Xiang
Ren‡
a,
Ruyu
Zhou
b,
George
Ronan
a,
S. Gulberk
Ozcebe
a,
Jiaying
Ji
a,
Satyajyoti
Senapati
c,
Keith L.
March
d,
Eileen
Handberg
d,
David
Anderson
d,
Carl J.
Pepine
d,
Hsueh-Chia
Chang
c,
Fang
Liu
b and
Pinar
Zorlutuna
*a
aDepartment of Aerospace and Mechanical Engineering, University of Notre Dame, Notre Dame, IN 46556, USA. E-mail: pinar.zorlutuna.1@nd.edu; Tel: +1 (574)631 8543
bDepartment of Applied and Computational Mathematics and Statistics, University of Notre Dame, Notre Dame, IN 46556, USA
cDepartment of Chemical and Biomolecular Engineering, University of Notre Dame, Notre Dame, IN 46556, USA
dDivision of Cardiology, Department of Medicine in the College of Medicine, University of Florida, Gainesville, FL 32611, USA
First published on 4th October 2024
Rapid diagnosis of acute myocardial infarction (AMI) is crucial for optimal patient management. Accurate diagnosis and time of onset of an acute event can influence treatment plans, such as percutaneous coronary intervention (PCI). PCI is most beneficial within 3 hours of AMI onset. MicroRNAs (miRNAs) are promising biomarkers, with potential of early AMI diagnosis, since they are released before cell death and subsequent release of larger molecules [e.g., cardiac troponins (cTn)], and have greater sensitivity and stability in plasma versus cTn regardless of timing of AMI onset. However, miRNA-based AMI diagnosis can result in false positives due to miRNA content overlap between AMI and stable coronary artery disease (CAD). Accordingly, we explored the possibility of using a miRNA profile, rather than a single miRNA, to distinguish between CAD and AMI, as well as different stages following AMI onset. First we screened a library of 800 miRNA using plasma samples from 4 patient cohorts; no known CAD, CAD, ST-segment elevation myocardial infarction (STEMI) and STEMI followed by PCI, using Nanostring miRNA profiling technology. From this screening, based on machine learning SCAD and Lasso algorithms, we identified 9 biomarkers (miR-200b, miR-543, miR-331, miR-3605, miR-301a, miR-18a, miR-423, miR-142, and miR-132) that were differentially expressed in CAD, STEMI and STEMI-PCI and explored them to identify a miRNA profile for rapid and accurate AMI diagnosis. These 9 miRNAs were selected as the most frequently identified targets by SCAD and Lasso, as indicated in the “drum-plot” model in the machine learning approach. We used age-matched patient samples to validate selected 9 miRNA biomarkers using a multiplexed ion-exchange membrane-based miRNA sensor platform, which measures specific miRNAs, and cTn as a control, simultaneously as a point-of-care device. Findings from this study will inform timely and accurate diagnosis of AMI and its stages, which are essential for effective management and optimal patient outcomes.
MicroRNAs (miRNAs) are small (17–22 nucleotides), non-coding RNAs that regulate gene expression post-transcriptionally21 found in various body fluids, including whole blood and plasma. They have recently emerged as promising tools involved in many pathophysiological conditions including cardiovascular diseases as they are produced by cells deliberately under pathophysiological conditions as a first-response as opposed to cTns that emerge because of cell death. Hence, miRNAs have been explored as early detection biomarkers for various diseases. Studies have reported that miRNAs are dysregulated in CAD and circulating miRNA profiles can serve as potential biomarkers for prognosis and rapid and accurate diagnosis of AMI and CAD.22,23 As miRNA turnover is much faster than proteins, information gathered from miRNAs can be used for early detection and severity assessment of the acute phase without need for additional testing. However, since the AMI often times develops as a result of chronic CAD, miRNA-based AMI diagnosis can result in false positives due to miRNA overlap between AMI and CAD. As such, current miRNA-based approaches have limitations in distinguishing between chronic stable CAD and AMI and have been deemed so far as clinically not reliable.
