Plasma kallikrein predicts primary graft dysfunction after heart transplant

Abstract

Background:

Primary graft dysfunction (PGD) is the leading cause of early mortality after heart transplant. Pre-transplant predictors of PGD remain elusive and its etiology remains unclear.

Methods:

Microvesicles were isolated from 88 pre-transplant serum samples and underwent proteomic evaluation using TMT mass spectrometry. Monte Carlo cross validation (MCCV) was used to predict the occurrence of severe PGD after transplant using recipient pre-transplant clinical characteristics and serum microvesicle proteomic data. Putative biological functions and pathways were assessed using gene set enrichment analysis (GSEA) within the MCCV prediction methodology.

Results:

Using our MCCV prediction methodology, decreased levels of plasma kallikrein (KLKB1), a critical regulator of the kinin-kallikrein system, was the most predictive factor identified for PGD (AUROC 0.6444 [0.6293, 0.6655]; odds 0.1959 [0.0592, 0.3663], Furthermore, a predictive panel combining KLKB1 with inotrope therapy achieved peak performance (AUROC 0.7181 [0.7020, 0.7372]) across and within (AUROCs of 0.66-0.78) each cohort. A classifier utilizing KLKB1 and inotrope therapy outperforms existing composite scores by more than 50 percent. The diagnostic utility of the classifier was validated on 65 consecutive transplant patients, resulting in an AUROC of 0.71 and a negative predictive value of 0.92-0.96. Differential expression analysis revealed a enrichment in inflammatory and immune pathways prior to PGD.

Conclusions:

Pre-transplant level of KLKB1 is a robust predictor of post-transplant PGD. The combination with pre-transplant inotrope therapy enhances the prediction of PGD compared to pre-transplant KLKB1 levels alone and the resulting classifier equation validates within a prospective validation cohort. Inflammation and immune pathway enrichment characterize the pre-transplant proteomic signature predictive of PGD.

Introduction

Primary graft dysfunction (PGD) after heart transplant is defined as idiopathic ventricular dysfunction during immediate post-transplant period (1). PGD can affect either or both ventricles simultaneously and is graded from mild to severe depending on the amount of compensatory support required. The International Society for Heart and Lung Transplantation reported that PGD is the leading cause of death within 30 days after transplant (2). Identifying predictive factors of PGD has the potential to improve risk stratification, organ allocation, and post-operative care as well as increase our understanding behind the etiology of PGD (3). While predictive models for PGD have been described, a risk model based solely on pre-transplant recipient factors remains elusive (4). Risk factors and biomarkers for PGD have been investigated (3) and recipient, perioperative and donor clinical measures have been linked with PGD risk (5–9). However, cross validation and generalizability of these risk factors is lacking across multiple institutions.

Molecular biomarkers have been shown to be predictive and robust for many diseases (10, 11). A rich and underexplored source of potentially prognostic biomarkers are contained in extracellular vesicles (12). In addition to diagnostic potential, extracellular vesicles are stable, easily extracted from patient blood, and have shown promise in the prediction of heart disease (13). Moreover, analysis of these vesicles has also revealed mechanisms underlying numerous disease processes (14–17). Here we report a multi-institutional cohort study to predict PGD using machine learning to identify novel combinations of serum microvesicle proteomics and clinical characteristics.

Materials and Methods

For the complete methodology details, please see Supplementary Materials.

Patient cohorts

Patient blood samples were prospectively recruited at Columbia University Irving Medical Center (Columbia) between 2014 and 2016. Patient blood samples were retrospectively collected from biobanks at Cedars-Sinai hospital (Cedars) and Pitié Salpêtrière University Hospital (Paris). Only severe PGD by ISHLT definition were included. Patients undergoing retransplant were excluded. The initial cohort for PGD prediction was comprised of PGD samples matched to non-PGD samples by age and gender. In order to calculate more clinically relevant predictive values, the validation ELISA cohort was comprised of consecutive patients undergoing transplant. Human subjects protocol was approved by each institution’s IRB and patients provided informed consent. Patient characteristics were collected including demographics, biometrics, labs, medications and hemodynamics. PGD status was defined per ISHLT guidelines (8).

