Dynamic Model for RNA-seq Data Analysis

TL;DR

This paper introduces a second-order ODE model for RNA-seq data analysis that effectively captures gene expression dynamics, enabling accurate prediction, classification, and biological insight extraction from transcriptomic data.

Contribution

The study develops a novel second-order ODE-based statistical framework for modeling RNA-seq data, improving accuracy in data fitting, prediction, and classification tasks compared to existing methods.

Findings

High accuracy in gene expression prediction across genes

Effective classification of normal vs. tumor cells using ODE features

Identification of genes related to cancer through response analysis

Abstract

The newly developed deep-sequencing technologies make it possible to acquire both quantitative and qualitative information regarding transcript biology. By measuring messenger RNA levels for all genes in a sample, RNA-seq provides an attractive option to characterize the global changes in transcription. RNA-seq is becoming the widely used platform for gene expression profiling. However, real transcription signals in the RNA-seq data are confounded with measurement and sequencing errors, and other random biological/technical variation. How to appropriately take the variability due to errors and sequencing technology variation into account is essential issue in the RNA-seq data analysis. To extract biologically useful transcription process from the RNA-seq data, we propose to use the second ODE for modeling the RNA-seq data. We use differential principal analysis to develop statistical methods for estimation of location-varying coefficients of the ODE. We validate the accuracy of the ODE model to fit the RNA-seq data by prediction analysis and 5 fold cross validation. We find the accuracy of the second ODE model to predict the gene expression level across the gene is very high and the second ODE model to fit the RNA-seq data very well. To further evaluate the performance of the ODE model for RNA-seq data analysis, we used the location-varying coefficients of the second ODE as features to classify the normal and tumor cells. We demonstrate that even using the ODE model for single gene we can achieve high classification accuracy. We also conduct response analysis to investigate how the transcription process respond to the perturbation of the external signals and identify dozens of genes that are related to cancer.