Most studies exploring miRNA expression in CAD patients use miRNAs isolated from cell-free plasma for convenience and availability of the samples.24,25 A qualitative study of human and rat serum and plasma samples highlights the preference for using plasma, especially in translational miRNA studies, due to higher aligned reads in plasma than in serum.26 In contrast, another study comparing miRNA content of plasma and whole blood samples reported that only a few miRNAs were differentially expressed in both sources, while most of the information was lost in plasma samples.27 Therefore, it is important to consider the source of miRNAs and the method used for the detection of the miRNA, especially plasma-derived exosomal miRNA, as the source and the methods can influence the results obtained. Recently, the Nanostring nCounter platform® has been introduced for the detection, quantification, and assessment of miRNA expression. This platform offers the advantage of direct screening of more than 800 miRNAs in patient samples without potential biases that might arise from RNA amplification, however, it is expensive and the sample preparation process for Nanostring requires lysing the exosome, which can cause miRNA loss due to additional freeze–thaw cycles. Additionally, traditional data analysis methods fall short to process the large datasets generated by Nanostring, especially when screening more than 800 miRNAs, which translates into over 800 dimensions. While both principal component analysis (PCA) and partial least-squares discriminant analysis (PLS-DA) can reduce dimensions in a limited space, they are not suitable for analyzing over 800 parameters with a limited sample size (n = 6 for each category). Therefore, we implemented smoothly clipped absolute deviation (SCAD) and least absolute shrinkage and selection operator (Lasso) methods to analyze bulky Nanostring miRNA expression data of the 4 different cohorts (NCAD, CAD, STEMI-pre and STEMI-PCI) have here.
Mainly due to the low efficiency of extraction and increased error margin from extensive sample preparation and processing miRNAs are at low in concentration. In addition, miRNA isolation from exosomes traditionally involves chemical lysing, causing further potential miRNA degradation or loss. Moreover, traditional protein (i.e., cTn) sensing methods use western blots or enzyme linked immunosorbent assay (ELISA) and have high uncertainties in sample preparations, antibody processing, and testing protocols. We previously demonstrated that our multiplexed ion-exchange membrane-based miRNA (MIX.miR) sensor platform can accurately measure miRNA concentrations at physiologically relevant picomolar (pM) levels.28 Increased detection sensitivity enabled MIX.miR to overcome the main limitation of current methods (i.e., RT-qPCR), and we identified miRNAs in plasma that had been previously reported for only whole blood samples from patients. Having used this membrane for other proteins before,29 we adapted it for cTn detection as an internal control. The MIX.miR has three main advantages over traditional methods, increased detection sensitivity, integrated chemical-free exosome lysing unit using surface acoustic waves (SAW), and integrated ion-exchange membrane (IEM) at a relatively low cost. While traditional biological methods of testing miRNAs require extensive work with sample preparation and miRNA duplication, resulting in miRNA loss and long assay time detrimental to this application, the MIX.miR sensor platform can accurately detect miRNAs with a concentration level of 1 pM within one hour. Additionally, traditional lysing involves mechanical or chemical interventions on the exosomes, which often interfere with miRNAs and potentially cause miRNA degradation or loss. With the SAW lysing unit relying on physical forces, MIX.miR breaks the exosomes and extracellular vesicles without introducing chemical impurities.28,30,31 Furthermore, the sensor constructed with IEM can easily be functionalized into protein sensors for untreated blood.29,32,33 The IEM-based sensor can utilize the principle of ELISA at a much lower cost with better detection sensitivity.29
In the current study, we used the MIX.miR platform to determine the differential expression of miRNAs in patients with CAD, STEMI, and reference patients without CAD. We then analyzed the data with SCAD and Lasso methods to assess the predictive value of miRNAs for CAD and STEMI. We hypothesized that miRNA profiles would represent the stage of STEMI and evaluate the efficacy of PCI better than the current reference standards. Point-of-care miRNA and high-sensitivity cTn testing platforms have potential to be within emergency medical services, such as ambulances, to facilitate efficient and rapid diagnosis of AMI at the pre-hospital stage.
