Systematic Development of Sandwich Immunoassays for the Plasma Secretome

The plasma proteome offers a clinically useful window into human health and disease. With recent progress made on the development of highly multiplexed immunoassays with high sample throughput, a remaining need is to establish a pipeline for validating the individual proteins that build such bio-signatures by using targeted assays. In order to streamline such efforts, we developed a workflow to build dual binder sandwich immunoassays (SIA) and chose to evaluate this on proteins predicted to be secreted form cells and tissues. Utilizing the multiplexing capacities of the bead array technology, we first screened ~ 1,800 unique antibody pairs against 209 protein targets and collected data from dilution series of recombinant proteins as well as EDTA plasma. Employing 624 unique antibodies from the Human Protein Atlas, we obtained dilution-dependent curves in plasma and concentration-dependent curves of full-length proteins for 102 (49%) of the targets. For 22 protein assays, the longitudinal, inter-individual and technical performance was determined in a set of plasma samples collected from 18 healthy subjects every third month over one year. Lastly, we compared 14 of these assays with SIAs composed of other binders, proximity extension assays and affinity-free targeted mass spectrometry. Our workflow provides a multiplexed approach to screen for SIA pairs that suggests using at least three antibodies per target. This design is applicable for a wider range of targets of the plasma proteome, while the assays can be applied for discovery but also to validate emerging candidates derived from other platforms.


Introduction
There is a continuously great interest in increasing our understanding about those proteins that are expected to be present in blood and found outside the intracellular space, and to apply appropriate tools to discover and validate these in a given study context [1]. Such efforts preferably target the proteins that are actively secreted in comparison to those that appear in blood due to leakage, cell death or cellular turnaround. Today, the human secretome can be defined by bioinformatics tools annotating our genome based on sequences found in the proteinencoding regions [2]. Using an updated annotation [3], more than 2,600 proteins were defined as the secretome. Of these only about 600 proteins are predicted to be actively secreted to the blood while another 1,000 proteins were localized to the membrane and/or the intracellular space [3].
The technically measurable content of the human plasma proteome currently contains nearly 5,000 proteins when combining the efforts conducted with mass spectrometry (MS) as well as immunoassay platforms [4]. It has been shown that while untargeted and MS-based approaches contribute to this list with primarily cellular components, immunoassays are often more sensitive to detect low abundant proteins linked to cytokines and inflammation processes.
However, and upon excluding the recent large-scale aptamer studies, only about a third of the currently annotated 2,600 proteins of the secretome [3] can be measured in plasma using other methods. Plasma profiling efforts using shotgun MS, such as those by Mann and co-workers, detected 1,200 proteins [5] in plasma. The latest versions of multiplexed immunoassay, not included in the above stated collection of plasma proteins, used 5,000 aptamers to profile 5,000 donors, as shown by Emilsson et al [6].
Here we present a complementary approach to multiplexed assays systems and systematically build sandwich immunoassays for the proteins of the plasma secretome. Our efforts are centered around the feasibility of screening and validating the antibody pairs for a variety of proteins at the same time, rather than focusing on only a few shortlisted targets. Hence it expands on previous workflows that primarily work on selected candidates [7]. Our approach was accelerated by accessing a large resource of antibodies from the Human Protein Atlas (HPA) and full-length proteins generated within the Human Secretome Project [3] (HSP) within the Wallenberg Centre for Protein Research (WCPR). The study was conducted on a multiplexed bead array platform and combined the assessment of antibody pairs using both recombinant proteins and EDTA plasma. We did not preselect the secreted proteins based on other prior

Results
We aimed to develop a multiplexed workflow (Figure 1) to search and select for antibody pairs for the analysis of proteins secreted into human plasma ( Figure 2). We combined the capabilities of the suspension bead array (SBA) technology with the resource of HSP's fulllength proteins and antibodies generated by the HPA project and investigated > 200 proteins as well as ~ 1,800 possible antibody pairs. The project was designed to be conducted in the following stages: (1) screening for possible antibody (Ab) pairs in dilution series of protein and plasma samples, (2) pre-selection of suitable pairs after assessing their apparent functionality, (3) annotation of pre-selected Ab pairs according to their binding area, (4) selection of antibody pairs for further investigations focusing on technical aspects, (5) preparation of duplex sets for plasma analysis, (6) quantification of plasma protein levels in a longitudinal sample set, and (7) compare these results with data from independent, orthogonal methods.