Mass spectrometry analysis

Briefly, patient samples from each site were collected for processing by the Proteomics Core at Columbia University Irving Medical Center (Figure 1A). Each patient cohort was processed independently. Total microvesicle was isolated from serum. Each sample was proteolytically cleaved with trypsin and chemically labeled with TMT10plex isobaric mass tags (18–20) separately. MS spectra were acquired with an Orbitrap Fusion Tribrid Mass Spectrometer (Thermo Scientific) and raw spectrometric data were analyzed using Proteome Discoverer 2.2.

Figure 1. Blood-derived micro-vesicle proteomics and identified proteins in the three patient cohorts.

A) Patients blood were drawn and serum was processed according to protocol at the clinical site (see Methods). Microvesicles from serum were pelleted, isolated, and lysed. The proteomic content was extracted and diluted within buffer. Samples were tagged by isobaric labeled TMT-tags to enable multiplexing. After processing, samples were ionized by liquid chromatography mass spectrometry (LCMS) and processed by the Thermo Fisher Tribrid MS Orbitrap mass spectrometer. Individual peptides and proteins from the quantified mass spectrum were identified and estimated by the Proteome Discoverer software (Thermo Fisher). B) Mass spectrometry identified microvesicle proteins of patient from each cohort. Proteins were identified from spectral images at a False Discovery Rate of 1%. C) Protein filtering steps from all 345 identified and quantified proteins to 181 eligible proteins for downstream analysis.

Protein expression analysis

We calculated a differential protein expression signature between PGD and non-PGD patient samples (Figure S5). The protein association calculated was used as the differential rank statistic for pathway analysis using gene set enrichment analysis (GSEA) (21).

PGD prediction

We used a Logistic Regression model with L1 regularization for each marker to determine their predictive performance and association to PGD (see Figure S2). To estimate the prediction variance, we employed Monte Carlo cross validation (MCCV) (22). The PGD prediction probabilities were compared to the true PGD status to compute the area under the receiver operating characteristic curve (AUROC), and other metrics. Bootstrapping analysis (samples with replacement) resulted in a population distribution for prediction performances, and a permutation analysis was similarly performed, with random labeling of PGD status in patients, to generate and test prediction metrics from random PGD assignment. We tested differences in the bootstrap and permutation distributions, as well as between the two bootstrap distributions, with the 2-sample Kolmogorov-Smirnov test. Statistics followed by the use of bracket notation indicated reporting of the average statistic and its 95% confidence interval. The average statistic and standard errors were noted when reporting Student-t test results.

KLKB1 ELISA assay heart transplant patients

Enzyme linked immunosorbent assay (ELISA) (Abcam) was used to assess KLKB1 protein concentration in a validation cohort of pre-transplant serum prospectively collected in 65 consecutive patients at CUIMC. To be able to compare ELISA and mass spectrometry derived protein expression, we first minimum-maximum normalized the patient cohort data before our MCCV strategy as per our normalization strategy for all predictions. The putative PGD classifier equation shown in Figure 3F.

Data and materials availability:

All data associated with this study are available in the main text or the supplementary materials. All code for the analyses can be found on the following website: https://github.com/ngiangre/clinical_klkb1_analysis (23).

Results

Patient clinical characteristics

In total, 88 patients who underwent heart transplantation between 2014 and 2016 at Cedars Sinai Medical Center (n=43), Pitié Salpêtrière University Hospital (n=29) and Columbia University Irving Medical Center (n=16) were used for the initial proteomic and clinical characteristic analysis (Table 1). There are 37 different pre-transplant clinical characteristics across all the patients including PGD status (Table S1). Prior inotrope therapy significantly differed (linear model with and without site-of-origin p-values=0.002 and 0.003) between PGD and non-PGD (Table 1). In a multivariate model including all characteristics, only pre-transplant inotrope therapy associated with PGD (Table S3).

