Development and validation of a diagnostic model for ischemic cardiomyopathy with Artificial Neural Network by bioinformatic analysis

doi:10.21203/rs.3.rs-3787156/v1

Download PDF

Research Article

Development and validation of a diagnostic model for ischemic cardiomyopathy with Artificial Neural Network by bioinformatic analysis

https://doi.org/10.21203/rs.3.rs-3787156/v1

This work is licensed under a CC BY 4.0 License

Version 1

posted

You are reading this latest preprint version

Background

Ischemic cardiomyopathy(ICM) is a significant global health concern caused by high morbidity and mortality.In addition, no previous study has reported the diagnostic biomarkers in ICM.

Objective

The presentstudy is aimed at establishing and validating a diagnostic model for ICM with Artificial Neural Network (ANN) by screening key potential biomarkers using bioinformatic analysis.

Method

Through searching the Gene Expression Omnibus(GEO) database, three gene expression datasets were downloaded and merged. Differentially expressed genes(DEGs) in the mergeddatasetswere detectedusing R software and subject to Gene Ontology(GO) and Kyoto Encyclopedia of Genes and Genomes(KEGG) enrichment analysis. Then, Lasso regression analysis and random forest (RF) wereapplied to identify critical genes based on DEGs. Afterwards, we intersected the key genes screened from Lasso regression analysis and RF. An ICM diagnostic model was constructed by ANN. Based on a validation dataset, the diagnostic model was assessed, whereasits diagnostic performance was assessed usingarea under curve(AUC) values.

Results

Totally 18 ICM-related DEGs were detected. Then, six hub genes(COL1A1, FCN3, GLUL, MYOT, SERPINA3, and SLC38A2) were identified by intersecting the key genes filtered out by Lasso regression analysis and Random forest(RF). In the end, a diagnostic model for ICM was successfully designed by ANN, obtaining an AUC of 0.907 and 0.745 in training datasets, separately.

Conclusion

this study detected several potential genetic biomarkers and successfully developed an early predictive model with high diagnostic performance for ICM. In addition, the obtained findings offer a significant guidance for the early diagnosis as well as screening of ICM in the future.

Ischemic Cardiomyopathy

hub genes

bioinformatic analysis

Cardiovascular disease (CVD) is still a main reason for death across the world^[1]. Ischemic cardiomyopathy(ICM), in particular, is the main factor causing global prevalence and mortality^[2]. Furthermore, ICM has been detected as the leading reason for CVDs in the United States and the most commonly seen risk factor for HF^[3]. In accordance with the global pandemic, approximately26 million ICM cases have cardiac insufficiency, which induces over $30 billion burdens on the health system worldwide^{[4, 5]}. Moreover, the mortality rate for cardiac disease cases has been as high as 50% over the last five years^{[6, 7]}.

The initial reason for ICM is the development of atherosclerosis in multi-coronary arteries, especially the diffusive lesions, and reduced or ceased myocardial blood flow that can generate severe myocardial dysfunction, causing heart muscle injury^{[8, 9]} and persisting injury.

According to previous studies, ICM has high incidence and mortality rates, and induces great economic burdens on the global health system; thus, early diagnosis and treatment are needed for reducing the re-hospitalization and mortality rates. NT-proBNP combined with coronary angiograph remains the gold standard for ICM diagnosis. However, some patients do not accept the coronary angiograph due to its invasive nature, making it challenging to diagnose and treat ICM accurately.

Recently, microarray technique has been developed and applied to help for reveal pathogenic factors and pathogenic mechanisms associated with ICM. However, the identification of key feature genes for ICM by microarray gene expression profiles is still challenging for the development of diagnosis models. Therefore, different machine learning algorithms, including Lasso regression analysis, artificial neural networks(ANN), and random forest(RF), have been used for dealing with the above-mentioned issue^{[10, 11]}. Furthermore, machine learning methods have been extensively applied in in medical application and gesture recognition^[12]. Combining the above technologies has greatly contributed to diagnosing diseases and predicting their prognosis.

In the current work, we aimed to establish an early diagnostic model of ICM through microarray gene expression data from the Gene Expression Omnibus(GEO) database with the combination of Lasso regression analysis, RF and ANN because of the extreme computational power. In addition, we employed an other GEO dataset, aiming to confirm the accuracy of the diagnostic model(as shown in the design process presented in Fig. 1).

