基于随机森林变量重要性评分的变量筛选方法及其在肿瘤分型诊断中的应用

王文杰; 马金沙; 高倩; 王彤

doi:10.16462/j.cnki.zhjbkz.2023.03.005

基于随机森林变量重要性评分的变量筛选方法及其在肿瘤分型诊断中的应用

doi: 10.16462/j.cnki.zhjbkz.2023.03.005

030001 太原，山西医科大学公共卫生学院卫生统计学教研室

基金项目:

国家自然科学基金 81872715

国家自然科学基金 82073674

山西省科技重大专项项目 202005D121008

山西省重点研发计划项目 202102130501003

详细信息

通讯作者:
王彤，E-mail: tongwang@sxmu.edu.cn

中图分类号: R181.3; R733.4
计量
- 文章访问数: 572
- HTML全文浏览量: 237
- PDF下载量: 81
- 被引次数: 0
出版历程
- 收稿日期: 2022-02-18
- 修回日期: 2022-05-23
- 网络出版日期: 2023-04-04
- 刊出日期: 2023-03-10

Variable selection methods based on variable importance measurement from random forest and its application in diagnosis of tumor typing

Department of Health Statistics, School of Public Health, Shanxi Medical University, Taiyuan 030001, China

Funds:

National Natural Science Foundation of China 81872715

National Natural Science Foundation of China 82073674

Major Science and Technology Project of Shanxi Province 202005D121008

Major Science and Technology Project of Shanxi Province 202102130501003

More Information

Corresponding author: WANG Tong, E-mail: tongwang@sxmu.edu.cn

摘要

摘要: 目的探究高维组学数据中结局为二分类时基于随机森林(random forest, RF)变量重要性评分的变量筛选方法，并选择合适方法构建结局预测模型。方法首先根据不同的变量筛选目标，对最小优化变量筛选类RF算法[递归特征消除(recursive feature elimination, RFE)-RF、biosigner]与全部相关变量筛选类RF算法(Boruta、vita、altmann、r2vim)在高维数据中识别重要变量的能力进行了模拟比较。然后结合不同方法优势用于弥漫大B细胞淋巴瘤(diffuse large B-cell lymphoma, DLBCL)分型相关基因的筛选，并构建DLBCL分型诊断模型。结果模拟研究表明，vita方法的灵敏度较高，biosigner方法的阳性预测值较高。实例分析表明，经vita方法筛得1 019个与DLBCL分型相关的基因，后经biosigner方法筛得77个与DLBCL分型相关的基因。所建DLBCL分型诊断模型的受试者工作特征(receiver operating characteristical, ROC)曲线下面积(area under the ROC curve，AUC)为0.910。结论 vita及biosigner方法可用于DLBCL分型相关基因的初步和最终筛选阶段。由最终筛得基因所建立的模型可有效实现DLBCL的分型诊断。
- 随机森林 /
- 变量筛选 /
- 弥漫大B细胞淋巴瘤
Abstract: Objective To explore the variable selecting methods based on variable importance measurement from random forest (RF) for binary outcome in the high-dimensional omics data, and to choose the appropriate methods to construct the outcome prediction model. Methods First, according to the different variable selection objectives, we simulated and compared the ability of minimum optimized variable selection RF methods [recursive feature elimination (RFE)-RF, biosigner] and all relevant variable selection RF methods (Boruta, vita, altmann and r2vim) to identify important variables in high-dimensional data. Then we combined different methods to select genes related to diffuse large B cell lymphoma (DLBCL) classification and constructed the model for diffuse large B cell lymphoma classification diagnosis. Results Simulation study showed that vita had higher sensitivity, and biosigner had higher positive predictive value. Empirical study showed that a total of 1 019 genes related to DLBCL classification were obtained by vita method, and 77 genes related to DLBCL classification were obtained by biosigner method. The area under the receiver operating characteristical (ROC) curve (AUC) of the DLBCL typing diagnostic model was 0.910. Conclusions Vita and biosigner can be used in the preliminary and final selecting stages of genes related to DLBCL classification. The model we developed can effectively distinguish the different subtypes of DLBCL.
- Random forest /
- Variable selection /
- Diffuse large B-cell lymphoma