Table 1:

Clinical Characteristics

		PGD = No	PGD = Yes	P-value
	N	46	42
Patient characteristics	Age (mean (SD))	54.90 (13.43)	58.43 (10.13)	0.171
	BMI (mean (SD))	25.11 (4.07)	26.46 (5.07)	0.171
	Blood Type (%)
	A	19 (41.3)	15 (35.7)
	AB	5 (10.9)	3 (7.1)
	B	8 (17.4)	5 (11.9)
	O	14 (30.4)	19 (45.2)
	Donor Age (mean (SD))	39.98 (13.92)	40.95 (13.44)	0.74
	Sex = F (%)	14 (30.4)	13 (31.0)	1
	History of Tobacco Use = Y (%)	15 (32.6)	16 (38.1)	0.753
	Autoimmune Diseases =Y(%)	2 (4.4)	1 (2.2)
	Diabetes = Y (%)	14 (30.4)	15 (35.7)	0.765
Cardiomyopathy	Ischemic = Y (%)	14 (30.4)	18 (42.9)	0.323
	Non-Ischemic (%)			0.271
	Adriamycin	0 (0.0)	1 (2.4)
	Amyloid	2 (4.3)	0 (0.0)
	Chagas	0 (0.0)	1 (2.4)
	Congenital	0 (0.0)	1 (2.4)
	Hypertrophic cardiomyopathy	1 (2.2)	0 (0.0)
	Idiopathic	28 (60.9)	19 (45.2)
	Myocarditis	0 (0.0)	1 (2.4)
	Valvular Heart Disease	0 (0.0)	1 (2.4)
	Viral	1 (2.2)	0 (0.0)
Transplant factors	Ischemic Time (minutes (SD))	156.50 (62.97)	169.09 (53.02)	0.315
	Ventricular Assist Device = Y (%)	8 (17.4)	13 (31.0)	0.215
Hemodynamics	PA Diastolic (mean (SD)) mmHg	20.73 (6.76)	20.38 (7.79)	0.823
	PA Systolic (mean (SD)) mmHg	42.85 (12.46)	45.48 (15.07)	0.373
	PA Mean (mean (SD)) mmHg	29.31 (8.31)	31.28 (9.04)	0.289
	CVP (mean (SD)) mmHg	9.49 (5.28)	9.97 (5.18)	0.671
	PCWP (mean (SD)) mmHg	19.81 (7.78)	20.08 (8.88)	0.877
Lab values	Creatinine (mean (SD)) mg/dL	1.36 (1.14)	1.25 (0.43)	0.558
	INR (mean (SD))	1.48 (0.49)	1.65 (0.74)	0.196
	TBili (mean (SD)) mg/dL	0.93 (0.48)	0.79 (0.50)	0.205
	Sodium (mean (SD)) mEq/L	137.50 (4.64)	136.90 (5.08)	0.565
Medications	Antiarrhythmic Use = Y (%)	22 (47.8)	25 (59.5)	0.376
	Beta Blocker = Y (%)	24 (52.2)	30 (71.4)	0.102
	Inotrope = Y (%)	30 (65.2)	14 (33.3)	0.006
Composite Scores	CVP/PCWP (mean (SD))	0.49 (0.25)	0.55 (0.28)	0.273
	MELD (mean (SD))	13.57 (4.74)	14.31 (5.18)	0.483
	Radial Score (mean (SD))	2.57 (1.28)	2.24 (0.98)	0.185

Recipient characteristics at the time of transplant unless otherwise specified. Significance evaluated with a continuity-corrected chi-squared test for categorical characteristics and t-test for continuous characteristics. Abbreviations: Primary Graft Dysfunction, PGD; Body Mass Index, BMI; Pulmonary Artery, PA; Central venous pressure, CVP; Pulmonary capillary wedge pressure, PCWP; International Normalized Ratio, INR; Total bilirubin, TBili; Model for End Stage Liver Disease Score, MELD.

