Characterization and statistical modeling of glycosylation changes in sickle cell disease

Sickle cell disease (SCD), the most common hemoglobinopathy, affects up to 100 000 people in the United States and 13 million people worldwide.¹  The inheritance of a homozygous mutation from valine to glutamic acid in the hemoglobin HbS β chain causes polymerization of deoxy sickle hemoglobin within red blood cells (RBCs).²  In an oxygen-deprived state, RBCs take on a sickled shape and occlude blood vessels. Individuals are afflicted with anemia, pain crises, organ infarction, and infections; however, clinical phenotypes vary and remain unpredictable. Additionally, a heterozygous mutation results in the sickle cell trait (SCT), with predominantly silent features.³ 

SCD requires a multifaceted approach for long-term treatment.^4,5  Current SCD therapies remain limited, usually comprising hydroxyurea therapy,⁶  and increasingly gene therapy and stem cell transplants to correct hemoglobin mutations.^7,8  However, the repertoire of potential therapeutic targets continues to grow.^9-11  For example, rivipansel, a glycomimetic pan selectin antagonist, which targeted E-selectin, showed reduced resolution times of vaso-occlusive episodes¹²  but ultimately failed to meet its treatment goals.¹³ 

Glycans (carbohydrates), biologically diverse cell surface molecules,¹⁴  are often overlooked as potential mediators of vaso-occlusive crises in SCD.¹⁵  Sialic acid (SA; the ultimate “do not eat me” signal) containing glycan motifs on both N- and O-linked glycans contributes to diverse aspects of immunity, cell–cell interactions, and cell signaling.¹⁶  For example, siglecs drive B-cell development and humoral response using sialyl-glycan recognition by their receptors to modulate immune response.^17,18  Another sialyl-glycan motif, sialyl-Lewis X, binds selectins and promotes leukocyte trafficking to lymph nodes and sites of inflammation.^19,20 

Changes in sialylation and fucosylation are well documented in chronic inflammation.^21,22  Although changes in SA levels are associated with RBC sickling, there are conflicting reports of decreased²³  and increased^24,25  sialylation compared with healthy individuals. This aberrant glycan clustering and exposure of cryptic antigens and new epitopes in sickling RBCs results in antibody formation and cell destruction.^26-28  For example, the anti-Galili (galactose-α-1,3-galactose [α-Gal]) immunoglobulin G (IgG) is increased in SCD, promoting macrophage phagocytosis of sickled RBCs.²⁹  Although the inflammatory milieu in SCD is expected to increase antibody formation globally,^30,31  little is known with regard to specific glycan motifs and resultant antibodies present in this disease.

To address this knowledge gap and identify potential new therapeutic targets, we analyzed plasma and RBCs using a combination of lectin and glycan microarrays and mass spectrometry and integrated this glycan and antibody data using statistical modeling. Our data show that SCD RBCs have the following: (1) increased α2,6 sialylation; (2) decreased α2,3 sialylation and blood group antigen expression; and (3) increased glycan binding by plasma proteins in SCD to immunogenic asialo and sialyl core 1, Lewis A, and Lewis Y structures, suggesting a heightened anti-glycan immune response. Computational modeling (MixOmics) affirms changes in glycan expression and glycan-binding protein (GBP) binding, including immunoglobulins, of immunogenic glycan structures in SCD, but reveals additional changes, including ABO blood group expression.

Whole blood was prospectively obtained with an institutional review board–approved protocol as deidentified residual samples.

Data were collected from a total 50 blood group O and 37 group A African Americans with demographics distributed as in Table 1. SCT samples were collected from donors positive by SICKLEDEX (Somagen Diagnostics, Edmonton, Canada).³²  Hemoglobin levels were between 12.5 and 20.0 g/dL for healthy and SCT samples as per guidelines of blood donation (see https://www.accessdata.fda.gov/scripts/cdrh/cfdocs/cfcfr/cfrsearch.cfm?fr=630.10). SCD samples were acquired before transfusion, and hemoglobin levels were between 0.8 and 2.0 g/dL determined using a Sysmex 2000v. Samples were separated and run based on blood group and disease state, as previous research concluded that sex was not a significant contributor to differential glycosylation.³³ 

Whole blood was collected in EDTA and separated by centrifugation (3500g, 10 minutes) to remove plasma. All whole blood samples were processed 7 days after collection. Based on previous validation, this allowed for uniformity of sample age without significant loss in glycosylation (supplemental Table 1). Plasma was further centrifuged to remove residual RBC cells and fibrinogen (10 000 rpm, 20 minutes) and stored at −20°C.