HTML全文

图 1 各模拟情景下筛得总变量数目

Figure 1. The total number of selected variables in each simulated scenario

下载: 全尺寸图片幻灯片

图 2 各模拟情景下筛得变量强相关灵敏度

Figure 2. Sensitivity of selected strongly relevant variables in each simulated scenario

下载: 全尺寸图片幻灯片

图 3 各模拟情景下筛得变量弱相关灵敏度

Figure 3. Sensitivity of selected weakly relevant variables in each simulated scenario

下载: 全尺寸图片幻灯片

图 4 各模拟情景下筛得变量的阳性预测值

Figure 4. Positive predictive value of selected variables in each simulated scenario

下载: 全尺寸图片幻灯片

图 5 原始真实结局及模型预测结局情况下的Kaplan-Meier曲线

Figure 5. Kaplan-Meier curves in the case of the original true outcome and the predicted outcome

下载: 全尺寸图片幻灯片

参考文献(26)

[1]	Kohavi R, John GH. Wrappers for feature subset selection[J]. Artif Intell, 1997, 97(1-2): 273-324. DOI: 10.1016/S0004-3702(97)00043-X.
[2]	Tang YC, Zhang YQ, Huang Z. Development of two-stage SVM-RFE gene selection strategy for microarray expression data analysis[J]. IEEE/ACM Trans Comput Biol Bioinform, 2007, 4(3): 365-381. DOI: 10.1109/TCBB.2007.70224.
[3]	Stańczyk U, Jain LC. Feature selection for data and pattern recognition[M]. Berlin: Springer, 2015: 12-14.
[4]	Nilsson R, Peña JM, Björkegren J, et al. Detecting multivariate differentially expressed genes[J]. BMC Bioinformatics, 2007, 8: 150. DOI: 10.1186/1471-2105-8-150.
[5]	Ein-Dor L, Kela I, Getz G, et al. Outcome signature genes in breast cancer: is there a unique set?[J]. Bioinformatics, 2005, 21(2): 171-178. DOI: 10.1093/bioinformatics/bth469.
[6]	张鼎, 赵亚双. 生物信息学在分子流行病学中的应用[J]. 中华疾病控制杂志, 2021, 25(1): 20-24. DOI: 10.16462/j.cnki.zhjbkz.2021.01.005. Zhang D, Zhao YS. Applications of bioinformatics in molecular epidemiology[J]. Chin J Dis Control Prev, 2021, 25(1): 20-24. DOI: 10.16462/j.cnki.zhjbkz.2021.01.005.
[7]	Breiman L. Random forests[J]. Mach Learn, 2001, 45(1): 5-32. DOI: 10.1023/A:1010933404324.
[8]	Nicodemus KK, Malley JD, Strobl C, et al. The behaviour of random forest permutation-based variable importance measures under predictor correlation[J]. BMC Bioinformatics, 2010, 11: 110. DOI: 10.1186/1471-2105-11-110.
[9]	Degenhardt F, Seifert S, Szymczak S. Evaluation of variable selection methods for random forests and omics data sets[J]. Brief Bioinform, 2019, 20(2): 492-503. DOI: 10.1093/bib/bbx124.
[10]	Wu XY, Wu ZY, Li K. Identification of differential gene expression for microarray data using recursive random forest[J]. Chin Med J (Engl), 2008, 121(24): 2492-2496. DOI: 10.3238/arztebl.2008.0900a.
[11]	Rinaudo P, Boudah S, Junot C, et al. Biosigner: a new method for the discovery of significant molecular signatures from omics data[J]. Front Mol Biosci, 2016, 3: 26. DOI: 10.3389/fmolb.2016.00026.
[12]	Kursa MB, Jankowski A, Rudnicki WR. Boruta-a system for feature selection[J]. Fundam Informaticae, 2010, 101(4): 271-285. DOI: 10.3233/fi-2010-288.
[13]	Altmann A, Toloşi L, Sander O, et al. Permutation importance: a corrected feature importance measure[J]. Bioinformatics, 2010, 26(10): 1340-1347. DOI: 10.1093/bioinformatics/btq134.
