Best Practices of Feature Selection in Multi-Omics Data

Abstract

With the recent advances in molecular biology techniques such as next-generation sequencing, mass-spectrometry, etc., a large omic data is produced. Using such data, the expression levels of thousands of molecular features (genes, proteins, metabolites, etc.) can be quantified and associated with diseases. The fact that multiple omics data contains different types of data and the number of analyzed variables increases the complexity of the models created with machine learning methods. In addition, due to many variables, the investigation of molecular variables associated with diseases is very costly. Therefore, selecting the informative and disease-related molecular features is applicable before model training and evaluation. This feature selection step is essential for obtaining accurate and generalizable models in minimum time with minimum cost. Some current methods used for feature selection are as follows: recursive feature elimination, information gain, minimum redundancy maximum relevance (mRMR), boruta, altmann, and lasso.