2) for cataract (HR 1.41, p-value 3.8×10 −97 ). Less-established links with strong effect sizes even after accounting for all selected phenotypic characteristics included a 2.3-fold increased risk (p-value 1.1×10 −38 ) for abdominal aortic aneurysm (AAA) per 1 s.d. increase in plasma levels of matrix metalloproteinase 12 (MMP12) ( Fig. 4b , Fig. EV5 ). A putative role of MMP12 in the progressive degradation of the ECM at the aortic wall, a hallmark of AAA, was thereby supported by multiple lines of evidence. Firstly, we obtained evidence that the same genetic variant (rs17368814) that increases plasma MMP12 levels by acting on its protein coding gene also increased the risk for AAA (Posterior probability shared genetic signal: 97.5%; Mendelian randomization estimate: hazard ratio=1.15; p-value=3.3×10 −16 ; Fig. EV6 ), providing support that changes in plasma MMP12 precede AAA onset. Secondly, MMP12 was elevated among patients diagnosed with AAA even after adjustment for all factors explaining MMP12 levels (beta=1.09 s.d. units, p-value<2.1×10 −21 ), potentially indicating ongoing degradation of the ECM. Thirdly, our observational findings are in line with experimtental evidence linking matrix metalloproteinases to the onset and progression of AAA 26 , with MMP12 being highly expressed at disease sites and suggested to contribute to the degradation of elastic fibres and hence weakening and dilation of the aortic wall 27 . These findings seem to be context-specific, since pharmacolocigal inhibition of MMP12 protected Apoe −/− deficient mice from AAA 28 , while, paradoxically, Mmp12 −/− / Apoe −/− mice were more susceptible to AAA and subsequent rupture. Identification of putative protein biomarkers has previously been proposed using genetic imputation 14 , suggesting that measuring common genomic variation can proxy measurement of plasma protein levels and hence guide tailored biomarker assessement. However, even scaled to the power of the entire UK Biobank cohort, we observed little overlap between associations with genetically imputed and measured plasma protein levels, including those that we robustly linked to disease outcomes ( Fig. 4c-d ; Tab. EV10 ). For example, the five most significant genetically proxied protein – disease association were not among the twenty most strongly associated measured protein – disease associations for two-thirds of all diseases considered (268 out of 419). This included not only strongly differing effect estimates, and hence power to identify people at high risk early, but also discordant results whether considering cis-or trans-pQTLs for imputing plasma protein levels ( Fig. 4c-d ). A protein foundation community resource Most plasma proteomic studies are done at small scale, with often incomplete metadata on participants, adding to the complexity of the unknown sources of variation of most plasma proteins. We therefore developed a phenotype enrichment framework that allows to test for significant enrichment (correcting for multiple testing using the Bonferroni procedure) of participant characteristics otherwise hidden in differentially expressed plasma proteomic signatures ( https://github.com/comp-med/r-prodente ). As a proof of principle, we observed proteins associated with the UKB characteristic ‘fasting time’ to be more than 90-fold enriched (odds ratio: 91.0; p-value<1.6×10 −5 ; Fig. 5a ) among proteins significantly different following one day of complete caloric restriction in a well-controlled intervention study 29 . However, a similar enrichment of a protein signature linked to plasma bilirubin was not reported due to missing measurements but resembles the well-known increased reuptake of bilirubin via the enterohepatic cycle due to lower gut motility during prolonged fasting 30 , 31 . Another important application of phenotype enrichment is the discovery of unknown confounders or imperfect matching in biomarker studies. For example, we observed a more than 2-fold enrichment of proteins associated with smoking among those differentially expressed between patients with ovarian carcinoma and selected controls 32 ( Fig. 5a ). Accordingly, smoking status was among the top three variables explaining plasma variation in the most differentially expressed proteins in patients with cancer, Kunitz-type protease inhibitor 1 (SPINT1) (3.0%; Fig. 5b ). We observed similar residual phenotypic enrichments, including smoking or socioeconomic factors, in a plasma proteomic model to predict future coronary artery disease 33 ( Fig. 5b ), or among recently proposed proteomic organ clocks 34 . Those proxying pancreatic or intestinal age being enriched for established markers of muscle mass (odds ratio: 28.3; p-value:3.6×10 −10 ) or fasting time (odds ratio: 28.2; p-value<1.6×10 −9 ), respectively ( Fig. 5c ). Download figure Open in new tab Figure 5 Phenotype enrichment of plasma proteomic signatures. a Phenotype enrichment for plasma proteomic signatures that 1) differed after one day of complete caloric restriction, 2) differentiated healthy women from women with ovarian cancer 32 , and 3) a proteome score to predict onset of coronary artery disease CAD 33 . Enrichment was computed using Fisher’s exact test and only factors passing corrected statistical significance are shown (Bonferroni-correction). b Factors explaining variance in plasma levels of serine peptidase inhibitor, Kunitz type 1 (SPINT1), one of the most differential plasma proteins described for ovarian cancer. c Phenotype enrichment among plasma proteomic signatures proposed to track organ age 34 . Data information: In (A and B), enrichment statistics were derived based on Fisher’s exact test; In B; partial variance estimates were derived from a multivariable linear regression model Most notably, all studies showed strong depletion of proteins associated with technical characteristics, such as study centre, suggesting that protein biomarkers emerging from well-controlled settings are less likely to be driven by analytical artefacts. However, in study designs with separate blood sampling of patients and controls, such as in the blood disease atlas from the Human Protein Atlas project 3 , we observed strong enrichment of such characteristics, dominating biologically plausible findings ( Fig. EV7 ). A knowledge graph to triangulate evidence for clinical impact We finally integrated results from the multivariable variance estimation with drug target annotations, pQTLs and potential effector genes, along with genetic disease associations within a shallow knowledge graph to visually and dynamically illustrate examples of potential clinical impact ( Fig. 6a ; Tab. EV11-12 ). This included 34 examples in which plasma protein levels might act as readouts of successful drug target engagement ( Fig. 6b ). We identified those as subnetworks that link high-confidence drug – protein associations from the present study with genetic variants that mimick drug target modulation by changing the expression or function of the drug target and are associated with the same protein(s) ( Fig. 6b ). For example, plasma levels of desmoglein 2 were explained by pioglitazone intake (0.2%, beta=1.02 s.d. units, p-value=3.7×10 −39 ) and the corresponding pQTL, rs1801282 (beta=0.06; p-value<1.4×10 −12 ), maps to PPARG encoding the pioglitazone target peroxisome proliferator-activated receptor gamma. Other such examples included plasma oxytocin as readout for oral contraceptives or adiponectin as read out for atenolol treatment. The latter had previously been suggested from small-scale trials in a specific target population 35 , whereas a role of desmoglein 2 in regulation of beta-cell activity has been proposed only recently 36 . Download figure Open in new tab Figure 6 Integrated knowledge graph of gene, protein, drug, and disease relationships. a Full knowledge graph connecting single nucleotide polymorphisms (SNPs – pQTLs) to proteins and diseases from previous publications 5 and the GWAS catalog 54 . Each SNP association represents a genome-wide significant finding. Drugs were mapped to target genes based on Open Targets 51 . Protein – disease and protein – drug associations were derived from the present study. The graph was visualised using cytoscape v.3.10.3. b-c Specific subnetworks derived from the knowledge graph illustrating evidence for drug target engagement markers ( b ) by linking pQTLs to drug target genes, and non-random pQTL – drug protein profiles ( c ). An interactive version of ( a ) can be found at https://omicscience.org/apps/prot_foundation/ . In general, enrichment of proteins differentially expressed in disease states for associations with germeline genetic variants, i.e., pQTLs, can point to a causal relationships. We observed 595 examples with significant support from enrichment analysis ( Tab. EV13-14 ). For example, cystatin E/M (CST6), endonuclease, poly(U) specific (ENDOU), and galanin (GAL) were associated with acne in our study (e.g., galanin: beta=0.27 s.d. units, p-value<1.2×10 −13 ) and six independent pQTLs, which themselves have been reported to increase the risk for acne ( Fig. 6c and Tab. EV13 ). The robust link between acne and the neuropeptide galanin aligns with its expression in non-neural tissues such as skin 37 where it can modulate inflammation in a context specific manner 38 . Although no link to acne has been made so far. Similar triangulation with medication intake provided further strong support for the role of platelet modulation on the plasma proteome with 41 genetic loci being significantly enriched for proteins associated with the intake of the antiplatelet medication clopidogrel, almost all of which have been previously linked to platelet characteristics ( Tab. EV14 ). We observed similar convergence of protein profiles for genetic predisposition to and effects of drugs prescribed for osteoporosis (e.g., bisphosphonates like alendronic acid linked to 5 independent pQTLs) or gastric ulcers (e.g., proton pump inhibitors like rabeprazole linked to 3 independent pQTLs) ( Fig. 6c ), providing new avenues to understand drug effects. An interactive version of the knowledge graph can be explored in a customized fashion to identify other such examples ( https://omicscience.org/apps/prot_foundation/ ). DISCUSSION Deep phenotyping of biobank-sized cohorts now provides unprecedented opportunities to identify novel protein biomarkers for common and rare diseases 39 and to inform drug target discovery 5 . However, emerging production-type protein biomarker association analyses are already repeating failures of decades of biomarker research as they rarely translate into actionable knowledge. We created a data-driven model of plasma protein variation that revealed cell type-specific, but also systemic effects and provides a foundation to understand how variation in blood protein levels is linked to human health. We demonstrated its importance in several ways, including biomarker identification for the onset of severe diseases, drug target engagement markers, and as a community resource to complement pathway enrichment analysis providing otherwise unattainable insights. We created a community resource to enable interactive exploration ( https://omicscience.org/apps/prot_foundation/ ) and incorporation in plasma proteomic workflows ( https://github.com/comp-med/r-prodente ). Although not directly comparable, we prioritized ten-times less protein – phenotype combinations (∼50,000 vs >500,000) compared to similar efforts within UKB 7 , 15 emphasising the need to consider and quantify the effect of different factors simultaneously rather than in isolation to overcome confounding. Only such increase in specificity of phenotype – protein associations enabled tangible phenotype enrichment analysis, and subsequently the identification of disease-specific biomarkers that are not only statistically significant, but relevant. The same applies to the considerable amount of sharedness of protein – disease associations proposed previously 15 , that we demonstrated to be largely driven by known phenotypic characteristics. Notably, phenotype enrichment analysis of sparse predictive biomarker signatures previously derived by our group 39 showed no evidence for widespread enrichment of phenotypic characaterisitcs. It rather pointed to tissue damage markers or refinement of imprecise phenotypic risk factors such as smoking or socioeconomic status ( Tab. EV15 ). At the same time, we provided evidence that purely genetically-anchored strategies 14 are unlikely to discover powerful protein biomarkers as they lack the dynamic component. Including >1,800 participant and technical characteristics explained on average about a quarter of the variance in plasma protein levels across the population. Based on our findings, we hypothesize that factors we had not considered or were not assessed in UK Biobank will likely substantially increase the explained variance for at most few targets. Significant improvements in the explained variance across all protein targets are more likely to be achieved by improving assay performance and sample handling, but also by measuring exposures more precisely. For example, quantifying the individual exposure to environmental pollutants instead of regional measures or quantifying drug exposure using dosages instead of self-report. While we were underpowered to quantify ancestral-differential effects at scale, the proteins we identified were either explained by varying effect allele frequencies or differential effect sizes of the same genetic variant across ancestries. The