2.1 Data acquisition and preprocessing. We acquired three microarray expression datasets from the National Center for Biotechnology Information Gene Expression Omnibus database(NCBI-GEO; https://www.ncbi.nlm.nih.gov/geo/). In addition, three GEO datasets(GSE1869, GSE5406 and GSE42955) were merged to be the training dataset with the purpose of obtaining a larger sample size. GSE116250, which is “fpkm” format, served as the validation dataset. Using the “limma” package, all datasets are standardized and normalized. The "sva" package was applied to conduct the batch effect and to combine the three data sets of training set.

2.2 DEGs Analysis. The training dataset comprised 130 ICM and 42 normal samples. DEGs from training set were explored by R software “limma” package. Besides, genes were regarded DEGs if they satisfied the double-filtering criterion: absolute log2-foldchange >1.5 and adjusted P-value ≤ 0.05. DEGs were visualized with the “ggplot2” and “heatmap” package of R software^{[13, 14]} in order to yield volcano plots and heatmaps, separately.

2.3 Functional enrichment analysis. With the purpose of further revealing the characteristic biological properties of DEGs, the “cluster profiler” package in R software^[15] was adopted for performing functional enrichment analysis, containing cellular component(CC), biological process(BP), molecular function(MF) and Kyoto Encyclopedia of Genes and Genomes(KEGG) pathway analysis, and visualized by “GOplot” package^[16]. In addition, with the use of the“HALLMARK” gene set in “MSigDB” database(http://www.gsea-msigdb.org/gsea/msigdb), GSEA analysis was conducted by “clusterprofiler” package.

2.4 Screening of key genes. Lasso regression was conducted on DEGs using the “glmnet” Package through ten-foldcross validation. Based on the minimum lambda value, the best model effect was determined and the hub genes were selected. Then, the random forest (RF) analysis was carried out on DEGs by the “randomforest” package^[11], the Gini coefficient method was applied during modeling in order to measure the variable importance, and hub genes were selected from DEGs based on the Gini coefficient score. The final key genes were identified by intersecting the hub genes obtained from LASSO and RF.

2.5 Construction and validation of ANN. The “neuralnet” and “NeuralNetTools” Package^[17] was used. ANN can simulate neural activity with mathematical model, which refers to an information processing system by imitating the structure and function of brain neural network. The network structure is classified into three layers. The input layer neurons are the previously screened genes which can be used to build the diagnostic model. The output layer contains two neurons, namely, control group and ICM group. Then umber of neurons in the hidden layer was adjusted based on model. The model is regarded as the best model when the area under the ROC curve is the largest. To draw the ROC curve, the “pROC” package^[18] is employed. In addition, the model was validated by plotting the ROC and calculating the AUC using the validation dataset.

2.6 Co-expression and Protein Interaction of Diagnostic Genes. By “gghalves”

package plotted a violin map to investigate and compare the level of diagnostic genes in the control group and the ICM group. The “gghalves” package was used. With the use of the “RCircos” package, the location map circle of diagnostic gene on the chromosome was plotted. Proteins potentially interacting with the diagnostic gene were predicted using the online tool GENEMANIA(http://genemania.org/).

2.7 Diagnostic gene and immune infiltration. The “GSVA” and “GSEABase” packages were used to estimate the enrichment scores of immune cells and immune processes for all samples in the training set. To evaluate the connection of diagnostic genes with immune infiltration, we analyzed the association of diagnostic genes with immune cell abundance and immune process enrichment fraction. Furthermore, the co-expression of diagnostic genes and chemokines, and the co-expression of diagnostic genes and HLA family were analyzed. In addition, the correlation heat map was visualized using the “Hmisc” package.

Our study gathered and merged three ICM expression profiling datasets (GSE1869, GSE5406 and GSE42955) as a training set. In addition, GSE116250 dataset was collected as a validation set. All datasets are standardized and normalized using the “limma” package. Using the “sva” package, batch effects were performed on the training set for subsequent analyses (Fig. 2A,B)

3.1. Screening of DEGs. Based on the “limma” package of the R software, totally 18 DEGs were detected between 130 ICM and 42 normal samples in the training dataset. DEGs contained 10 lowly expressed genes and 8 highly expressed genes, with |FC| > 1.5 & adjusted P value < 0.05.In addition, the findings of DEGs were visualized in the volcano map and heatmap (Fig. 2C, D).

