Plasma proteome signatures are predictive of mortality in sickle cell disease

Key Points

• •Proteins easily measured in the plasma of patients with SCD predict mortality and are associated with complications.
• •Plasma proteome signatures emerge as a new promising tool to monitor severity in SCD.

Visual Abstract

Abstract

Sickle cell disease (SCD) is associated with severe systemic complications and increased mortality risk. Predicting SCD severity is currently difficult because of a lack of biomarkers. Here, we measured 5411 plasma proteins in 376 patients with SCD and 103 participants without SCD to find new predictors of SCD mortality. We used protein signatures of mortality that were developed in non-SCD populations to calculate predicted mortality risk scores in our SCD data set. The mortality scores were higher in patients with SCD than in individuals without SCD (P = 3.7 × 10^˗10) and were associated with increased mortality in patients with SCD (risk factor–adjusted hazard ratio, 2.2; 95% confidence interval, 1.3-3.6; P = .0032). The mortality scores correlated with several clinical variables (eg, white blood cell count and hemoglobin concentration) and complications (eg, leg ulcers and stroke) that are clinically relevant yet insufficient individually to predict SCD mortality. In addition to the protein signatures, we found 499 plasma proteins that associate with mortality in patients with SCD (false discovery rate of ≤5%), including many proteins involved in inflammatory responses, such as the interleukin-18 signaling cascade. Finally, we estimated biological age in patients with SCD and individuals without SCD using the plasma proteome data. We confirmed that SCD patients age prematurely (+6.0 ± 5.4 years older than their chronological age) and found that brain biological age positively associates with past occurrences of stroke. Altogether, our results support the use of the plasma proteome to monitor and predict clinical severity in SCD.

Introduction

Sickle cell disease (SCD) is a recessive hematological disorder caused by mutations in the β-globin gene. It affects ∼8 million individuals worldwide, mostly in sub-Saharan Africa and the Indian subcontinent.¹ SCD is characterized by acute and chronic multiorgan complications of variable penetrance and expressivity (eg, painful crises, stroke, neurocognitive impairment, renal failure, and pulmonary hypertension),² a poorer quality of life, and a life expectancy that is 20 to 30 years shorter than for individuals without SCD.³ In the United States, a recent study based on mortality data from the National Center for Health Statistics (2008-2023) showed that the mortality rate for SCD is increasing.⁴ Many biomarkers have been proposed to predict SCD-related complications and mortality, although their utility remains limited because of poor statistical performance, a lack of validation, or challenges associated with their implementation in the clinic.^5,6

Omics research provides new opportunities to understand clinical severity in SCD, expand the list of biomarkers that can predict SCD-related complications, and prioritize new drug targets. As an example, human genetic studies of the regulation of fetal hemoglobin (HbF) levels identified the transcription factor BCL11A and paved the way to an approved gene therapy for SCD.^7,8 The development of affinity-based methods that can accurately measure thousands of proteins simultaneously has recently enabled large-scale proteomic studies.^9,10 When combined with well-phenotyped and prospectively followed large human cohorts, these untargeted proteomic approaches can pinpoint unanticipated individual proteins or groups of proteins (ie, protein signatures) that are predictive of human diseases. For instance, using large plasma proteome data sets available in the UK Biobank (UKBB) and other cohorts, investigators have developed and optimized protein signatures that predict chronic diseases, mortality, and biological age.11, 12, 13, 14, 15, 16, 17 Because many of the proteins detected in the plasma likely originate from a specific tissue or organ, it has even been possible to generate organ-specific predictive models of mortality and biological age from profiles of the plasma proteome.^11,14

Previous proteomic studies in SCD have used mass spectrometry in small patient cohorts (n = 15-51), but there have been few follow-up studies.18, 19, 20 Here, we used the Olink technology to measure 5411 proteins in the plasma of 479 participants from the Genetic Modifier (GEN-MOD) cohort, including 376 patients with SCD (HbSS), 51 individuals with sickle cell trait (HbAS, non-SCD), and 52 individuals without the sickle cell mutation (HbAA, non-SCD). We validated protein signatures of mortality and biological age in patients with SCD and explored how the proteomic data associate with baseline clinical variables and complications. Our results support an expanded role for the study of the plasma proteome in SCD research.