former explains ancestral-enriched or even –specific phenotypic consequences, such as for beta-thalassemia, whereas the latter likely indicated the presence of yet to be identified gene – environment interactions. Larger studies are needed to establish transferability of our results to different population subgroups to ensure equitable biomarker research. Pathway enrichment has been transformative in genome-wide differential gene expression studies to enable interpretation of otherwise massive gene lists and has been almost seamlessly adopted for proteomic studies. While such analyses can guide tissue or single cell studies, proteins found in plasma are of diverse origin making the interpretation of such findings difficult and possibly even misleading. We demonstrated across different applications, how our results can guide the interpretation of plasma protein biomarker studies, including novel insights but also pitfalls when differences in sample handling coincide with case – control comparison. Future extensions may incorporate different proteomic technologies and biofluids to enable widespread application. The advances of our study have to be considered according to a number of limitations. Firstly, our results remain restricted to the selection and precision of measurements of participant and sample characteristics. Other data modalities, such as imaging of organs or exposure to different environments or controlled challenge studies, will likely reveal additional contributions to the plasma proteome. Secondly, we cannot entirely rule out that selected participant characteristics may only approximate truly underlying reasons for variation in plasma protein levels, and that observed associations, even if pertaining statistical significance in multivariable linear regression models, do not necessarily represent causal associations. Thirdly, we observed evidence that variation in plasma levels of proteins close to the limit of detection were less well explained suggesting that more sensitive techniques for such protein targets may improve estimations of explained variance. Lastly, while cis-pQTLs provide reassurance in assay specificity, we cannot completely rule out that comparatively high amounts of variance explained by such cis-pQTLs might be the result of measurement artefacts. This may imply that variance estimates for such targets therefore relate to a specific isoform of the protein target. However, our approach to mutually model genetic and non-genetic effects delivered reliable estimates for the latter, even if the genetic part might have been inflated. MATERIALS & METHODS Study participants UKB is a large-scale, population-based cohort with deep genetic and phenotypic data with the full cohort consisting of approximately 500,000 participants 40 . The participants were recruited across centres in United Kingdom and were aged 40 to 69 years at the time of recruitment 40 . Ethics approval for the UKB study was obtained from the North West Centre for Research Ethics Committee as a Research tissue biobank (REC reference 11/NW/0382) 40 and all participants provided informed consent. We used the subset of individuals from UKB where both genotype and proteomic measurements were available after excluding ancestry outliers or samples which have failed genomic or proteomics quality control (n=43,240). This research has been conducted using the UK Biobank Resource under Application Number 44448. Proteomic measurements The UKB proteomic measurements were conducted for plasma samples of ∼54,000 participants using the antibody-based Olink technology, Explore 3072 platform which uses the Proximity Extension Assay technology 41 . In summary, each protein is targeted by two unique antibodies with unique complimentary oligonucleotides, which only hybridize when they come into close proximity. This is subsequently quantified by next-generation sequencing. Normalized protein expression (NPX) units, which are reported on a log2 scale, were generated by normalization of the extension control and further normalization of the plate control. Further details about antibody-based proteomic measurements and QC have been described elsewhere, including the exclusion of samples due to poor quality and selective measurements with assay warnings 5 . After initial quality control checks (e.g., principal component analysis), we only retained individuals with at least 50% valid protein measurements (n=43,240). Notably, UKB provided proteomic data with values flagged as assay warnings by Olink as ‘NA’, which meant that each protein measurement had differing numbers of valid values and we decided not to impute values for the purpose of this study to minimize skewed estimates of features. We included a total of 2919 unique protein targets in our analysis. Phenotyping We collated a large set of phenotypic and sample characteristics available for at least half of UKB participants ( Tab. EV1 ). Those included information on genetically inferred ancestry (see below), basic demographics (e.g., age and sex, n=5), blood-based biomarkers (e.g., low-density lipoprotein cholesterol, n=28), blood cell counts (n=29), body composition (n=20), indicators of bone health (n=7), cardiovascular risk factors (n=3), diet (n=23), pre-existing diseases (n=1,198), self-reported medication intake (n=704), generic indicators of health (n=4), indicators of pulmonary health (n=8), information on air pollution (n=10), socioeconomic indicators (n=9), and technical variables (n=5), following pruning as described below. Measurement of blood-based biomarkers had been described in detail previously 42 , and we further implemented standard quality control procedures to avoid strong influences of single values. We removed values more than 5 times the median absolute deviation away from the respective sample median (median of ≤0.05% values removed). Biomarkers with strong indications of skewed distributions were log-normalized (similar to protein values). We applied a similar quality control workflow to blood cell counts. We obtained information of when the blood sample was taken during the day, age of the sample when proteomics measurements were done, and duration since last meal (‘fasting time’) as technical variables. We compiled additional measures of body composition by augmenting Dual-X ray predictions of lean and fat mass using equations provided by Powell et al. 43 for each sex separately. In addition, we computed the waist-to-hip ratio as a common, readily available measure of body composition. To identify participants with pre-existing diseases, we collated information from ICD-coded self-report, hospital episode statistics, cancer registry, and primary care (for 45% of the population). We parsed all records to exclude codes with a