3.2. GO and KEGG Enrichment Analyses of the DEGs.

In order to investigate the biological significance of these DEGs in the pathogenesis of ICMs, this study carried out GO and KEGG pathway enrichment analyses via the R package “clusterProfiler”. GO terms were divided into the following three types including biological process(BP), cellular component(CC), and molecular function(MF). In addition, the top ten GO terms of genes with notably up regulated and down regulated expression levels were visualized (Fig. 3A,B,C). KEGG pathway enrichment analyses were carried out via the R package “clusterProfiler”, namely, the top ten genes (Fig. 3D). GSEA enrichment analysis was carried out using the HALLMARK gene set from MSigDB database. The top ten signaling pathways were presented, all of which were down-regulated in ICM(Fig. 3E).

3.3. Screening for Key DEGs by LASSO and Random Forest (RF). By performing Lasso regression analysis, 11 genes were filtered out from 18 DEGs(Fig. 4A,B).18 DEGs were put into the RF classifier, and the 123456 was set as the random seeds. Through referring to the correlation between the model error and number of decision trees(Figue3 C), we chose 152 trees to be the parameter of the random forest model, representing a stable error in the model. During the modeling process, the Gini coefficient method was adopted for measuring the importance of all variables in accordance with reduced mean square error and model accuracy (Fig. 4D). This study chose the top ten DEGs with a mean decrease of the Gini index. After intersecting11 DEGs from LASSO and 10 from RF, six characteristic genes were finally determined, including FCN3, MYOT, SERPINA3, COL1A1, SLC38A2 and GLUL(Fig. 4E). Then, these six hub genes were used to perform multivariable logistic regression analysis, the OR value was calculated and the forest map was drawn (Fig. 4F).

3.4. Construction and Verification of the ANN Model. By intersecting 11 DEGs filtered out from LASSO and 10 DEGs from RF, six Vital genes that best differentiate ICM from normal samples were detected. To construct ANN model, firstly, the data needs to be normalized. Secondly, based on ANN analysis, gene weights were determined to construct the ICM scoring model. Based on ANN topology of training set, there were two out-put layers, four hidden layers, and six input layers. According to the above-mentioned discussion, we constructed an ANN model of ICM for the classification of gene expression data between ICM and normal samples using the “neuralnet” package of R (Fig. 5A). Meanwhile, AUC was used to evaluate model classification ability. Our model achieved the AUC of 0.907 for training set (Fig. 5B) whereas 0.745 AUC for test set (GSE116250) (Fig. 5C), representing excellent classification ability of our model. ANN model made high classification performance. In addition, the findings revealed that a diagnostic model for ICM was successfully constructed from the differential gene expression of ICM and normal samples.

3.5 The co-expression and PPI map of hub genes. To explore the hub gene levels in ICM group and control group, the violion chart was plotted (Fig. 6A)using the “gghalves” package. Then, expression correlation analysis between six diagnostic genes was conducted using the “corrplot” package(Fig. 6B). By using the “RCirco” package, the circle map that described the location of hub genes on the chromosome was developed (Fig. 6C). Further, proteins potentially related to the diagnostic genes were predicted (Fig. 6D) by the online-tool GENEMANIA(http://genemania.org/).

3.6 Furthermore, immune abundance related analysis of the six hub genes was conducted, the plot was drawn (Fig. 7,A), and immune process correlation analysis was carried out with six hubgenes(Fig. 7,B). Besides, a heatmap that described the correlation between 6 hub genes and expression levels of chemotactic factors was developed (Fig. 7C). Meanwhile, a heatmap describing the relation between six hubgenes and expression levels of HLA family members was also plotted (Fig. 7D).

Increasing the morbidity and mortality of ICM has become the main health concern. Therefore, predicting and diagnosing ICM can increase the early intervention chance. However, no previous study has reported the diagnostic biomarkers in ICM. As a result, it is essential to construct the model for early diagnosis in order to identify representative biomarkers for ICM, and maybe it can predict the ICM morbidity before the development of ICM. Thus, we can intervene before the development of ICM. Recently, machine learning technologies have been greatly improved, and gene expression profiles are available from public databases, which can also provide the novel options for diagnosing and predicting ICM.

The present study gathered microarray expression profiling datasets from GEO database(GSE1869, GSE5406 and GSE42955).Meanwhile, 18 DEGs between ICM and normal samples were detected. Furthermore, GO enrichment analysis (BP, CC, MF) and KEGG analysis were conducted. Meanwhile, GSEA was performed, and the top ten signaling pathways were presented (Fig. 3).

