Identification and Verification of Novel Biomarkers Involving Rheumatoid Arthritis with Multimachine Learning Algorithms: An In Silicon and In Vivo Study

Abstract

Background: Rheumatoid arthritis (RA) remains one of the most prevalent chronic joint diseases. However, due to the heterogeneity among RA patients, there are still no robust diagnostic and therapeutic biomarkers for the diagnosis and treatment of RA.

Methods: We retrieved RA-related and pan-cancer information datasets from the Gene Expression Omnibus and The Cancer Genome Atlas databases, respectively. Six gene expression profiles and corresponding clinical information of GSE12021, GSE29746, GSE55235, GSE55457, GSE77298, and GSE89408 were adopted to perform differential expression gene analysis, enrichment, and immune component difference analyses of RA. Four machine learning algorithms, including LASSO, RF, XGBoost, and SVM, were used to identify RA-related biomarkers. Unsupervised cluster analysis was also used to decipher the heterogeneity of RA. A four-signature-based nomogram was constructed and verified to specifically diagnose RA and osteoarthritis (OA) from normal tissues. Consequently, RA-HFLS cell was utilized to investigate the biological role of CRTAM in RA. In addition, comparisons of diagnostic efficacy and biological roles among CRTAM and other classic biomarkers of RA were also performed.

Results: Immune and stromal components were highly enriched in RA. Chemokine- and Th cell-related signatures were significantly activated in RA tissues. Four promising and novel biomarkers, including CRTAM, PTTG1IP, ITGB2, and MMP13, were identified and verified, which could be treated as novel treatment and diagnostic targets for RA. Nomograms based on the four signatures might aid in distinguishing and diagnosing RA, which reached a satisfactory performance in both training (AUC = 0.894) and testing (AUC = 0.843) cohorts. Two distinct subtypes of RA patients were identified, which further verified that these four signatures might be involved in the immune infiltration process. Furthermore, knockdown of CRTAM could significantly suppress the proliferation and invasion ability of RA cell line and thus could be treated as a novel therapeutic target. CRTAM owned a great diagnostic performance for RA than previous biomarkers including MMP3, S100A8, S100A9, IL6, COMP, LAG3, and ENTPD1. Mechanically, CRTAM could also be involved in the progression through immune dysfunction, fatty acid metabolism, and genomic instability across several cancer subtypes.

Conclusion: CRTAM, PTTG1IP, ITGB2, and MMP13 were highly expressed in RA tissues and might function as pivotal diagnostic and treatment targets by deteriorating the immune dysfunction state. In addition, CRTAM might fuel cancer progression through immune signals, especially among RA patients.

Figure 1

Workflow and batch removal. (a) Overall workflow of this study. (b) PCA plot illustrating the efficiency of batch effect removal (left: before; right: after batch effect removal).

Figure 2

Differentially expressed signatures between RA and normal tissues. (a) Volcano plot of differentially expressed genes. Red represents upregulated genes; blue represents downregulated genes. (b, c) Bar plot of BP and KEGG pathway enrichment analysis of all DEGs. (d, e) GSEA indicates the most activated and inhibited pathways in RA compared with normal tissues.

Figure 3

Identification and verification of RA-related signatures. (a) LASSO, (b) SVM–RFE, (c) RF, and (d) XGBoost algorithms were applied to identify RA-related biomarkers based on DEGs in the discovery cohort. (e) Intersections of features from the four machine learning algorithms in the discovery cohort. (f, g) ROC curves were used to evaluate the specificity and sensitivity of the four intersection signatures to distinguish RA and normal tissues in the discovery and validation cohorts.

Figure 4

Differences in immune components and signatures between RA and normal tissues. (a) Differences in ESTIMATE score, immune score, and stromal score between RA and normal tissues. (b, c) Different infiltration degrees of 28 types of immune cells and immune signatures between RA and normal tissues.  ^∗P < 0.05,  ^∗∗P < 0.01,  ^∗∗∗P < 0.001,  ^∗∗∗∗P-value is too small, close to zero.

Figure 5

Correlation of four biomarkers and immune signatures in RA. (a) Different expression levels of four biomarkers between normal and RA tissues. (b) Spearman and Pearson correlations of four biomarkers in the RA expression matrix. (c–e) Relationship of four biomarkers and estimated related scores, immune-related signature scores, and immune cell infiltration scores in RA.  ^∗∗∗P < 0.001,  ^∗∗∗∗P-value is too small, close to zero.

Figure 6

Construction and verification of the susceptibility quantification system for RA. (a) Nomogram based on the expression of four biomarkers to predict the susceptibility scores of RA arthritis patients. (b, c) ROA curve and calibration curve of the prediction system in the training and testing cohorts.

Figure 7

Identification of two distinctive subtypes in the RA groups. (a) Consensus cluster matrix of RA patients when k turns to 2. (b) The cumulative distribution function curves suggested k2 as the optimal cluster number in RA patients. (c) The relative change in area under the CDF curve. (d) 2D principal component plot by the matrix derived from the four signatures. The blue dots represent C1, and the red dots represent C2. (e) Heatmap illustrating the different expression levels of four biomarkers between C1 and C2. (f) Differences in ESTIMATE score, immune score, and stromal score between C1 and C2. (g) Difference in immune infiltration score between C1 and C2.  ^∗P < 0.05,  ^∗∗P < 0.01,  ^∗∗∗P < 0.001,  ^∗∗∗∗P-value is too small, close to zero.

Figure 8

Characteristics of CRTAM across cancers. (a) Expression level of CRTAM between cancer and normal tissues. (b) Differential expression of CRTAM in different stages of pan cancer. (c) Univariable Cox analysis of CRTAM on overall survival. (d) Correlation of CRTAM and immune-related signatures.  ^∗P < 0.05,  ^∗∗P < 0.01,  ^∗∗∗∗P-value is too small, close to zero.

Figure 9

Biological influence of CRTAM across cancers. (a) Correlation of CRTAM and immune infiltration based on the Xcell algorithm. (b) GSEA results based on high CRTAM expression group vs. low CRTAM expression group. (c, d) Correlations of CRTAM expression with the expression levels of MMR and DNA methyltransferase signatures.  ^∗P < 0.05,  ^∗∗P < 0.01,  ^∗∗∗P < 0.001.