Methods

Study participants

Sample collections and procedures were in accordance with the institutional and national ethical standards of the responsible committees, and proper informed consent was obtained. The project was approved by the Montreal Heart Institute ethics committee (project no. 09-1137). GEN-MOD has been briefly described previously.²¹ It comprises a cohort of patients with SCD and family members recruited at Hôpital Henri-Mondor in Créteil, France. For each participant, a wealth of detailed clinical information was collected in a centralized database (eg, blood parameters and complication status). However, many other relevant variables are not available in the centralized GEN-MOD database (eg, albuminuria, iron overload, direct measures of hemolysis, and socioeconomic status). Biospecimens (eg, blood and plasma) were also collected at steady state. None of the GEN-MOD participants with SCD were treated with hydroxyurea at recruitment. Survival status was ascertained using information from the French national registry (Institut National de la Statistique et des Études Économiques).

Ethics approval statement

We collected data according to the Declaration of Helsinki, and the study was approved by the Montreal Heart Institute ethics committee, project no. 2009-106 (09-1137).

Plasma protein quantification

The Olink experiment was performed at the McGill Genome Centre. Plasma samples (40 μL) from 504 participants of the GEN-MOD cohort, including 396 patients with SCD (HbSS) and 108 individuals without SCD (56 HbAS and 52 HbAA), were analyzed using the Olink Explore 5K platform, which quantifies the expression of 5416 proteins. To minimize batch effects, samples were randomized with the Well Plate Maker R package before transfer into 96-well Eppendorf twin.tec polymerase chain reaction plates.

Quality control

Quality control was performed as described previously.¹⁰ Briefly, proteins with “warn” or “fail” quality control flags were excluded before principal component analysis, resulting in the removal of apolipoprotein E. Principal component 1 (PC1) and PC2 were used to identify outliers among participants and assays; points with values of >5 standard deviations (SD) from the mean of PC1 or PC2 flagged 2 participants and 3 proteins (STK24, GAS8, and PHLDA3) as outliers. Additional filtering was performed using the OlinkAnalyze R package with interquartile range plots, identifying 10 participant outliers. Missingness was evaluated for both participants and assays, and data with >11% missing values were excluded from downstream analyses, resulting in the removal of 14 participants and 1 protein (FN1). In total, 25 of the 504 participants and 5 of the 5416 plasma proteins were excluded from subsequent analyses. A flow diagram summarizing the initial sample size, quality control steps, and reasons for participant exclusion is shown in supplemental Figure 1.

Prediction of biological age and mortality

The predicted biological age and mortality risk for GEN-MOD participants were calculated using models developed by Goeminne et al.¹¹ In the aforementioned study, plasma expression levels of 3072 proteins measured with the Olink Explore platform in ∼50 000 UKBB participants were used to develop organ-specific prediction models for 18 organs as well as a whole-organism model referred to as “conventional.” We applied the coefficients provided in supplemental Table 1A (for biological age) and supplemental Table 1C (for mortality) to GEN-MOD by summing the protein-specific coefficients from each organ model multiplied by their corresponding Olink-normalized expression values. For biological age predictions, intercept values of each model were added.

Because Olink provides relative quantification of protein expression, which can vary across platforms, we accounted for potential differences between the Olink Explore 5K platform used in GEN-MOD and the Olink Explore 3K platform used by Goeminne et al¹¹ by recalibrating the predictions using UKBB data. Specifically, we fitted separate linear regression models of predicted biological age or mortality risk on chronological age for GEN-MOD and UKBB. We then extracted the slopes and intercepts from each regression and recalibrated the GEN-MOD predictions as follows:

recalibrated GEN-MOD predicted biological age or mortality risk = a + b × GEN-MOD predicted biological age or mortality risk, in which:

• •b = slope for UKBB / slope for GEN-MOD
• •a = intercept for UKBB − b × intercept for GEN-MOD

Finally, we evaluated the performance of each organ-specific biological age and mortality model in GEN-MOD by calculating the Pearson correlation coefficient (r) between predicted values and chronological age in individuals without SCD. Only organ models with r >.3 and included by Goeminne et al¹¹ were retained for downstream analyses, resulting in the inclusion of 8 organ models and the conventional model. The DeltaAge for each organ was calculated as the difference between the predicted biological age and the chronological age. Comparisons between 376 patients with SCD and 103 individuals without SCD were performed using the Wilcoxon test.

