Screening Biomarkers for Severe Lupus Based on Multi-omics Studies

Overview

In order to achieve accurate diagnosis and treatment of the disease, we performed RNA sequencing and ATAC(Assay for Transposase Accessible Chromatin) chromatin open sequencing in lupus patients in the early stage. By comparing with normal controls, other rheumatic immune diseases (rheumatoid arthritis), and before and after treatment, dozens of disease-causing genes independently associated with the disease were identified. Based on the previous omics results, this project will analyze its changes in different outcomes of lupus patients, and use machine learning methods to establish an optimal severe prediction model, so as to build an early diagnosis system based on novel biomarkers and reduce all-cause mortality in patients with treatment failure rate. It is expected to produce good social and economic benefits.

The results of the previous study found that the gene expression value in the active stage of lupus disease was more than 1.5 times higher on average than in the inactive stage. PASS（Power Analysis and Sample Size） 15.0 software was used for analysis, and the sample size was calculated by t-test between groups. Two-sided α=0.05 test level was used, and the statistical power was 80% (β=0.2). According to the 3-fold difference in sample size between the groups, a total of 60 samples were needed. According to the 20% dropout rate, the total number of cases is 76. In the second stage, the validation set was allocated in a ratio of 1:1, the statistical power was 90% (β=0.1), and other conditions remained unchanged. An additional 76 lupus patients were recruited for model validation. The patients were visited in the 0th, 1st, 3rd, 6th and 12th months, and the error within 7 days was allowed for each time. The contents include the medical history, blood routine testing, urine routine testing, vital organ function testing, and BLIAG（British Isles Lupus Assessment Group index）evaluation. If any patients died, the time and cause of death were recorded. The effect evaluation index mainly include the following four aspects: (1) Epidemiological questionnaire: face-to-face inquiries by uniformly trained health professionals, mainly including the socio-demographic information of the research subjects (including gender, date of birth, marital status, education level, occupation, income level, height and weight, etc.), environmental exposure factors (including smoking, alcohol and other drug use, etc.), medical history (hypertension, diabetes, liver and kidney and other important organ diseases) and historical medical information. (2) Clinical features: the clinical manifestations (buccal erythema, skin erythema, rash, oral ulcer, arthritis, digestive, nervous and blood system involvement) of the patients were observed and examined face to face by clinicians in the department of rheumatology. (3) Test items: including general blood routine examination, urine routine (urine protein, urine protein-creatinine ratio, etc.), biochemical indexes (total bilirubin, creatinine, C-reactive protein, etc.) detection, immune routine (complement and immune globulin, etc.) and echocardiography, chest CT and other data. (4) Gene expression detection: According to the previous multi-omics research, screen out the gene sets that are different from normal people and have significant changes before and after treatment (see the research basis for details). The patients' RNA was extracted and reverse transcribed into cDNA（complementary deoxyribonucleic acid）, and the expression levels of related genes at different periods were detected by PCR（Polymerase Chain Reaction） array technology. The electronic medical record report form is used uniformly for data management. In the early stage, we have established the Jiangsu Provincial Lupus Research Database Entry System to store the data of this study. The data entry and modification shall be completed by the researcher, the data shall be traceable and consistent with the original documents. Any observation and inspection results in the trial should be timely, correct, complete, clear, standardized and true. The data administrator (a member of the team statistics) is responsible for reviewing and managing the entered data. For questions about the data, the data administrator will send corresponding questions to the researcher, and the researcher will respond to the questions sent by the data administrator in time. The data administrator can question again when necessary. All subjects' information will be kept strictly confidential. Research data are also confidential. SAS（STATISTICAL ANALYSIS SYSTEM）9.4 and R software were used to process and analyze the data. The cleaning of the dataset mainly includes: a) For covariates with missing values in the dataset, excluding covariates with missing values greater than 30% and using bagging trees to fill in missing values; b) Two variables with a high degree of correlation (correlation coefficient \> 0.9), excluding variables with more missing values; c) Excluding variables whose variance is 0 or close to 0, the rough calculation principle is that the frequency of the unique value is too small relative to the whole (10% in this study), and the ratio of the most frequent value to the frequency of the sub-multiple value is greater than 20; d) Box-Cox transformation was performed for non-normally distributed continuous variables. Different algorithms in machine learning are used to select features and construct models, and the prediction ability of different models is compared to obtain the optimal model. In addition to the classical logistic regression method, we also used some common methods to deal with high-dimensional data, such as linear discriminant analysis considering the linear relationship between covariates and outcomes, partial least squares regression, multiple adaptive regression spline method and elasticity network (EN). Considering that many clinical features and outcomes in clinical medicine show nonlinear correlation, we also use k-Nearest neighbors, Adaptive Boosting, support vector machine, random forest and neural network method to build a predictive model. The variance inflation factor is used to judge the collinearity problem. Different indicators are used to reflect the predictive ability of the model from multiple perspectives, and the C statistic is calculated to evaluate the predictive ability of the built model. The comprehensive judgment improvement index is used to judge the improvement of the model after the introduction of new variables. And the decision curve is drawn to find a model that predicts the largest net benefit. Before the start of the trial, the trial staff explained the informed consent form to each participant participating in the trial in an easy-to-understand manner, and obtained the written informed consent form of the participant voluntarily participating in the trial. It is guaranteed that participants can refuse to participate in this trial or withdraw from this trial at any time during the progress of the trial, and the rights and interests of subjects will not be affected in any way.