Scalable FRaC Variants: Anomaly Detection for Precision Medicine

摘要

The FRaC anomaly detection algorithm has been previously used to identify anomalous mRNA expression patterns, and has served as the core of an approach that characterizes individual anomalies by identifying dysregulated molecular functions. However, FRaC operates by training supervised models for each feature in a data set. Thus, scaling to substantially larger data sets, such as those reflecting common sequence variants, would require prohibitive amounts of computation time and memory. Additionally, although FRaC is designed to be relatively robust to irrelevant variables, it is not perfectly so; due to the low sample sizes and large number of variables inmolecular data sets, substantially increasing the number of features beyond those in gene expression data sets raises the possibility of overwhelming the signal with noise. In this paper, we examine the scalability of FRaC variants using different feature reduction methods. We demonstrate that it is possible to preserve the anomaly detection accuracy of the original FRaC algorithm while requiring considerably fewer computational resources, allowing these methods to scale to handle other types of genomic data.