Survival analyses

Survival analyses of patients with SCD were conducted using Cox proportional hazards regression models implemented in the survival R package. The analysis included 40 events (deaths) and 270 censored participants (controls). Associations between plasma proteins and survival were assessed using Normalized Protein eXpression values for each protein, which were rank-based inverse normalized and standardized to have a mean of 0 and an SD of 1 before analysis, with models adjusted for chronological age and sex. For the association between predicted mortality and survival, the predicted mortality was also standardized to have a mean of 0 and an SD of 1 before analysis. These models were adjusted either for chronological age and sex for all the organ models or for chronological age, sex, white blood cell (WBC) count, HbF, estimated glomerular filtration rate (eGFR), and acute chest syndrome (coded as 0/1) for the conventional model. The Kaplan-Meier survival curve was generated using the ggsurvplot function of the survminer R package, comparing the lowest (Q1) and highest (Q4) quartiles of the conventional predicted mortality score. P values were calculated with the log-rank test. Forest plots were generated with the forestplotter R package. We used the statistical method developed by Clogg et al²² to assess the heterogeneity between the GEN-MOD and UKBB hazard ratios (HRs).

Associations with traits and complications

Associations of predicted mortality, DeltaAge, or plasma protein levels with clinical traits and complications in patients with SCD were evaluated using regression models adjusted for chronological age and sex. Linear regression was applied to continuous variables (WBC, platelet, reticulocytes, lactate dehydrogenase [LDH], dense red blood cells [RBC], bilirubin, mean corpuscular hemoglobin, mean corpuscular volume, RBC counts, hemoglobin, hematocrit, HbF, and eGFR), whereas logistic regression was used for dichotomous outcomes (osteonecrosis, cholecystectomy, leg ulcer, priapism, and stroke, coded as 0/1). The z scores for each association were calculated as the regression coefficient estimate divided by its standard error for heat map visualization. Associations of predicted mortality or DeltaAge with the traits and complications were plotted using the pheatmap R package. We used the ComplexHeatmap R package to plot the data (unsupervised clustering).

LASSO-based Cox regression

Prioritization of a set of proteins associated with survival was performed using a Cox proportional hazards model with least absolute shrinkage and selection operator (LASSO) penalization, implemented in the glmnet R package. Age and sex were included as unpenalized covariates, whereas penalization was applied to the protein variables. Normalized Protein eXpression values for each protein were rank-based inverse normalized and standardized to have a mean of 0 and an SD of 1 before analysis. The optimal penalty parameter (λ) was selected via 10-fold cross-validation. Survival analyses, including computation of the concordance index (c-index) and visualization of survival curves, were performed using the survival and survminer R packages. Replication in the UKBB cohort was performed by applying the coefficients derived from GEN-MOD for 4 of the 5 proteins in the model that were available in UKBB (interleukin-18 receptor 1 [IL-18R1], LGMN, MMP7, and VEGFA).

Comparison with survival data in the UKBB

Comparisons of proteins associated with survival in GEN-MOD and UKBB were performed using data from supplemental Table 4 of Gadd et al,²³ which reported Cox proportional hazard models for death adjusted for age and sex. Comparisons were restricted for proteins measured on both the Olink Explore 5K platform (GEN-MOD) and the Olink Explore 3K platform (UKBB).

Differential expression analyses

Differential expression analyses of plasma proteins between SCD and individuals without SCD were performed using the limma R package. Linear models were fitted for each protein and chronological age and sex were included as covariates. P values were corrected for multiple testing using the Benjamini-Hochberg false discovery rate method.

Prediction of biological age using the ProtAge model

ProtAgeGaps were calculated using the model developed by Argentieri et al.¹² Of the 204 proteins in the published model, 196 were measured in the GEN-MOD cohort and included in this analysis. Protein expression values were normalized as described by Argentieri et al, and the pretrained model was used to calculate protein-predicted ages (ProtAges). ProtAges in GEN-MOD were recalibrated using the ProtAges calculated in the UKBB, based on the slopes and intercepts of regression models of ProtAge on chronological age, as we have described earlier. ProtAgeGaps were then defined as the difference between the recalibrated ProtAges and chronological age.

