Integration of machine learning to identify diagnostic genes in leukocytes for acute myocardial infarction patients

Acute myocardial infarction (AMI), the most severe form of cardiovascular disease, is associated with [1, 2] millions of deaths annually around the world [3, 4]. Generally, the diagnosis of AMI includes clinical syndrome, electrocardiogram, and serum changes in enzyme levels [5]. However, AMI is easily misdiagnosed because of the following three aspects: nonclassic clinical symptoms [6, 7], atypical underappreciation [8], and an untimely serum peak. Because of the above three problems, a previous study [9] reported that the missed diagnosis rate of AMI is higher than 0.9%. The diagnosis and treatment of AMI must be prompt; otherwise, it may trigger irreversible results. Therefore, exploring new markers of AMI to decrease missed diagnoses is essential and urgent.

Leukocytes play an important and varied role in the entire evolution of AMI. During the acute injury phase of AMI, leukocytes promote a severe inflammatory cascade response through the polarization of M1 macrophages [10]. During the repair phase of AMI, M2 macrophages in leukocytes suppress inflammation and mediate the repair of injured myocardium [11]. Furthermore, leukocyte alteration positively correlates with AMI severity and, inversely, with patient survival [12, 13].

RNAs are involved in the evolution of AMI. For example, miR-155 correlated positively with the concentration of inflammatory cytokines, such as IL-6 and TNF-α [14], in AMI. Neutrophil-derived S100A8/A9 amplify granulopoiesis and cardiac injury in AMI mice [15]. Conversely, M2 macrophage-derived exosomes carry miR-1271-5p [16] to alleviate AMI-related cardiac injury. In conclusion, RNA on leukocytes plays a different role in the evolution of AMI, possibly related to different leukocyte subtypes. However, numerous studies have focused on integrating target interventions [12, 17] and leukocyte complications [17, 18]. Few studies have focused on the diagnostic value of leukocytes' RNA. Because the leukocytes' RNA is involved in the evolution of AMI, these RNA might have diagnosing value for AMI patients. The diagnosis value might be related to various leukocyte subtypes.

Machine learning (ML) helps humans learn patterns from complex data to predict future behavioural outcomes and trends. ML was widely utilized in variable filtering. A previous study used a single ML algorithm or two integrated ML algorithms (e.g., support vector machine [18] or least absolute shrinkage and selection operator [19]) to optimize variables. Still, these approaches may have missed potential genes [20]. Compared with a single ML algorithm, the integrated ML (IML) approach [21‐23] we developed is more advantageous in variable screening and model building. IML helps identify potential genes mistakenly deleted by a single ML and find more meaningful variables [21]. IML integrates the advantages of a single ML, and its predictive classification value is better [23]. Based on a favourable filtration value in transcriptomics, IML might be used to comprehensively explore the diagnostic value in AMI patients.

In summary, we aim to explore the potential diagnostic value of transcriptome within leukocytes for identifying AMI patients. Because of IML's good variable screening and excellent predictive value, IML was first used to mine diagnostic genes in AMI leukocytes with multiple microarrays. Single microarray data might have inherent biases in capturing the entire transcriptomic landscape, so multiple microarrays are integrated after resolving batch effects to reduce bias and validate each other. And clinical validation was added to confirm the result. The relationship between transcriptome and leukocyte subtypes was unclear, so the correlation between immune cells and target transcriptome was subsequently accomplished. We expect to explore the functional roles of the identified genes in AMI pathophysiology, investigating their potential as therapeutic targets.

To our knowledge, our work is the first to filter AMI diagnosis genes based on the overall normalized weights of IML. Four microarrays with 220 samples were adopted for data analysis, and further clinical studies were performed to validate the results. Two genes, AQP9 and SOCS3, showed an AUC > 0.9 in both the training set and testing set (Fig. 6). Both genes showed a typical and highest correlation coefficient (Fig. 7) in monocytes. The clinical study verified the significance between AMI and SCAD controls, indicating a potential diagnostic value of AQP9 and SOCS3. Compared with previous studies, we reached similar conclusions that AQP9 presented diagnostic value for AMI [34, 35], and we further explored the immune correlation of AQP9. Additionally, Prof. Zhu [36] identified SOCS3 as an immune-related gene in AMI, and we expanded it to have diagnostic value. More importantly, this study is the first to reveal the RNA correlation of AQP9 and SOCS3, especially SOCS3, between the number of stenotic coronary arteries and the Killip classification.

AQP9, a cell membrane protein, transports water down the concentration gradient. ERK1/2 can be reversed in AMI rats by silencing AQP9, attenuating cardiomyocytes' inflammatory response and apoptosis and upregulating cardiac function [37]. The above research indicated the crucial role of AQP9 in the pathogenesis of AMI. In human polymorphonuclear leukocytes, AQP9-related inflammation may result from the NK-κB [38] and F-actin polymerization [39]. In our work, the ROC curve of AQP9 was > 0.9. Therefore, AQP9 might be a potential genetic marker for diagnosing AMI with SCAD.

SOCS3 is increased in AMI mice [29] and regulates the T-cell repertoire with STAT3/SOCS3 signalling [40]. More importantly, cardiac-specific silencing of SOCS3 triggers sustained STAT3 and decreases myocardial apoptosis [41]. Therefore, SOCS3 is the dominant negative modulator [42] of Th17 via STAT3 [43]. Apoptosis regulates the pathophysiological evaluation of AMI [44]. In vitro, SOCS3 can trigger the apoptosis of mammary cells [45], and knocking out SOCS3 regulates the expression of apoptosis in 3T3-L1 preadipocytes [46]. The above research emphasized the immune regulation of SOCS3 and the regulation of apoptosis with STAT3. In our work, the ROC curve of SOCS3 was > 0.9. Therefore, SOCS3 might be an effective genetic marker for diagnosing AMI.