Lasso represents the regression method using regularization to improve prediction ability^[19]. RF refers to a reliable feature selection approach which can determine the optimal variables by the removal of the feature vectors yielded by RF^[20]. Thus, we conducted Lasso regression on 18 DEGs. As a result,11 key genes between ICM and normal samples were obtained. Additionally, RF screening was also carried out on 18 DEGs, which detected 10 hub genes. Afterwards, genes identified from Lasso regression analysis and RF were intersected by venin, and six hub genes were eventually obtained. Then, we employed a characteristic gene-based ANN model to calculate predictive weights for six hub genes. Then, a diagnostic model for ICM was established by ANN. Additionally, we compared the established diagnostic model regarding predictive accuracy in the training and validation datasets with the use of the AUC of ROC curves. According to the obtained findings, the established model exhibited high diagnostic power. Moreover, the present study is the first attempt to construct a diagnostic model for ICM.

At first, our research identified 6 hub genes through bioinformatic analysis of 130 ICM and 42 normal samples in the GEO database. Apparently, the expression levels of COL1A1 genes in the ICM group are notably higher when compared with those in the normal group(log2FC > 1.5). Other five(FCN3, GLUL, MYOT, SERPINA3, SLC38A2) genes are significantly lower expressed in ICM group than normal group (log2FC<1.5)(Fig. 6A). COLIA1 is mapped to the long arm of chromosome 17 (17q21.33). More than 400 mutations related to human disease are identified in COL1A1 gene, and many of them are associated with osteoporosis^[21]. In addition, COLIA1 is highly denoted in diverse cancers and controls a variety of cellular processes, containing cell proliferation, metastasis, apoptosis, and cisplatin resistance. COLIA1 is also correlated with cancer progression and prognosis; besides, elevated COLIA1 expression relates to poor prognosis in cancer patients^[22].

Although COL1A1 and other five DEGs were correlated with various disease, no previous study reported the correlation between ICM and six DEGs. Thus, this is the new finding concerning this notion. This is the novel gene-based model to diagnose ICM. Furthermore, we presented violion chart (Fig. 6A), predicting the expression level of six hub DGEs in the ICM and normal subjects. It can be observed that all of the six genes are significantly different in ICM and normal subjects. In addition, it explains that every gene independently has the diagnostic value of ICM. Although we find the change of six hub gene levels within ICM, the causal relationship between differential expression of six hub genes in ICM and its mechanism remains unclear. Thus, we need to clarify whether changing six gene expression induced ICM or six gene levels resulted from self-protection reaction after the development of ICM. Moreover, we need to further investigate the mechanism of the existing relationship. ICM can be resulted by various risk factors including hypertention, diabetes, drinking alcohol and smoking.

Furthermore, immune abundance related analysis of the six hub genes was conducted, the plot was drawn (Fig. 7,A), and immune process correlation analysis was performed with six hub genes(Fig. 7,B). A heatmap that described the correlation between 6 hub genes and expression levels of chemotactic factors was developed (Fig. 7C). At the same time, a heatmap describing the relationship between six hub genes and expression levels of HLA family members was also plotted (Fig. 7D).

However, this study still has the following limitations. First, three datasets with small sample size were combined into training set. As a result, this diagnostic model requires to be re-established one independent datasets with large sample size. Secondly, our ICM prediction model was constructed on the basis of datasets from the GEO database. Furthermore, we need to perform further in vitro and in vivo experiments to practice and confirm the diagmostic model.

To sum up, six genetic biomarkers closely related to ICM were screened by bioinformatic analysis, and used to construct the highly-efficient ICM diagnosismodel. In addition, this study lays the foundation for early diagnosis of ICM, which also provides reliable biomarkers for predicting ICM. However, further researches on the molecular mechanisms where six genetic biomarkers are involved are still required with the purpose of confirming the roles of these genes in ICM.

Ethics Statement

The present work gained approval from Ethics Committee of the First Affiliated Hospital of Xinjiang Medical University (Approval No. 2021D01D17). All data utilized in this study is obtained from GEO database，thus there is no informed consents.

Conflict of Interest

All authors claimed that there existed no competing interest.

Author Contributions

All authors contributed to this work equally.

Found

This study was supported by the grant from the “Key Project of Natural Science Foundation of Xinjiang Uygur Autonomous Region” (No.2021D01D17).

Data availability statement

All datasets utilized during the presentstudy can be obtained in GEO (http://www.ncbi.nlm.nih.gov/geo/).

