Plasma Insulin-like Growth Factor Binding Protein 7 Contributes Causally to ARDS 28-Day Mortality

Abstract

Background

ARDS is a devastating syndrome with heterogeneous subtypes, but few causal biomarkers have been identified.

Research Question

Would multistage Mendelian randomization identify new causal protein biomarkers for ARDS 28-day mortality?

Study Design and Methods

Three hundred moderate to severe ARDS patients were selected randomly from the Molecular Epidemiology of ARDS cohort for proteomics analysis. Orthogonal projections to latent structures discriminant analysis was applied to detect the association between proteins and ARDS 28-day mortality. Candidate proteins were analyzed using generalized summary data-based Mendelian randomization (GSMR). Protein quantitative trait summary statistics were retrieved from the Efficiency and safety of varying the frequency of whole blood donation (INTERVAL) study (n = 2,504), and a genome-wide association study for ARDS was conducted from the Identification of SNPs Predisposing to Altered Acute Lung Injury Risk (iSPAAR) consortium study (n = 534). Causal mediation analysis detected the role of platelet count in mediating the effect of protein on ARDS prognosis.

Results

Plasma insulin-like growth factor binding protein 7 (IGFBP7) moderately increased ARDS 28-day mortality (OR, 1.11; 95% CI, 1.04-1.19; P = .002) per log2 increase. GSMR analysis coupled with four other Mendelian randomization methods revealed IGFBP7 as a causal biomarker for ARDS 28-day mortality (OR, 2.61; 95% CI, 1.33-5.13; P = .005). Causal mediation analysis indicated that the association between IGFBP7 and ARDS 28-day mortality is mediated by platelet count (OR, 1.03; 95% CI, 1.02-1.04; P = .01).

Interpretation

We identified plasma IGFBP7 as a novel causal protein involved in the pathogenesis of ARDS 28-day mortality and platelet function in ARDS, a topic for further experimental and clinical investigation.

Take-home Points.

Study Question: Would multistage Mendelian randomization identify new causal protein biomarkers for ARDS 28-day mortality?

Results: Plasma IGFBP7 moderately increased ARDS 28-day mortality, and the association between IGFBP7 and ARDS 28-day mortality is mediated by platelet count.

Interpretation: Plasma IGFBP7 may involve in the pathogenesis of ARDS 28-day mortality and platelet function in ARDS, a topic for further experimental and clinical investigation.

ARDS is characterized by acute, diffuse inflammatory lung injury and affects nearly 200,000 patients annually in the United States.^1,2 The mortality rate ranges from 35% to 46% despite decades of research into effective treatments.³ A promising pharmacotherapy for ARDS is neuromuscular blockade, but the efficacy of this approach has been brought into question. The latest clinical trial shows that in a cohort of moderate to severe ARDS patients, the addition of early continuous neuromuscular blockade with concomitant deep sedation did not result in lower mortality than a usual-care approach to mechanical ventilation.⁴ Because ARDS has heterogeneous lung pathophysiologic characteristics, identifying clinical and biological features to discriminate ARDS patients into subphenotypes may be helpful to categorize those who might be more responsive to different therapies.⁵

Proteomics technology applied to investigate the pathogenesis and progression of ARDS resulted in the identification of many plasma protein biomarkers. Plasma tumor necrosis factor-α, IL-1β, IL-6, and IL-8 levels were reported as significantly higher in patients dying of ARDS.⁶ Similarly, soluble receptor for advanced glycation end-products (sRAGE) in plasma independently is associated with death in ARDS patients, mainly because of its critical role in inflammatory reactions.⁷ Other evidence supports that a high level of angiopoietin 2 (ANG-2) leads to increased microvascular permeability in critically ill patients with ARDS and is associated with worse outcome.⁸ Although these studies provide promise, insufficient causal validity means biomarkers seldom are incorporated into clinical practice for ARDS treatment. Thus, although many potential biomarkers have been identified by observational studies,⁹ a gap in understanding their clinical applicability exists. Theoretically, causal biomarkers independent of confounding factors have more potential in pharmacologic responses to a therapeutic intervention.