recorded date before or within the year of birth of the participant to minimize coding errors from electronic health records. We mapped different coding systems to a total of 1,198 ‘phecodes’ that represent medical ontology terms intended to reduce redundancy among ICD-10 coding systems 44 . We recorded the earliest occurrence of each ‘phecode’, referred to hereafter as ‘disease’ from simplicity, and created binary variables indicating whether it had been recorded before the baseline examination of the volunteer. We obtained mappings of self-reported brand names to ATC codes from previous work to generate corresponding medication variables 45 . We kept a total of 704 unique ATC codes mapping to medications taken by 10 or more participants. If not otherwise indicated information was derived from basic questionnaires or the data portal (e.g., dietary habits and blood pressure readings) with minimal cleaning. We obtained regional lead variants explaining variation in plasma protein levels from the trans-ancestral meta-analysis performed by Sun et al. 5 to comput cis– and trans-pQTL scores. We used the regression estimates provided to compute a weighted gentic risk score for each individual for each protein with at least one reported pQTL. After initial data collation, we filtered for variables with a correlation coefficient <0.85 retaining among highly correlated variables the one with the least missing values or manually curated to maintain interpretability. This left us with a total of 1876 variables. Since feature selection required a complete data matrix, we imputed missing values in phenotypic data using Random-forest models that can cope with a mixture of different data types better than usual linear regression techniques 46 . We used all selected phenotypic features for imputation even if they had no missing values to ensure best predictive models for imputation. We further imputed the entire data set before splitting into sets for feature selection and variance estimation to ensure coherent estimates. We implemented these using the R package miceRanger (v.1.5.0). To ensure reasonable computation times, we performed imputation only once. Genotyping and ancestral assignment The UK Biobank samples were genotyped using the Affymetrix UK BiLEVE or the Affymetrix UK Biobank Axiom arrays. The following QC criteria was applied to the genotyping data (a) routine quality checks carried out during the process of sample retrieval, DNA extraction, and genotype calling; (b) checks and filters for genotype batch effects, plate effects, departures from Hardy Weinberg equilibrium, sex effects, array effects, and discordance across control replicates; and (c) individual and genetic variant call rate filters as previously described 47 . We used the ancestry assignments as defined by the pan-UKB 48 , and further assigned unclassified individuals to their respective ancestries based on a k-nearest neighbour approach using genetic principal components. Statistical analysis To select approximately independent participant and sample characteristics that may collectively best explain variation in plasma protein levels, we implemented stability selection 19 and subsequent variance partition 49 . We first split UKB into a set for feature selection (70%) and used the R package stabs (v.0.6.4) to implement feature selection using regularized linear regression with the least absolute shrinkage and selection operator (LASSO). We controlled the upper bound per-family error rate at 1, used a cut-off of 0.75, and a maximum of 500 iterations. We controlled the output from stability selection by artificial introducing 10 random variables, but observed consistent behaviour across the study (i.e., random variables were not selected as important). Based on the features selected, we next performed variance decomposition on 200 bootstrap samples of the remaining UKB set not used for feature selection (30%) and computed the partial explained variance using the etasq() function of the heplots R package (v.1.7.5) as recommended previously 16 . We finally used the median of variance estimates across all 200 bootstrap samples as an estimate for the explained variance in plasma protein levels and provide the 2.5 th and 97.5 th percentils as confidence intervals. We repeated the same procedure using different strata of UKB (biological sexes and genetically inferred ancestry). To establish whether differentially selected features indeed corresponded to differential effects across strata, we additionally implemented interaction testing by performing linear regression analysis with the respective protein target as outcome and including the strata, the differentially selected feature, as well as an interaction effect among both as explanatory variables. We subsequently report strata-differential effects if the p-value for the interaction term passed multiple testing correction. This test was only done for variables not specific to only one of the strata, e.g., no interaction analysis was done for the selection of oral contraceptive across the genetically inferred sexes. We further created four different subsamples of the White-European cohort to match sample sizes of British African and British Central South Asian participants accounting for age and sex distributions. We created four matched European subsamples for each ancestry and repeated the entire feature selection. We took the median of the explained variance across all four subsamples as summary measure. To understand whether protein characteristics may have contributed to different levels of explained variance, we collated information for all protein targets from UniProt via the R package queryup (v.1.0.5) as well as the Human Protein Atlas ( Fig. EV1 ). We binarized information of protein characteristics from UniProt to facilitate numeric analysis. We further computed summary measures of plasma protein level distributions based on untransformed NPX values. We then used Boruta feature selection 50 to identify protein characteristics significantly explaining variation in the levels of explained variance across protein targets based on the entire UKB cohort. We used standard parameters apart from setting the maximum number of iterations to 500 to ensure decisions on most characteristics. We used Uniform Manifold Approximation and Projection (UMAP) to visualize any potential underlying structure in the plasma proteome according to factors contributing to the explained variation in plasma protein abundances. UMAP is a dimensionality reduction technique which enables better preservation of the data’s global structure compared to similar methods such as t-SNE. We applied UMAP on the matrix of explained variation by all participant, environmental, and technical characteristics on