A Precision Medicine Approach to Stress Testing Using Metabolomics and Microribonucleic Acids

Background Acute coronary syndrome (ACS) is a growing global health problem, and precision medicine techniques hold promise for the development of diagnostic indicators of ACS. In this pilot, we sought to assess the utility of an integrated analysis of metabolomic and microRNA data in peripheral blood to distinguish patients with abnormal cardiac stress testing from matched controls. Methods We used prospectively collected samples from emergency department (ED) patients placed in an ED-based observation unit who underwent stress testing for ACS. We isolated microRNA and quantified metabolites from plasma collected before and after stress testing in patients with myocardial ischemia on stress testing versus those with normal stress tests. The combined metabolomic and microRNA data were analyzed jointly for case (ischemia) and 1:1 matched control patients in a supervised, dimension-reducing discriminant analysis. Two integrative models were implemented: a baseline model utilizing data collected prior to stress-testing (T0) and a stress-delta model, which included the difference between post-stress test (T1) and pre-stress test (T0). Results Seven case patients with myocardial ischemia on ED cardiac stress testing (6 females, 85% Caucasian, mean Thrombolysis In Myocardial Infarction Score=3, 4 patients ultimately received percutaneous coronary intervention) were 1:1 age and sex-matched to controls. Several metabolites and microRNAs were differentially expressed between cases and controls. Integrative analysis of the baseline levels of metabolites and microRNA expression showed modest performance for distinguishing cases from controls with an overall error rate of 0.143. The stress-delta model showed worse performance for distinguishing cases from controls, with an overall error rate of 0.500. Conclusions Given our small sample size, results are hypothesis-generating. However, this pilot study shows a potential method for a precision medicine approach to cardiac stress testing in patients undergoing workup for ACS.


86
Metabolomics is another promising modality for the characterization of function in high-87 energy utilization organs such as the heart. Such analyses examine a wide range of 88 fundamental biological molecules, many of which are connected to underlying metabolic 89 processes in the body such as fatty acids and oxidation products. The concentration of these 90 molecules can also rapidly change in response to acute disease states. It has been shown that 91 certain amino acids and acylcarnitines levels in peripheral blood are associated with long-term 92 risk of cardiovascular disease, particularly coronary related(16-18). We previously reported the 93 analysis of stress-induced changes in selected metabolites including amino acids and 94 acylcarnitines (19).

96
In current clinical practice, the information gathered from stress testing is usually 97 reduced down to a single data dimension to simplify decision-making. However, it is widely 98 recognized that a precision medicine strategy for chest pain evaluation will require expanding 99 the number of biomarkers (whether blood-based, imaging, or in other forms) and to integrate 100 information to provide a more accurate answer. The challenge associated with expanding the 101 number of biomarkers is that very large datasets can be problematic for biostatistical analysis.
102 However, several dimension-reducing biostatistical approaches now allow the integration of 103 large datasets from different categories of molecules(20).

105
In this paper, we used transcriptomic and metabolomic approaches to demonstrate the 106 feasibility of a biomarker-based stress test using precision medicine techniques. We also sought 107 to develop the technical capabilities and protocols to study serially measured microRNAs and 108 metabolites in patients undergoing cardiac stress testing for symptoms of ACS. We believe this 109 novel model of stress testing represents an exciting opportunity to apply a precision medicine 110 approach to cardiac disease diagnosis and prognosis. We conducted a pilot study to determine whether serial microRNA and metabolomic 115 data could be combined to enhance the diagnostic performance of cardiac stress testing. We 116 used peripheral blood samples in EDTA collection tubes from a biorepository created to study 117 changes in high-sensitivity troponin and B-type natriuretic peptide during stress testing. This 118 biorepository has been previously described (21,22). Briefly, samples were collected from adult 119 emergency department (ED) patients who had symptoms of ACS and who underwent stress 120 testing in our observation unit. All patients, as a condition of enrollment, underwent standard 121 symptom-limited Bruce Protocol exercise echocardiogram tests as part of their usual care.
122 These tests reported the presence or absence of inducible myocardial ischemia, defined as 123 stress-induced regional wall motion abnormality in at least one segment. All tests were 124 interpreted by board-certified cardiologists who were blinded to any biomarker data. Two 125 reviewers independently confirmed the accuracy of the reports for this study.  SmRNA-seq data were processed using the Trim Galore toolkit(24), which employs 157 Cutadapt(25) to trim low-quality bases and Illumina sequencing adapters from the 3' end of the 158 reads. Only reads that were 18-28 nucleotides in length after trimming were kept for further 159 analysis. Reads were mapped to the hg19 version of the human genome using the Bowtie 160 alignment tool(26). Reads were kept for subsequent analysis if they mapped to no more than 13 161 genomic locations. Gene counts were compiled using custom scripts that compare mapped 162 read coordinates to the miRbase microRNA database(27). Reads that match the coordinates of 163 the known mature microRNAs were kept if they perfectly matched the coordinates of the miRNA 164 seed while not varying by more than 2 nucleotides on the 3' end of the mature miRNA. Only 165 mature miRNAs that had at least 10 reads in any given sample were used in subsequent 166 analysis. Normalization was performed using the DESeq2 Bioconductor package from the R 167 statistical programming environment applying the 'poscounts' approach to eliminate systematic 168 differences across the samples(28). The normalized data were log-transformed and differential 8 169 expression was tested using linear regression. For the stress-delta model, we employed a 170 mixed-effects model with the patient ID as a random effect. The false discovery rate was used 171 to adjust for multiple hypothesis testing.