Screening for antibody pairs Experimental study design
We studied a total of 209 full length proteins and used a pool of EDTA plasma samples to determine and to develop SIAs. The screening was conducted in two rounds of 109 and 124 proteins, where we aimed at replicating the findings from the first round by also including all targets with an apparent functional antibody pair, corresponding to 23 proteins, in screening round 2. For finding antibodies from the HPA resource, we chose a concentration cut-off of 0.05 mg/ml for protein capture, and found 624 antibodies for all proteins (see Supplementary   Table 1). This meant that assays could be developed using an average of 3-4 antibodies per protein, and the coverage ranged from 1-8 HPA antibodies per protein (see Supplementary Figure 2). We chose to combine 49-88 different Abs in one SBA. Among all 624 Abs, we selected those with concentrations ≥ 0.1 mg/ml and an available volume ≥ 0.5 ml as detection agents. This means that an average of 2-3 Abs were biotinylated per target protein, and the coverage ranged from 1-7 detection Abs (detAbs). Hence a total of 2,170 antibody pairs were investigated, of which 1,791 were unique.
The screening rounds were conducted by grouping the proteins into sets of 6 per assay batch and SBA. Each protein assay contained an SBA of the corresponding Abs as well as those targeting the other five proteins. Each SBA also contained control beads to judge the unspecific binding to the beads, and was distributed into 2x 384 well plates. The total number of assays per protein was defined by the number of available detAbs, and each detAb was used in 8 concentration levels of proteins and 8 dilutions steps for EDTA plasma. In total, we conducted 9,264 assays and generated 553,056 data points for the 209 proteins.

Reproducibility of screening results
To assess the reproducibility between the two performed screening rounds, 23 proteins corresponding to 200 unique and target matched antibody pairs were included in the second screenings. The assay conditions were in terms of dilution factor, number of dilution steps and the starting concentration for both plasma and protein maintained in the second screening. Out of the 23 targets, assay pairs for 16 targets (= 70%) revealed reproducible binding curves for both protein and EDTA plasma, as exemplified by ANGPTL3 shown in Figure 3. For each target, at least one unique antibody pair had a correlation of R 2 > 0.92 for the protein as well as R 2 > 0.86 in plasma. In total, 73% of all overlapping pairs had R 2 > 0.95 in protein assays, while 59% were > 0.95 in assays with plasma. The pairs towards the additional 6 targets were regarded as reproduced for the protein assay, however the detectability in plasma was lower compared to the first screening. A higher degree of reproducibility was observed for targets that provided signal intensity levels further away from the apparent limit of detection (LOD).
Additionally, using two different batches of EDTA plasma pool for the two screening rounds could have influenced the reproducibility.

Pre-selection of antibody pairs
From the generated data, binding curves were manually annotated for plasma and protein in order to classify each antibody pair and select those for further optimization as illustrated in Figure 2. A summary of the outcome of the selection process is shown in Table 1. To select pairs based on their apparent functionality, we assessed the shape and concentration dependency of the curve for the expected antibody pair with the used assay conditions. We also considered all other Abs (off-target antibody pairs) included in each SBA as a background measure and noted if unexpected pairs were found in either the protein or plasma samples. All possible pairs were then assigned to one of the following four classifications according to their functionality: 1. Dilution and concentration dependent curves plasma and protein, respectively 2. Concentration dependent curves with protein only 3. Dilution dependent curves with plasma only 4. No dilution or concentration dependent curves As also summarized in Supplementary Figure 3, from almost 1,800 possible Ab pairs there were 170 from screening #1 and 247 from screening #2 that detected their target in plasma and recombinant protein. In total, 361 unique pairs were consequently annotated as "passed" and considered for further assessment and optimization. The remaining pairs did not show a concentration dependent curve for both sample types and may require further time to develop, hence were not considered for further sample analysis.

Annotation of pre-selected antibody pairs
As an additional assessment, we investigated the location on the protein to which the pairs of capture antibody (capAb) and detAb bound their respective target. We chose to approximate the binding areas of the Abs by using their immunogens aminoacidic sequences (22-151 residues in length) and mapping these to the sequences of the canonical protein. Here, we segmented each protein sequence into three equally long parts (N-terminal, middle and C-terminal). As shown in Table 2, there were generally more pairs for constellations that targeted the same region, which was due to using antibodies for capture and detection, generated towards an epitope located in the same region. Also, there were more pairs targeting the middle and C-terminal region than in combination with N-terminal binders. Considering the success rate for building SIAs from the screening assessment criteria, we found that on average about 12% of all pre-selected pairs passed these. A slightly higher success rate of 16% was found for purely N-terminal targeted antibody pairs as well as pairs built with a capAb and detAb targeting the C-terminal and middle, respectively (18%). The lowest success rate of ~ 6% was related to pre-selected binder pairs targeting each one of the termini. Out of the total 1791 pairs, there were 138 pairs passing the selection process of which both Abs targeted the same region.