all protein targets. We used default values for most parameters used by the algorithm in the umap R package. Custom configuration was done for the following parameters as follows: random_state (seed for random number gnerator) = 10, metric = “pearson”, n_epochs = 1000, init = “random”. To establish clusters of protein targets, we implemented a k-nearest neighbour clustering based on the full variance explained matrix, but excluding features selected ≤10 times to reduce dimensional burden. We selected the number of clusters by visual inspection of the UMAP mapping and corroborating with marker analysis using hurdle models. The latter was implemented using the zeroinfl() function of the pscl R package (v.1.5.9) and we modelled the cluster assignement as exposure and the explained variance by phenotypic characterisitcs as count data to account for thresholding of the explained variance between 0 and 100 and an inflation of zeros (i.e., characteristics not being selected). Ancestral differential variants For each cis-pQTL with significant evidence for ancestral-differential effects, we obtained ancestral-specific GWAS results 5 and retained the strongest regional variant for further testing. We declared and effect as ancestry-specific, if the regional lead variant within the ancestry was not included among a list of highly correlated genetic variants with the other ancestral regional lead variants (r 2 >0.1), basically testing for overlapping haplotypes across ancestries. We subsequently tested each ancestral regional lead variant for effects in other ancestries and computed the explained variance. We further tested, whether the association between the variant and the corresponding protein differed by ancestry (significant interaction effects). The latter was important to establish whether different amounts of explained variants across ancestries were due to different allele frequencies or a differential effect sizes of similarly frequent alleles. Generation of a knowledge graph We generated a shallow knowledge graph to facilitate triangulation of evidence for potential causal relationships. The graph contained a total of 30,068 edges and 6,591 nodes. Nodes included protein targets, pQTLs from Sun et al. 5 , genes (either drug targets or effector genes for pQTLs), drugs (based on ATC codes in the present study), and diseases. We linked proteins to diseases (protein – disease) and drugs (protein – drug) based on evidence from the present study. We included edges linking pQTLs to proteins, and further linked pQTLs to diseases based on overlap with variants with genome-wide significance (r 2 ≥0.8; p<5×10 −8 ) reported in the GWAS catalog (download: 22/05/2025). We retained only entries from the GWAS catalog that we could match to diseases defined in UK Biobank by linking Experimental Factor Ontology (EFO) terms to ICD-10 codes using the EMBL-EBI API. We obtained information from Open Targets 51 to map drugs to targets (genes) and restricted drugs to the once available in UK Biobank by mapping CHEMBL identifiers to ATC codes. We finally used effector gene assignments from Sun et al. 5 to link pQTLs to effector genes. We implemented a user-friendly interactive version of the knowledge using Python. More specifically, we implemented the knowledge graph using the NetworkX Python package 52 (v.3.3), and we enabled the knowledge graph to be interactively accessible using the Pyvis Python package 53 (v.0.3.1). Notably, we customised the JavaScript generated by Pyvis to enhance the functionality of the interactive knowledge graph, allowing the users to obtain information about the neighborhood of the selected node and the shortest cycle path associated with it, computed using the breadth-first search algorithm. Protein – incident disease analyses We systematically investigated prospective associations of protein targets with 390 diseases with more than 200 incident cases during a follow-up time of 10 years. For each protein – disease pair, we employed separate Cox proportional hazards models to model the associations between baseline plasma protein levels and the time-to-event, adjusting for age and sex of the participant. In case of sex-specific diseases, we used only the relevant subset of the population. We computed and compared Cox models with protein levels in two different forms, i.e., crude protein levels and residuals of proteins after regressing out identified explanatory factors. For each protein, we ran a multivariable linear model including all selected characteristics as explanatory variables and subsequently used the residuals (the fraction of variation in protein levels not explained) from this model as updated exposure in Cox-models. We computed similar residual plasma protein levels accounting for the top 3, 5, 10, and 15 selected features.We entered the protein levels as continuous variables after performing rank inverse normal transformation for each predictor to ensure comparability of results. Hence, we report hazard ratios along with standard errors per increment in one standard deviation for each protein. To account for potential reverse causality, we excluded cases that occurred during the first 6 months. We considered associations significant when p< 4.4 x 10 −8 =0.05/(2919 proteins x 390 diseases). We computed time varying effects for each protein – disease associations by computing Schoenfeld residuals, as well as by restricting follow-up time to 2, 5, and 10 years. We used the R package survival (version 3.7.0) for these analyses. Mapping to tissues and cell-types We programmatically downloaded tissue expression data from the Human Protein Atlas (HPA) for all Olink proteins in JSON format (on 20.03.2024). We performed a two-sided Fisher’s exacttest to determine for any participant characteristic selected >5 times across all proteins, whether there was an enrichment of associated protein targets for those showing specific expression of corresponding mRNA levels in certain tissues or cell-types. We defined tissue-or cell-type specific as ‘enhanced’, ‘enriched’, or ‘group enriched’ according to HPA classification. We further performed two-sided Fisher’s exact test to determine whether proteins predicted to be secreted according to the HPA annotation were significantly enriched among any of the clusters of proteins explained by the same biological influence. A phenotype enrichment framework To enable phenotype enrichment, we first collated all results across all strata (i.e., all participants, the sexes, and the three ancestries) tested into one results file ( Tab. EV3 ). For each stratum, we then implemented Fisher’s exact test to test for an enrichment of a specified list of protein targets among protein signatures associated with participant and sample characteristics tested