The RBC fraction was further processed using a Ficoll-Paque (GE Healthcare) separation and a dextran gradient to purify erythrocytes.³⁴  Six percent dextran was mixed (1 part) with 1 part RBCs and 2 parts Dulbecco’s phosphate-buffered saline (PBS; Gibco/Thermo Fisher Scientific). Erythrocytes were allowed to sediment at room temperature for 15 minutes. The supernatant was removed, and RBC ghosts were obtained as described in Anani et al.^33,35  Patient HbS cells were morphologically counted using a Sysmex 2000v to guarantee a population of 60% or greater.

To reduce nonspecific binding, lipids were removed from RBCs with a triton extraction buffer.³⁶  Extraction buffer (500 µL) was added to the RBC pellet, incubated at 4°C with rotation (30 minutes), and centrifuged (16 000g, 15 minutes at 4°C), and supernatant collected for analysis.

Protein quantification was performed, and 0.125 µg/mL labeled protein was incubated on a GlycoTechnica LecChip microarray as previously described in Anani et al.³³  The array was scanned at 532 nm with a GlycoStation 2200 (GlycoTechnica), and fluorescence values were analyzed using SignalCapture 3.0 and GlycoStationToolsPro 3.0 (GlycoTechnica).

The RayBiotech label-based protocol (https://www.raybiotech.com/glycan-array-100/) for whole plasma was followed. Plasma was diluted 1:4 in Dulbecco’s PBS and dialyzed for 3 hours in PBS at 4°C before samples (35 µL) were biotin labeled. RayBiotech Glycan Array 100 slides were blocked for 30 minutes before 80 µg/mL sample was incubated on the arrays with agitation for 2 hours. After washing, the array was dried by centrifugation (1000 rpm, 3 minutes) and scanned in a GenePix 4000B microarray reader (Molecular Devices). Glycan arrays were scanned at 532 nm (450 photomultiplier tubes [PMT], 30% power), and spots were selected using GenePixPro 3.0 software. The median sample-specific background subtracted data for each glycan was achieved by averaging the 3 highest replicate fluorescent spot values.

Lipid extracted RBCs were prepared for mass spectrometry from erythrocyte ghosts, as prepared above. Proteins were reduced with 5 mM final concentration of tris(2-carboxyethyl) phosphine at 37°C for 45 minutes and alkylated with a 10 mM final concentration of iodoacetamide for 45 minutes in the dark at room temperature, followed by Qubit analysis for protein quantitation (Thermo Fisher Scientific).

Protein (25 μg) was immobilized on a polyvinylidene difluoride membrane with subsequent N- and O-glycan release as previously described.^37,38  Membrane spots were excised and washed in wells of a flat bottom polypropylene 96-well plate. N-glycans were released from the membrane-bound protein using 2 U PNGase F (Promega) with overnight incubation (37°C). Following N-glycan removal, 500 mM NaBH₄ in a 50 mM KOH solution was added to the membrane spots for 16 hours to release reduced O-linked glycans by reductive β-elimination. Released N-glycans were reduced (1 M NaBH₄ in 50 mM KOH solution) for 3 hours (50°C), after which the reaction was neutralized with equimolar glacial acetic acid. Both N- and O-glycans were desalted and enriched offline using strong cation exchange resin (Dowex 50WX8 [200-400 mesh], Millipore Sigma) followed by porous graphitic carbon (PGC) solid phase extraction micro-columns (Thermo Fisher Scientific).

Samples were dissolved in 65 μL of 10 mM NH₄HCO₃ containing 1 μL of dextran ladder internal standard (26 ng, centrifuged to remove particulates),³⁹  and then 60 μL of resulting solution was transferred to autosampler vials. Samples were randomized before injection. An internal standard-only sample was analyzed between each replicate block to verify no carryover. PGC-liquid chromatography-electrospray ionization- mass spectrometry/mass spectrometry (PGC-LC-ESI-MS/MS) experiments were performed using a nanoLC-2D high-performance liquid chromatography system (Eksigent) interfaced with an LTQ Orbitrap Velos mass spectrometer (Thermo Fisher Scientific). Glycans were separated on a PGC-LC column (3 μm, 100 mm × 0.18 mm, Hypercarb, Thermo Fisher Scientific) maintained at 80°C for N-glycans and 40°C for O-glycans. To enhance ionization and improve detection of low intensity glycans, postcolumn make-up flow consisting of 100% methanol was used.

