Proteomic Profiling to Identify Prognostic Biomarkers in Heart Failure

Patients and Methods

Patient population. All patients were recruited from those attending the Southampton University Hospital device service. Two different patient populations were enrolled: 1) Control patients. This group comprised of consecutive patients with a permanent pacemaker and preserved LVEF. Patients with a high percentage of right ventricular pacing (>30%), or history, signs or symptoms of heart failure, were excluded. This group were chosen as controls to avoid the introduction of potential bias related to pre-analytical factors due to sample collection and patient preparation; they attended the same hospital clinic, in the same fashion (same time of day, no specific food requirements i.e. not nil by mouth), as patients with ICDs (the LVSD group). 2) Left ventricular systolic dysfunction patients. This group comprised of patients with LVSD, and an ICD or cardiac resynchronisation defibrillator (CRT-D). All patients were on optimal medical therapy. None had heart failure admissions or therapy changes in the six weeks prior to enrolment. Based on New York Heart Association (NYHA) functional class and left ventricular ejection fraction (LVEF), two further subgroups of patients with LVSD were identified: (i) Patients with SHF, defined as LVEF ≤40% and NYHA ≥II. (ii) Patients with asymptomatic LVSD, defined as LVEF ≤40% and NYHA I.

Additional exclusion criteria for both groups were pregnancy, an acute coronary syndrome or surgery of any type within the preceding 6 weeks.

At enrolment, baseline demographic and clinical data were recorded, a 12-lead resting ECG performed, and NYHA functional class assessed. All patients had a transthoracic echocardiogram prior to study entry. Blood was drawn from a forearm vein and collected in serum separator (serum) and EDTA tubes (plasma). The collection and handling of all sera samples was applied in accordance to the recommendations of the Standard Operating Procedure Integration Working Group (SOPIWG) (10). Briefly, serum samples were allowed to clot for 30 minutes and then centrifuged at 3000 rpm for 10 minutes. Samples were then divided into aliquots and frozen within 1 hour of sampling. Samples were finally stored at −80°C prior to analysis and underwent no more than two freeze-thaw cycles. The study complied with the Declaration of Helsinki and was approved by the local research ethics committee. Written informed consent was obtained from all patients.

Study end points and follow-up. Control patients were not followed-up. LVSD patients were followed up at 3-6 months with a hospital visit or via a remote patient management system. Patients under remote follow-up also attended the hospital every 6 months. At each hospital visit the patient was clinically assessed and the device interrogated. The occurrence of any ICD therapy was recorded.

Appropriate ICD therapy was defined as: (i) Antitachycardia pacing therapy (ATP) for ventricular tachycardia (VT). (ii) Shock therapy for VT or ventricular fibrillation (VF).

Correct arrhythmia detection/discrimination was confirmed by analysis of stored electrograms by two electrophysiologists blinded to the biomarker analysis.

For the prospective part of the study two study end-points were chosen to enable exploration of the utility of the serum proteomic biomarkers in defining outcomes. These were: (i) All-cause mortality (as an approximate surrogate for non-sudden cardiac death). (ii) Patient survival with appropriate ICD therapy (as an approximate surrogate for preventable sudden cardiac death).

NT-proBNP analysis. In view of its established diagnostic and prognostic role in LVSD, N-terminal pro-brain natriuretic peptide (NT-proBNP) was measured as a comparator to proteomic biomarkers. The NT-proBNP assay was based on a two-site non-competitive assay format, using in-house antibodies, as previously described (11).

SELDI-TOF MS analysis. Serum samples were analysed on the weak cationic exchange (CM10) ProteinChip array (BioRad, California, USA), chosen on the basis of greater peak abundance relative to other surface chemistries. To aid reproducibility, a BioMek3000 (Beckman Coulter, California, USA) liquid handling robot was used. Samples were analysed in duplicate on a bioprocessor (BioRad). Samples were randomly assigned to bioprocessor wells to minimise bias. All samples were run over a 1 week period to reduce potential error due to variation in instrument performance.

Serum samples (10 μL) were denatured at pH 9 for 60 min on ice in 90 μl of U9 buffer (9 mol/L urea, 2% CHAPS, 50 mM Tris-HCl, pH 9) (Bio-Rad). Wells were pre-incubated with 5 μl of buffer (CM10 low-stringency, Bio-Rad), and then incubated twice with 200 μl of buffer at room temperature for 10 min. A 10-μl aliquot of the resulting serum mixture was further diluted in 90 μl sample buffer and applied to the wells. Samples were further incubated for 60 min at room temperature (23°C) on a Micromix5 platform shaker (Diagnostic Products Corporation, California, USA), using a form of 20 and amplitude of 7. Following incubation, unbound proteins were removed by washing with two volumes of 200 μl of each buffer, and di-ionised water. Each washing step was performed with horizontal shaking for 10 min. The ProteinChips were removed from the bio-processor and allowed to air-dry for 45 min. Two applications of 1 μl of 50% sinapinic acid in 50% acetonitrile/0.5% trifluoroacetic acid were delivered to each spot and allowed to air-dry for 15 min.