Selection of antibody pairs
In a third step, we aimed at further shortlisting those pre-selected Ab pairs. As a primary criterion, the generated level of intensity (reported as MFI) was chosen as an additional cut-off in order to report signals that were > 5x above an average background level determined by the assay controls (MFI = 30 AU). Those pairs that did not reach a maximal MFI from the protein assay curves of MFI > 150 were therefore excluded. Out of the initial 361 pairs, we used 221 for further studies, which corresponded to 70 out of the 102 initial proteins.
The subsequent investigations focused on technical aspects such assay reproducibility, the apparent LOD using proteins assays, as well as sensitivity of detecting the target protein in plasma samples. All analyses were conducted using triplicates of protein concentration series.
To further resemble a sample matrix of higher complexity, the buffer used for technical assessment of protein assays and plasma samples was supplemented by adding 1% BSA. For each protein, one detAb was prioritized to limit the number of total assays. In cases where several detAbs were available after pre-selection, additional criteria for prioritization were applied: Ab pairs with the widest range of detectable concentrations of proteins in buffer and plasma, an overall lower background level in antigen-free samples, and no previous indications about possible interferences or off-target recognition of other captured proteins. The latter was possible to be observed during the screening phase, where five other proteins were also tested in parallel, as each SBA was built with a common set of 49-88 Abs covering six proteins. For the detAbs, any concentration dependent binding for other Ab-coupled beads in the SBA, such as the internal controls, were added as exclusion criteria. Lastly, the available antibody volume was considered for the polyclonal binders.
For the selection processes of Ab pairs, target proteins were regrouped into new sets of five protein targets. Concentration of the proteins and the dilution of EDTA plasma were adapted for each individual target according to the data obtained during the screening. Each assay therefore covered a broader range of concentrations in order to determine the optimal dilution point for plasma analysis. Exemplified for SIA pairs targeting ANGPTL3, CHIT1, CPA1 and FGF21, shown in Figure 4A-D, protein detection was specific and accompanied by only very minor increase in signals from other beads. When analyzing plasma, we found that background signals from other antibodies arose when using plasma at a lower dilution than 1:50 dilutions.
Still, at a plasma dilution of 1:12, the intended signals were > 5-fold above any other binder pair.
Out of the 221 pairs targeting 70 proteins, we found 43 pairs for 27 proteins suitable for further analysis according to the criteria stated above. To further find the best performing Ab pairs for one protein, we determined the coefficient of variation (CV) by calculating the variance for each dilution point using log2 data, and then using the average across all dilutions within this range for ranking the pairs. As shown in the annotation table (Supplementary Table 2) the average CV using log-transformed data was 2.2% for plasma and 2% for protein standard and ranged from 0.4% to 7% in plasma and 0.9% to 4.8% in protein. A set of 10 Ab pairs showing averaged CVs for > 3.3% in protein assays and > 4.3% in plasma assays were excluded from further analysis. Prior to choosing the final set of Ab pairs for plasma profiling, sample and protein consumption was considered. Plasma assays requiring more samples > 12.5 µl per assay (representing a 1:4 sample dilution) and amounts of proteins exceeding 150 ng (= 3000 ng/ml as the highest concentration point) were deprioritized. This led to 32 Ab pairs against 22 proteins for further plasma analysis.

Preparation of Ab pairs for duplexed plasma analysis
To achieve a more efficient sample analysis, the data from each protein and plasma dilution curve were compared. The concentration levels for a 50% effective dose (ED50) were calculated and chosen as the optimal sample dilution point. In order to find the optimal plasma dilution factor per target protein, one Ab pair and ED50 had to be chosen per protein. In cases of similar performance assessment characteristics (see above), Ab pairs generated towards different binding regions were prioritized. Also, Ab pairs with the superior lower limit of quantification (LLOQ) were preferred as these generally allow us to cover a broader range of protein levels. To attempt for a higher protein throughput, improve time-and cost-efficiency of the assays and also reduce sample consumption, we searched for possible combinations of different Ab pairs with the same optimal sample dilution and limited us to assays in duplex.
Some combinations were directly excluded due to previously observed incompatibility, so that 4 duplex assays and 14 single-plex assays remained, as shown in in Table 3. We did not find a direct relation between the protein concentration values found in the literature and the degree of sample dilution (Supplementary Figure 4).