Normalized retention time by dextran ladder was determined as described previously using Skyline, and proprietary raw files were converted to an open and vendor neutral file format, mzML, using Proteowizard v3.0.8725.^39,40  MS2 scans, required for confirmation of glycan composition and structure, were filtered to include only precursor masses consistent with probable human glycan compositions (searched within 20-ppm error with GlycoMod).^41,42 

Skyline v4.2.1.19095 was used for all glycan analysis as we previously described,^37,43  and 99.999% of the isotopic envelope with a centroid mass accuracy value of 20 ppm was used for peak integration for glycan structure measurements. Peak picking was largely automated based on explicit retention time, and integration bounds were manually supervised. Peak areas for each measured glycan structure were exported and normalized to relative signal in Microsoft Excel. Glycan structure relative signal was calculated for the specific glycan structure’s peak area as a percentage of the total peak area for all glycan structures measured for each sample.

MS2 scans were assigned glycan structures based on the existence of A-, B-, C-, X-, Y-, and Z-product ions matched using GlycoWorkBench v2.1 (available from https://code.google.com/archive/p/glycoworkbench/, maximum of 3 glycosidic cleavages and 1 cross-ring cleavage with 0.6-Da mass accuracy).⁴⁴  For all MS/MS scans, at least 2 probable glycan structures were used for comparative annotation to assign the most suitable structure assignment. Diagnostic ions were used to confirm glycan motifs as described previously.^42,45  To promote utility of the glycan structure libraries generated here and to adhere to MIRAGE guidelines,^46,47  all data and metadata are available on Glycopost (https://glycopost.glycosmos.org/preview/15566942915f39ca7bd7b83, PIN: 6982)⁴⁸  and Panorama Public (https://panoramaweb.org/RBCSickleGlycomics.url).⁴⁹ 

Array fluorescence data were statistically analyzed as described in Anani et al.³³  Additional details are included in supplemental Methods.

Data Integration Analysis for Biomarker discovery using Latent variable approaches for ‘Omics studies (MixOmics or DIABLO) creates pathway linkages between apparently disparate datasets to find correlated probes.^50,51  Data from the same RBC and plasma samples analyzed using lectin and glycan arrays were loaded as individual blocks, normalized as above, and annotated as healthy, SCT, or SCD. The weighted correlations were set to 0.1 between RBC lectin (RBCs analyzed with the lectin array) and plasma glycan (plasma analyzed with the glycan array) data with the remaining associations set to 0. Performance of the data were run with the following parameters: seed 123, Mfold cross-validation, 5 folds, 10 components, and 1000 repeats. The performance was modeled as centroids, mahalanobis, and maximum distances.⁵²  The smallest overall error rate was selected to choose the most influential probes in distinguishing between cohorts. Probe selection (tuning) was run with the following parameters: seed 123, Mfold cross-validation, 5 folds, 5 components, centroids distance, and 1000 repeats. Once the appropriate dimensions of data and probes were selected, the data were visualized by heat map.

Lectins are typically used for cell typing, histochemical staining, and glycoprotein fractionation.⁵³  Immobilized lectins on microarrays can rapidly screen samples for glycan changes by comparing differences in expression.⁵⁴  We used GlycoTechnica lectin arrays, containing 45 diverse lectins, to determine glycan changes in RBC glycoprotein structures between healthy donors, SCT donors, and SCD patients (Figure 1A). Using sparse partial least squares discriminant analysis (sPLS-DA), lectin binding combinations unique to each cohort were calculated and visualized with a prediction map (Figure 1B). With 1 exception, healthy and SCD samples were separated easily from one another, whereas the SCT cohort expressed glycan moieties observed in both healthy and SCD groups, precluding a complete separation from both (see supplemental Figure 1 for all lectin contributions).

Pairwise moderated t tests identified significant changes in lectin binding between groups (Figure 1C). Healthy donors expressed more terminal fucose (LTL), mannose (NPA), and galactose (ABA and DBA) motifs compared with SCD patients. LTL (blood group O), DBA (blood group A₁), and NPA identify with blood group expression,³³  indicating blood group loss in SCD RBCs. SCD patients expressed more α2,6 SA moieties (SNA, SSA), in addition to terminal GalNAc (PTL-I) and complex bisecting N-linked glycans (Calsepa) compared with healthy donors. Healthy and SCT donors differed by increased expression of asialo-galactose (BPL) and terminal mannose/GalNAc (UDA). The statistically significant differences between SCT and SCD included increased terminal GalNAc (PTL-I) and decreased α2,6-sialylation and polylactosamine in SCT.