[14]	Szymczak S, Holzinger E, Dasgupta A, et al. r2VIM: a new variable selection method for random forests in genome-wide association studies[J]. BioData Min, 2016, 9: 7. DOI: 10.1186/s13040-016-0087-3.
[15]	Janitza S, Celik E, Boulesteix AL. A computationally fast variable importance test for random forests for high-dimensional data[J]. Adv Data Anal Classif, 2018, 12(4): 885-915. DOI: 10.1007/s11634-016-0276-4.
[16]	Shin M, Bhattacharya A, Johnson VE. Scalable Bayesian variable selection using nonlocal prior densities in ultrahigh-dimensional settings[J]. Stat Sin, 2018, 28(2): 1053-1078. DOI: 10.5705/ss.202016.0167.
[17]	Yan WH, Jiang XN, Wang WG, et al. Cell-of-origin subtyping of diffuse large B-cell lymphoma by using a qPCR-based gene expression assay on formalin-fixed paraffin-embedded tissues[J]. Front Oncol, 2020, 10: 803. DOI: 10.3389/fonc.2020.00803.
[18]	Alizadeh AA, Eisen MB, Davis RE, et al. Distinct types of diffuse large B-cell lymphoma identified by gene expression profiling[J]. Nature, 2000, 403(6769): 503-511. DOI: 10.1038/35000501.
[19]	Reif DM, Motsinger-Reif AA, McKinney BA, et al. Integrated analysis of genetic and proteomic data identifies biomarkers associated with adverse events following smallpox vaccination[J]. Genes Immun, 2009, 10(2): 112-119. DOI: 10.1038/gene.2008.80.
[20]	Salzer U, Chapel HM, Webster AD, et al. Mutations in TNFRSF13B encoding TACI are associated with common variable immunodeficiency in humans[J]. Nat Genet, 2005, 37(8): 820-828. DOI: 10.1038/ng1600.
[21]	Kralickova P, Milota T, Litzman J, et al. CVID-associated tumors: Czech nationwide study focused on epidemiology, immunology, and genetic background in a cohort of patients with CVID[J]. Front Immunol, 2019, 9: 3135. DOI: 10.3389/fimmu.2018.03135.
[22]	Inamo J, Suzuki K, Takeshita M, et al. Identification of novel genes associated with dysregulation of B cells in patients with primary Sjögren's syndrome[J]. Arthritis Res Ther, 2020, 22(1): 153. DOI: 10.1186/s13075-020-02248-2.
[23]	Liu JQ, Yao YL, Hu ZY, et al. Transcriptional profiling of long-intergenic noncoding RNAs in lung squamous cell carcinoma and its value in diagnosis and prognosis[J]. Mol Genet Genomic Med, 2019, 7(12): e994. DOI: 10.1002/mgg3.994.
[24]	Blenk S, Engelmann J, Weniger M, et al. Germinal center B cell-like (GCB) and activated B cell-like (ABC) type of diffuse large B cell lymphoma (DLBCL): analysis of molecular predictors, signatures, cell cycle state and patient survival[J]. Cancer Inform, 2007, 3: 399-420. DOI: 10.1177/117693510700300004.
[25]	Wood O, Woo J, Seumois G, et al. Gene expression analysis of TIL rich HPV-driven head and neck tumors reveals a distinct B-cell signature when compared to HPV independent tumors[J]. Oncotarget, 2016, 7(35): 56781-56797. DOI: 10.18632/oncotarget.10788.
[26]	Coutinho R, Clear AJ, Owen A, et al. Poor concordance among nine immunohistochemistry classifiers of cell-of-origin for diffuse large B-cell lymphoma: implications for therapeutic strategies[J]. Clin Cancer Res, 2013, 19(24): 6686-6695. DOI: 10.1158/1078-0432.CCR-13-1482.