Analysis of protein levels in a longitudinal sample set.
Lastly, the selected 22 Ab pairs were used in SIAs for the determination of protein levels in a collection of longitudinal plasma samples. The study set was built of 18 individuals that donated plasma every third month over one year. Using the 72 samples collected from four visits each subject allowed us on the one hand to determine the technical suitability of the selected Ab pairs for analysis of proteins, on the other hand we could illustrate how protein levels of individuals vary longitudinally and between sample collections.
We quantified 21 of the tested 22 proteins and listed the performance of the assays in Table 3, where the stated protein concentration for each target was calculated from the average concentration over all samples per donor. The protein levels determined here generally agreed well with those found in the literature (see Supplementary Figure 5). The standard curves from the new assays are shown in Figure 5A-D and relate to those introduced in Figure 4A-D.
In Figure 6 we further compared three layers of variance: technical precision, inter-individual differences as well as longitudinal changes. The ternary plot showed that several proteins, such as CHIT1, CPA1 or FGF21, were stable over time and could be accurately measured, while differing in levels between the donors. The data for ANGPTL3 however was less conclusive due to an elevated technical variance (CV > 21%, using raw data).
In addition to variance analysis, we used distances from clustering analysis to compare the inter-individual differences and the intra-individual differences. Our analysis reveals an average intra-individual Euclidean distance of 3.4 compared to 5.4 between the individuals. This is in line with other observations that protein levels in plasma remain constant over the course of one year and that each person has a unique profile.

Comparison to orthogonal plasma assays
Lastly, we aimed to confirm the data obtained by the selected Ab pairs though using additional analyses. This assessment was based on comparing our data with results from targeted plasma MS analysis [8] and solution-based proximity extension assays (PEA) [9]. For above methods, we obtained data sets generated in previous studies of the longitudinal sample analyzed in the application phase (Fagerberg et al, unpublished). Using direct correlation analysis as a proxy to determine the similarity between the generated data sets, protein levels from 14 targets were studied. Of the alternative methods, data for 10 proteins only was available for PEA and for 4 proteins from MS only. As shown in Table 4 and Figure 7A-D correlations between our protein levels and another affinity-based method, PEA, reached R 2 = 0.6 ± 0.2 while correlations with peptide abundance from MS were R 2 = 0.3 ± 0.2. This illustrates that it was possible to obtain supportive evidence for some of the target proteins, but differences between the assay types in terms of sensitivity and assay interference between the technologies may have contributed to a reduction in concordance. It is worth noting that the assays were performed in different labs and at different timepoints, too. When choosing other capAbs of the SBAs than those shortlisted for the preferred pairs, an additional set of 10 capAbs were available to compare the data from the primary Ab pairs with. Since the data from these assays was obtained from the same sample incubation and used the same detAb, it was less surprising but reassuring to find a high correlation between the primary and additional Ab pairs of R 2 = 0.9 ± 0.1 (Table 4).