in our atlas work. In other words, for a given list of proteins, we ask the question whether there are participant characteristics more frequently selected than by chance to explain variation in their plasma levels. We restricted enrichment tests to participant and sample characteristics with at least 5 associated, e.g., selected, protein targets. We used dot plots to visualize results. All analyses have been implemented in the R package prodente ( https://github.com/comp-med/r-prodente ) for dissemination. For plasma proteomic studies not using the Olink Explore 3072 platform, we used a specific background list of shared protein targets to minimize bias from differences in proteomic coverage. For the purpose of enrichment tests, we obtained the lists of differentially expressed or selected proteins as proposed by the authors of respective studies. DISCLOSURE and COMPETING INTEREST STATEMENT None of the authors declare a conflict of interest. DATA AVAILABILITY The datasets and computer code produced in this study are available in the following databases: UK Biobank data: All individual-level data is publicly available to bona fide researchers from the UK Biobank ( https://www.ukbiobank.ac.uk/ ). GWAS statistics: GWAS catalog ( https://www.ebi.ac.uk/gwas/ ) RNA expression: Human Protein Atlas ( https://www.proteinatlas.org/ ) Drug characteristics: Open Targets ( https://platform.opentargets.org/downloads ) Modelling computer scripts: GitHub ( https://github.com/comp-med/protein-foundations-ukb-olink-50k ) Results can be obtained from https://omicscience.org/apps/prot_foundation/ EXPANDED VIEW FIGURE LEGENDS Figure EV1 Summary of protein and assay characteristics associated with the variance explained achieved in plasma protein levels . a Variable importance of protein and assays characteristics based on Boruta feature selection predicting the variance explained achieved for each protein target. Boxplots indicate the distribution of the variable importance across 500 iterations. Darker colours indicate features passing corrected statistical significance (p<0.01). b Scatterplot opposing the skewness of individual plasma protein distributions with the variance explained. c – h Variance explained according to different criteria deemed important by the Boruta feature selection. Figure EV2 Major foundations of plasma protein variation . a Uniform Manifold Approximation and Projection (UMAP) mapping of the variance explained matrix across 2853 protein targets for which we identified at least one feature explaining the variance in plasma levels. Each protein has been assigned a cluster based on k-means clustering and is coloured accordingly. b Number of protein targets included in each cluster. c – l Same UMAP plot but coloured according to the variance explained by the factor given on top of each plot. Proteins with strong contributions (>1%) are highlighted. pQTL = protein quantitative trait loci. Figure EV3 Chord diagram of phenotype associated protein signature enrichment among protein coding genes with enhanced tissue expression. Each line represents a significant enrichment (p<6.9×10 −6 ) of proteins associated with a given participant characteristic among protein coding genes with enhanced expression in a given tissue. Enhanced expression estimates were derived from the Human Protein atlas. Corresponding statistics can be found in Table EV7. Colouring was done according to phenotype categories as introduced in Figure 1 in the main text. Figure EV4 Summary of sex-differential feature selection and contribution . a-b Scatterplots comparing the variance explained achieved for plasma protein levels among the entire UK Biobank population (x-axis) compared to what was achieved in females (left) and males (right) alone. Proteins that deviated the most (>4 s.d.) were annotated. c Comparison of cumulative variance explained when stratifying the UK Biobank population by sex. Proteins are coloured by cluster assignments as introduced in Fig. 1 . Proteins with strong differences are annotated with gene names. The inlet depicts the distribution of the Jaccard index of overlapping features for the same protein across the sexes. d Protein – Feature combinations with significant evidence for sex-differential effects (p<9.0×10 −7 ). Only combinations with a difference of more than 10% are shown. e Individual variance explained estimates for plasma oxytocin levels by sex, ordered by the variance explained among females. Proteins in b and c are coloured according to feature categories as introduced in Fig. 1 . Figure EV5 Volcano plots for protein – abdominal aortic aneurysm associations, with adjustment for age and sex (left panel), or comprehensive adjustment based on results from the feature selection algorithm for each protein target (right panel). Figure EV6 Regional association plot of the MMP12 locus for MMP12 protein plasma levels (top) and abdominal aortic aneurysm (AAA, bottom). One SNP (rs17368814) have been prioritized as shared genetic signal. Summary statistics from logistic regression models for AAA are publicly available from the AAAgen consortium and summary statistics from linear regression models for MMP12 plasma levels are based on UK Biobank. Coloring indicates linkage disequilibrium to the lead variant of the same color code. Probability for a shared genetic signal (PP-H 4 ) is given as PP rs17368814 =97.5%. Figure EV7 Phenotype enrichment among differentially expressed plasma proteins between a random set of controls and different patient groups from the blood protein atlas section of the Human Protein Atlas. Each row refers to a protein signature significantly differential (corrected p-value 0.5) in plasma of diseased patients. Each column refers to phenotype associated protein signature from the protein atlas that was significantly (p<4.7×10 −6 ) enriched among the respective disease – protein signatures. Significant findings are highlighted by dots, with colour representing transformed p-values and size reflecting odds ratios. ACKNOWLEDGEMENT The authors acknowledge the Scientific Computing of the IT Division at the Charité – Universitätsmedizin Berlin for providing computational resources that have contributed to the research results reported in this paper ( https://www.charite.de/en/research/research_support_services/research_infrastructure/science_it/#c30646061 ). This work was supported by the DZHK (German Centre for Cardiovascular Research) and the BMBF (German Ministry of Education and Research) to C.L. (Grant Number: 81X2100281), as well as by the DFG (German Research Foundation) to K.D. (Walter Benjamin Fellowship, Grant Number: 547107463). Co-funded by the European Union (ERC, GenDrug, 101116072) to M.P.. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council. Neither the European Union nor the granting authority can be held responsible for them. The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript. This work was supported by the de.NBI Cloud within the German Network for Bioinformatics Infrastructure (de.NBI) funded by the German Federal Ministry of Education and Research (BMBF) (031A532B, 031A533A, 031A533B, 031A534A, 031A535A, 031A537A, 031A537B, 031A537C, 031A537D, 031A538A). Footnotes There was a major update in some of the underlying data resources, including self-reported drug ascertainment and disease coding in UK Biobank, as well as the use of an updated version of the GWAS Catalog to generate the knowledge graph. REFERENCES 1. ↵ Topol , E. J . The revolution in high-throughput proteomics and AI . Science 385 , eads5749 ( 2024 ). OpenUrl CrossRef PubMed 2. ↵ Ioannidis , J. P. A. & Bossuyt , P. M. M. Waste, Leaks, and Failures in the Biomarker Pipeline . Clin. Chem . 63 , 963 – 972 ( 2017 ). OpenUrl Abstract / FREE Full Text 3. ↵ Uhlén , M. et al. Proteomics. Tissue-based map of the human proteome . Science 347 , 1260419 ( 2015 ). OpenUrl Abstract / FREE Full Text 4. ↵ Uhlén , M. et al. The human secretome . 0274 , 1 – 9 ( 2019 ). OpenUrl 5. ↵ Sun , B. B. et al. Plasma proteomic associations with genetics and health in the UK Biobank . Nature ( 2023 ) doi: 10.1038/s41586-023-06592-6 . OpenUrl CrossRef 6. Dhindsa , R. S. et al. Rare variant associations with plasma protein levels in the UK Biobank . Nature 622 , 339 – 347 ( 2023 ). OpenUrl CrossRef PubMed 7. ↵ Sveinbjornsson , G. , Magnusson , M. I. , Helgason , A. & Oddsson , A . Large-scale plasma proteomics comparisons through genetics and disease associations . ( 2023 ) doi: 10.1038/s41586-023-06563-x . OpenUrl CrossRef 8. ↵ Pietzner , M. et al. Mapping the proteo-genomic convergence of human diseases . Science 374 , eabj1541 ( 2021 ). OpenUrl CrossRef PubMed 9. Sun , B. B. et al. Genomic atlas of the human plasma proteome . Nature 558 , 73 – 79 ( 2018 ). OpenUrl CrossRef PubMed 10. Suhre , K. , McCarthy , M. I. & Schwenk , J. M . Genetics meets proteomics: perspectives for large population-based studies . Nat. Rev. Genet . 22 , 19 – 37 ( 2021 ). OpenUrl PubMed 11. ↵ Emilsson , V. et al. Co-regulatory networks of human serum proteins link genetics to disease . Science 361 , 769 – 773 ( 2018 ). OpenUrl Abstract / FREE Full Text 12. ↵ Zheng , J. et al. Phenome-wide Mendelian randomization mapping the influence of the plasma proteome on complex diseases . Nat. Genet . 52 , 1122 – 1131 ( 2020 ). OpenUrl CrossRef PubMed 13. Zuber , V. et al. Combining evidence from Mendelian randomization and colocalization: Review and comparison of approaches . Am. J. Hum. Genet . 109 , 767 – 782 ( 2022 ). OpenUrl CrossRef PubMed 14. ↵ Xu , Y. et al. An atlas of genetic scores to predict multi-omic traits . Nature 616 , 123 – 131 ( 2023 ). OpenUrl CrossRef PubMed 15. ↵ Deng , Y.-T. et al. Atlas of the plasma proteome in health and disease in 53,026 adults . Cell 0 , ( 2024 ). 16. ↵ Garcia , T. et al. Relationship of proteins and subclinical cardiovascular traits in the population-based LIFE-Adult study . Atherosclerosis 398 , ( 2024 ). 17. ↵ Meyer , D. H. & Schumacher , B . Aging clocks based on accumulating stochastic variation. Nat . Aging 4 , 871 – 885 ( 2024 ). OpenUrl PubMed 18. ↵ Carrasco-Zanini , J. et al. Mapping biological influences on the human plasma proteome beyond the genome . Nat. Metab . 1 – 14 ( 2024 ) doi: 10.1038/s42255-024-01133-5 . OpenUrl CrossRef 19. ↵ Meinshausen , N. & Bühlmann , P. Stability selection . J R Stat. Soc B 72 , 417 – 473 ( 2009 ). OpenUrl 20. ↵ Geyer , P. E. , et al. Plasma Proteome Profiling to detect and avoid sample-related biases in biomarker studies . EMBO Mol. Med . 11 , e10427 ( 2019 ). OpenUrl CrossRef PubMed 21. ↵ Yunga , S. T. et al. Effects of ex vivo blood anticoagulation and preanalytical processing time on the proteome content of platelets . J. Thromb. Haemost . 20 , 1437 – 1450 ( 2022 ). OpenUrl PubMed 22. ↵ Korff , K. et al. Pre-Analytical Drivers of Bias in Bead-Enriched Plasma Proteomics . 2025.05.07.652495 Preprint at doi: 10.1101/2025.05.07.652495 ( 2025 ). OpenUrl Abstract / FREE Full Text 23. ↵ Karlsson , M. et al. A single-cell type transcriptomics map of human tissues . Sci. Adv . 7 , ( 2021 ). 24. ↵ Kalman , S. M . Effects of Oral Contraceptives . Annu. Rev. Pharmacol. Toxicol . 9 , 363 – 378 ( 1969 ). OpenUrl CrossRef 25. ↵ Ikeda , K. et al. Serum surfactant protein D predicts the outcome of patients with idiopathic pulmonary fibrosis treated with pirfenidone . Respir. Med . 131 , 184 – 191 ( 2017 ). OpenUrl PubMed 26. ↵ Hellenthal , F. A. M. V. I. , Buurman , W. A. , Wodzig , W. K. W. H. & Schurink , G. W. H . Biomarkers of AAA progression. Part 1: extracellular matrix degeneration . Nat. Rev. Cardiol . 6 , 464 – 474 ( 2009 ). OpenUrl CrossRef PubMed Web of Science 27. ↵ Curci , J. A. , Liao , S. , Huffman , M. D. , Shapiro , S. D. & Thompson , R. W . Expression and localization of macrophage elastase (matrix metalloproteinase-12) in abdominal aortic aneurysms . J. Clin. Invest . 102 , 1900 – 1910 ( 1998 ). OpenUrl CrossRef PubMed Web of Science 28. ↵ Di Gregoli , K. , Atkinson , G. , Williams , H. , George , S. J. & Johnson , J. L . Pharmacological Inhibition of MMP-12 Exerts Protective Effects on Angiotensin II-Induced Abdominal Aortic Aneurysms in Apolipoprotein E-Deficient Mice . Int. J. Mol. Sci . 25 , 5809 ( 2024 ). OpenUrl CrossRef PubMed 29. ↵ Pietzner , M. et al. Systemic proteome adaptions to 7-day complete caloric restriction in humans . Nat. Metab . 1 – 14 ( 2024 ) doi: 10.1038/s42255-024-01008-9 . OpenUrl CrossRef 30. ↵ Gambino , S. R. SERUM-BILIRUBIN AND FASTING . The Lancet 300 , 1369 ( 1972 ). OpenUrl 31. ↵ Barrett , P. V. D . Hyperbilirubinemia of Fasting . JAMA 217 , 1349 – 1353 ( 1971 ). OpenUrl CrossRef PubMed 32. ↵ Qian , L. et al. Proteomic landscape of epithelial ovarian cancer . Nat. Commun . 15 , 6462 ( 2024 ). OpenUrl CrossRef PubMed 33. ↵ Helgason , H. et al. Evaluation of Large-Scale Proteomics for Prediction of Cardiovascular Events . JAMA 330 , 725 – 735 ( 2023 ). OpenUrl CrossRef PubMed 34. ↵ Oh , H. S.-H. et al. Organ aging signatures in the plasma proteome track health and disease . Nature 624 , 164 – 172 ( 2023 ). OpenUrl CrossRef PubMed 35. ↵ Pöyhönen-Alho , M. K. et al. Central sympatholytic therapy has anti-inflammatory properties in hypertensive postmenopausal women . J. Hypertens . 26 , 2445 ( 2008 ). OpenUrl PubMed 36. ↵ Myo Min , K. K. , et al. Desmoglein-2 is important for islet function and β-cell survival . Cell Death Dis . 13 , 1 – 15 ( 2022 ). OpenUrl CrossRef PubMed 37. ↵ Kofler , B. et al. Expression of Neuropeptide Galanin and Galanin Receptors in Human Skin . J. Invest. Dermatol . 