AGA: Interactive pipeline for reproducible gene expression and DNA methylation data analyses

Introduction

While high dimensional genetic data have increased in availability at reduced cost, robust analyses remain labor intensive and costly. Numerous automated software pipelines have been developed in an effort to increase the rate and decrease the costs at which analyses can be completed, including SVAw¹¹, Partek⁴, InSilicoDB¹⁸. Automated Genomics Analysis (AGA) provides a more dynamic experience, allowing the user to start with raw data and a text file containing corresponding sample annotations from either a single or multiple studies. AGA performs all necessary normalization and batch correction, and then enables the user to interactively determine the samples to contrast in the analysis based on the sample annotations. AGA is implemented in R to facilitate adaptation of state-of-the art genomics analysis techniques. Linking R to a web browser-based interface through RStudio’s shiny also facilitates collaborative analyses in research teams with diverse bioinformatics expertise.

AGA bridges the gap between interactive and reproducible analyses for several platforms, including expression arrays, methylation arrays, and processed RNAseq data. Through the interface, the user determines the size and scope of the analyses. AGA first performs data normalization, including the ComBat⁷ and SVA⁹ batch correction algorithms to enable comparison across multiple datasets for non-methylation platforms. The software then performs differential analysis¹⁶, and gene set analyses^2,17 based upon defined sample groups. Users obtain standard visualization of genomics data, including hierarchical clustering, boxplots and heatmaps as part of the default analysis. Plots and tables summarizing the results from each analysis are customizable through the interface. The figures and tables in AGA are interactive and customizable. In contrast to other point and click software, AGA logs the R code, and exports the workspace with each figure and table, ensuring that each analysis can be reproduced and further customized. The runtime of analyses will depend largely on the desktop hardware, but also on the data platform and optional analyses selected. On a Mac Pro workstation, containing a 3.2 GHz Quad-Core Intel Xeon processor and 10Gb 1066 MHz DDR3 RAM, analyses containing under 100 samples were completed in under 30 minutes.

Methods

The AGA application is run through R and interactive through web browsers. AGA is implemented with RStudio’s shiny¹³, integrating the R code used in the analysis with HTML and JavaScript, for the interactive user interface. Usage requires R version 3.0.1 or higher, and either Mozilla Firefox or Google Chrome, and R packages described in the AGA User’s Manual. The program is divided into seven tabs. Clicking the respective Update button generates the results to be displayed in each tab and clicking the Download buttons save the plots and data.

Example

As an example, we perform analysis on sample datasets containing gene expression of primary head and neck squamous cell carcinoma (HNSCC) tumors. We downloaded measurements from a combination of frozen tumor samples from two distinct studies in GEO available under accession numbers GSE10300³ and GSE6791¹², representing two distinct batches. Raw CEL files and annotation csv files were obtained as described in the User’s manual. We initialize AGA by selecting the directory containing these data. Once loaded, we check the HPV and Tumor.Source.Type columns to group the samples into primary HPV-positive and HPV-negative tumors for differential expression analysis. We then click “Run the Analysis” to normalize the CEL files with RMA⁶, batch correct the data with ComBat and SVA, and perform differential expression analysis. The plot in the Dendrogram Plot tab confirms that the batch effects are apparent between these datasets but removed after batch. The heatmap generated in the Heatmap Plot tab (Figure 1) demonstrates that the batch correction nonetheless preserves gene expression difference between HPV-positive and HPV-negative tumors. Moreover, performing differential expression analysis comparing HPV-positive and HPV-negative HNSCC in the “Differential Analysis” tab confirms the well-established overexpression (p=8.74e-9) of CDKN2A (p16) in HPV-positive HNSCC^14,15.

Figure 1. Heatmap displaying the relative expression of the probes with p values below 0.0001 from the example analysis, including CDKN2A.

We note that sample names are truncated in the heatmap, but users can reduce the lengths of sample names or ensure that sample identity can be determined by the final characters in the name to associate specific samples with the heatmap.

Discussion

AGA provides an interface to enable users who may be unfamiliar with R to perform reproducible genomics class comparison analysis. Unlike other automated pipelines, experienced R users can reproduce, extend or modify preliminary analyses. Thus, AGA facilitates collaborations between novice and expert R users for genomics analysis. Future work will extend the AGA pipeline to encode normalization routines to DNA methylation, and analysis routines for other genomics platforms, including copy number data.

Software availability