Results

Profiling the plasma proteome of patients with SCD

To identify new biomarkers of SCD-related mortality and complications, we measured 5411 plasma proteins in participants from GEN-MOD, a prospective cohort of African-ancestry patients with SCD and their relatives recruited at Hôpital Henri-Mondor (Créteil, France).^21,24 We included in our proteomic study 376, 51, and 52 GEN-MOD participants with HbSS, HbAS, and HbAA, respectively; patients with HbSS were not treated with hydroxyurea at recruitment (supplemental Table 1). As expected, the β-globin genotype influenced the structure of the proteomic data, with patients with SCD (HbSS) separating from the participants without SCD (HbAS + HbAA) along PC1 (supplemental Figure 2). We found 598 proteins that were differentially expressed between those with and without SCD (|log₂(fold-change)| ≥ 1, and false discovery rate [FDR] of ≤5%), including many proteins that are highly relevant to SCD biology, hemolysis, and stress erythropoiesis (eg, HMOX1, HBG2, TFRC, EPO, AHSP, and CAT; supplemental Figure 3 and supplemental Table 2). We replicated findings from a recent study that targeted 92 proteins in 57 patients with SCD and 13 healthy controls (98% validation rate; supplemental Table 2)²⁵ and confirmed 3 proteins (HMOX1, AHSP, and UMOD) that are also differently expressed between participants with HbAS and HbAA (supplemental Table 3).²⁶ We also found that many of the proteins previously implicated in SCD complications are differentially expressed between patients with SCD and sickle cell trait carriers in GEN-MOD (supplemental Table 4).^20,27, 28, 29 Overall, these results validate the quality of our proteomic data set.

Protein signatures predict SCD mortality

During a mean ∼12-year follow-up period, the survival rate among the GEN-MOD HbSS probands was 87% (40 events, 310 participants; supplemental Table 1). We tested whether existing protein signatures of mortality and aging, developed in non-SCD cohorts, could be good prognostic biomarkers in SCD. We calculated conventional (ie, organismal, that is, all plasma proteins available) and organ-specific mortality and biological age predictions in GEN-MOD using the models developed by Goeminne et al¹¹ in the UKBB and our Olink plasma proteomic data. For downstream analyses, we only kept models that show good performance to predict chronological age, as previously reported (Pearson r > 0.3; “Methods”; supplemental Table 5).¹¹

The predicted mortality risk by the conventional model was positively associated with chronological age in both patients with SCD and individuals without SCD from GEN-MOD (Figure 1A). It was also significantly higher in patients with SCD than in individuals without SCD (Figure 1B). In patients with SCD, we found that the predicted mortality risk calculated with the conventional model was associated with mortality in GEN-MOD (HR, 1.7; P = .013), and this association was stronger when we corrected for known SCD mortality risk factors such as baseline WBC counts, HbF levels, eGFR, and acute chest syndrome (HR, 2.2; P = .0032; Figure 1C).³⁰ Patients with SCD with a predicted mortality score in Q1 of the distribution were more likely to survive than patients in Q4 (log-rank P = .016; Figure 1D). After 5, 10, and 15 years of follow-up, the survival probability was 100%, 96.2%, and 91.2%, respectively, in the Q1 group, and 93.6%, 84.6%, and 70.3%, respectively, in the Q4 group (Figure 1D).

Figure 1.