Discussion
This study describes a workflow based on a multiplexed bead-based platform and high-quality reagent resources to systematically screen, select and apply pairs of antibodies for the quantification of secreted proteins in human plasma. Starting from 1,791 unique antibody pairs built on 624 unique antibodies, two rounds of screening were conducted on parallel dilution series of EDTA plasma and the recombinant full-length target protein. We found that 20% or 361 of all possible pairs detected the recombinant as well as the plasma protein in a concentration dependent manner. We applied a selected set of 32 SIA pairs to study the protein levels in plasma collected from 18 subjects every third month over one year and lastly confirmed these findings by using orthogonal assays for 14 targets. For 6 of these, the Pearson correlation between the orthogonal assay and the validated SIA pair was R 2 > 0.6.
Our study was based on the use of available HSP proteins and HPA antibodies. The polyclonal binders undergo a stringent quality assessment for the use of immunoassays such as Western blot, immunohistochemistry and confocal microscopy. However, the functionality in other types of assays and samples require separate efforts and we did not find a direct link between building functional SIA pairs with pre-assessing these binders in Western blots on plasma (data not shown). Further to this did we not include affinity reagents generated by other providers, which may have limited us in providing larger number of assays in the end. In addition, we estimate that many of the capAbs that are currently part of "non-functional" pairs could indeed enrich the target protein of interest, but the tested detAb was not suitable in combination with these. We also acknowledge the fact that polyclonal antibodies need to be regenerated for The proteins that we quantified here are important indicators of health status relevant for different diseases: FGF21 for example is a known key regulator in lipid and glucose metabolism, which is increased in conditions such as type 2 diabetes, obesity and nonalcoholic fatty liver disease [10]. CHIT1 serves as a neuroinflammatory marker and shows increased concentrations in sALS (sporadic amyotrophic lateral sclerosis) [11]. ER stress is hypothesized to lead to hereditary pancreatitis may promote the development of pancreatic ductal adenocarcinoma. CPA1 is among the highest expressed genes in acinar cells, thus CPA1 proteins are expected to cause more ER stress than lower expressed pancreatic enzymes and are indicated to be associated not only with pancreatic cancer development, but also its susceptibility [12]. ANGPTL3 is a novel factor modulating the plasma lipoprotein metabolism [13] and may additionally contribute to uremic dyslipidemia [14]. Plasma levels of NPPA have been described as prognostic predictors in patients with chronic heart failure, but are also known to reflect the severity of left ventricular hemodynamic dysfunction [15]. It has been suggested that FGL1 plays a role in liver protection and liver regeneration, but it also has the potential to serve as a target for the treatment of gastric cancer and to predict gastric cancer prognosis [16]. In addition, CCL16, a human CC chemokine, has been shown to be differentially expressed in ovarian cancer [17,18].
The study presented here uses pooled plasma collected from non-diseased subjects. This sample source may have limited the possibility to detect those proteins that increase with inflammation, infection or other diseases. In addition, more assay optimization may have been necessary to rescue some of those pairs that detected their target protein but did not reveal signals above background in plasma. It may further be necessary to choose other, more sensitive detection systems and thereby sacrificing some of the SBA's capabilities in terms of target and sample throughput. Some Abs showed binding to other than their intended targets once higher amounts of plasma (in particular 1:2). This points at further optimization of the assays are needed in terms of blocking agents and that not all Ab pairs are compatible with another. Additionally, the apparent LOD and LLOQ levels may increase once more complex buffer solutions are used.
To our current knowledge, this is one of the largest and first systematic study to screen for SIAs.
We focused on the plasma secretome, as plasma is an important sample for clinical routine analysis and a highly interesting source for searching for disease related proteins. We assumed that the proteins secreted into the blood stream remain detectable in solution when plasma is being prepared. In comparison to studying proteins in cell lysates, proteins found in plasma do not require to be extracted, hence the need to apply strong detergents can be omitted. However, we acknowledge that proteins may precipitate or denature during sample processing. Some proteins are known to be unstable and degrade over time, hence the possibility to detect these decreases with the age of the sample.
In addition to expand our possibilities to measure proteins actively secreted into blood, highly multiplexed immunoassays as well as MS-based approaches require more targeted assays to validate and quantify potential findings in larger number of samples as well as using orthogonal 13 methods. It holds a great value to expand this list of assays for the plasma secretome to measure and quantify the plasma components of interest on a protein level. Considering the targets included in our study, this suggests that the detectability of proteins in plasma is predominantly depended on the interplay between the available reagents, the technology and protein stability.
We are, however, well aware that many other, such as the low abundant cytokines [19], require minimal sample dilutions to detect proteins of pg/ml concentrations, and we suggest to include necessary controls and considerate elevated and unspecific background binding.
In summary, multiplexed bead arrays were used to screen for functional antibody pairs in proteins and plasma samples. With a success rate of 20% we found that investing at least three different antibodies per target protein and assessing different capture-detection combinations was necessary to obtain antibody pairs for protein quantification. While further assay optimization, additional antibodies and target-centric studies will be needed to assess the utility of these antibody pairs, we could not observe a trend towards any binding-site preference for building functional assays for secreted proteins. Considering the need to generate renewable reagents for extended use, our study provides valuable leads on selecting and building antibody pairs even with other reagents than those used here.

Plasma samples
All methods were conducted according to the Declaration of Helsinki, which establishes the regulations and guidelines for research project execution for human health. CArdioPulmonary bioImage Study (SCAPIS), which is a prospective observational study of randomly selected subjects aged 50-64 years from the general Swedish population. Within this study, all subjects have been extensively phenotyped before entering the S3WP program [20].
Several exclusion criteria were applied for choosing study participants: 1) previously received health care for myocardial infarction, stroke, peripheral artery disease or diabetes, 2) presence of any clinically significant disease which, in the opinion of the investigator, may interfere with the results or the subject´s ability to participate in the study, 3) any major surgical procedure or trauma within 4 weeks of the first study visit, or 4) medication for hypertension or hyperlipidemia. The study was approved by the Ethical Review Board of Göteborg, Sweden.
All participants provided written informed consent. The S3WP program has the aim to collect longitudinal data in a community-based cohort and is non-interventional and observational. A total of 4 examinations were performed every third month (+/-2 weeks). All subjects were fasting overnight (at least 8 hours) before the visits. Subjects underwent the same examinations at each visit, answered a selection of questions to note changes in health and life-style factors between each visit and blood, urine and stool for subsequent clinical chemistry and omics analyses was collected at each visit (Fagerberg et al, unpublished).