Patient blood microvesicle proteomic characteristics

Serum microvesicle protein spectra were obtained in at least triplicate for each patient (322 total replicates) (Figure 1A). As expected, the identified proteins were enriched in microvesicle and extracellular components (Table S4). Protein expression in the three patient cohorts (Figure S1) does not follow a normal distribution (Omnibus test of normality p-values << 0.001). The Columbia cohort was significantly different from Cedars-Sinai (Kolmogorov Smirnov test p-value < 3.19E-08) and from Pitíe Salpetriere (p-value = 8.70E-06). Protein expression was significantly different between the two retrospective patient cohorts (p-value = 0.030).

In total, 681 unique proteins were identified with 345 identified proteins present in every cohort the patient cohorts and 80 proteins were not identified in at least one patient (Figure 1B). There were 81 identified immunoglobulin proteins which were not included in the analysis.

Additionally, three proteins did not have corresponding gene name annotations. A final set of 181 proteins, which were identified in every patient across all patient cohorts, were used in downstream analyses (Figure 1C).

Prediction of post-transplant PGD using pre-transplant clinical and protein markers

The prediction of post-transplant PGD in patients was investigated using clinical and protein markers derived prior to transplant. Monte Carlo cross validation (MCCV; Figure S2) and permutation analysis (See Methods) was employed to calculate the prediction and significance of each clinical and protein marker in predicting PGD.

Overall, the expression of all protein markers did not significantly outperform (AUROC 0.4119 ± 0.05473 vs. 0.3751 ± 0.04712 independent 2-sample t-test p-value=0.9147) nor were more influential (odds 1.3477 ± 1.3324 vs. 1.0544 ± 0.2115 p-value=0.1819) than all clinical characteristics in predicting the post-transplant occurrence of PGD (Figure 2). Individually, 16 proteins and one clinical characteristic were significantly predictive of PGD occurrence (AUROC > 0.5, Bonferroni-corrected p-value < 0.001, beta coefficient 95% CI not including the null association, and permutation beta coefficient 95% CI including the null association; Table S5). The most predictive protein marker was plasma kallikrein (KLKB1) (AUROC 0.6444 [0.6293, 0.6655]; odds 0.1959 [0.0592, 0.3663]) where decreased expression of KLKB1 was significantly predictive of PGD status. The next most predictive markers (AUROC > 0.6) were the proteins peroxeridoxin 2 (PRDX2), tropomyosin alpha-4 (TPM4), and myeloperoxidase (MPO), where increased expression of each was significantly predictive of PGD status (Table S5). With respect to clinical factors, the absence of pre-transplant inotrope therapy was significantly predictive of PGD on its own, albeit modestly. (AUROC 0.5618 [0.5387, 0.5800]; average odds 0.4342 [0.3043, 0.6033]). Notably the presence of mechanical support was not predictive (AUROC 0.4753 [0.4395, 0.4741], odds 1.192 [1.000, 1.781],) nor did it attenuate the predictive performance of pre-transplant inotrope therapy towards PGD (Figure S8).

Figure 2. PGD prediction by clinical and protein markers.

Post-transplant PGD prediction, by Monte Carlo Cross Validation (Figure S2; see Methods), of the 181 proteins and 37 binarized clinical characteristics. Protein (crosses) and clinical (circles) marker association (beta coefficient; a measure of influence towards PGD occurrence) versus predictive performance (AUROC). A L1-regularized logistic regression model estimated the association (beta coefficient) of each marker to post-transplant PGD in patients. The diverging color palette indicates the negative log10 of the feature importance significance or p-value, after Bonferroni correction, from a permutation analysis.

Two marker panel PGD predictions

We next investigated 136 pairwise combinations of the 17 significantly predictive clinical and protein markers (Figure 3A). Overall, panels of inotrope therapy and a protein had significantly increased performance than combinations of two proteins (AUROC 0.6505 ± 0.02980 vs. 0.6070 ± 0.0454 p-value=2.123E-4; Figure 3B). For combinations involving pre-transplant inotrope therapy, the addition of KLKB1 outperformed all other protein combinations (AUROC 0.7181 [0.7020, 0.7372]). We also found that protein-protein marker panels containing KLKB1 outperformed all panels composed of other protein markers (t-test p-values 2.193E-13 to 6.634E-02; Figure 3C). The best performing panel overall and for each patient cohort was a combination of pre-transplant inotrope therapy and expression of KLKB1 protein (Figure 3D). There were 5 two marker panels significantly more predictive than the most predictive marker KLKB1 on its own (Table S6). The panel of inotrope therapy and KLKB1 showed the least variation while maintaining high performance across all cohorts (95% AUROC CI above 0.7; Figure 3E).