Additionally, the CIBERSORT algorithm showed that the proportion of neutrophils and monocytes in the AMI group was higher than in the control group. The progression of AMI is correlated with immune disorder. For example, the white blood cell count correlates highly with in-hospital mortality after AMI [47]. Neutrophils are increased in peripheral blood, and researchers have emphasized that neutrophils-lymphocytes [48, 49] and monocytes/macrophages [50] can be easily acquired factors for the prognosis of AMI. Macrophages were dominant in infarcted myocardium, especially over the first week of AMI [51]. However, NK cells have diminished cytotoxic function [52], and the targeted regulation of NK cells may indicate a dominant role in the cure of AMI. At the beginning of AMI, inflammation deteriorates with increased neutrophils and monocytes [53], and inflammation decreases over time with the reduced function of NK cells. Innate immunity is a vital regulatory factor in the inflammatory, proliferative, and maturation phases [3, 54, 55]. AMI leads to a deteriorated inflammatory process. Currently, novel therapeutic interventions targeting the immune system may regulate slant inflammation, which is conducive to resolving pathological conditions. In a previous clinical trial of 182 NSTEMI patients (a subtype of AMI), the patient's intake of IL-1 blockers decreased acute inflammation [56]. Another immune study showed that short-term blockade of S100A9 downregulates inflammation [57] in permanent coronary ischemia mice. However, the above immune interventions are still experimental and not in the clinic. In summary, regulating immune cells along with the progression of AMI and immune intervention in AMI might be a potential target.

AQP9 expression was highest in human polymorphonuclear leukocytes [39] compared with the spleen and liver, suggesting a possible correlation between AQP9 and immunity or inflammation. AQP9 regulates water flow on leukocytes [58], which regulates cellular morphology and motility, a change that facilitates the migration of leukocytes to inflammatory sites. Similar to our result, Hawang [59] indicated the correlation between AQP9 and neutrophile granulocytes. Research [29, 60, 61] emphasizes the correlation between SOCS3 and neutrophils in inflammation. In our research, both genes had a higher correlation with two immune cells, neutrophils and monocytes. The immune cell correlation indicated that the targeted gene therapy of immune cells may benefit the course of AMI—potential feasibility of using AQP9 and SOCS3 as therapeutic targets or predictors of treatment response.

ML algorithms are widely performed for various cardiovascular diseases, such as optimizing variables, classification, and congression. For variable filtration, numerous studies take only single or double ML algorithms (e.g., weighted gene coexpression network analysis [60], LASSO, and SVM). However, only the single or double ML algorithms might unconsciously delete the potential genes. For example, AQP9 will be ignored if we only take DT because the weights of AQP9 were zero in DT (Table 4). Taking only a single ML might miss some potential genes. For example, although LASSO can detect candidate genes with big data when highly correlated features exist, the LASSO regression method tends to select one of them and ignore all the other features, leading to the instability of the results [61]. In pigmented skin lesions [62], SVM and NN displayed their talent classification value. In preoperative postsurgical mortality [63], GBM is optimized rather than DT, RF, and SVM. Various ML algorithms may show different weights even in the same variable (Table 4). Necessarily, the overall normalized weights of IML were taken to filter genes. Surprisingly, IML explores two potential, unreported diagnostic genes in AMI. In our study, IML has good value in both variable screening and model prediction.

Inevitably, four limitations exist in this work, although the best efforts were taken to eliminate them. Primarily, small sample size verification might possess some bias. So, multicentre collaborations or leveraging larger external datasets is crucial for further verification. Although testing sets and clinical validation were developed to assess the stability of the diagnostic value, the bias of single-centre validation might exist. More confirmation, clinical trials and animal experiments are indispensable for solid verification. Next, the ML algorithms contained limitations (e.g., the black box phenomenon [64]), especially NN, which has numerous layers [65]. The set of operations an ML performs in making a prediction is unknown, even if a human knows precisely what the model is doing at each step of the decision-making process. The operations performed cannot be described in terms of human-understandable semantics. And the interpretability techniques for ML models always catch the eye of developers, which enhances the transparency and reliability of the ML. Thirdly, because of the limitations of our laboratory extraction process, clinically validated acquired leukocytes are predominantly lymphocytes and monocytes. Finally, limited clinical features were obtained (e.g., age [66], ethnicity, and race [67]). Clinical features could potentially enhance the predictive accuracy of the diagnostic model and provide a more comprehensive understanding of AMI. For example, various combinations (e.g., sex, smoking or not, and laboratory indicators) of clinical variables [68] are calibrated to analyze the relationship between the target variable and the outcome.

Based on the overall normalized weights of IML, the research successfully merges four microarrays and uncovers hidden diagnostic genes AQP9 and SOCS3 for leukocytes of AMI patients. AQP9 and SOCS3 are closely associated with monocytes and neutrophils, which might contribute to advancing AMI diagnosis and shedding light on novel genetic markers, including AMI pathogenesis, targeted therapies, and potential precision medicine. Although clinical validation copies the result again. Multiple clinical characteristics, multicenter, and large-sample relevant trials are still needed to confirm its clinical value.