Target selection and generation
Protein targets for the secretome were selected according to availability of full-length proteins within the Human Secretome Project (HSP) and antibodies from the Human Protein Atlas (HPA) as well as considering the recombinant protein concentration. HPA antibodies needed to have a concentration of > 0.05 mg/ml for being chosen as capAb and of > 0.1 mg/ml for being considered as detAb.

Production and purification of secreted proteins
Secreted proteins were defined based on data in the Uniprot database as well as signal peptide and transmembrane region predictions made for the transcripts in the Ensembl database. A generic expression cassette, based on the CMV promoter and with an N-terminal CD33 signal peptide for secretion of all produced proteins and a C-terminal human protein C tag for purification, was used. All secreted proteins were produced by using the transient Icosagen Cell Tris, 1 M NaCl, 2 mM CaCl2, pH 7.5) the protein was eluted using a mild elution with EDTA (20 mM Tris, 100 mM NaCl, 2 mM EDTA, pH 7.5) prior a buffer exchange into 1xPBS. After desalting the protein concentration was determined (Abs). Each purified protein was identified by MS/MS and the purity was analyzed using SDS-PAGE and Western Blot. Primary antibody for western blotting was a rabbit Anti-C tag polyclonal (GTX18591, Genetex). Glycosylation patterns of the purified proteins were also analysed using SDS-PAGE.

Antibodies
Overall, we included 624 antibodies targeting 209 unique secreted human proteins, as well as 11 assay specific controls. Majority of the antibodies used polyclonal rabbit antibodies generated within Human Protein Atlas project (HPA) (www.proteinatlas.org) [21]. The assay specific controls included affinity purified rabbit IgG (P120-301, Bethyl laboratories) in order to control for background binding to rabbit IgG molecules and a blocked bare bead (without coupled antibody) to monitor background binding to the beads. These two will from now on be referred to as "assay controls". In addition, a set of 10 monoclonal mouse antibodies from BioSystems International [22], targeting plasma proteins commonly enriched by immunocapture assays were included [23]. These will be referred to as "internal controls". All antibodies used are listed in (Supplementary Table 1).

Coupling of antibodies to beads
Bead arrays were created as previously described [24]. Antibodies were diluted to 17.5 µg/ml ImmunoResearch). 100 µl RPE-conjugated antibodies diluted to 0.5 µg/ml in PBS-T were added to 5 µl antibody-coupled bead stock in different wells, followed by incubation for 20 min at 650 rpm at room temperature. After incubation, wells were washed three times with 100 µl PBS-T before analyzed on a Flexmap 3D instrument (Luminex corp.). Signals for the coupling efficiency was reported in terms of median fluorescence intensities (MFI). Coupled beads were regarded as a failed coupling if the signals obtained were lower than 2x SD than the mean value for the bead stock. In case of failed coupling, a recoupling of this specific antibody was performed.

Biotinylation of detection antibodies
Antibodies used as detAbs were biotinylated as described previously [25]. In short, 2 µg of each were added and incubated for 20 min, room temperature, 650rpm. After incubation, wells were washed three times with 100 µl PBS-T, before analyzed on a Flexmap 3D instrument (Luminex corp.). Antibodies with signals 50 x above background were considered successfully biotinylated.

Assay procedure and read out
For assay performance two different batches of commercially available human K2 EDTA Life Science) and supplemented with 0.5 mg/ml rabbit IgG. A spike-in serial dilution of standard proteins in plasma dilution buffer was performed, covering a concentration range of 1 µg/ml -1 ng/ml. Blanks of both assay buffer and plasma dilution buffer were added and will be referred to as "blank sample".
The pre-diluted plasma samples and the pre-diluted protein standards (45 µl) were transferred to 5 µl bead stock in an assay plate (Greiner 384-well assay plate) using a liquid handler (SELMA, CyBio) before overnight incubation at 650 rpm at room temperature.
After incubation, the beads were washed three times with 60 µl PBS-T. The biotinylated detAbs were diluted to 1 µg/ml in PBS-T. Subsequently the beads were incubated for 1.5 h at room temperature at 650 rpm with 25 µl of the respective pre-diluted detAb. Beads were washed three times with 60 µl PBS-T before incubation with 50 µl of a 1:750 dilution of RPE-labelled streptavidin for 20 min at room temperature at 650 rpm. Finally, beads were washed three times with 60 µl PBS-T, before they were re-suspended in 60 µl PBS-T and analyzed on a Flexmap 3D instrument (Luminex corp.). Binding events were displayed as MFI where at least 50 beads per bead ID were counted.
Each assay plate represents one experimental assay run combined of 2x 96-well plates containing serial dilutions of human plasma pool as well as 2x 96-well plates containing serial diluted standard curves of the proteins investigated. Protein and plasma dilution series for the same detAbs were placed on the same 384-plate (Supplementary Figure 1). Additionally, interfering plate-effects were avoided by running all measurements for one target protein on the same 384-plate. In total, 26 384-well plates were measured, containing between 2 and 15 proteins each. A detailed plate layout can be found in the Supplementary Figure 1.

