Standard machine learning approaches outperform deep representation learning on phenotype prediction from transcriptomics data

摘要

The potential to tailor therapies for individual patients rests on the ability to accurately diagnose disease and predict outcomes under various treatment conditions. Predictors based on high-throughput ’omics technologies hold great promise, but a number of technical challenges have limited their applicability [ ]. Phenotypes may be complex—involving contributions from large numbers of genes—but ’omics data are so high-dimensional that exploring all possible interactions is intractable. This situation is further complicated by the small sample sizes of typical biological studies and by large systematic sources of variation between experiments [ , ]. However, recent developments in machine learning have raised hopes that new computational methods integrating data from many studies may be able to overcome these difficulties. Accurate prediction of phenotype or endpoint(s) from ’omics data would usher in an era of molecular diagnostics [ , ].