Time-of-flight (TOF) spectra were generated using an Enterprise 4011 mass spectrometer (Bio-Rad). Each spot was analysed using laser settings optimised for both low- and high-mass proteins. For the low-mass range (mass range 0-20 kDa, focus mass 10 kDa, matrix attenuation 1 kDa) spectra were generated with 17 shots per position, at laser intensities of 2000, 3000 and 4000, preceded by 2 warming shots at 2200, 3300 and 4400, respectively. For the high mass range (mass range 18-200 kDa, focus mass 20 kDa, matrix attenuation 1 kDa), TOF spectra were generated with 17 shots at a laser intensity of 4000, preceded by 2 shots at 4400.

Spectra were externally calibrated using a protein standard calibration kit composed of recombinant hirudin (6.96 kDa), equine cytochrome C (12.23 kDa), equine myoglobin (16.95 kDa) and carbonic anhydrase (29.0 kDa) (Bio-Rad). Following baseline subtraction, spectra were normalised against the total ion current at the 2-20kDa region for the low mass range and the 18-200 kDa region for the high mass range. Spectra were visually inspected and those exhibiting a poor signal-to-noise (S/N) ratio (approx. <5:1) were excluded for further processing. For each sample, one spectrum for each mass range was used, so as to minimise deviation in total ion current to within 0.4-2.5 times the mean of all patients.

The PCDM software (Biomarker Wizard version 3.1, Bio-Rad) was used to identify peaks. The low m/z range between 0 and 2000 was excluded from the analysis to avoid chemical noise interference from the UV matrix species. Peaks with a S/N ratio of ≥5 for the first pass and ≥2 for the second pass, and a valley depth greater than or equal to 3, were considered for clustering if present in ≥10% of spectra. The mass window for each cluster was 0.3% of the peak mass. To avoid analysis of artifact signals all peaks were visually inspected and relabelled as required prior to statistical analysis. The high mass range was run in parallel.

Every other ProteinChip contained a control sample taken from a single pooled mixture of 50 case and control patients to assess assay variability. Co-efficient of variation (CV) for peak intensity for these spectra, derived from the pooled sample, run 34 times (five assays) was 34%. These data were obtained by averaging values for 17 randomly selected peaks spread across the analysed mass range (2-200kDa), using the formula CV=√ ((CV₁²+CV₂²+CV₃²…+CV_n²)/n), where n represents the number of included peaks.

Statistics. Categorical variables are expressed as percentages (numbers). Normally-distributed continuous variables are expressed as mean±standard deviation. Variables not normally distributed are expressed as median (lower quartile to upper quartile) and compared using the Mann-Whitney U-test.

To identify potential protein biomarker peaks associated with LVSD, the serum proteomic profiles were compared between patients with SHF and controls. To achieve this analysis objective, patients were randomly separated into two equal-sized discovery and validation sets, each containing equal numbers of SHF and control patients. For each proteomic peak, the p value was calculated, using the Mann-Whitney U-test, to indicate its ability to differentiate SHF patients from controls. Peaks differentially expressed (p<0.05) in both sets were considered significant. For each of the significant peaks, Receiver Operating Characteristic (ROC) curves were constructed, to assess accuracy in distinguishing SHF from control patients. The Jonckheere-Terpstra test was used to assess the trend in peak intensity levels across patient groups categorised by functional class.

The 6 peaks that demonstrated an association with SHF were then further evaluated in a prospective study of ICD recipients to assess their ability to predict prognosis in patients with LVSD. Univariate Cox proportional hazards analyses were used to investigate biomarker predictors of the two prospective end-points (all-cause mortality and survival with appropriate ICD therapy). For the end-point of all-cause mortality, in view of the small number of patients reaching the end-point (n=11), multivariable analysis was not performed. As NT-proBNP was not normally-distributed and its normality improved with log transformation the log transformed values were used for analysis. The proportional hazards assumption was checked using Schoenfeld residuals (12).

For the end-point of all-cause mortality, for selected biomarker peaks Kaplan-Meier survival analysis was performed to compare survival between patient groups stratified according to ROC curve-derived cut-off points. Survival between groups was compared using the log-rank test.

Statistical analyses were performed on SPSS Version 17 (SPSS Inc., Chicago, IL, USA). In all analyses p<0.05 was considered significant.