Assay optimization and validation
The assay design was modified and optimized each time between two phases. After the screening phase, the buffer matrix for the protein standards was changed by adding 1% BSA (A7030, Sigma Life Science) to achieve a higher matrix complexity. Additionally, the length of the dilution curves for the protein standard was extended from a 7-step concentration to a 14-step concentration series in triplicates covering a range of 1 µg/ml to 1 pg/ml when evaluating the reproducibility of the assays. The dilution points for plasma were also adapted to the signals achieved during screening, to both cover a broader measuring range, but also to be more suitable for the obtained signals. Thus a 7-step dilution series of human EDTA plasma pool with a consistent step size of 3 starting between 1:2 and 1:36 in plasma dilution buffer was conducted for each antibody pair. The SBAs for the selection process were composed with different capAbs for further technical investigations in order to exclude additional off-target interactions. The remaining target proteins were grouped as sets of five into SBAs containing [8][9][10][11][12][13][14][15][16][17][18] Abs, based on their alphabetical order. Each set of SBAs was supplemented with the assay controls, to record possible binding to the beads. For protein quantification and application of the SIA pairs, assays were run on a longitudinal sample set as well as an 8-step protein concentration series, covering the optimized signal range and measured in triplicates.

Selection criteria
After selecting protein targets for the secretome according to availability, and HPA antibodies both according to availability and concentration (> 0.05 mg/ml for being chosen either as capAb or detAb), antibody pairs had passed several selection rounds in order to achieve reliably functioning antibody pairs. This process was divided into three phases: an initial screening phase, which was sub-divided into two rounds, a selection phase and an application phase. For all phases MFIs were registered for each bead ID and sample. Pairs being processed to be further tested had to be assigned to group 1 as well as reach a maximum signal intensity of at least 150 MFI in order to report only signals above an average background. To limit the number of total assays, one detAb was chosen per protein. After an additional testing round of the chosen pairs in triplicates of plasma dilution and protein standard dilution the CV was determined for each dilution point averaging it across all dilution steps to find the best performing Ab pairs (see Figure 4A-D). As cut-off criteria a CV of 3.3% in protein assays and 4.3% in plasma assays was defined. In addition, pairs requiring more than 12.5 µl sample (representing a 1:4 sample dilution) or more than 150 ng (= 3000 ng/ml as the highest concentration point) per assay were excluded in respect to sample and protein consumption. For the technical replicates during the selection phase the upper limit of quantification (ULOQ), LLOQ, LOD and ED50 were calculated. One pair per target was prioritized and the calculated ED50 point was chosen as the optimal sample dilution point. In cases of similar performance, Ab pairs generated towards different binding regions were prioritized. Also, Ab pairs with the superior LLOQ were preferred (see Supplementary   Table 2).
Before processing the remaining Ab pairs into the final application phase and measuring them on a selection of 72 samples from a healthy longitudinal cohort, we combined different Ab pairs into possible duplex combinations with the same determined optimal sample dilution. The final concentration measured for each sample was calculated by transforming the measured MFI signal intensities into a concentration value according to the dilution curve obtained from the 5-parametic fit and multiplying it with the applied dilution factor of the sample (see Figure   5A-D). Samples with protein concentration below the calculated LLOQ or above ULOQ were excluded from further analysis.

Orthogonal assays
For orthogonal comparison of our targets, we correlated the overlap of our chosen 21 targets with data achieved by two independent experimental setups for the same sample selection: (1) a recently published targeted MS approach [8] and (2) multiplex proximity extension assays (Olink Bioscience, Uppsala Sweden) [9].
For the targeted MS approach 432 samples were prepared semi-automatically using the Bravo liquid handler and subsequently measured using a combination of Ultimate 3000 binary RS nano liquid chromatography (LC) system (Thermo Scientific) with an EASY-Spray ion source connected to an on-line Q Exactive HF (Thermo Scientific) MS. All plasma samples were Samples were incubated in the presence of proximity antibody pairs tagged with DNA reporter molecules. When the antibodies pair binds to their corresponding antigens, the corresponding DNA tails form an amplicon by proximity extension, which can be quantified by highthroughput real-time PCR [9,26]. To minimize inter-and intra-run variation, the data are normalized using both an internal control (extension control) and an interplate control, and then transformed using a pre-determined correction factor. The pre-processed data were provided in the arbitrary unit Normalized Protein eXpression (NPX) on a log2 scale. A high NPX presents high protein concentration [26].