We employed the LASSO (least absolute shrinkage and selector operation)34 and the SCAD (smoothly clipped absolute deviation),35 to regularize the variable selection (technical details on LASSO and SCAD are provided on the ESI†). We implemented the leave one out cross validation (LOOCV) to tune the penalty parameter for LASSO and SCAD. In addition to running the above logistic regression models with LASSO and SCAD penalties on the full dataset with 809 miRNAs, we also run the models on three subsets of the 809 biomarkers after manual screening based on domain knowledge. The first subset contains purely cardiac specific miRNAs (450); the second subset contains cardiac specific and vasculature miRNAs (565); the third group contains cardiac specific, vasculature, cerebral vasculature, and rare cardiac disease miRNAs (576). Among 11 logistic regression models, no biomarker was selected in 3 regression models regardless of the penalty term or the tuning parameters, which are NCAD vs. MI-Pre, CAD vs. MI-PCI, and MI-Pre vs. MI-PCI. Taken together, the mRNA selection results were obtained from 8 regression models with LASSO or SCAD penalty, respectively, among the full set of biomarkers and 3 subsets, leading to a total of 64 regressions models. Nine frequently selected miRNAs biomarkers among the 64 regression models, taken together with domain knowledge, were chosen to move onto the next stage, which is validation in the lab.
The analysis was performed in R. For LASSO regression, R package “glmnet v4.1.4” was used; for SCAD, R package “ncvreg v3.13.0” was employed.
The LASSO (least absolute shrinkage and selector operation) and the SCAD (smoothly clipped absolute deviation) are popular choices of P(β). Specifically,
The tuning parameter λ controls the model complexity and the tradeoff between bias and variance in parameter estimation and prediction. When λ = 0, lp(β0, β; X, Y) reduces to l(β0, β; X, Y), so are the parameter estimates. When λ → ∞, all β estimates would be 0 (no variable is selected). When λ is in between 0 and ∞, some β estimates will be 0 and some are not (the larger λ, the more β estimates will be 0). The miRNAs with non-zero β estimates are “selected” and deemed relevant in the prediction of Y. In practice, the value of λ is commonly chosen by a cross validation (CV) procedure – λ that yields the minimum mean CV error or the CV error within one standard error of the minimum. We implemented the leave one out CV (LOOCV) procedure due to the small sample sizes.
The functionalization of miRNAs on MIX.miR sensors is described in our previous work.28 Briefly, the MIX.miR sensor was fabricated using the standard replica molding of silicone rubber and polyurethane (PU) molding. The ssDNA of specific miRNA is added onto the AEM after 3,3′,4,4′-benzophenonetetracarboxylic acid (BPDA) and EDAC carboxylation procedures. The antibody for cTnT is added onto the AEM after 0.4 M EDAC and 0.4 M sulfo-NHS (EDAC/NHS) treatment. The silica beads are 50 nm in diameter with detection antibody after EDAC/NHS treatment. To achieve better attachment of the antibody, beads were centrifuged, sonicated (Elmasonic S30h, Elma Schmidbauer GmbH, Singen, Germany) and re-suspend multiple times.32
The current–voltage curve (CVC) of the sensors can reveal the concentrations of the targeted biomarkers, either the specific miRNA or the cTnT. Due to the ion-depleting action of the membrane on the side of the functionalized oligonucleotide, the conductivity near that surface membrane is 3 orders of magnitude lower than the bulk or within the charged membrane. The surface layer hence controls the voltage drop and the additional surface charge of the hybridized duplex can sensitively gate the ion current, producing a voltage signal much larger than those from electrochemical sensors. The CVC voltage shifts were correlated with miRNA concentration through calibration curves.
