Abstract
The development of multigene classifiers for cancer prognosis, treatment prediction, molecular subtypes or clinicopathological groups has been a cornerstone in transcriptomic analyses of human malignancies for nearly two decades. However, many reported classifiers are critically limited by different preprocessing needs like normalization and data centering. In response, a new breed of classifiers, single sample predictors (SSPs), has emerged. SSPs classify samples in an N-of-1 fashion, relying on, e.g. gene rules comparing expression values within a sample. To date, several methods have been reported, but there is a lack of head-to-head performance comparison for typical cancer classification problems, representing an unmet methodological need in cancer bioinformatics. To resolve this need, we performed an evaluation of two SSPs [k-top-scoring pair classifier (kTSP) and absolute intrinsic molecular subtyping (AIMS)] for two case examples of different magnitude of difficulty in non-small cell lung cancer: gene expression–based classification of (i) tumor histology and (ii) molecular subtype. Through the analysis of ~2000 lung cancer samples for each case example (n = 1918 and n = 2106, respectively), we compared the performance of the methods for different sample compositions, training data set sizes, gene expression platforms and gene rule selections. Three main conclusions are drawn from the comparisons: both methods are platform independent, they select largely overlapping gene rules associated with actual underlying tumor biology and, for large training data sets, they behave interchangeably performance-wise. While SSPs like AIMS and kTSP offer new possibilities to move gene expression signatures/predictors closer to a clinical context, they are still importantly limited by the difficultness of the classification problem at hand.
Author Correction: Molecular subtypes of small cell lung cancer: a synthesis of human and mouse model data
