Machine Learning Identifies Metabolic Signatures that Predict the Risk of Recurrent Angina in Remitted Patients after Percutaneous Coronary Intervention: A Multicenter Prospective Cohort Study

Abstract Recurrent angina (RA) after percutaneous coronary intervention (PCI) has few known risk factors, hampering the identification of high‐risk populations. In this multicenter study, plasma samples are collected from patients with stable angina after PCI, and these patients are followed‐up for 9 months for angina recurrence. Broad‐spectrum metabolomic profiling with LC‐MS/MS followed by multiple machine learning algorithms is conducted to identify the metabolic signatures associated with future risk of angina recurrence in two large cohorts (n = 750 for discovery set, and n = 775 for additional independent discovery cohort). The metabolic predictors are further validated in a third cohort from another center (n = 130) using a clinically‐sound quantitative approach. Compared to angina‐free patients, the remitted patients with future RA demonstrates a unique chemical endophenotype dominated by abnormalities in chemical communication across lipid membranes and mitochondrial function. A novel multi‐metabolite predictive model constructed from these latent signatures can stratify remitted patients at high‐risk for angina recurrence with over 89% accuracy, sensitivity, and specificity across three independent cohorts. Our findings revealed reproducible plasma metabolic signatures to predict patients with a latent future risk of RA during post‐PCI remission, allowing them to be treated in advance before an event.


Study population
The inclusion criteria were hospitalized CAD patients with stable angina. All patients had at least one lesion, and each patient received one second-generation rapamycin-eluting stent (Firebird 2, Microport, Shanghai, China) ( Figure S1). In addition to percutaneous coronary intervention (PCI), patients were given optimal medical therapy (OMT). The symptom of angina was dramatically alleviated before the patients were discharged from the hospital. The exclusion criteria were as follows: patient with the lesion of > 90% stenosis or total occlusion, active inflammation (High sensitive C-reactive protein [hsCRP] >1.0 mg/dl), renal disease (serum creatinine > 3.0 mg/dl or ongoing hemodialysis), the onset of cancer in the previous five years, coronary artery bypass grafting, contraindications to iodine media, and severe hepatic insufficiency.
Finally, three cohorts were obtained in this study. Total 750 patients in Beijing Anzhen Hospital were enrolled between December 2015 and May 2016 as the discovery cohort (cohort 1) and another 775 patients from February 2017 to August 2017 for additional independent discovery cohort (cohort 2). The external validation cohort (cohort 3) was recruited in Qufu Hospital, Shandong Province from January 2018 through June 2018. The written informed consent was obtained from all subjects. The patient characteristics are listed in Table 1 (Discovery cohort), Table S1 (Additional discovery cohort), and Table S2 (External validation   cohort).

Blood collection and follow-up for angina recurrence
After an overnight fast, venous blood was collected from each patient at 48 h after PCI. Plasma was separated by centrifugation at 900 g for 10 minutes at room temperature within one hour of collection. The resulting fresh lithium-heparin plasma was transferred to new tubes and stored at -80 ˚C for further analysis.
The patients were discharged after the blood draw. They were followed up every 30 days for angina recurrence up to 270 days (9 months). Seattle angina questionnaire-7 (SAQ-7) is a validated, self-administered, disease-specific measure for patients with CAD to evaluate the severity of angina. The SAQ-7 was conducted at each follow-up for all patients.

Extraction of metabolites from plasma for metabolomic analysis
Ninety-five (95) µL of plasma was thawed, mixed with 5 µL of custom-synthesized stable isotope standards and extracted with four volumes (400 µL) of prechilled (-20 ˚C) extraction buffer containing acetonitrile and methanol (50: 50, v/v). The mixture was incubated on crushed ice for 10 min and then centrifuged at 16,000 g for 10 min at 4 ˚C. The supernatant containing the extracted metabolites was transferred to labeled cryotubes and stored at -80 ˚C for metabolomic analysis.

