A Method for Massively Parallel Analysis of Time Series

TL;DR

The paper introduces MPATS, a scalable method for analyzing large-scale time series -omic data by quantifying perturbations through pairwise L1 distances, validated on immune response data, outperforming traditional methods.

Contribution

It presents MPATS, a novel parallel approach for transcriptome-wide time series analysis, capable of identifying biologically relevant perturbations more effectively than existing methods.

Findings

MPATS successfully identified immune response gene sets.

MPATS outperformed EDGE and GSEA-TS in relevant signature detection.

The method performs well with small sample sizes in simulations.

Abstract

Quantification of system-wide perturbations from time series -omic data (i.e. a large number of variables with multiple measures in time) provides the basis for many downstream hypothesis generating tools. Here we propose a method, Massively Parallel Analysis of Time Series (MPATS) that can be applied to quantify transcriptome-wide perturbations. The proposed method characterizes each individual time series through its $\ell_1$ distance to every other time series. Application of MPATS to compare biological conditions produces a ranked list of time series based on their magnitude of differences in their $\ell_1$ representation, which then can be further interpreted through enrichment analysis. The performance of MPATS was validated through its application to a study of IFN$\alpha$ dendritic cell responses to viral and bacterial infection. In conjunction with Gene Set Enrichment Analysis (GSEA), MPATS produced consistently identified signature gene sets of anti-bacterial and anti-viral response. Traditional methods such as EDGE and GSEA Time Series (GSEA-TS) failed to identify the relevant signature gene sets. Furthermore, the results of MPATS highlighted the crucial functional difference between STAT1/STAT2 during anti-viral and anti-bacterial response. In our simulation study, MPATS exhibited acceptable performance with small group size (n = 3), when the appropriate effect size is considered. This method can be easily adopted for other -omic data types.