Data mining and machine learning approaches for the integration of genome-wide association and methylation data: methodology and main conclusions from GAW20

Abstract

BackgroundMultiple layers of genetic and epigenetic variability are being simultaneously explored in an increasing number of health studies. We summarize here different approaches applied in the Data Mining and Machine Learning group at the GAW20 to integrate genome-wide genotype and methylation array data.ResultsWe provide a non-intimidating introduction to some frequently used methods to investigate high-dimensional molecular data and compare the different approaches tried by group members: random forest, deep learning, cluster analysis, mixed models, and gene-set enrichment analysis. Group contributions were quite heterogeneous regarding investigated data sets (real vs simulated), conducted data quality control and assessed phenotypes (eg, metabolic syndrome vs relative differences of log-transformed triglyceride concentrations before and after fenofibrate treatment). However, some common technical issues were detected, leading to practical recommendations.ConclusionsDifferent sources of correlation were identified by group members, including population stratification, family structure, batch effects, linkage disequilibrium and correlation of methylation values at neighboring cytosine-phosphate-guanine (CpG) sites, and the majority of applied approaches were able to take into account identified correlation structures. The ability to efficiently deal with high-dimensional omics data, and the model free nature of the approaches that did not require detailed model specifications were clearly recognized as the main strengths of applied methods. A limitation of random forest is its sensitivity to highly correlated variables. The parameter setup and the interpretation of results from deep learning methods, in particular deep neural networks, can be extremely challenging. Cluster analysis and mixed models may need some predimension reduction based on existing literature, data filtering, and supplementary statistical methods, and gene-set enrichment analysis requires biological insight.

Personalized prediction of delayed graft function for recipients of deceased donor kidney transplants with machine learning

“…After removal of records with missing values and/or outliers, RFE-RF was applied to the processed dataset to reveal the predictors for DGF. In brief, RFE-RF ranks features by VIS, which is calculated based on final predictive accuracy and determines the optimal number of predictors in an arbitrarily pre-defined search space 61 . In this study, we tested 5, 10, 15, 20, 25, 30, and 126 features with tenfold cross-validation and used mean ROC-AUC as the performance metric.…”

Identification of m6a-related signature genes in esophageal squamous cell carcinoma by machine learning method

“…Meanwhile, it is prone to underfitting, the accuracy is not very high, and it is unable to handle a large number of multi-class features or variables well. Based on these, more and more machine learning methods are used in the field of medical statistics ( Watson et al, 2019 ), ( Darst et al, 2018 ). In this study, we also directly chose machine learning instead of logistic regression analysis as the main method to screen ESCC-related m6A regulators.…”

Fibrate Pharmacogenomics: Expanding Past the Genome

“…As the focus of the workshop, simulated data based on the GOLDN cohort was used to rigorously evaluate the statistical methods used, and the real data were used to deliver new insights into fibrate response [95]. While a review of all the statistical methods used in the workshop is beyond the scope of this paper, Cherlin et al [96] and Darst et al [97] provide excellent perspectives on the diversity of methods employed. The following outlines some of the major findings of the 20th GAW.…”