The detection board can detect the miRNA and the cTnT at the same time. The sample was introduced from the inlet after SAW lysing.30 The exosomal miRNAs were released to the plasma and flowed through the channel with the MIX.miR sensor. The preconcentration unit was balanced with the flow and miRNAs were kept around the MIX.miR sensing area, enhancing the chance of specific miRNAs attaching to the ssDNA. The remaining sample passed through the channel and tubing toward the cTnT sensor on the secondary device. The sample with cTnT was incubated in the cTnT sensing device for 20 min, followed by another 20 min incubation with silica beads with detection antibody. After additional high (4× PBS) and low ionic (2× PBS) wash, the CVC of miRNA and cTnT were obtained separately. We added a microfluidic mixing unit, a customized Tesla valve, between the miRNA sensing device and the cTnT sensing device, such that it has one inlet of 4× PBS and another inlet of DI water with the same flow rate. The 4× PBS was diluted into 2× PBS at the outlet of the Tesla valve, as illustrated in the COMSOL® simulation in Fig. 1. The cTnT in the plasma sample was detected via the antibody bound to silica beads in the protein sensing unit (Fig. 1(3)).30 Once measured, the voltage shift was used to calculate the miRNA concentration detected in the plasma sample.
![]() | ||
Fig. 2 (a) Drum plot on miRNA biomarkers selected from penalized regression models. Each node at the bottom of the drum represents a model and each node at the top of the drum represents a miRNA biomarker selected at least once by the models. A connected line between a top node and a bottom represents the former is selected by the latter. An orange-colored line represents a positive coefficient, and a blue-colored line represents a negative coefficient in the regression. The thicker the line is and the more intense its color is, the larger the magnitude of a coefficient is. The solid nodes at the top of the drum with the red large-font biomarker names are the biomarkers validated subsequently in lab experiments. These biomarkers are selected at least 10 times and most, if not all, are associated with positive coefficients. (b) miRNA biomarker selection frequency heat-map by penalized regression. The mRNAs that moved onto the next stage of lab validation are annotated. (c) Selection frequency of the miRNA biomarker was selected at least 6 times by penalized regression. The mRNAs that moved onto the next stage of lab validation are annotated are those associated with the bar graphs. (d–g) The miRNA detection voltage shifts by MIX.miR sensors on random samples of NCAD (blue), CAD (green), STEMI-pre (red), and STEMI-PCI (magenta), respectively. The inner colored line is the average value; while the outer colored line is the max value; the light blue lines in the center labels the limit of detection (LOD). The LOD is calculated from the calibration curves of each miRNA with linear fittings (Fig. 5). (h) The scatter plot of NCAD, CAD, STEMI-pre and STEMI-PCI random samples with different clustering parameters and objectives; (i) the ROC of different clusters based on the scatter plots with different parameters in (h). |
The association between the selected miRNA and the patient condition such as NCAD, CAD, STEMI-pre, and STEMI-PCI, was further validated with the voltage shifts by MIX.miR sensors (Fig. 2d–g). The studies in miRNA selection by SCAD and Lasso suggested that multiple miRNAs are not completely independent. Fig. 2h presented four different selections of the miRNA combinations in differentiation NCAD, CAD, STEMI-pre, and STEMI-PCI. Based on the scatter plots, we generated the ROC of each combination. The ROC indicated that the 9 miRNA combinations has promising outcomes in differentiating the four groups, which almost reaches 1 in ROC. Some of the ROCs have curves with AUC < 0.5 because not all of the miRNA markers are significant in differentiating among the four groups. However, if consider the 9 miRNA combinations as a spectrum, we are able to differentiate a specific group from NCAD, CAD, STEMI-pre, and STEMI-PCI. In order to validate the selected miRNAs and the robustness, we selected the Matched Samples with age between 60–70 years old, without diabetes, and AMI onset hours below 5–6 hours (for STEMI patient samples) and repeated the detection with MIX.miR sensors.