Metabolomic analysis
Metabolomic analysis was performed on a Shimadzu LC-20A ultra-high-performance liquid chromatography coupled with SCIEX QTRAP 6500 triple quadrupole mass spectrometer (LC-MS/MS) using a broad-spectrum targeted metabolomic approach. A total of 606 metabolites covering the major metabolic pathways were targeted by scheduled multiple reaction monitoring (sMRM) under Analyst v1.6.2 software control in both negative and positive mode with rapid polarity switching (20 ms). Nitrogen was used for curtain gas (set to 30), collision gas (set to high), ion source gas 1 and 2 (set to 35). The source temperature was set to 500 ˚C. Spray voltage was set to -4500 V for negative mode and 5500 V for positive mode. Compound-dependent MRM parameters were optimized using the purified standards. Ten µL of the extract was injected into a 250 mm × 2.0 mm, 4 µm polymer-based NH 2 HPLC column (Asahipak NH 2 P-40 2E, Showa Denko America, Inc., NY) held at 25 °C for chromatographic separation. The mobile phase was solvent A: 95% water, 5% acetonitrile with 20 mM NH 4 OAC, and 20 mM NH 4 OH (pH 9.4); solvent B: 100% acetonitrile. Separation was achieved using the following gradient: 0- The pooled plasma samples were used as quality control (QC) and injected four times in each batch. Metabolites with the inter-day coefficient of variations (CVs) > 25% in QC samples were excluded.

External validation using the patients from another hospital and a different analytical approach
To validate the selected metabolite biomarkers for the prediction of future angina recurrence, we quantified these six metabolites in the external validation cohort (cohort 3) plasma samples by and Cer(d18:1/18:1)-13 C 18 . The extraction method was the same as described in the metabolomic analysis section.
Quantitative analysis of the target compounds was performed on a Shimadzu LC-20A UHPLC system coupling with a QTRAP 6500 triple quadrupole mass spectrometer (SCIEX, USA). Ten μL of the extract was loaded onto ACQUITY UPLC BEH C 18 column (1.7 μm particle size, 2.1 × 150 mm, Waters) held at 35 ºC for separation. Solvent A consisted of water/IPA/methanol (90:5:5, v/v/v) with 20 mM NH 4 OAC and solvent B was isopropanol/ methanol (50:50, v/v/) with 20 mM NH 4 OAC. The flow rate was 0.3 mL/min. The analytes were separated using linear gradient elution to 100% B over 25 min and isocratic 100% B for 4 min. The chromatographic system was returned to the initial conditions in 1 min, followed by a 2-min equilibration before the subsequent injection.
MS/MS analysis was performed in both negative and positive model with rapid polarity switch (3 ms). The ion spray voltage was -4500 V for negative mode and 5500 V for positive mode. The source temperature was set at 500 ºC. The MRM transitions were optimized using the purified standards, and two transitions were developed for each analyte.
The matrix-specific calibration curves for the targeted analytes were generated by plotting the peak area ratios (analyte/internal standard) versus the concentration ratios (analyte/internal standard) and fitting to linear regression.

Metabolomic method reproducibility
Out of 606 targeted metabolites, 458 metabolites were detected in all the samples without missing values. Forty-eight (48) metabolites with inter-day CVs >25% were excluded, and 407 metabolites were finally included in the analysis. The representative chromatogram is shown in Figure S2. The intra-day and inter-day batch correlation r values were 0.995 and 0.981, respectively, for theses metabolites in QC samples. The median intra-day and inter-day coefficient of variation (CVs) were 10.2% and 11.5%, respectively, for QC samples (Table S3).
These metrics suggested excellent reproducibility for the included metabolites.

LC-MS/MS
The optimal compound-dependent parameters for LC-MS/MS analysis of six metabolic predictors are described in Table S10. The representative chromatogram is shown in Figure S9.
All six metabolites were chromatographically separated with the optimized gradient on an ACQUITY UPLC BEH C 18 column. Figure S10 shows the standard curves and linear ranges for the quantification of these metabolites. The method is reproducible with the median intra-day CV of 2.14% and inter-day CV of 6.73%, which are below the acceptance limit (20% CV) of FDA's guideline for bioanalytical method validation using LC-MS. Notes: Values are mean ± SD, n (%), or median (interquartile range). Differences between recurrent angina and angina-free groups were analyzed using the Student's t-test (parametric distribution), Mann-Whitney U test (nonparametric distribution), or two-proportion z-test (Categorical or proportional data). The effect size between two groups was calculated by Hedge's statistic. Notes: Values are mean ± SD, n (%), or median (interquartile range). Differences between angina and angina-free groups were analyzed using Student's t-test (parametric distribution), Mann-Whitney U test (nonparametric distribution), or two-proportion ztest (Categorical or proportional data). The effect size between the two groups was calculated by Hedge's statistic.         A multi-biomarker model for predicting future angina recurrence was built using six selected metabolites by multivariate logistic regression in the external validation cohort (n = 130). The model was adjusted for baseline age, BMI, hsCRP and LDL.