Figure 3. Performance of two-marker panels for predicting post-transplant PGD.

A) The AUROC for all pairwise marker panels. The rows and columns were sorted by the average performance across all panels per marker. B) The AUROC distribution for all panels per marker composition, either two clinical marker, one clinical and one protein marker, or two protein markers. C) The AUROC distribution for all two marker panels composed of at least one protein marker and all inotrope therapy panels. D) The AUROC performance of two marker panels comparison overall against the average of individual cohorts and the integrated cohort. E) The performance versus the variation of the performance between the three patient cohorts. F) KLKB1 and Inotrope therapy PGD classifier equation. The MCCV prediction scheme produces putative PGD classifier equations. The average of the feature importance bootstraps results in the coefficients for each marker. Computing the logit transformation on the sum of coefficients multiplied by patient data gives a probability of PGD occurrence. All panels were significantly predictive after Bonferroni correction at alpha 0.05.

PGD classifier performance

Each panel’s predictions form a two-marker classifier equation, as shown for KLKB1 protein and inotrope therapy panel in Figure 3F. The classifier equation for inotrope therapy and KFKB1 is the summation of multiplying −0.9946 by a binary value of pre-transplant inotrope therapy (0 or 1) and −2.140 by normalized pre-transplant KFKB1 expression. This equation demonstrates an inverse relationship between post-transplant PGD risk and either pre-transplant KFKB1 expression or inotrope therapy (or both). The PGD classifier has significantly increased performance compared to the markers on their own (Kolmogorov-Smirnov 2-sample test p-values<2.165E-23; Figure 4 and Table 2). We compared our prediction panel to existing PGD predictors: the radial score (7), the MEED score (3), and the CVP/PCWP ratio (24). Our two-marker panel significantly outperforms all composite scores by 50% on average (Figure 5 and Table 3; Kilogorov Smirnov 2-sample p-values<2.165E-23).

Figure 4. Pre-transplant KLKB1 protein expression and inotrope therapy predict post-transplant PGD.

A clinical and protein marker classifier improves PGD prediction over random predictions and predictions by individual markers alone. Out of all 171 combinations of two-marker panels, we optimized the across and within patient cohort PGD prediction where panels were significantly predictive from permutation analysis. Shown are the Receiver Operator Characteristic curves and Precision-Recall curves for the two-marker panel and each individual marker. The AUROC and AUPRC values are presented in the tables below the panels (the average AUROC and 95% confidence intervals). The prediction metrics and curves are the average of the 200 calculated MCCV bootstraps. Abbreviations: Area Under the Receiver Operating Characteristic curve (AUROC); Area Under the Precision-Recall Characteristic curve (AUPRC).

Table 2. Performance of KLKB1, inotrope therapy, and two-marker panel predictive performance across and within patient cohorts.

The clinical equations derived from MCCV applied onto KLKB1 expression and inotrope therapy patient data from each site and all sites combined.

	AUROC			AUPRC
Cohort	KLKB1 + Inotrope therapy	KLKB1	Inotrope therapy	KLKB1 + Inotrope therapy	KLKB1	Inotrope therapy
All	0.7181 [0.7020, 0.7372]	0.6444 [0.6293, 0.6655]	0.5618 [0.5387, 0.5800]	0.7322 [0.7092, 0.7486i	0.6659 [0.6434, 0.6968]	0.5213 [0.4974, 0.5410]
Columbia	0.7125 [0.6680, 0.7571]	0.6507 [0.5917, 0.6933]	0.6005 [0.5505, 0.6574]	0.7547 [0.7232, 0.8048]	0.7240 [0.6856, 0.7763]	0.5211 [0.4778, 0.5816]
Cedars-Sinai	0.7782 [0.7542, 0.7982]	0.6094 [0.5787, 0.6326]	0.6770 [0.6532, 0.7114]	0.7754 [0.7425, 0.8035]	0.6344 [0.6026, 0.6708]	0.6262 [0.5938, 0.6637]
Pitié Salpêtrière	0.6711 [0.6417, 0.7018]	0.6984 [0.6731, 0.7279]	0.4048 [0.3702, 0.4412]	0.7053 [0.6750, 0.7381]	0.6659 [0.6557, 0.6708]	0.4388 [0.4011, 0.4716]

