Environmental and omics-related marker panels for the prediction of autoantibody positivity through integrated machine learning feature selection

Backgroud: A significant gap in the study of the onset of the autoimmune response for Type 1 Diabetes (T1D) and its progression is the lack of biomarkers that can be used to accurately predict and monitor these processes. Therefore, two goal of The Environmental Determinants of Diabetes in the Young (TEDDY) study are to identify new biomarkers and obtain a mechanistic understanding of the autoimmune response leading to T1D. Methods: As a first step towards achieving these goals, we have performed integrative machine learning of multiple omics datasets (genomics, metabolomics and lipidomics), as well as associated meta-data, on a nested case-control cohort from TEDDY, a prospective study of children higher genetic risk of developing diabetes. We applied a pipeline that first identified the best machine learning classification algorithm per data type. We subsequently performed feature selection in the context of a Bayesian integration across the multiple data sources, cross-validation, and optimization via simulated annealing generating a likelihood that each feature is important to the panel predicting seroconversion. A hold-out set of 25% of the total number of available case-control pairs was used to validate the model. Results: We performed predictive modeling with the goal of identifying the features that separate cases from controls in increments of 3, 6, 9 and 12 months prior to seroconversion. The number of case-control pairs with adequate data across the multiple omics and demographic data ranged from 208 to 336 for the various time points. The matched cross-validated training sets returned an average area under the Receiver Operating Characteristic curve of over 0.71 and the feature sets included dozens of disparate markers that were relatively uniform across the various data sources. This proof-of-principle demonstrates the ability to identify candidate biomarker panels for prediction of time to seroconversion.