To fully assign a structure to a glycan, a combination of MS, biosynthetic rules, chemical and enzymatic treatment, and retention time is necessary.⁴⁸  Healthy and SCD RBCs (3 samples each, both blood group O) were analyzed by PGC-LC-MS to define glycan structures. From the mixture of N-glycans released from RBCs, 118 structures were quantified over 3 orders of magnitude. These glycans cover all major glycan classes with no clear relationship between class and relative signal. Paucimannose, hybrid, and complex mono-antennary classes were detected with relative intensities less than 1%.

To identify glycan structures specific to SCD, we compared the SCD RBC N-glycome to that of healthy donors (supplemental Table 4). Of the quantified glycan structures, 14 (12%) were significantly different (P < .05 and greater than twofold difference in sickle cell/healthy average signal; Figure 2) between SCD and healthy samples. Five glycans had a higher signal in SCD samples, with all structures modified with α2,6 SA (ranging from 2.5- to 18-fold differences; Figure 2). Of the 9 glycans with higher signals in healthy donors, 3 structures are classed as hybrid-type glycans (two- to fourfold differences), and 3e structures were confirmed to feature α2,3 SA (two- to threefold differences; Figure 2). O-glycans were similarly analyzed, but no significant differences were detected between healthy and SCD RBCs (supplemental Table 5). These results agree with those observed by lectin array, as increases in SSA and SNA binding denote increases in α2,6 SA. Conversely, mass spectrometry analysis points to a decrease in α2,3 SA in SCD compared with healthy donors, supporting the increased terminal galactose moieties identified by lectins.

As changes in glycan expression in RBCs between healthy donors, SCT donors, and SCD samples were identified, we hypothesized that these glycan changes would translate to downstream alterations of plasma GBPs. To determine differences in GBP binding between healthy and SCD samples, we applied plasma to a 100-glycan microarray (Figure 3A) featuring a broad array of blood group and Lewis antigens, milk oligosaccharides, gangliosides and glycosaminoglycans, and globo-series glycolipids. These broad glycan motifs can serve as a screening tool for GBPs, including antibodies formed during immune responses.

Samples were well separated based on disease state using sPLS-DA (Figure 3B). Moderated t tests of normalized data showed significant changes in glycan binding between healthy and SCD samples and SCT and SCD samples (P ≤ .05; for glycan structure numbers and identification, see supplemental Table 7).

Increased GBP binding to fucosylated glycan motifs in SCD was observed, including 3-sialyl-3-fucosyllactose (F-SL), Lewis Y, terminal Lewis A, and Gal-β-1,4-(Fuc-α-1,3)-GlcNAc-β-1,3-Gal-β) (glycans #64, 71, 73 and 63, respectively). GBP binding to asialo, α2,6-sialylated and di-sialylated (Neu5Ac-α-2,6-[Neu5Ac-α-2,3]-Gal-β-1,3-Gal​NAc-β) core-1 O-glycan structures (glycans 22, 44, and 48, respectively) was also increased in SCD samples. Additionally, GBP binding to chitin-trisaccharide and isomaltose was increased in SCD plasma, suggesting increased antimicrobial immune responses. GBP binding to glycans 39 (Neu5Ac-α-2,6-Gal-β-1,4-Glc-β), 74 (Gal-α-1,3-Gal-β-1,3-Glc​NAc-β), and 26 (tobramycin) (Figure 3C) was decreased in SCD plasma. The data show that SCD patients have increased GBPs to immunogenic mucin glycan motifs (Lewis A and Y, and asialo and α2,6-sialylated core 1 structures) often associated with cancer and chronic inflammation and suggests a heightened immune response to aberrant glycosylation.

To combine the analysis of our different data sets, we used MixOmics (DIABLO),⁵¹  an R package that uses a multivariate method for integration of biological datasets with a particular focus on variable selection measured on the same biological samples.^50,51  MixOmics analysis distinguished healthy and SCD samples based on a unique combination of glycans and lectins (Figure 4) but was not able to isolate SCT samples. Changes in glycan and lectin binding observed with individual analyses were reaffirmed in the integrated MixOmics analysis with the addition of new interarray associations missed by singular analysis. The combined heat map showed (1) higher fluorescence in 8 lectins and 13 glycans and (2) lower fluorescence in 8 lectins and 8 glycans in SCD patients.