Figure 5. Clinical and protein panel outperforms existing clinical predictors.

Our clinical and protein marker classifier outperforms existing clinical PGD predictors. Shown are the Receiver Operator Characteristic curves and Precision-Recall curves for the two-marker panel and existing clinical PGD predictors. The AUROC and AUPRC values are presented in the tables below the panels.

Table 3.

Performance comparison between existing PGD predictors and KLKB1 and inotrope therapy two-marker panel.

	AUROC				AUPRC
Cohort	KLKB1 + Inotrope therapy	CVP/PCWP	MELD	Radial Score	KLKB1 + Inotrope therapy	CVP/PCWP	MELD	Radial Score
All	0.7181 [0.7020, 0.7372]	0.3868 [0.3709, 0.4026]	0.3759 [0.3615, 0.3922]	0.3917 [0.3760, 0.4070]	0.7322 [0.7092, 0.7486]	0.4266 [0.4034, 0.4439]	0.4257 [0.4040, 0.4417]	0.4283 [0.4069, 0.4439]
Columbia	0.7125 [0.6680, 0.7571]	0.4189 [0.3753, 0.4644]	0.4175 [0.3832, 0.4586]	0.4893 [0.4502, 0.5313]	0.7547 [0.7232, 0.8048]	0.4563 [0.4192, 0.5151]	0.4411 [0.4037, 0.4870]	0.4815 [0.4430, 0.5296]
Cedars-Sinai	0.7782 [0.7542, 0.7982]	0.3951 [0.3728, 0.4147]	0.3531 [0.3333, 0.3687]	0.3830 [0.3546, 0.4017]	0.7754 [0.7425, 0.8035]	0.4282 [0.4041, 0.4569]	0.4122 [0.3841, 0.4372]	0.4205 [0.3925, 0.4439]
Pitié Salpêtrière	0.6711 [0.6417, 0.7018]	0.3619 [0.3409, 0.3877]	0.3847 [0.3648, 0.4053]	0.3551 [0.3335, 0.3863]	0.7053 [0.6750, 0.7381]	0.4194 [0.3844, 0.4419]	0.4466 [0.4093, 0.4685]	0.4197 [0.3831, 0.4431]

Whole serum KLKB1 ELISA in PGD

A validation cohort of 65 consecutive patients’ serum samples was prospectively collected on the day prior to heart transplant at CUIMC. Whole serum was used for KLKB1 ELISA to test the feasibility of a clinical test without microvesicle purification. Patients who had RV PGD or mechanical support for reasons other than PGD were excluded from the analysis. Potentially due to the small number of severe PGD (n=3), there was no significant difference in average KLKB1 levels when comparing patients with severe PGD to no PGD levels (Mann-Whitney U test 19.81 ± 6.248 vs. 45.796 ± 32.54 p-value=0.0511). However, by adding patients with moderate PGD (n=4), defined per ISHLT guidelines as moderate LV dysfunction requiring pharmacologic but not mechanical support, KLKB1 levels were significantly lower (Mann-Whitney U test 20.44 ± 11.40 vs. 45.796 ± 32.54 015 p-value=0.0128; Figure 6A). The putative PGD classifier from the original proteomic data produces an AUROC of 0.7143 to predict moderate/severe PGD compared to patients who did not have PGD (Figure 6B). The incidence of PGD in this cohort more closely approximates the national PGD rate of 7.4% (3) and in this setting, the classifier was marked by a high sensitivity and negative predictive value but a low specificity and positive predictive value (Figure 6C and Table S7).

