Session Type: ACR Poster Session C
Session Time: 9:00AM-11:00AM
Background/Purpose: Diagnosis and monitoring the disease progression of RA is challenging requiring a combination of imaging techniques and blood tests. There is currently no biochemical test for detection of early-stage disease. In this study, we aimed to define a Rheumatoid Arthritis meta-profile and identify biomarkers by leveraging publicly available gene expression data with machine learning approaches.
Methods: We carried out a comprehensive search for publicly available microarray data at NCBI GEO database for whole blood and synovial tissue in Rheumatoid Arthritis and health controls. For the synovium, we collected 13 datasets with 312 biopsy samples. Among them, there were 276 RA samples and 36 healthy tissue biopsies. For whole blood data, we collected 11 datasets with 2,153 samples: 1,394 RA and 759 healthy controls. We computed differential expression using Significance Analysis of Microarrays (SAM) approach. We applied the cutoff of FDR < 0.05 and abs(FC) > 1.2 to the results to identify significant differentially expressed genes. For pathway analysis we leveraged the gene list enrichment analysis tool ToppGene.
Results: As a result of our analysis we were able to identify 882 genes that were significantly differentially expressed in the synovium between RA patients and healthy controls. Among them we recognized 502 up-regulated and 380 down regulated genes. We confirmed the gene regulation of the immune system process and response, and cell activation and aggregation in both innate and adaptive immune system pathways were involved in RA. As for the whole blood data, we identified 339 significantly differentially expressed genes with 166 up-regulated and 173 down-regulated genes among them. Aiming to determine RA biomarkers we performed a machine learning feature selection procedure to sets of significant genes for both tissues. First, we filtered out genes that cumulatively contribute to the biological variance less than 5%. Then we applied a Variable Selection Using Random Forests (VSURF) approach to the leftovers. Next, we performed a hypergeometric test and found 12 common genes with p = 0.001 with 3 common up-regulated genes: Antigen peptide transporter 1 (TAP1), Matrix Metallopeptidase 9 (MMP9), and DNA Damage Regulated Autophagy Modulator 1 (DRAM1), and 2 common down-regulated genes: DDX3Y, MYC. Finally, we built a Random Forest classification model on the synovium data with these 5 genes. We applied 5-fold cross-validation with 10 repeats technique and used Cohen’s Kappa statistic as a metric. We obtained Kappa equals 0.61 with sensitivity 0.86 and specificity 0.9 on the testing set. In the final step, we validated the prediction model on the whole blood data, resulting kappa of 0.57 with sensitivity 0.54 and specificity 0.98.
Conclusion: Our computational analysis of public data allowed us to perform a comprehensive in-silico search for biomarkers in Rheumatoid Arthritis. We found three protein coding genes that have the strongest association with RA. Identification of extensive proteins secretion in blood could allow precision phenotyping on even early stages of the disease which could have a positive impact on monitoring disease progression and patient treatment.
To cite this abstract in AMA style:Rychkov D, Sirota M, Lin C. Leveraging Publicly Available Gene Expression Data and Applying Machine Learning to Identify Novel Biomarkers for Rheumatoid Arthritis [abstract]. Arthritis Rheumatol. 2018; 70 (suppl 10). https://acrabstracts.org/abstract/leveraging-publicly-available-gene-expression-data-and-applying-machine-learning-to-identify-novel-biomarkers-for-rheumatoid-arthritis/. Accessed April 13, 2021.
« Back to 2018 ACR/ARHP Annual Meeting
ACR Meeting Abstracts - https://acrabstracts.org/abstract/leveraging-publicly-available-gene-expression-data-and-applying-machine-learning-to-identify-novel-biomarkers-for-rheumatoid-arthritis/