Sialyation changes were again observed with increased α2,6 sialylation (SSA, SNA) and decreased α2,3 sialyation (WGA, MAH) on RBCs, distinguishing SCD patients from healthy donors. However, new increased GPB binding to an α2,8 SA baring oligosaccharide (Neu5Ac-α-2,8-Neu5Ac-α-2,6-Gal-β-1,4-Glc, 54), attributed to chronic inflammation,²²  was also detected by MixOmics.

Group O and A individuals display Fucα1-2Galβ1-4GlcNAc and GalNAcα1-3(Fucα1-2)Galβ1-4GlcNAc glycan structures, respectively.⁵⁵  The RBC i/I blood group system is represented by unbranched and branched polylactosamine glycans, respectively.^56-58  Lectins able to distinguish blood group related antigens including LTL (group O) and both DBA and PTL-I (group A) showed decreased recognition in SCD patients, but LEL and DSA (i/I blood group) were observed to increase binding, suggesting terminal blood group antigen alterations on RBCs in SCD. Notably, MixOmics detected new increases in GBP binding to blood group A and H antigens (65 and 66) in SCD, with a reciprocal decrease of group A– and O–specific lectin binding (LTL and DBA) to RBCs. Additionally, decreased GBP binding to a sulfated lactosamine structure (Gal-β-1,4-(6S)GlcNAc-β, #21) was detected, further reinforcing the notion that blood group antigen expression differs in SCD. Using MixOmics we also observed increases of PWM and GNA binding and decreases of NPA binding to SCD RBCs, suggesting alterations in mannose moieties, and decreased binding of Galactose and GalNAc recognizing lectins (ABA, Jacalin).

GBPs showed increased recognition of immunogenic mucin-associated glycans in SCD, including Lewis Y, Lewis A, asialo, α2,6-sialylated and di-sialylated core 1 structures (glycans 71, 73, 22, 44, 48). GBPs binding to other fucosylated oligosaccharides (Gal-β-1,4-[Fuc-α-1,3]-GlcNAc-β-1,3-Gal-β and F-SL) was also increased. Decreased binding to aminoglycoside antibiotics (18, gentamicin; 26, tobramycin; 31, neomycin) was measured, except for kanamycin (29), which may have shown an overall increase because of heavily increased fluorescence from a subset of 6 SCD patients.

MixOmics analysis reiterated our findings obtained using sPLS-DA and regular t test methods, reinforcing observations of changes in (1) blood group expression and antibodies directed to blood group antigens and (2) binding of GBPs to immunogenic mucin-associated glycans, including sialyl core 1, Lewis A and Y structures, and sialylation in SCD often associated with chronic inflammation and cancer.

Previous SCD research has mainly focused on differences in receptor binding to RBCs and plasma components brought on by changes in glycosylation.^22,59,60  However, little research has been performed to identify the structural glycosylation changes on RBCs and in plasma. Using a unique approach of pairing lectin and glycan microarrays with mass spectrometry to analyze healthy, SCT, and SCD plasma and RBCs, we found significant changes in both glycosylation of red cells, and glycan recognition of plasma GBPs between healthy and SCD samples.

Typically, RBCs in SCD have altered surface sialylation,²⁵  displaying an uneven distribution of SA compared with healthy cells.⁶¹  We found consistent changes in sialylation between healthy and SCD RBCs using both lectin array and mass spectrometry. All structures that were significantly increased in SCD samples contained α2,6-Neu5Ac, consistent with increases in SSA and SNA lectin binding. Mass spectrometry data suggest that α2,6-Neu5Ac is either replacing α2,3-Neu5Ac structures or being added to terminal galactose observed on healthy RBCs. One reason for changes in surface SA could be because of differences in the age of circulating RBCs. Aging RBCs see reductions of surface sialylation between 10% and 30%,^61-64  and in some hemolytic anemias, decreases of surface SA by 50% have been observed after release of RBCs into circulation.⁶⁵ 

Sickled RBCs are more susceptible to endothelial adhesion because of these changes in terminal SA, leading to vaso-occlusive episodes.^66,67  E- and P-selectins are expressed on activated endothelium and platelets.²²  Sialyl-Lewis X binds selectins and promotes leukocyte trafficking to lymph nodes and sites of inflammation.^19,20  In addition to SA modifications, selectin binding depends on fucose and sulfated glycan decorations.^59,60  Fucosylated glycans adorned the top 5 structures for which GBPs in SCD had higher affinity, including the immunogenic Lewis Y and Lewis A, suggesting that fucosylated structures and reciprocal GBPs are generated in SCD. Increased fucosylation is also observed in inflammatory conditions,^21,68  such as fucosylated IgG in rheumatoid arthritis.^69,70  Thus, changes in sialylation and fucose-containing structures on sickled RBCs could alter selectin and RBC–ligand interactions to drive inflammation in SCD.

