Background: In the typical setting of gene-selection problems from high-dimensional data, e.g., gene expression data from microarray or next-generation sequencing-based technologies, an enormous volume of high-throughput data is generated, and there is often a need for a simple, computationally-inexpensive, non-parametric screening procedure than can quickly and accurately find a low-dimensional variable subset that preserves biological information from the original very high-dimensional data (dimension p > 40,000). This is in contrast to the very sophisticated variable selection methods that are computationally expensive, need pre-processing routines, and often require calibration of priors.
Results: We present a tree-based sequential CART (S-CART) approach to variable selection in the binary classification setting and compare it against the more sophisticated procedures using simulated and real biological data. In simulated data, we analyze S-CART performance versus (i) a random forest (RF), (ii) a fully-parametric Bayesian stochastic search variable selection (SSVS), and (iii) the moderated t-test statistic from the LIMMA package in R. The simulation study is based on a hierarchical Bayesian model, where dataset dimensionality, percentage of significant variables, and substructure via dependency vary. Selection efficacy is measured through false-discovery and missed-discovery rates. In all scenarios, the S-CART method is seen to consistently outperform SSVS and RF in both speed and detection accuracy. We demonstrate the utility of the S-CART technique both on simulated data and in a control-treatment mouse study. We show that the network analysis based on the S-CART-selected gene subset in essence recapitulates the biological findings of the study using only a fraction of the original set of genes considered in the study's analysis.
Conclusions: The relatively simple-minded gene selection algorithms like S-CART may often in practical circumstances be preferred over much more sophisticated ones. The advantage of the "greedy" selection methods utilized by S-CART and the likes is that they scale well with the problem size and require virtually no tuning or training while remaining efficient in extracting the relevant information from microarray-like datasets containing large number of redundant or irrelevant variables.
Availability: The MATLAB 7.4b code for the S-CART implementation is available for download from https://neyman.mcg.edu/posts/scart.zip.