122 , 1050 – 1053 ( 2004 ). OpenUrl CrossRef PubMed Web of Science 38. ↵ Lang , R. et al. Physiology, Signaling, and Pharmacology of Galanin Peptides and Receptors: Three Decades of Emerging Diversity . Pharmacol. Rev . 67 , 118 – 175 ( 2015 ). OpenUrl CrossRef PubMed 39. ↵ Carrasco-Zanini , J. et al. Proteomic signatures improve risk prediction for common and rare diseases . Nat. Med . 30 , 2489 – 2498 ( 2024 ). OpenUrl CrossRef PubMed 40. ↵ Sudlow , C. et al. UK biobank: an open access resource for identifying the causes of a wide range of complex diseases of middle and old age . PLoS Med . 12 , e1001779 ( 2015 ). OpenUrl CrossRef PubMed 41. ↵ Assarsson , E. et al. Homogenous 96-plex PEA immunoassay exhibiting high sensitivity, specificity, and excellent scalability . PLoS ONE 9 , ( 2014 ). 42. ↵ Sinnott-Armstrong , N. et al. Genetics of 35 blood and urine biomarkers in the UK Biobank . Nat. Genet . ( 2021 ) doi: 10.1038/s41588-020-00757-z . OpenUrl CrossRef PubMed 43. ↵ Powell , R. et al. Development and validation of total and regional body composition prediction equations from anthropometry and single frequency segmental bioelectrical impedance with DEXA . medRxiv 44 , 1 – 15 ( 2020 ). OpenUrl 44. ↵ Bastarache , L . Using Phecodes for Research with the Electronic Health Record: From PheWAS to PheRS . Annu. Rev. Biomed. Data Sci . 4 , 1 – 19 ( 2021 ). OpenUrl CrossRef PubMed 45. ↵ Wu , Y. et al. Genome-wide association study of medication-use and associated disease in the UK Biobank . Nat. Commun . 10 , ( 2019 ). 46. ↵ Stekhoven , D. J. & Bühlmann , P. MissForest – nonparametric missing value imputation for mixed-type data . Preprint at doi: 10.48550/arXiv.1105.0828 ( 2011 ). OpenUrl CrossRef 47. ↵ Bycroft , C. et al. The UK Biobank resource with deep phenotyping and genomic data . Nature 562 , 203 – 209 ( 2018 ). OpenUrl CrossRef PubMed 48. ↵ Karczewski , K. J. et al. Pan-UK Biobank GWAS improves discovery, analysis of genetic architecture, and resolution into ancestry-enriched effects . 2024.03.13.24303864 Preprint at doi: 10.1101/2024.03.13.24303864 ( 2024 ). OpenUrl Abstract / FREE Full Text 49. ↵ Friendly , M . HE Plots for Multivariate Linear Models . J. Comput. Graph. Stat . 16 , 421 – 444 ( 2007 ). OpenUrl 50. ↵ Kursa , M. B. & Rudnicki , W. R . Feature selection with the boruta package . J. Stat. Softw . 36 , 1 – 13 ( 2010 ). OpenUrl CrossRef PubMed 51. ↵ Ochoa , D. et al. The next-generation Open Targets Platform: reimagined, redesigned, rebuilt . Nucleic Acids Res . 51 , D1353 – D1359 ( 2023 ). OpenUrl CrossRef PubMed 52. ↵ Hagberg , A. A. , Schult , D. A. & Swart , P. J. Exploring Network Structure, Dynamics, and Function using NetworkX . in 11 – 15 (Pasadena, California, 2008 ). doi: 10.25080/TCWV9851 . OpenUrl CrossRef 53. ↵ Perrone , G. , Unpingco , J. & Lu , H. Network visualizations with Pyvis and VisJS . Preprint at doi: 10.48550/arXiv.2006.04951 ( 2020 ). OpenUrl CrossRef 54. ↵ Buniello , A. et al. The NHGRI-EBI GWAS Catalog of published genome-wide association studies, targeted arrays and summary statistics 2019 . Nucleic Acids Res . 47 , D1005 – D1012 ( 2019 ). OpenUrl CrossRef PubMed View the discussion thread. Back to top Previous Next Posted July 15, 2025. Download PDF Supplementary Material Email Thank you for your interest in spreading the word about medRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Machine learning-guided deconvolution of plasma protein levels Message Subject (Your Name) has forwarded a page to you from medRxiv Message Body (Your Name) thought you would like to see this page from the medRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Machine learning-guided deconvolution of plasma protein levels Maik Pietzner , Carl Beuchel , Kamil Demircan , Julian Hoffmann Anton , Wenhuan Zeng , Werner Römisch-Margl , Summaira Yasmeen , Burulça Uluvar , Martijn Zoodsma , Mine Koprulu , Gabi Kastenmüller , Julia Carrasco-Zanini , Claudia Langenberg medRxiv 2025.01.09.25320257; doi: https://doi.org/10.1101/2025.01.09.25320257 Share This Article: Copy Citation Tools Machine learning-guided deconvolution of plasma protein levels Maik Pietzner , Carl Beuchel , Kamil Demircan , Julian Hoffmann Anton , Wenhuan Zeng , Werner Römisch-Margl , Summaira Yasmeen , Burulça Uluvar , Martijn Zoodsma , Mine Koprulu , Gabi Kastenmüller , Julia Carrasco-Zanini , Claudia Langenberg medRxiv 2025.01.09.25320257; doi: https://doi.org/10.1101/2025.01.09.25320257 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Epidemiology Subject Areas All Articles Addiction Medicine (568) Allergy and Immunology (863) Anesthesia (299) Cardiovascular Medicine (4422) Dentistry and Oral Medicine (443) Dermatology (382) Emergency Medicine (607) Endocrinology (including Diabetes Mellitus and Metabolic Disease) (1507) Epidemiology (15217) Forensic Medicine (30) Gastroenterology (1123) Genetic and Genomic Medicine (6584) Geriatric Medicine (667) Health Economics (997) Health Informatics (4524) Health Policy (1367) Health Systems and Quality Improvement (1611) Hematology (540) HIV/AIDS (1264) Infectious Diseases (except HIV/AIDS) (15909) Intensive Care and Critical Care Medicine (1103) Medical Education (622) Medical Ethics (145) Nephrology (667) Neurology (6583) Nursing (345) Nutrition (998) Obstetrics and Gynecology (1143) Occupational and Environmental Health (956) Oncology (3330) Ophthalmology (970) Orthopedics (369) Otolaryngology (420) Pain Medicine (435) Palliative Medicine (129) Pathology (663) Pediatrics (1690) Pharmacology and Therapeutics (691) Primary Care Research (710) Psychiatry and Clinical Psychology (5437) Public and Global Health (9218) Radiology and Imaging (2194) Rehabilitation Medicine and Physical Therapy (1369) Respiratory Medicine (1195) Rheumatology (593) Sexual and Reproductive Health (709) Sports Medicine (529) Surgery (709) Toxicology (99) Transplantation (289) Urology (265) (function(){function c(){var b=a.contentDocument||a.contentWindow.document;if(b){var d=b.createElement('script');d.innerHTML="window.__CF$cv$params={r:'9ff728b959001b23',t:'MTc3OTQwNDAyNQ=='};var a=document.createElement('script');a.src='/cdn-cgi/challenge-platform/scripts/jsd/main.js';document.getElementsByTagName('head')[0].appendChild(a);";b.getElementsByTagName('head')[0].appendChild(d)}}if(document.body){var a=document.createElement('iframe');a.height=1;a.width=1;a.style.position='absolute';a.style.top=0;a.style.left=0;a.style.border='none';a.style.visibility='hidden';document.body.appendChild(a);if('loading'!==document.readyState)c();else if(window.addEventListener)document.addEventListener('DOMContentLoaded',c);else{var e=document.onreadystatechange||function(){};document.onreadystatechange=function(b){e(b);'loading'!==document.readyState&&(document.onreadystatechange=e,c())}}}})();