A potential way to identify causal biomarkers independent of confounders is through Mendelian randomization. Mendelian randomization studies often are described as naturally occurring randomized controlled trials in which genetic factors are assigned randomly by nature,¹⁰ with this random assortment of genetic variants used to infer causal relationships between exposures and outcomes.¹¹ In the context of ARDS, Mendelian randomization analysis linked two causal protein biomarkers (ANG-2 and sRAGE) with development of sepsis-associated ARDS.^12,13 We hypothesized that additional plasma proteins can serve as causal biomarkers for ARDS. To identify markers that ultimately may have clinical usefulness, we used generalized summary data-based Mendelian randomization (GSMR) with protein quantitative trait loci data from the Efficiency and safety of varying the frequency of whole blood donation (INTERVAL) study and genome-wide association study (GWAS) data from the Identification of SNPs Predisposing to Altered Acute Lung Injury Risk (iSPAAR) consortium study to infer the causal association between protein levels and ARDS 28-day mortality.^14,15

Methods

Study Populations and Plasma Samples

Study design is illustrated in Figure 1. ARDS patients were enrolled at the ICUs of Massachusetts General Hospital and Beth Israel Deaconess Medical Center as part of the Molecular Epidemiology of ARDS (MEARDS) cohort between 1998 and 2014. Detailed information on this cohort was published previously.^16,17 Patients were eligible for the study if they did not have any of the following: age younger than 18 years, HIV infection, diffuse alveolar hemorrhage, chronic lung diseases other than COPD or asthma, directive to withhold intubation, immunosuppression not secondary to corticosteroid, treatment with granulocyte colony-stimulating factor, cytotoxic therapy, or solid organ or bone marrow transplant. We collected demographics, medical history, vital signs, hematologic characteristics, and biochemical indicators and performed frequent arterial blood gas analysis and chest radiography within 24 h of admission. All plasma samples also were collected within the 24 h of ICU admission, which were used for proteomics profiling. Presence of ARDS was adjudicated in accordance with Berlin criteria requiring that chest radiograph and oxygenation criteria be met on the same calendar day while invasively ventilated.¹⁸ Detailed biospecimen collection information was described previously,¹⁹ and 300 plasma samples (obtained from the MEARDS cohort) were quantified precisely by the Olink platform, including 13 panels (1,161 proteins) followed by quality control.²⁰ The quality control process was sample control and variable control.^21,22 All proteins were log₂ transformed to obey normal distribution, and values of less than the detection limit of the assay were replaced by the lower limit of detection. Samples were removed if the control sample deviated more than ± 0.3 from the median value of all samples on the run. The batch effect in Olink data was adjusted by sva package from R version 3.6.1 software (R Foundation for Statistical Computing).²³

Figure 1.

Study design flowchart. ARDSNet = ARDS Clinical Trials Network; GSMR = generalized summary data-based Mendelian randomization; GWAS = genome-wide association study; IGFBP7 = insulin-like growth factor binding protein 7; INTERVAL = Efficiency and safety of varying the frequency of whole blood donation; MEARDS = Molecular Epidemiology of ARDS; OPG = osteoprotegerin; OPLS-DA = orthogonal projections to latent structures discriminant analysis; pQTL = protein quantitative trait loci; SNP = single nucleotide polymorphism; t-PA = tissue-type plasminogen activator; vWF = von Willebrand factor; WPB = Weibel-Palade body.

GWAS Data Source and Quality Control

GWAS data for ARDS were obtained from the Identification of SNPs Predisposing to Altered Acute Lung Injury Risk (iSPAAR) consortium study. iSPAAR samples were from the MEARDS study enrolled at the Massachusetts General Hospital and Beth Israel Deaconess Medical Center and from the ARDS Clinical Trials Network (ARDSNet). In the current study, we selected ARDS cases with complete clinical information, including: status at 28 days, age, sex, ethnicity, BMI, Acute Physiology and Chronic Health Evaluation III score, vasopressors within 24 h, prednisone within 21 days, smoking status, alcohol abuse, and predisposing conditions such as sepsis, pneumonia, multiple trauma, or aspiration. Clinical variables were adjusted in GWAS multilogistic regression, Mendelian randomization, and causal mediation analyses. GWAS data and quality control procedures are described in e-Appendix 1.

Summary statistics of protein quantitative trait loci were obtained from the INTERVAL study, which analyzed 2,994 proteins via the SOMAScan assay from 3,301 European people.¹⁵ Selection of proteins on the platform reflects both the availability of purified protein targets and a focus on proteins suspected to be involved in the pathophysiologic characteristics of human disease.

