Multivariate genome-wide association study models to improve prediction of CROHN’S disease risk and identification of potential novel variants

Crohn's disease (CD) is a type of inflammatory bowel disease (IBD) that affects the gastrointestinal tract with diverse symptoms. At present, genome-wide association studies (GWAS) has discovered more than 140 genetic loci associated with CD from several datasets. Using the usual univariate GWAS methods, researchers have discovered common variants with small effects. Univariate methods assume independence among the variants that miss subtle combinatorial signals. Multivariate approaches have improved risk prediction and have complemented univariate methods for elucidating the etiology of complex traits and potential novel associations. However, the current multivariate models for CD have been assessed for three datasets (published from 2006 to 2008) under unrelated methodological settings showing a broad performance spectrum. Notably, these multivariate studies do not analyze potential novel variants. Here, we aimed to perform a robust multivariate analysis of a CD dataset different from the one commonly used, and we used the information yielded by the models to identify whether the generated models could provide additional information about the potential novel variants of CD. Therefore, we compared different multivariate methods and models, LASSO (least absolute shrinkage and selection operator), XGBoost, random forest (RF), Bootstrap stage-wise model selection (BSWiMS), and LDpred, using a strict random subsampling approach to predict the CD risk using a recent GWAS dataset, United Kingdom IBD IBD Genetics Consortium (UKIBDGC), made available in 2017, that had not been used for CD prediction studies. In addition, we assessed the effect of common strategies by increasing and decreasing the number of single-nucleotide polymorphism (SNP) markers (using genotype imputation and linkage disequilibrium (LD)–clumping). We found that the LDpred model without any imputation was the best model among all the tested models for predicting the CD risk (area under the receiver operating characteristic curve (AUROC) = 0.667 ± 0.024) in this dataset. We validated the best models using a second dataset (National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) IBD Genetics Consortium, which was previously used in CD prediction studies) in which LDpred was also the best method with a similar performance (AUROC = 0.634 ± 0.009). Based on the importance of the variants yielded by the multivariate models, we identified an unnoticed region within chromosome 6, tagged by SNP rs4945943; this region was close to the gene MARCKS, which appeared to contribute to CD risk. This research is the first multivariate prediction analysis applied to the UKIBDGC dataset. Our robust multivariate setting analysis enabled us to identify a potential variant that contributed to the CD risk. Multivariate methods are valuable tools for identifying genes that contribute to disease risk. • Multivariate models allow ranking variants, according to their contribution to the disease-risk prediction. • LDpred performed better to predict CD-risk, compared with other multivariate and the common polygenic risk score (PRS) analysis. • The LDpred model without imputation was the best model to predict CD risk (AUROC = 0.667 ± 0.024) in the UKIBDGC dataset. • An unnoticed region was identified, within chr 6 (SNP rs4945943) close to gene MARCKS, likely to contribute to CD risk.