To exclude the effect of diabetes on STEMI and CAD patients, the testing groups were selected as 55–70 years old patients without diabetes. Additionally, as the best time window of PCI is within 3–5 h of AMI onset, the STEMI patients were selected with the onset time within 5 hours. The MIX.miR sensor measurement results are presented in Fig. 4 and 6. Fig. 5 illustrates the calibration curves of the 9 miRNA biomarkers. The red colored lines are the detection of limit for each miRNA. The CVC voltage shifts were correlated with miRNA concentration through calibration curves. The correlations in the linear region of the Langmuir isotherm are described by
![]() | ||
Fig. 5 Calibration curves of new miRNA markers and the concentration results of random samples and matched samples of NCAD, CAD, STEMI-pre, and STEMI-PCI. The red regions are the LOD for each miRNA biomarker and also indicated in the radar maps in Fig. 3d–g and 4a–d. |
Fig. 6 summarizes the 9 miRNA concentrations in both random samples and matched samples. The dark read labels the limit of detection (LOD) in each miRNA. The LOD indicates the limit of the MIX.miR sensor in detecting different specific sequences of the miRNA biomarkers. The LODs are calculated from the calibration curves in Fig. 5. According to the Langmuir isotherm above, the LOD of each miRNA shows minor differences.
The MIX.miR sensor platform captured different miRNA profiles of NCAD, CAD, STEMI-pre and STEMI-PCI samples, supported by the current literature on the selected miRNAs as biomarkers. The CAD patients had high levels of miR-3605 and miR-301a (Fig. 4b), which is in agreement with a previous study that reported differentially expressed miR3605-3p and miR301a-3p in whole blood samples of CAD patients with recurrent MI events.27 STEMI-pre patients had high levels of miR-543, 301a and moderate levels of miR-331, 142, and 200b (Fig. 4c). miR-543, miR-301a-3p and miR-200b have been previously associated with CAD presentations including AMI.36–38 Additionally, miR-331 was reported to be a significant STEMI biomarker detected before myocardial necrosis markers (i.e., cTnI, miR-208 and -499)39 and miR-142-3p elevation was suggested as a potential biomarker for detection and diagnosis of STEMI.40 The elevated miR-543 and 301a in STEMI samples dropped after PCI, miR-543 levels even reached the control NCAD levels (Fig. 4d). miR-543 downregulation was reported to mitigate inflammatory response and MI-induced cardiomyocyte apoptosis via SIRT-1,41–43 reflecting beneficial effects of PCI. The miR-142 and 200b levels in STEMI patients were sustained in STEMI-PCI while miR-423, 18a, 331 and 132 levels increased moderately. miR-18a was previously reported in AMI, where its downregulation promoted autophagy of cardiac cells via inactivating Akt/mTOR axis.44 Elevated miR-18a in STEMI-PCI samples might indicate suppressed cell apoptosis and potentially improved matrix remodeling, inflammation inhibition and as previously reported in other model systems (i.e., osteoarthritis).45 It was reported that plasma levels of miR-423-5p in AMI patients increased prior to PCI and then returned to normal within 6 h after PCI.46 Additionally, contrasting our STEMI-PCI results, miR-142-3p was previously found to be decreased in plasma of STEMI patients undergoing PCI with no-reflow.47 Such differences in miRNA profile were expected due to the fast miRNA turnover and varying sample timing. Overall, it is clear that miRNA profile rather than a single miRNA level is more informative and MIX.miR can rapidly and accurately measure any miRNA combination corresponding to the disease condition of interest.