Mendelian Randomization Analysis

We conducted Mendelian randomization analysis to infer the causal effect of genetically predicted proteins on ARDS 28-day mortality using GSMR. All single nucleotide polymorphisms (SNPs) significantly associated with each protein (P < 5 × 10^-8) were selected as candidates for genetic instrumental variables (IVs). The effects of those SNPs on ARDS 28-day mortality risk were calculated separately from MEARDS and ARDSNet. Summarized results of the two parts were integrated separately with the INTERVAL study to evaluate the causal effect of the proteins on ARDS 28-day mortality. High linkage disequilibrium (r² > 0.1) SNPs were removed in the IVs. Then, pleiotropic SNPs were removed by Heterogeneity in dependent instrument (HEIDI)-outlier detection methods with a P < .05 / (number of IVs).¹⁴ Using all markers on the genotyping platform, we performed principal component analysis to identify 10 principal components allowing for adjustment of genetic population stratification. Four Mendelian randomization methods—sample median, penalized weighted median, inverse-variance weighted, and robust inverse-variance weighted—were used to perform sensitivity analysis.^{24, 25, 26, 27} The statistical power of Mendelian randomization in this study was calculated by mRnd, an online tool calculating the statistical power given the chosen sample size, type I error (α value), causal effect of protein on ARDS 28-day mortality, and the proportion of death within 28 days.²⁸

Statistical analysis is described in e-Appendix 1.

Results

Baseline characteristics of patients are described in Table 1. Ninety-nine ARDS patients died within 28 days (33.00%), which is similar to the mortality reported in a previous systematic review (33.53%).²⁹ Briefly, patients of older age, with low BMI, high Acute Physiology and Chronic Health Evaluation III score, low platelet count, low Pao₂/Fio₂, high creatinine, high bilirubin, sepsis, or bacteremia had higher 28-day mortality.

Table 1.

Description of ARDS Patient Characteristics

Baseline Characteristics	Survival Within 28 Days (n = 201)	Nonsurvival Within 28 Days (n = 99)	All Patients (N = 300)	P Value
Age, y	53.71 ± 18.77	68.27 ± 15.50	58.51 ± 19.01	1.33 × 10-11
Sex
Female	78 (38.81)	36 (36.36)	114 (38.00)	.77
Ethnicity
White	190 (94.53)	95 (95.96)	285 (95.00)	.69
BMI, kg/m2	29.18 ± 8.78	26.76 ± 7.58	28.39 ± 8.47	.01
APACHE III score	71.71 ± 21.02	92.26 ± 20.46	78.49 ± 22.94	5.10 × 10-14
Vasopressin within 24 h
Yes	137 (67.18)	73 (73.74)	210	.39
Prednisone within 21 d
Yes	20 (9.95)	15 (15.15)	35	.25
Smoking status
Current	68 (33.83)	21 (21.21)	89 (29.67)	.06
Former	44 (21.89)	36 (36.36)	80 (26.67)
Alcohol abuse
Yes	25 (12.44)	16 (16.16)	41 (13.67)	.48
ARDS risk factors
Sepsis	172 (85.57)	94 (94.95)	266 (88.67)	.03
Septic shock	126 (62.69)	74 (74.75)	200 (66.67)	.05
Pneumonia	81 (40.30)	38 (38.38)	119 (39.67)	.84
Bacteremia	25 (12.44)	27 (27.27)	52 (17.33)	.002
Multiple fractures	13 (6.47)	1 (1.01)	14 (4.67)	.07
Aspiration	17 (8.46)	11 (11.11)	28 (9.33)	.59
Pulmonary contusion	10 (4.98)	2 (2.02)	12 (4.00)	.36
Platelet count, 109/L	228.56 ± 147.22	167.26 ± 123.43	226.01 ± 208.36	.006
Pao2/Fio2	109.04 ± 42.44	98.08 ± 37.83	105.43 ± 41.24	.02
Creatinine, mg/dL	1.53 ± 1.23	1.95 ± 1.33	1.67 ± 1.28	.01
Total bilirubin, mg/dL	0.90 (0.20-9.07)	1.55 (0.20-21.11)	1.00 (0.20-16.25)	.004
Ventilator-free days	24 (15.35-27.83)	22 (2.48-27.53)	19 (0.00-27.00)	< 2.20 × 10-16

Data are presented at No. (%), mean ± SD, or median (95% CI), unless otherwise indicated. Bold represents P value < .05. APACHE = Acute Physiology and Chronic Health Evaluation.

