Statistical Methods Inspired by Challenges in Pediatric Cancer Multi-omics

Abstract

Pediatric cancer multi-omics is a uniquely rewarding and challenging domain of biomedical research. Public generosity bestows an abundance of resources for the study of extremely rare diseases; this unique dynamic creates a research environment in which problems with high-dimension and low sample size are commonplace. Here, we present a few statistical methods that we have developed for our research setting and believe will prove valuable in other biomedical research settings as well. The genomic random interval (GRIN) method evaluates the loci and frequency of genomic abnormalities in the DNA of tumors to identify genes that may drive the development of malignancies. The association of lesions with expression (ALEX) method evaluates the impact of genomic abnormalities on the RNA transcription of nearby genes to inform the formulation of biological hypotheses on molecular mechanisms. The projection onto the most interesting statistical evidence (PROMISE) method identifies omic features that consistently associate with better prognosis or consistently associate with worse prognosis across multiple measures of clinical outcome. We have shown that these methods are statistically robust and powerful in the statistical bioinformatic literature and successfully used these methods to make fundamental biological discoveries that have formed the scientific rationale for ongoing clinical trials. We describe these methods and illustrate their application on a publicly available T-cell acute lymphoblastic leukemia (T-ALL) data set. A companion github site (https://github.com/stjude/TALL-example ) provides the R code and data necessary to recapitulate the example data analyses of this chapter.