Figure 6. PGD classifier validation with KLKB1 ELISA.

Applying the classifier equation to predict PGD in the assessment heart transplant patients with blood serum-derived KLKB1 ELISA protein concentrations and pre-transplant inotrope therapy. A) Normalized (between 0 and 1) ELISA KLKB1 concentrations comparison of no PGD versus moderate and severe PGD patients, as included in the assessment set to validate our putative PGD classifier (for a total of 65 patients). B) Applying the putative PGD classifier onto the assessment data, PGD prediction AUROC. C) Performance metrics of the applied classifier at the highest sensitivity.

Primary Graft Dysfunction pathway analysis and clinical tests in patients

To investigate PGD pathogenesis, a differential expression signature was calculated from proteomic data (262 proteins, including immunoglobulins, identified in all patients with corresponding gene names) (Figure S4). Gene set enrichment analysis (GSEA) was used to investigate enriched pathways and functions from the differential protein signature (21). Six pathways were significantly enriched (Table S8; FDR < 0.2), and 3 pathways were depleted in patients with PGD (Table S9; FDR < 0.2; Figure 7A).

Figure 7. Gene set enrichment analysis and validation using clinical diagnostic assays.

A) Enrichment and depletion of pathways using differential protein expression. B) Significant protein marker predictors also significantly predictive within significantly enriched and depleted pathways. Proteins are significantly predictive as previously described (AUROC > 0.5, Bonferroni-corrected p-value < 0.001, beta coefficient 95% CI not including the null association, and permutation beta coefficient 95% CI including the null association). Clinical diagnostic ELISA tests for C) ESR and D) hsCRP. The units of ESR are mm/hr, CRP are mg/dL, C3 are mg/dl, C4 are mg/dl, and total complement are U/ml.

We considered the sets of proteins involved within each pathway and function in combination to predict PGD in patients. The same MCCV methodology and the prediction significance thresholds defined above were used for this analysis. Out of 196 proteins, 8 proteins were found to be significantly predictive within at least one of the 136 pathways and functions: KLKB1, PRDX2, TPM4, MPO, CAT, HSPA5, IGHD and IGLV2-11 (Table S9). Significant protein predictions within these pathways and functions (Figure 7B) revealed enrichment of processes related to inflammation, coagulation, and activation of the innate immune system. Downregulation of KLKB1 was identified in the activated complement and immune response pathways.

Markers of inflammation were also analyzed in the validation cohort. We found a trend towards increased erythrocyte sedimentation rate (66.0 ± 43.20 versus 33.70 ± 26.86 Mann-Whitney U test p-value=0.07; Figure 7C). C-reactive protein (27.24 ± 19.51 versus 11.27 ± 25.54 p-value=0.16; Figure 7D) and complement levels were not significantly altered (Figure S5). This analysis was hampered by a small number of severe PGD patients and wide confidence intervals. However, there does appear to be some laboratory trend towards increased inflammation corresponding with the results of the GSEA analysis.

Discussion

In this study, we identify pre-heart transplant recipient clinical and proteomic markers predictive of post-transplant PGD using a data-driven methodology to generate a clinically interpretable PGD classifier. Machine learning and statistical techniques were used to mitigate confounding in biological enrichment analyses and improve predictive accuracy with a modest population size. Reduction in KLKB1 was the strongest predictor of PGD both by itself and in combination with other markers. KLKB1 is a serine protease that controls the activation of both inflammation and coagulation in what is known as the kallikrein-kinin-system (KKS) (25). In the inflammatory response, KLKB1 converts high molecular weight kininogen into bradykinin stimulating the release of nitric oxide and prostacyclin causing vasodilation and increased vascular permeability (26). It also acts as a neutrophil chemoattractant, causing degranulation (27–29). Studies of the KKS system in patients with sepsis, a markedly inflammatory state, demonstrated increased KKS activity, characterized by decreased levels of plasma kallikrein, likely due to consumption (30). Decreases in KLKB1 have been noted in typhoid fever (31), ARDS (32), cardiopulmonary bypass (27) and in normal volunteers infused with gram negative endotoxin (33). Similarly, in animal models of inflammatory bowel disease (34) and inflammatory arthritis (35), plasma kallikrein levels were markedly reduced.