Heterogeneous Proteins in Mortality and Survivor Groups

Overall, plasma protein levels differentiated between patients who died and those who survived, with R²Y at 0.60 and Q² at 0.50 (Fig 2A). Three hundred ninety-three proteins with variable influence on projection of more than 1.00 were considered potential biomarkers (Fig 2B).³⁰ The function of potential biomarkers was identified by gene ontology annotation and Kyoto Encyclopedia of Genes and Genomes pathway analysis. Gene ontology analysis categorized potential biomarkers into several essential biological processes, including inflammatory response, immune response, and cell adhesion, which are all critical processes in ARDS (Fig 2C). Kyoto Encyclopedia of Genes and Genomes analysis associated those 393 proteins mainly with 20 signaling pathways (Fig 2D). Two hundred thirty-seven of 393 potential biomarkers were identified as associating with ARDS mortality by single logistic regression without adjustment for covariates (e-Fig 1A, 1C, e-Table 1), and 31 of 237 were confirmed by multilogistic regression analysis with adjustment for critical covariates (e-Fig 1B, 1D, e-Table 2). Statistically significant protein biomarkers separately demonstrated a considerable discrimination ability, with areas under the receiver operating characteristic curve (AUCs) ranging from 0.64 to 0.74 (e-Fig 2). Predictive ability, evaluated by AUC, for the clinical variables (critical covariates) combination was 0.81 and increased to 0.92 with the contribution of 31 proteins (e-Fig 3A). Protein-protein interaction analysis showed that the 31 proteins participated in platelet-related, extracellular matrix-related, and proteinaceous extracellular matrix-related pathways critical in ARDS development and prognosis (e-Fig 3B).

Figure 2.

Orthogonal projection to latent structure-discriminant analysis (OPLS-DA) and protein pathway analysis. A, OPLS-DA score plots comparing deaths to survival within 28 days. Plasma protein levels differentiated between patients who died and those who survived, with R²Y at 0.60 and Q² at 0.50. B, Feature selection results based on variable importance in projection (VIP). Three hundred ninety-three proteins with VIP of more than 1.00 were considered potential biomarkers. C, Gene ontology analysis for potential biomarkers. D, Kyoto Encyclopedia of Genes and Genomes analysis for potential biomarkers. ECM = extracellular matrix; HIF = hypoxia-inducible factor; NF = nuclear factor; PI3K-Akt = phosphatidylinositol 3-kinase/protein kinase B; PPAR = peroxisome proliferator-activated receptors; TNF = tumor necrosis factor.

Mendelian Randomization Analysis to Infer the Causal Effect of Proteins on ARDS 28-Day Outcome

In total, 12 of the 31 proteins (CD163, FCGR2A, GDF15, IGFBP7, INHBC, MET, PAPPER, PROC, SIGLEC10, SPARCL1, VCAM, and TNC) were identified from the INTERVAL study, and the candidate IVs are described in e-Tables 3-14. Mendelian randomization analysis in MEARDS showed that two genetically predicted proteins (predicted by SNPs without confounder interference) significantly increased the risk for ARDS 28-day mortality: IGFBP7 (OR, 2.61; 95% CI, 1.33-5.13; P = .005) and GDF15 (OR, 1.80; 95% CI, 1.38-2.34; P = 1.61 × 10^-5) (Fig 3 In contrast, one genetically predicted protein decreased the risk of ARDS 28-day mortality: FCGR2A (OR, 0.71; 95% CI, 0.61-0.82; P = 6.07 × 10^-6). However, its effect on ARDS 28-day mortality was opposite to the proteomics analysis (Fig 3C), so FCGR2A was excluded from further analysis. Selected IVs, used in GSMR, are detailed in e-Tables 15 and 16 according to studies. Mendelian randomization analysis was applied in an independent ARDS GWAS data set from ARDSNet to replicate the results from MEARDS; analysis revealed that genetically predicted IGFBP7 significantly elevated the risk of ARDS 28-day mortality (OR, 7.61; 95% CI, 2.01-28.86; P = .003) (e-Fig 4). The statistical power of IGFBP7 detection in this study was calculated as 0.81 (sample size_MEARDS = 403; α = 0.05; K_{Proportion of 28-day death} = 0.3; OR_{Mendelian randomization} = 2.61). However, genetically predicted GDF15 in ARDSNet exhibited an opposite effect with the Mendelian randomization results in MEARDS, although the effect was significant (OR, 0.32; 95% CI, 0.19-0.53; P = 1.02 × 10^-5). Further, the four summary data-based Mendelian randomization methods used in sensitivity analysis support the adverse effect of IGFBP7 on ARDS 28-day mortality (e-Fig 5). Because most ARDS cases originate from sepsis, we also conducted a subgroup analysis restricted to sepsis patients. Consistently, multivariate logistic regression showed that IGFBP7 was associated with ARDS 28-day mortality (OR, 2.32; 95% CI, 1.05-1.19; P = .002) in plasma analysis and was associated causally with ARDS 28-day mortality (OR_MEARDS = 2.23; 95% CI, 1.10-4.51; P = .02; OR_ARDSNet = 6.69; 95% CI, 2.06-21.67; P = .002) in GSMR analysis.