Plasma protein signatures predict SCD mortality. (A) Mortality predictions based on the conventional plasma protein model increase with chronological age in GEN-MOD participants with SCD (HbSS) and without SCD (HbAA + HbAS). (B) The mortality prediction scores from the conventional model are higher in patients with SCD than in individuals without SCD. (C) The conventional model predicts mortality in patients with SCD. The 2 models are adjusted for biological sex and age at baseline; the adjusted conventional model is further adjusted for baseline WBC counts, HbF levels, eGFR, and acute chest syndrome. The forest plot shows the HR per SD increase in the normalized prediction scores. (D) Kaplan-Meier plot comparing the survival probability in patients with SCD with mortality prediction scores from the conventional model in the bottom (Q1) or top (Q4) quartile of the distribution. (E) Heat map that summarizes the association between mortality predictions from the conventional or organ-specific models and baseline clinical variables. (∗nominal P < .05; ∗∗FDR < 0.05; ∗∗∗Bonferroni-corrected P < .05; correction applied separately for each model). CI, confidence interval; Cox PH, Cox proportional hazards; HCT, hematocrit; HGB, hemoglobin; MCH, mean corpuscular hemoglobin; MCV, mean corpuscular volume; PLT, platelet; Retic, reticulocyte.

The mortality prediction scores by the conventional model were associated with many clinical variables measured at baseline: positively with dense RBC, mean corpuscular hemoglobin, mean corpuscular volume, reticulocyte counts, WBC counts, and LDH levels, and negatively with RBC counts, hemoglobin levels, and hematocrit (Figure 1E; supplemental Table 6). Patients with SCD with leg ulcers, osteonecrosis, or stroke also had higher mortality prediction scores based on the conventional model (Figure 1E; supplemental Table 6). However, none of these baseline variables were individually associated with mortality in GEN-MOD, indicating that the plasma protein signature from the conventional model captures more information than routinely measured clinical biomarkers (supplemental Table 7; supplemental Figure 4).

We explored whether the organ-specific mortality predictions added additional information to the conventional mortality model. Except for the artery model, the mortality scores for all other organ models were higher in patients with SCD than in individuals without SCD (supplemental Figure 5A). The lung, skin, and artery models were associated with increased mortality in GEN-MOD (supplemental Figure 5B). When testing against baseline clinical variables and SCD-related complications, we found that the organ-specific mortality predictions were redundant with the mortality scores by the conventional model, with some noteworthy exceptions: (1) eGFR was correlated with mortality scores from many organ-specific models, including the kidney and intestine models. (2) Platelet counts and bilirubin levels were associated with predictions from the liver and brain models, respectively. and (3) The presence of leg ulcers was more strongly associated with predictions from the skin and immune models than the conventional model (Figure 1E; supplemental Table 6).

Various proteins associate with mortality in patients with SCD

Next, we tested the association between single proteins and mortality. Using Cox proportional-hazards regression, we identified 499 proteins associated with mortality in GEN-MOD at an FDR of ≤5% (Figure 2A; supplemental Table 8). Only 2 proteins, CBLN4 and SIRT1, were associated with increased survival. Of the 497 proteins associated with increased mortality, many are involved in inflammatory responses, including 3 members of the IL-18 signaling cascade (IL-18R1, IL-18BP, and IL-18; supplemental Table 8).^31,32 Using a recently released proteomic atlas that assigns plasma proteins to their tissue (or cell type) of origin,³³ we determined that proteins associated with mortality in GEN-MOD are enriched for proteins common to most tissues (odds ratio [OR], 1.8; Fisher exact test P = 2.0 × 10^˗6) and proteins from the spleen (OR, 1.8; P = .0013), and are depleted for proteins from the brain (OR, 0.4; P = 7.3 × 10^˗6; supplemental Table 9).

Figure 2.

Plasma proteins associate with mortality in GEN-MOD patients with SCD (HbSS). (A) Volcano plot with 5411 plasma proteins tested against mortality in patients with SCD from GEN-MOD. We found 499 proteins that are significant after accounting for multiple testing (FDR of ≤5%; horizontal dashed line). The vertical dashed line corresponds to an HR of 1. We labeled proteins of interest, including 3 members of the IL-18 cascade. (B) Mortality HR in GEN-MOD (x-axis) and the UKBB (y-axis) are correlated for 1421 proteins present in both data sets. For GEN-MOD and UKBB, the HR are per normalized SD units of protein levels. We labeled in red proteins that have significantly different HR (heterogeneity FDR of ≤5%) between GEN-MOD and UKBB. We provided protein name labels for 7 proteins with a nominal heterogeneity P < 1 × 10^˗4. (C) Forest plot for 15 mortality-associated proteins with the smallest heterogeneity P values between GEN-MOD and the UKBB. GALNT2 is a risk protein in GEN-MOD but protective in the UKBB. DAPP1 is associated with increased mortality in GEN-MOD but is not significant in the UKBB. HetP, heterogeneity P value.