Overexpression of sialyl Lewis Y and sialyl Lewis A is also associated with cancers and chronic inflammation. In support of a heightened immune response to aberrant glycans, GBP binding to asialo and sialyl core 1 structures (glycans 22, 44, and 48; Figure 3C) was increased in SCD, with these glycans often expressed on mucin-rich glycoproteins, including CD59, that help to signal RBC hemolysis.⁷¹  Sickling of RBCs could increase the expression of aberrant glycans resulting in lysis.⁷²  A recent report shows that cholesterol-dependent lysins (intermedilysin) bind to sialyl-core 1 O-glycans using CD59 to promote RBC lysis,⁷³  further supporting the role of O-glycans in RBC destruction. The role of RBC glycolipid expression of antigenic glycan structures needs to be elucidated. However, our data highlight mucin O-glycans as potential therapeutic targets in SCD.

A limitation of our results is that, although SCD samples were collected before transfusion, it is impossible to discern transfused from endogenous patient RBCs because of repeat transfusions. Transfusions can be episodic or timed, ranging from every other week to monthly, therefore, based on an RBC lifetime of ∼120 days, contributing significantly to the data acquired. Another limitation is that we cannot correlate our results to patient treatment regimes.

MixOmics analysis confirmed trends we observed with the lectin and glycan arrays individually. However, it also brought to light subtle blood group changes. Lectins known to bind both group O and A antigens (LTL, DBA, and PTL-I) showed decreased affinity to SCD RBCs, whereas those lectins with affinity for i/I blood group antigens (LEL and DSA) showed increased binding. The i blood group antigen is converted to I antigen within 18 months of birth, and therefore an increased binding of lectins to i/I antigens suggests more immature RBCs in SCD, which lack ABO antigens and the complete conversion of i to I antigens.⁷⁴  Previous studies have observed that RBC maturation in individuals with hemoglobinopathies decreases from 3 to 1.5 days with increased reactivity to i antibodies.⁷⁵  Additionally, increased erythropoiesis, as in SCD and thalassemia, can weaken the expression of ABO blood group antigens.^76,77  Loss of ABO antigens is also seen on the RBCs of myeloid malignancies⁷⁸  and on carcinoma tumor cells, suggesting that ABO antigen expression change contributes to the immune response in disease,^79-81  including SCD. Because of sample size limitations, only blood groups A and O were analyzed. Further investigation is needed to understand the impact of SCD on blood group expression.

In conclusion, our multipronged approach shows changes to both glycosylation on RBCs and binding of plasma proteins to aberrant glycans in SCD. Specifically, our data showed loss of α2,3-Neu5Ac and surprisingly revealed a reciprocal increase of α2,6-Neu5Ac on SCD RBCs, in addition to changes in blood group antigen expression consistent with more immature RBCs. Our data further show increased binding of plasma GBPs to potentially immunogenic glycans, including mucin core 1 and Lewis A and Lewis Y structures, suggesting these glycan structures as potential therapeutic targets.

This work was supported by National Institutes of Health, National Heart, Lung, and Blood Institute grants R01-HL126785 and R01-HL134010 (R.L.G.) and K12-HL141954 and R01-HL089224 (K.M.H.) and The Jacquelyn & Arlyn Fredrick Endowment Fund for Clinical Research (W.Q.A.). The Core Labs at the Blood Research Institute and the Medical College of Wisconsin Genomic Sciences and Precision Medicine Center and Mass Spectrometry Center provided key resources.

Contribution: H.E.A. performed research, collected, analyzed, and interpreted results, and cowrote the manuscript; C.A. performed all mass spectrometry research, analyzed and interpreted the resulting data, and cowrote the manuscript; A.P.S. helped to collect results; R.L.G. helped to fund the research; K.M.H. helped to design the study and analyzed and interpreted results; W.Q.A. designed and helped perform research for the study and performed computational biological analysis; and all authors read, contributed to, and approved the writing of the manuscript.

Conflict-of-interest disclosure: The authors declare no competing financial interests.

Correspondence: Karin M. Hoffmeister, Versiti Blood Research Institute, 8727 W Watertown Plank Rd, Milwaukee, WI 53226; e-mail: khoffmeister@versiti.org.