Yang L, Wang L, Deng Y, et al. Serum lipids profiling perturbances in patients with ischemic heart disease and ischemic cardiomyopathy[J]. Lipids Health Dis, 2020, 19(1): 89.
Moroni F, Gertz Z, Azzalini L. Relief of Ischemia in Ischemic Cardiomyopathy[J]. Curr Cardiol Rep, 2021, 23(7): 80.
Divoky L, Maran A, Ramu B. Gender Differences in Ischemic Cardiomyopathy[J]. Curr Atheroscler Rep, 2018, 20(10): 50.
Braunwald E. The war against heart failure: the Lancet lecture[J]. The Lancet, 2015, 385(9970): 812-824.
Virani S S, Alonso A, Benjamin E J, et al. Heart Disease and Stroke Statistics-2020 Update: A Report From the American Heart Association[J]. Circulation, 2020, 141(9): e139-e596.
Lund L H, Savarese G. Global Public Health Burden of Heart Failure[J]. Cardiac Failure Review, 2017, 03(01).
Roth G A, Mensah G A, Johnson C O, et al. Global Burden of Cardiovascular Diseases and Risk Factors, 1990-2019: Update From the GBD 2019 Study[J]. J Am Coll Cardiol, 2020, 76(25): 2982-3021.
Hobby A R H, Berretta R M, Eaton D M, et al. Cortical bone stem cells modify cardiac inflammation after myocardial infarction by inducing a novel macrophage phenotype[J]. Am J Physiol Heart Circ Physiol, 2021, 321(4): H684-H701.
Zhao T T, Wasti B, Xu D Y, et al. Soluble epoxide hydrolase and ischemic cardiomyopathy[J]. Int J Cardiol, 2012, 155(2): 181-7.
Chen Y C, Ke W C, Chiu H W. Risk classification of cancer survival using ANN with gene expression data from multiple laboratories[J]. Comput Biol Med, 2014, 48: 1-7.
Kursa M B. Robustness of Random Forest-based gene selection methods[J]. BMC Bioinformatics, 2014, 15: 8.
Lin S, Chen C, Cai X, et al. Development and Verification of a Combined Diagnostic Model for Sarcopenia with Random Forest and Artificial Neural Network[J]. Comput Math Methods Med, 2022, 2022: 2957731.
Maag J L V. gganatogram: An R package for modular visualisation of anatograms and tissues based on ggplot2[J]. F1000Res, 2018, 7: 1576.
Yu B, Tao D. Heatmap Regression via Randomized Rounding[J]. IEEE Trans Pattern Anal Mach Intell, 2022, 44(11): 8276-8289.
Yu G, Wang L G, Han Y, et al. clusterProfiler: an R package for comparing biological themes among gene clusters[J]. OMICS, 2012, 16(5): 284-7.
Walter W, Sanchez-Cabo F, Ricote M. GOplot: an R package for visually combining expression data with functional analysis[J]. Bioinformatics, 2015, 31(17): 2912-4.
Bianconi A, Von Zuben C J, Serapiao A B, et al. Artificial neural networks: a novel approach to analysing the nutritional ecology of a blowfly species, Chrysomya megacephala[J]. J Insect Sci, 2010, 10: 58.
Robin X, Turck N, Hainard A, et al. pROC: an open-source package for R and S+ to analyze and compare ROC curves[J]. BMC Bioinformatics, 2011, 12: 77.
Xing Q, Jiaochen L, Shouyong L, et al. Six RNA binding proteins (RBPs) related prognostic model predicts overall survival for clear cell renal cell carcinoma and is associated with immune infiltration[J]. Bosn J Basic Med Sci, 2022, 22(3): 435-452.
Zhao E, Xie H, Zhang Y. Predicting Diagnostic Gene Biomarkers Associated With Immune Infiltration in Patients With Acute Myocardial Infarction[J]. Front Cardiovasc Med, 2020, 7: 586871.
Yu K H, Tang J, Dai C Q, et al. COL1A1 gene -1997G/T polymorphism and risk of osteoporosis in postmenopausal women: a meta-analysis[J]. Genet Mol Res, 2015, 14(3): 10991-8.
Li X, Sun X, Kan C, et al. COL1A1: A novel oncogenic gene and therapeutic target in malignancies[J]. Pathol Res Pract, 2022, 236: 154013.

No competing interests reported.

Download PDF

Version 1

posted

You are reading this latest preprint version

Development and validation of a diagnostic model for ischemic cardiomyopathy with Artificial Neural Network by bioinformatic analysis

Status:

Version 1

Abstract

Background

Objective

Method

Results

Conclusion

Figures

1. Introduction

2. Materials and Methods

3. Results

3.2. GO and KEGG Enrichment Analyses of the DEGs.

Discussion

Declarations

References

Additional Declarations

Status:

Version 1