Figure 3.

Generalized summary data-based Mendelian randomization (GSMR) analysis for ARDS 28-day mortality. Relationship between the effect size estimates on proteins (x-axis) and the effect size estimates on ARDS 28-day mortality (y-axis) for all single nucleotide polymorphisms (SNPs) that served as instrumental variables. The 95% CIs for the estimated SNP effect sizes on proteins are shown as vertical orange lines, whereas the 95% CIs for the estimated SNP effect sizes on ARDS 28-day mortality are shown as horizontal orange lines. A, Insulin-like growth factor binding protein 7 (IGFBP7). B, Growth differentiation factor 15 (GDF15). C, Plasma proteins and GSMR analysis from the Molecular Epidemiology of ARDS (MEARDS) studies.

Causal Mediation Analysis for IGFBP7 on ARDS openbrace28-Day Outcome

Many indices at ICU admission reflect the severity of ARDS, including platelet count, Pao₂/Fio₂, bilirubin, and creatinine. Platelet count is believed to contribute to ARDS pathogenesis and prognosis through platelet involvement in the inflammatory response and disseminated intravascular coagulation.^31,32 We previously showed that early-stage thrombocytopenia in critically ill patients is associated with the development of mortality in ARDS.³³ Previous studies also examined the relationship between ARDS mortality and markers such as Pao₂/Fio₂, bilirubin, and creatinine.^{34, 35, 36} We therefore conducted a causal mediation analysis to identify the mechanism by which protein biomarkers relate to ARDS 28-day mortality via explanatory variables or mediators (eg, platelet count).^37,38

IGFBP7 is a critical component of Weibel-Palade bodies (WPBs) along with von Willebrand factor (vWF), tissue-type plasminogen activator (t-PA), and osteoprotegerin.^{39, 40, 41, 42} The negative associations between each of vWF and t-PA and ARDS 28-day mortality also are replicated in this study (OR_vWF = 1.06 [95% CI, 1.01-1.11; P = .02]; OR_t-PA = 1.08 [95% CI, 1.04-1.13; P = 3.77 × 10^-4]). Causal mediation analysis showed that the effects of IGFBP7 and t-PA on ARDS 28-day mortality were mediated significantly by platelet count, with average causal mediated effects up to an OR of 1.03 (95% CI, 1.02-1.04; mediated proportion, 23.67%; P = .01) for IGFBP7 and an OR of 1.02 (95% CI, 1.01-1.03; mediated proportion, 17.20%; P = .046) for t-PA (e-Table 17).

Clinical Application

Clinical decision-making analysis revealed that the receiver operating characteristic curve of IGFBP7 was significantly different from clinical features and increased the AUC to 0.85 (P = .03) (Fig 4A). The receiver operating characteristic curve evaluates an overall discernment ability (sensitivity and specificity), but is insufficient in accounting for the clinical usefulness of a specific model.⁴³ The decision curve analysis allows threshold probability to vary to examine whether the IGFBP7 model is superior to another at a certain range of threshold with respect to net benefits (NBs).⁴⁴ The trapezoidal numerical integration method confirmed that the NBs of the IGFBP7 model were superior to those of the basic model (NB_{IGFBP7 model} = 0.35 vs NB_{basic model} = 0.26; P = .005), crossing a wide range of threshold from 0.25 to 0.73 (Fig 4B).