In the UKBB, Gadd et al analyzed 1468 plasma proteins measured with Olink and found 652 proteins significantly associated with mortality.²³ Although the UKBB and GEN-MOD are fundamentally different, the UKBB is a population-based cohort of mostly European-ancestry individuals not ascertained for a specific disease,³⁴ we tested whether the same proteins associate with mortality in both cohorts. Of the 499 proteins significantly associated with mortality in GEN-MOD, 253 proteins were tested in the UKBB, and 208 (84%) were associated with mortality with a consistent direction of effect (binomial P < 2.2 × 10^˗16; supplemental Table 10). We found that the protein-mortality HRs were correlated between GEN-MOD and the UKBB but were larger in GEN-MOD (Figure 2B-C). However, this was not true for all proteins. For instance, GDF15, a protein strongly associated with mortality in the UKBB (HR, 2.2; P < 1 × 10^˗300)²³ and previously implicated in human longevity³⁵ and chronic diseases36, 37, 38 was only marginally associated with mortality in patients with SCD (HR, 1.5; nominal P = .020; FDR, 0.10; supplemental Table 10). Together, our results suggest that a similar plasma proteome is associated with mortality in individuals without SCD and patients with SCD, but the magnitude of the association of the proteins with survival is different, presumably because of the unique pathophysiological environment encountered in SCD (eg, chronic hemolysis, systemic inflammation, and high reactive oxygen species production).³⁹

Clinical variables and mortality-associated proteins

To better understand how the plasma proteome can predict SCD mortality, we tested in GEN-MOD the associations between the levels of the 499 mortality-associated proteins with 13 clinical parameters and 5 SCD-related complications (supplemental Tables 1 and 7; all data collected at baseline). We found that increased levels of many of these proteins were positively correlated with WBC counts, reticulocyte counts, LDH levels, dense RBC percentage, and mean corpuscular hemoglobin, and negatively correlated with RBC counts, hemoglobin levels, hematocrit, HbF levels, and eGFR (Figure 3A; supplemental Tables 12 and 13). We noted a unique pattern of platelet counts, with significant enrichments for both positively and negatively correlated proteins (Figure 3A; supplemental Table 12).

Figure 3.

Associations between baseline clinical variables and mortality-associated plasma proteins. Associations between 13 blood-based biomarkers routinely measured in patients with SCD (A) or 5 SCD-related complications with the levels of 499 plasma proteins (y-axis) associated with mortality in GEN-MOD (B). We used |z score| ≥ 1.96 to define nominal significance. The red and blue colors indicate positive and negative associations, respectively. denseRBC, dense RBC; HCT, hematocrit; HGB, hemoglobin; MCH, mean corpuscular hemoglobin; MCV, mean corpuscular volume; PLT, platelet; Retic, reticulocyte.

For the SCD complications, we observed an enrichment of mortality-associated proteins that are positively correlated with leg ulcers (Figure 3B; supplemental Tables 11 and 12). The causes of leg ulceration in SCD are not fully understood but likely involve vasculopathy defects and systemic inflammation.⁴⁰ Thus, the enrichment observed with leg ulcer may reflect the abundance of positive correlations between high mortality–associated protein levels and WBC counts (Figure 3). It is interesting to note that although the presence of leg ulcers at baseline is not significantly associated with mortality in GEN-MOD (supplemental Table 7), 107 proteins positively associated with leg ulcers are also strongly and positively associated with mortality prospectively (supplemental Tables 11 and 12). This result highlights an advantage of an unbiased proteomic approach, which can nominate biomarkers that can predict future events (ie, deaths) even in the absence of clear clinical manifestations at baseline.

Developing a simpler protein signature of SCD mortality

The association between protein signatures developed in non-SCD populations and SCD mortality is promising, but we wondered whether we could use the GEN-MOD plasma proteome data to develop a more SCD-specific signature of mortality. Ideally, this new signature would also include fewer proteins to facilitate future clinical implementation.