172
173 Targeted metabolite data were log-transformed prior to analysis and a PCA was conducted to 174 assess for the presence of outliers and confounding demographic factors. MicroRNA-Seq count 175 data were also log-transformed. For the integrative analysis, microRNAs that were missing in 176 half or more of the samples were removed from the data set. All integrative analyses were 177 conducted with baseline ("T0", pre-stress test) and delta ("T1" -"T0") data sets, where T1 178 corresponds to post-stress test samples.

180 Regularized Canonical Correlation Analysis
181 Regularized canonical correlation analysis (rCCA) seeks to extract latent variables that 182 maximize the correlation between the two data sets, but with an additional regularization step 183 that reduces the number of variables contributing to each component. An initial leave-one-out 184 cross-validation step can be performed to select the regularization parameters for each data set.
185 To explore correlation between the metabolomic and microRNA baseline and delta datasets, a 233 These microRNAs are listed in S2 File. We constructed heat maps (Figs 1 and 2)   259 an overall error rate of 0.143 (Table 3). Using stress-delta data actually led to a worse error 260 (0.500) for distinguishing cases from controls.  294 which combines exercise tolerance information and electrocardiogram characteristics to make a 295 prediction of future risk. However, these multi-modal stress tests do not take advantage of the 296 large amounts of data that we are currently capable of collecting from patients' blood samples.

262
14 297 Although our current modalities of stress testing are sensitive for obstructive coronary disease, 298 their accuracy can be further improved, particularly for identifying specific high-risk phenotypes 299 that benefit from emerging therapies.

301
In contrast to the blood-based biomarker approach for evaluation of acute myocardial 302 infarction, currently assessment for myocardial ischemia and/or obstructive coronary artery is 303 heavily imaging-dependent. Use of serial biomarkers is not routine practice for assessing 304 myocardial ischemia, especially in the context of a stress test. Thus, this study presents a novel 305 paradigm for assessing patients for myocardial ischemia. Our current ability to serially measure 306 multiple blood-based molecules presents an opportunity to develop more sophisticated multi-307 modal stress tests that incorporate large amounts of data.

309
We have previously examined the utility of blood-based biomarkers to enhance cardiac 310 stress testing (19,21,22). In this pilot study, we outline the methodology to further develop a 311 biomarker-augmented stress test using a precision medicine approach. First, we identified 312 several differences at baseline, post-stress, and stress-delta between cases and controls, 313 largely as a consequence of the large number of analytes we assessed. While a great deal of 314 prior literature has examined baseline (resting) biomarkers for prediction of coronary heart 315 disease, stress-delta biomarker assessments give us the ability to assess acute changes in 316 response to a controlled ischemic event, with the benefit of within-patient control for baseline 317 values. We were able to assess a large number of potential biomarkers in each blood sample, 318 creating the possibility of a systems biology approach to biomarker discovery.

366
The dataset(s) supporting the conclusions of this article is(are) included within the article 367 (and its additional file(s)).