Machine learning-based integration develops biomarkers initial the crosstalk between inflammation and immune in acute myocardial infarction patients

FIGURE 1. The workflow of this study.

FIGURE 2. Identification of DEGs and functional annotation. (A) Gene expression level statistics of the integrated dataset after removed batch effect. (B) The heatmap of AMI-related DEGs expression levels: blue-low gene expression; red-high gene expression. (C) The volcano plot of AMI-related DEGs expression. (D) Ridgeline plot of GSEA results. (E,F) The main signaling pathways that are significantly enriched in the AMI group (E), and in the control group (F).

FIGURE 3. Weight correlation network analysis. (A) Sample clustering dendrogram with tree leaves corresponding to individual samples. (B) Analysis of the scale-free fit index (R2) and the mean connectivity for various soft-thresholding powers. (C) The original and combined modules under the clustering tree with cut-off values height of 0.25. (D) Collinear heat map of module feature genes. Red color indicates a high correlation, blue color indicates the opposite results. (E) Clustering dendrogram of module feature genes. (F) Heat map of module–trait correlations. Red represents positive correlations and blue represents negative correlations. (G) MM vs. GG scatter plot of AMI for cyan module.

FIGURE 4. Identification and functional enrichment analyses of overlapping candidate genes. (A) Venn diagram showed the intersection of DEGs and module genes of WGCNA. (B–F) GO (B), KEGG (C–E), and DO (F) enrichment analysis of the overlapping candidate genes. (G) Protein-Protein Interaction (PPI) network of overlapping candidate genes. (H) The co-expression network showing correlation intensity of hub genes from overlapping candidate genes.

FIGURE 5. Three machine learning algorithms were integrated to identify the optimal feature genes. (A) LASSO coefficient profiles of the candidate optimal feature genes and the optimal lambda was determined when the partial likelihood deviance reached the minimum value. Each coefficient curve in the left picture represents a single gene. The solid vertical lines in right picture represent the partial likelihood deviance, and the number of genes (n = 30) corresponding to the lowest point of the cure is the most suitable for LASSO. (B) The SVM-RFE algorithm was used to further candidate optimal feature genes with the highest accuracy and lowest error obtained in the curves. The x-axis shows the number of feature selections, and the y-axis shows the prediction accuracy. (C) Relative importance of overlapping candidate genes calculated in random forest (Top 10 genes’ importance > 2). Importance indexes on the x-axis and genetic variables are plotted on the y-axis. (D) Random forest for the relationships between the number of trees and error rate. The x-axis represents the number of decision trees and the y-axis is the error rate. (E) Venn diagram showing the seven optimal feature genes shared…

FIGURE 6. Verification of expression and diagnostic efficacy in predicting AMI progression of optimal feature genes. (A–G) Box plots showing the expression of ACSL1 (A), GABARAPL1 (B), IL1R2 (C), IRAK3 (D), MCEMP1 (E), NFIL3 (F), and THBD (G) in control and AMI samples. Statistic tests: Wilcoxon rank-sum test. (H–O) Roc curves (H) estimating the diagnostic performance of ACSL1 (I), GABARAPL1 (J), IL1R2 (K), IRAK3 (L), MCEMP1 (M), NFIL3 (N), and THBD (O).

FIGURE 7. Verification of expression and diagnostic efficacy for optimal feature genes using external validation dataset. (A–G) Box plots showing the expression of ACSL1 (A), GABARAPL1 (B), IL1R2 (C), IRAK3 (D), MCEMP1 (E), NFIL3 (F), and THBD (G) in control and AMI samples. Statistic tests: Wilcoxon rank-sum test. (H–O) Roc curves (H) estimating the diagnostic performance of ACSL1 (I), GABARAPL1 (J), IL1R2 (K), IRAK3 (L), MCEMP1 (M), NFIL3 (N), and THBD (O).

FIGURE 8. Gene sets enrichment analysis (GSEA) identifies signaling pathways in the optimal feature genes. (A–G) Top five signaling pathways that are significantly enriched in the high expression of ACSL1 (A), GABARAPL1 (B), IL1R2 (C), IRAK3 (D), MCEMP1 (E), NFIL3 (F), and THBD (G).

FIGURE 9. Visualization of immune cell infiltration. (A) The relative proportions of 22 immune cells types between control samples and AMI samples. Panel (B) representative boxplot shows the differences of infiltrated immune cells between control samples and AMI samples. Statistic tests: Wilcoxon rank-sum test. (P < 0.05*; P < 0.01**; P < 0.001***; ns, no significance). (C–I) Correlation between immune cells and optimal feature genes ACSL1 (C), GABARAPL1 (D), IL1R2 (E), IRAK3 (F), MCEMP1 (G), NFIL3 (H), and THBD (I); the larger the dots, the stronger the correlation. The color of the dots represents the P-value; the greener the color, the lower the P-value. (J,K) Correlation analysis of seven optimal feature genes in AMI samples.

FIGURE 10. Analysis of hallmark gene sets. (A) The specific distribution of the 50 hallmark gene sets in AMI and control samples. (B) Correlation analysis of the 50 hallmark gene sets with seven optimal feature genes. Statistic tests: Wilcoxon rank-sum test (P < 0.2#; P < 0.05*; P < 0.01**; P < 0.001***; ns, no significance).

FIGURE 11. The relative expressions of optimal feature genes were validated by qRT-PCR. (A–G) The expressions of ACSL1 (A), GABARAPL1 (B), IL1R2 (C), IRAK3 (D), MCEMP1 (E), NFIL3 (F), and THBD (G) between AMI patients and healthy subjects. Statistic tests: Student’s t-test (P < 0.01**; P < 0.001***; P < 0.0001****).