A two-dimensional genome scan for rheumatoid arthritis susceptibility loci

Supervised machine learning and logistic regression identifies novel epistatic risk factors with PTPN22 for rheumatoid arthritis

Investigating genetic interactions (epistasis) has proven difficult despite the recent advances of both laboratory methods and statistical developments. With no 'best' statistical approach available, combining several analytical methods may be optimal for detecting epistatic interactions. Using a multi-stage analysis that incorporated supervised machine learning and methods of association testing, we investigated epistatic interactions with a well-established genetic factor (PTPN22 1858T) in a complex autoimmune disease (rheumatoid arthritis (RA)). Our analysis consisted of four principal stages: Stage I (data reduction)-identifying candidate chromosomal regions in 292 affected sibling pairs, by predicting PTPN22 concordance using multipoint identity-by-descent probabilities and a supervised machine learning algorithm (Random Forests); Stage II (extension analysis)-testing detailed genetic data within candidate chromosomal regions for epistasis with PTPN22 1858T in 677 cases and 750 controls using logistic regression; Stage III (replication analysis)-confirmation of epistatic interactions in 947 cases and 1756 controls; Stage IV (combined analysis)-a pooled analysis including all 1624 RA cases and 2506 control subjects for final estimates of effect size. A total of seven replicating epistatic interactions were identified. SNP variants within CDH13, MYO3A, CEP72 and near WFDC1 showed significant evidence for interaction with PTPN22, affecting susceptibility to RA.

Discussing gene-gene interaction: warning--translating equations to English may result in jabberwocky

Interest in mapping susceptibility alleles for complex diseases, which do not follow a classic single-gene segregation pattern, has driven interest in methods that account for, or use information from one locus when mapping another. Our discussion group examined methods related to epistasis or gene x gene interaction. The goal of modeling gene x gene interaction varied across groups; some papers tried to detect gene x gene interaction while others tried to exploit it to map genes. Most of the 10 papers summarized here applied newly created or newly modified statistical methods related to gene x gene interaction, while two groups primarily examined computational issues. As is often the case, comparisons are complicated by little overlap in the data used across the papers, and further complicated by the fact that the available data may not have been ideal for some gene x gene interaction methods. However, the main difficulty in comparing and contrasting methods across the papers is the lack of a consistent statistical definition of gene x gene interaction. But despite these issues, two clear trends emerged across the analyses: First, the methods for quantitative trait gene x gene interaction appeared to perform very well, even in families initially ascertained as affected sib pairs; and second, dichotomous trait gene x gene interaction methods failed to produce consistent results. The difficulty of using (primarily) affected sib pair data in a gene x gene interaction analysis is explored.