To prioritize a set of plasma proteins associated with survival in patients with SCD, we applied a Cox proportional hazards model with LASSO penalization (“Methods”). This approach identified 5 proteins (IL-18R1, LGMN, MEX3B, MMP7, and VEGFA), which were combined into a Cox LASSO mortality risk score associated with mortality (HR, 5.70; P = 1.8 × 10^˗11). With a c-index of 0.81, the model provided better mortality prediction than IL-18R1 protein expression alone (c-index = 0.76), the most significant protein associated with survival in GEN-MOD (supplemental Table 8). In comparison, the c-index for the conventional and conventional adjusted models was 0.66 and 0.69, respectively (Figure 1C). We validated the GEN-MOD–based Cox LASSO mortality score in the non-SCD UKBB cohort (4401 events among 39 306 participants) using the expression of 4 of the 5 proteins available in this data set (IL-18R1, LGMN, MMP7, and VEGFA). Applying the coefficients derived from the GEN-MOD cohort, this score demonstrated good predictive performance for mortality in the UKBB, with a c-index of 0.72. The mortality HR associated with this 4-protein score was higher in GEN-MOD (HR, 5.4; P = 2.8 × 10^˗10) than in the UKBB (HR, 4.1; P < 2.2 × 10^˗16). Importantly, we acknowledge that the performance of this simple 4-protein signature is not immune to overfitting issues and will require validation in external SCD cohorts with proteomic data available.

The 4-protein mortality score was strongly associated with several clinical variables and complications in directions consistent with SCD pathophysiology: a significant negative association with eGFR (P = 9.4 × 10^˗12), and positive associations with leg ulcers (P = 1.6 × 10^˗5), dense RBC (P = 1.3 × 10^˗4), and osteonecrosis (P = 2.4 × 10^˗4; supplemental Table 13). Notably, none of the 4 proteins used in the model were individually associated with leg ulcers or osteonecrosis at baseline. In addition, although the 4 proteins were not included in the mortality models from Goeminne et al,¹¹ the resulting 4-protein signature was positively associated with mortality predictions from those models (supplemental Table 14).

The plasma proteome captures accelerated aging in SCD

Using different modalities (eg, imaging and DNA methylation), previous studies have found that patients with SCD age prematurely.^41,42 We calculated conventional and organ-specific predicted biological age in GEN-MOD using plasma proteome–based models and compared it with chronological age (ie, DeltaAge; supplemental Table 5).¹¹ For the conventional age model, the difference between biological and chronological age was larger in patients with SCD (+6.0 ± 5.4) than in individuals without SCD (+2.0 ± 5.0; Figure 4A). We validated this result using an independent protein-based aging clock model (supplemental Figure 6A).¹² For the organ-specific biological age models, we also observed that DeltaAge was larger in patients with SCD when compared with individuals without SCD (Figure 4A). Mortality in GEN-MOD was associated with DeltaAge calculated using the skin and intestine models but not with the conventional model (supplemental Figure 6B). DeltaAge across all models was not as strongly associated with baseline variables as the predicted mortality scores described earlier, with the exception of DeltaAge calculated with the brain model, which was strongly associated with stroke events that occurred retrospectively (Figure 4B; supplemental Table 15).

Figure 4.

Plasma proteome–based biological clocks capture accelerated aging in SCD. (A) The difference between the predicted biological age calculated with the plasma proteome and chronological age (y-axis, DeltaAge) is larger in GEN-MOD patients with SCD than in individuals without SCD from the same cohort. We calculated P values using the Wilcoxon test. (B) A heat map that summarizes the association between DeltaAge from the conventional or organ-specific models and baseline clinical variables. (∗nominal P < .05; ∗∗FDR < 0.05; ∗∗∗Bonferroni-corrected P < .05; correction applied separately for each model). denseRBC, dense RBC; HCT, hematocrit; HGB, hemoglobin; MCH, mean corpuscular hemoglobin; MCV, mean corpuscular volume; PLT, platelet; Retic, reticulocyte.