Cardiac troponin (cTnT or cTnI) assays have been recognized for improving diagnostic accuracy in early detection of AMI. The normal ranges for cTnT is 0.02–0.13 μg L−1, and >0.2 μg L−1 is considered an indication of MI. Elevated cTnT levels can be detected as early as 2–4 h after onset of AMI symptoms, and peaking at 12–48 h, and remaining elevated for 4–10 days.10 Dynamic changes of cTnT provide information about the onset time and severity of MI, the effectiveness of PCI, and the possibility of reperfusion injury. Despite these benefits, early studies showed that the slow increase and late peak of cTnT levels can lead to missed diagnoses of early-evolving AMI.52,53 Therefore, cardiac troponin elevation without a secondary diagnostic measure (CT imaging and ECG) is insufficient for complete diagnosis.54 Here, we showed that miRNA measurement together with cTnT provides a multi-dimensional approach for MI diagnosis. The integration of miRNA and troponin testing board provides an ability of rapid “point-of-care” MI diagnosis.
Our measurement of cTnT utilizes the principle of ELISA. The capture antibody is attached on the AEM, and the detection antibody is attached on the silica beads. Commercially available ELISA kits are build-in 48- or 96-well plates with pre-coated antibody. However, our AEM sensor-based cTnT measurement can tune the quantity of testing ports with as many sensors as needed for the patient samples. In addition, the turn-over time of our measurement (t = 40 min) is much less than the time in ELISA (t = 3 h).
A major benefit of miRNA profiling is that cTnT are released into plasma following cardiomyocyte injury, while miRNAs are released into the plasma at earlier time points by exosomes. Studies have showed that miRNA reached the peak expression 3–12 h earlier than conventional biomarkers (i.e., cTnI, CKMB) in the early phases of AMI,55 thus being more sensitive. Here we tested the 9 cardiac-associated miRNAs, miR-200b, miR-543, miR-331, miR-3605, miR-301a, miR-18a, miR-423, miR-142, and miR-132 for STEMI and CAD samples. As demonstrated in the radar map and concentration results, miR-301a is elevated in CAD and STEMI, where miR-301a in STEMI might inherited from CAD conditions. miR-543 is uniquely elevated in STEMI-pre, and miR-200b, miR-331, miR-423, and miR-18a together with a gradual increase indicate the reperfusion injury or STEMI-PCI. To identify CAD from STEMI cases, the results indicate that miR-3605 and miR-301a are joint markers for CAD together when they are co-expressed. From the radar map by patients from Nanostring and MIX.miR sensor results, we are able to identify the cases between NCAD, CAD, STEMI-pre, and STEMI-PCI. Additional miRNA biomarkers can be applied on MI diagnosis and PCI efficacy evaluation, if needed. The problem we deal with has a small sample size (n = 24) and a large number of predictor variables (p = 809) is a typical small n and large p problem, SCAD and Lasso are well suited for feature selection and initial screening for this type of problem, compared to some other machine-learning methods that may require large datasets to train a complex model with good performance (e.g. deep-learning methods) and that are not suitable for interpretable variable selection problems. Based on SCAD and Lasso analyses of over 800 miRNAs, additional miRNAs could be included, but due to the workload, we only presented 9 miRNAs here. There are chances that some other miRNAs are also useful in diagnosis of STEMI status, such as the ∼20 additional miRNAs in by SCAD and LASSO selections. The SCAD and Lasso are well-regarded in the medical and biological literature for identifying potential biomarkers and have become standard tools in these fields.
The integrated board with miRNA and troponin measurement abilities can be further miniaturized, systemized, and commercialized as a POC device that can be applied in the ambulance for AMI emergency care as a rapid pre-hospital, diagnostic tool. The results can be used to evaluate the severity of AMI and the onset time for deciding the need of PCI.
Footnotes |
† Electronic supplementary information (ESI) available: Detailed protocols, integration board, and clinical sample information. See DOI: https://doi.org/10.1039/d4lc00640b |
‡ Current affiliation: Department of Microelectronics, Tianjin University, Tianjin, 300072, China. |
This journal is © The Royal Society of Chemistry 2024 |