Other predictive proteins identified are likewise involved in either inflammation or innate immunity including PRDX2, MPO (36, 37) PGLYRP2 (38) and DEFA1 (39). Similarly, enrichment analysis of protein expression differences demonstrated several upregulated biological processes including inflammatory and immune pathways in patients prior to PGD. Laboratory tests in the validation cohort trended towards increased inflammation though were not significant. It remains to be seen whether this inflammatory signature is purely a biomarker or contributes to PGD and importantly, whether modifying this state will have an impact on the evolution of PGD.

The lack of inotrope therapy was predictive of PGD and this stands in contrast to prior analyses which demonstrated that the presence of inotrope therapy was associated with PGD (9). Pre-transplant inotrope therapy and durable mechanical support (such as LVAD) are generally exclusive prior to transplant and mechanical support has been associated with PGD in prior studies (6). However, mechanical support was not significantly predictive of PGD in our analyses and did not interact with inotrope therapy in prediction models. Whether inotrope therapy itself is an actual driver of PGD protection versus an epiphenomenal marker remains to be explored. There are clear differences in medical therapy, anticoagulation and mechanical support between patients receiving and not receiving inotrope therapy (Table S2) and further investigation will be necessary to determine if these differences impact PGD risk.

Integrating both proteomic and clinical variables into one model demonstrated that combinations of proteins and clinical characteristics can yield increased classification power. KLKB1 combinations resulted in the greatest classification performance. Interestingly, though inotrope therapy alone demonstrated modest prediction, its combination with KLKB1 resulted in the greatest increase in classification power when compared to combination of KLKB1 and other top performing proteins. Notably, this panel outperforms other composite scores and clinical variables such as the Radial score which demonstrated low performance all three cohorts.

Whether the proteomic results were being driven by a specific microvesicular process or a reflection of the greater overall serum milieu was tested in the validation ELISA cohort. The ELISA samples themselves were not able to generate a classifier using KLKB1 and inotrope therapy due to the paucity of PGD samples in that cohort. However, the proteomics-derived classifier generated a similar AUROC on whole serum as it did in the original microvesicle proteomic cohort. At the whole serum level, in a population whose incidence mirrored closely to national PGD rates, the classifier performed essentially as a rule out test with a very high negative predictive value.

There are several key limitations to this study. Most importantly was the limited number of samples despite amassing three institutional cohorts. We were able to increase proteomic depth of the initial 88 patient by attaining at least triplicate spectra for each sample and then validating those findings in another 65 patients. Still, it remains to be seen whether the associations and predictions discovered are representative of all heart transplant patients and future studies are required for further validation in larger cohorts. Furthermore, the clinical data collected was limited to recipient. This analysis did not take into account donor data and potential interactions between recipient proteomics and donor variables. The hypothesis of this study is that PGD risk of the recipient can be detected prior to transplant. Our goal was to ascertain whether proteomics can be treated as clinical variables that contribute to risk. This would not only be calculated as part of a transplant evaluation, but also provide an actionable basis with which to reduce that risk. This would necessarily take place prior to knowledge of the donor and their clinical characteristics.

The results of isobaric TMT labeling are thus semiquantitative and are not reflective of an absolute quantity of KLKB1 found in microvesicles. This precludes optimizing a value for KLKB1 that would maximize either PGD classification sensitivity, specificity or both. Despite this, the classifier performed well when we normalized absolute values of KLKB1 in the serum by ELISA. With only 3 cases of severe PGD in this cohort, which approximates the normal incidence of PGD, KLKB1 trended towards a significant decrease in PGD patients (p=0.051).

Looking forward towards clinical utility, PGD risk stratification would be best served in the outpatient setting as part of an overall pre-transplant evaluation. It will also be important to understand if patient risk is static or evolves and whether changes in that risk are associated with clinical status. The optimistic potential here is to use this classifier to evaluate therapies that may alter future PGD risk and improve heart transplant outcomes.