Discussion

One of the hallmarks of SCD is its clinical heterogeneity: patients with the same β-globin genotype can have mild, almost asymptomatic disease, whereas others can experience life-threatening complications. Although research has identified causal mechanisms (eg, hemolysis and vaso-occlusion)³⁹ and pinpointed some of the severity modifiers (eg, HbF; α-thalassemia; pollution; and social determinants such as access to health care services, patient stigmatization, or even racism),43, 44, 45, 46, 47 it remains challenging to predict clinical outcomes in SCD. Identifying specific and sensitive biomarkers of SCD complications would greatly improve patient care. In this study, we showed that plasma proteome signatures developed in the non-SCD UKBB were predictive of mortality in the prospective GEN-MOD cohort. Although associated with many relevant baseline clinical variables (eg, WBC counts and dense RBC) and complications (eg, leg ulcers), plasma proteome signatures outperform these variables in predicting SCD mortality.

It remains unclear how plasma proteome signatures can predict chronic diseases, biological age, or mortality. Hemolysis has been linked to SCD mortality,⁴⁸ but we do not yet have direct measures of hemolysis in GEN-MOD to test this hypothesis (eg, free hemoglobin). We note that some of the clinical variables implicated in the SCD hemolytic phenotype (eg, leg ulcer, reticulocyte count, and LDH) are associated with the plasma proteome signatures (Figure 1E), although they are not associated with mortality in GEN-MOD (supplemental Table 7). Even if the proteome signatures are partially correlated with hemolysis, they are likely to also capture other processes that are predictive of mortality, because these signatures were developed in the non-SCD, nonhemolytic UKBB. A recent study suggested that these signatures mostly capture environmental risk factors and do not causally influence disease risk or mortality.⁴⁹ Because the same signatures (Figure 1) and a large extent of the same proteins (Figures 2 and 3) predict mortality in the UKBB and GEN-MOD, it appears likely that the plasma proteome in patients with SCD captures extrinsic factors that contribute to disease severity (eg, socioeconomic status, pollution, and stress). Indeed, although the mortality proteome signatures were developed in the UKBB, they are correlated with dense RBC and leg ulcers (Figure 1E), 2 characteristic features of SCD. Therefore, the plasma proteome signatures are biomarkers of general biological processes (eg, increased systemic inflammation) that are influenced by the environment and predictive of disease onset or mortality. However, our results suggest that SCD creates a physiological environment that magnifies the predictive value of some of these proteins (Figure 2B-C).

Our study has several limitations. First, we only measured proteins in the plasma of adult patients with SCD, such that our conclusions cannot be extended to pediatric SCD-related complications and mortality. Second, we did not analyze data from patients with other β-globin genotypes (eg, HbSC and HbSβ⁰). Third, we did not assess the impact of current SCD treatments (eg, hydroxyurea) on the plasma proteome; this will need to be evaluated in future studies. Fourth, we do not know the cause of death in GEN-MOD, information which would have enabled stratified analyses. For instance, it would be important to validate previous results that implicated IL-18 signaling in SCD-related cardiomyopathy.³¹ Fifth, the predictive models that we used are likely suboptimal for patients with SCD because they were developed in relatively healthy individuals of a mostly different ancestry (European vs African ancestry for patients with SCD). Ideally, we would have access to proteomic data sets from large SCD cohorts to develop and calibrate SCD-specific predictive models, but such data sets are currently unavailable. However, using models developed in a completely different population offers the advantage of reducing the chance of data overfitting and false-positive results.

The implementation of using complex plasma proteome signatures in SCD clinics presents with many technological and analytical challenges. For this reason, we explored the possibility of using a small number of proteins to predict SCD mortality. We found that a 4-protein signature (IL-18R1, LGMN, MMP7, and VEGFA) could predict mortality in patients with SCD and in UKBB participants. It is critical to validate this simpler signature in external SCD cohorts.

In the clinic, we anticipate that plasma proteome signatures could be used together with existing strategies to identify adult patients with SCD who would benefit from more frequent follow-up visits or more aggressive therapeutic interventions (eg, cell or gene therapy). In the context of clinical trials, plasma proteome signatures could also be used as unbiased biomarkers to monitor the efficacy of new SCD treatments.⁵⁰ Overall, our results support plasma proteome signatures as new biomarkers of SCD severity and add proteomics to the toolkit to bring SCD 1 step closer to precision medicine.

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