This repository includes code for processing RNA-Seq FASTQ files and clinical data from The Cancer Genome Atlas. In addition, we have included the code used for analyzing data in our manuscript, "Alternative preprocessing of RNA-Sequencing data in The Cancer Genome Atlas leads to improved analysis results" (Rahman, Mumtahena, et al. Bioinformatics 2015:10.1093/bioinformatics/btv377.

What is this repository for?

• We used the 'Rsubread' R package to align and summarize reads at the gene level for 9264 tumor and 741 normal TCGA RNA-Seq samples. The R scripts we provide here can also be used to process samples that did not come from TCGA. We have also included the code for compiling clinical data available for these tumors into a matrix format and matching the clinical IDs with the RNA-Seq IDs.
• We have provided the code and various intermediate data files that we produced in performing the analyses we describe in the manuscript.

How to normalize raw RNA-Seq data and process clinical data from TCGA

This pipeline is designed to be executed on Unix-based systems. Most of the code is written in the R programming language. But it also requires "bash" scripts to be executed at the command line.

1. Install the R statistical package. We used version 3.1.0.

2. Install the following R packages, which can be obtained using either the install.packages function in R or via the Bioconductor framework:

   • Rsubread
   • limma
   • edgeR
   • tools

3. Clone this git repository to your local computer.

4. Via dbGAP, obtain access to the raw TCGA data. Then obtain a private key that allows you download raw data via the Cancer Genomics Hub. Store this key file as cghub.key in the current directory.

5. In the Genome directory, store the reference genome file and GTF file that can be obtained from here. We used version hg19. After extracting these files, you will find the reference genome in Homo_sapiens/UCSC/hg19/Sequence/WholeGenomeFasta/genome.fa and the GTF file in Homo_sapiens/UCSC/hg19/Annotation/Archives/archive-2012-03-09-03-24-41/Genes/genes.gtf. Move these directly to the local Genome directory. Update [08/04/2020]: These files are no longer available. You can find a copy of them here. You will need to decompress the files using the gunzip utility. However, if you are going to run this pipeline now, you might consider using a newer version of the human reference genome.

6. Execute Scripts/process_tcga_rsubread at the command line to begin downloading and normalizing samples.

All the RNA-Seq and clinical data files that we have processed are available from Gene Expression Omnibus (accession numbers: GSE62820 and GSE62944).

For informational purposes, we have also provided a bash script (Scripts/process_tcga_level_3) that contains the steps for producing "Level 3" values using the same steps that are performed by the TCGA consortium. These steps are described in more detail here: https://cghub.ucsc.edu/docs/tcga/UNC_mRNAseq_summary.pdf.

Process clinical data

1. Install R package 'plyr' using the install.packages function in R.

2. Download the Clinical data for individual cancer type from [TCGA Data Portal] (https://tcga-data.nci.nih.gov/tcga/dataAccessMatrix.htm) in Biotab format.

3. Download GSE62944_06_01_15_TCGA_24_CancerType_Samples.txt.gz from GEO (Accession number GSM1536837) and save the unzipped file to 'Datasets' folder.

4. Set directory to where all the clinical data folders for each cancer type is located.

5. Run the R script at Codes/ProcessClinicalData.R.

How to reanalyze our findings

We also provide an R Markdown file (Analysis/TCGA_24_manuscript_analysis.Rmd) that contains the analysis code that we used for our manuscript. If you desire to reexecute this analysis, please complete the following steps:

1. Install the R statistical package. We used version 3.1.0.

2. Install the following R packages, which can be obtained using either the install.packages function in R or via the Bioconductor framework:

   • stats
   • ROCR
   • pROC
   • caret
   • knitr
   • data.table
   • heatmap3
   • RColorBrewer

3. We used the BinReg 2 algorithm to make HER2 signature predictions on TCGA breast cancer samples. BinReg 2 runs on the MatLab platform. We used our HER2 signature datasets as training samples and the TCGA breast cancer datasets as test samples. We used the following parameters: 200 genes, 2 metagenes, quantile normalization (-g 200 -m 2 -q) to minimize the batch effects between training and test samples. The original outputs from BinReg2 are located within the Analysis_datasets/10_14_predictions_raw directory. Rerun of the HER2 pathway prediction excluding the two less consistent HER2 training samples is located at ``Analysis_datasets/5_01_predictions_raw``` .These output predictions are summarized in the Analysis_datasets directory folder for further evaluation.

4. The code we used to classify TCGA lung adenocarcinoma and squamous carcinoma samples is in Code/Classify_luad_vs_lusc.R. The outputs of this analysis are located in the Analysis_datasets directory. The bash script describing additional analysis to identify discordant LUAD samples and differentially expressed gene is located at Code/LUSC_LUAD_discordant_analysis.

5. Use the knitr package to compile Analysis/TCGA_24_manuscript_analysis.Rmd. (It is convenient to complete this step within the RStudio environment.) Also be sure to set the working directory to Analysis_datasets. Our results are stored in the TCGA_24_manuscript_analysis.html file.

6. Our analysis datasets and outputs are available [here] (https://www.dropbox.com/sh/4e0c8u7jke694tu/AADEQnB5LbCWihb3A5f04O9va?dl=0).

Contact information

• Mumtahena Rahman. [email protected]
• Stephen R Piccolo. https://piccolo.byu.edu