Figure 4.

Clinical application of insulin-like growth factor binding protein-7 (IGFBP7). A, Receiver operating characteristic (ROC) curve, which was used to evaluate the performance of IGFBP7 and clinical variables for ARDS 28-day mortality prediction in plasma samples. B, Decision curve analysis. Comparison of standard net benefits among treat all (gray line), treat none (black line), IGFBP7 model (red line), and basic model (green line) groups. C, Nomogram constructed with clinical findings (green) and IGFBP7 (red) for ARDS 28-day mortality risk. The probability of each predictor can be converted into the points axis in the top of the nomogram. The summary of the points of each predictor corresponded to the total points at the bottom of the nomogram. After adding the points of each predictor in the total points axis, a patient’s probability of mortality can be found at the bottom of the nomogram. For example, if a patient’s score was 250, the ARDS 28-day mortality risk probability corresponds to 0.85. APACHE = Acute Physiology and Chronic Health Evaluation; AUC = area under the receiver operating characteristic curve.

To apply IGFBP7 easily in clinical practice, we combined clinical information and log2-transformed IGFBP7 of patients to develop a nomogram, making intuitive graphical and individualized predictions. A weighted score calculated using all predictors was used to predict 28-day mortality (Fig 4C). Discrimination and calibration methods were applied in both the discovery and validation phases. The c-index was calculated as 0.80, referring to a relatively good prediction of the nomogram. Calibration plot, drawing from a 1,000-time bootstrap for correcting overfitting, showed high-level accordance between predicted and actual probability (e-Fig 6).

Discussion

Our comprehensive study incorporating both plasma proteomic analysis and two rounds of Mendelian randomization analysis demonstrated that genetically predicted IGFBP7 causally effected 28-day mortality in ARDS. We determined that plasma IGFBP7 associated with ARDS 28-day mortality partially mediated by platelet count at ICU admission. To the best of our knowledge, this is the first causal association study on ARDS 28-day mortality outcomes leveraging GWAS, proteomics, and novel causal inference methods combined.

In recent years, clinical research groups have performed clinical studies assessing protein biomarkers in ARDS. Calfee et al^45,46 used updated plasma biomarkers and proceeded to define a detailed endotype (hyperinflammatory and hypoinflammatory) of ARDS that may vary on the individual level. They described different ARDS subphenotypes and biomarker panels that may help clinicians to select patients who may benefit from different therapeutic strategies.⁴⁷ A large biomarker study of 1,341 patients enrolled in the Protocolized Care of Early Septic Shock trial found that endothelial cell permeability and hemostasis-related proteins were associated with increased mortality.⁴⁸ Unlike targeted proteomics study mentioned above, untargeted proteomics analysis in ARDS to date typically have examined a small number of samples.^{49, 50, 51, 52} Given the small sample size and potential environmental confounders, many proteomic biomarkers processed in other studies likely contributed to the heterogeneous conclusions. Causal biomarker studies benefit from interpretation of clinical outcomes, because disentangling correlation from causation is important in ARDS. For example, ANG-2 is an established biomarker of endothelial activation and permeability that is associated strongly with ARDS risk and outcome, but is limited in causality. Genetic variants exhibiting cis-protein quantitative trait loci with plasma ANG-2 were associated with ARDS risk from sepsis, and the risk is statistically determined by genetically predicted ANG-2 levels,⁵³ emphasizing the causal role of ANG-2 in inducing sepsis-associated ARDS. A Mendelian randomization approach confirmed that plasma sRAGE acts as a genetically regulated causal intermediate in sepsis-associated ARDS.¹² Subsequently, multiple clinical trials (clinicaltrials.gov Identifiers: NCT01600651, NCT00811629, NCT02070536, and NCT01270295) validated the usefulness of sRAGE in predicting ARDS development at ICU admission. Treatments based on ANG-2 and sRAGE have not been reported, and small sample sizes may limit clinical trial options. We used GSMR analysis in this study, a summary data Mendelian randomization framework, to expand current study sample size, resulting in higher power. One of the attractive features of our study is that the causal association between IGFBP7 and ARDS 28-day mortality risk is unlikely to be confounded with environmental factors and should be validated in follow-up clinical trials.