Data analysis (Data processing, Classification and Curve fitting)
Data analysis and visualizations were performed within R (www.rproject.org, version R 3.5.1) [27]. To assess reproducibility for overlapping targets between the two screening rounds, corresponding MFI values were log transformed and correlated using Pearson correlation with R 2 values. To assess the binding region for capAbs and detAbs on the screened proteins, the immunogens aminoacidic sequence for each HPA was mapped to the sequences of the corresponding canonical protein (www.proteinatlas.org). Protein sequences were exported from the Uniprot data base (release 2018_07) [28].
To evaluate the performance of Ab pairs during the selection phase, data was log10 transformed and visualized as dilution curves.
Data was log10 transformed and a 5-parametric log-logistic model was applied for the dilution curves in the application phase [29]. LOD levels were calculated as 3x SD of the blank sample above the average blank signal, LLOQ was defined as 10x SD of the blank sample above the average blank signal [30], ULOQ was defined as the averaged signal of the highest protein standard concentration point minus its SD. ED50 was calculated using the drc package [29]. In instances where the SD was small, leading to negative output of the 5-parametric fit for the LOD values, the MFI values for the blank were manually increase by adding 2 AU. No significant effect on the calculated protein concentrations for those targets could be observed.
Assay CVs within the selection phase were calculated between the duplicated dilutions steps of the plasma protein curve, while during the later application phase the assay variance was calculated with the triplicated sample pool. During this phase two additional layers of variance were calculated: Variances for each protein between the 18 individuals (per visit), which will be referred to as inter-individual CV, as well as the average variance between the 18 subjects over the course of one year (four samples per subject), which we will refer to as intra-individual variance. For visualizing different layers of variance (assay variance, inter-individual variance and intra-individual variance), CVs were calculated and ternary plots were generated using the ggtern package [31].
Euclidian distances for investigating the personal plasma profile differences were calculated using the daisy function [32,33]. Prior to the calculation, the data underwent an outlier removal process, meaning values above ULOQ and below LLOQ as well as NA values were removed from the data set. The data was then scaled before computing pairwise dissimilarities and Euclidean distances. Additionally, Pearson distance for inter-individual and intra-individual correlation was calculated, using the R 2 value.
For the correlation plots between different types of data were used: MS data (fmol/µl) was mean-centered and both MS data as well as SIA data (ng/ml) were log-transformed with the binary logarithm, while PEA data (NPX values) was used as provided and correlated using Pearson correlation with R 2 values.

Author Contributions Statement
RSH, AB, MJI, LS, MD, EB, SB, CF planned and performed immunoassays. RSH, AB, UQ and JMS analyzed data. FE performed MS assays and analysis. TDC analyzed data. LF provided Olink data. JR and HT supervised the protein production and HT provided the project with full-length proteins. MU and JMS conceived the study. UQ and JMS supervised the study.
RSH, AB and JMS wrote the manuscript with scientific input from all co-authors.

Additional information
UQ is employee of Atlas Antibodies AB. MU is co-founder of Atlas Antibodies AB. FE, LF, JR, HT and JMS acknowledge formal links to Atlas Antibodies AB.  For capturing, Abs (1) were immobilized onto magnetic color-coded beads (2) and combined into sets of suspension bead arrays (3). A dilution series of EDTA plasma (4a) and protein standard (4b) was performed. Beads were then combined with either EDTA plasma or protein standard (5). After washing off unbound proteins, the captured proteins were detected via biotinylated antibodies (6)(7). The readout occurred by the addition of a streptavidinfluorophore and using the Luminex systems (8).      Protein concentration achieved from the SIA pairs was compared with data from targeted plasma mass spectrometry analysis and solution-based proximity extension assays, which was generated on the same sample set. 14 out of 21 targets were possible to compare to either of the     The determined protein concentration was correlated to concentrations for the same protein according to literature. Axes are log-scaled and a line of identity was added to the plot. Ab pairs showing a signal intensity of MFI > 150 for the protein assay during the screening phase are listed as well as their exclusion criteria during the selection phase. Pairs in red were excluded before application phase, black and green pairs were processed to the subsequent application phase. Green pairs represent the final selected pairs for application phase according to ED50 and LLOQ. Exclusion reason for each antibody is indicated with x.