Previous studies strongly support that IGFBP7 plays a mechanistic role in pulmonary inflammatory response both in mammal models and population studies. For example, IGFBP7 levels were upregulated in experimental radiation-induced lung injury models⁵⁴ and were displayed synchronously in zinc chloride smoke inhalation lung injury models, suggesting IGFBP7 derangement with pulmonary inflammation.⁵⁵ A population-based retrospective study demonstrated that circulating IGFBP7 levels were increased during acute exacerbation of COPD and were reduced after convalescence, indicating the potential prognostic value of IGFBP7 for pulmonary inflammation.⁵⁶ Given its comparatively low affinity in a comparison with the insulin-like growth factor binding protein family, IGFBP7 functions not only as an insulin-like growth factor binding protein, but also as a direct growth-suppressing factor with an insulin-like growth factor-independent action similar to that of insulin-like growth factor binding protein 3.^57,58 Previous studies found that insulin-like growth factor 1, one of the IGFBP7 binding proteins, coupled with insulin-like growth factor binding protein 3 is involved in the pathogenesis of lung injury and is associated independently with ARDS risk and mortality.^52,59,60

WPBs, storage granules of endothelial cells, play a dual role in hemostasis and inflammation by capturing platelets during the onset of primary hemostasis.^61,62 IGFBP7 is a critical component of WPBs, along with vWF, t-PA, and osteoprotegerin.^{39, 40, 41, 42} The dysregulation of vWF and t-PA affect the development and prognosis of ARDS, mainly by regulating inflammatory response.^63,64 The negative associations between each of vWF and t-PA and ARDS 28-day mortality also are replicated in this study. Carboxy-terminal D4-C1-C2-C3-CK domains (vWF variants) are needed for cotargeting of IGFBP7 to pseudo-WPBs, which suggests that the binding of IGFBP7 to vWF is required for its targeting to WPBs.³⁹ However, little evidence indicates a potential interaction relationship between t-PA and IGFBP7. Platelets are believed to make an important contribution to ARDS among critically ill patients acting in conjunction with fibrinogen to mediate endothelial damage through multiple signal transduction pathways.⁶⁵ Our results indicate that the effects of IGFBP7 and t-PA on ARDS 28-day mortality were mediated partially by platelet count at ICU admission. Because of the physiological function of WPBs on platelets, theoretically, platelet count has the capacity to be an intermediator between IGFBP7 and ARDS 28-day mortality.

Our study has several strengths. First, the study design is based on both plasma proteomic detection and multistage causal inference analysis, which bolsters reliability. Second, previous simulation studies have proven that GSMR outperforms existing summary data-based Mendelian randomization methods because GSMR leverages power from multiple genetic variants, accounting for linkage disequilibrium between variants and for pleotropic effects.^25,66,67 Sensitivity analysis with the other four Mendelian randomization methods in both the MEARDS and ARDSNet studies also supports the adverse effect of IGFBP7 on ARDS 28-day mortality and further validates the robustness of the GSMR results. Third, the IV was selected based on a strict genome-wide significance threshold (P < 5 × 10^-8). Fourth, in the current study, we restricted the causal inference to non-Hispanic White people to minimize the population stratification bias, which is the most important confounder in genetic studies. However, we acknowledge that this is a limitation because recent studies showed that acute lung injury mortality differs among different races and ethnicities, indicating the importance of studying a diverse population.⁶⁸ Future studies with non-European populations are needed to replicate our findings. We also acknowledge some other limitations. First, the GWAS sample size arguably is small, decreasing the power of causal protein biomarker detection. Theoretically, more causal protein biomarkers will be discovered with an expansion of genotype ARDS cases. Second, our study did not elucidate underlying biological mechanisms linking these biomarkers to ARDS. Finally, although our study detected IGFBP7 as a causal protein biomarker by Mendelian randomization, the clinical benefits (AUCs, nomogram, and decision curve analysis) require external validation.

Interpretation

This study linking plasma IGFBP7 as a novel causal protein involved in the pathogenesis of ARDS 28-day mortality and in platelet function in ARDS provides a foundation for further experimental and clinical investigation of the mechanisms and usefulness of this biomarker.