1 Introduction

This document provides an introduction of the TCGAbiolinksGUI.data, which contains supporting data for TCGAbiolinksGUI (Silva et al. 2017).

This package contains the following objects:

  • For gene annotation
  • gene.location.hg38
  • gene.location.hg19
  • For Glioma Classifier function
    • glioma.gcimp.model
    • glioma.idh.model
    • glioma.idhmut.model
    • glioma.idhwt.model
  • For linkedOmics database:
    • linkedOmicsData

2 Contents

2.1 Creating list of GDC projects

The code below access the NCI’s Genomic Data Commons (GDC) and get the list of available datasets for The Cancer Genome Atlas (TCGA) and Therapeutically Applicable Research to Generate Effective Treatments (TARGET) projects.

This data is in saved in the GDCdisease object.

2.2 List of MAF files

The code below downloads a manifest of open TCGA MAF files available in the NCI’s Genomic Data Commons (GDC).

This data is in saved in the maf.tumor object.

2.3 Creating Training models

Based on the article data from the article “Molecular Profiling Reveals Biologically Discrete Subsets and Pathways of Progression in Diffuse Glioma” (www.cell.com/cell/abstract/S0092-8674(15)01692-X) (Ceccarelli et al. 2016) we created a training model to predict Glioma classes based on DNA methylation signatures.

First, we will load the required libraries, control random number generation by specifying a seed and register the number of cores for parallel evaluation.

The next steps will download the DNA methylation matrix from the article: the DNA methylation data for glioma samples, samples metadata, and DNA methylation signatures.

We will get metadata with samples molecular subtypes from the paper: (www.cell.com/cell/abstract/S0092-8674(15)01692-X) (Ceccarelli et al. 2016)

Probes metadata information are downloaded from http://zwdzwd.io/InfiniumAnnotation This will be used to remove probes that should be masked from the training.

Retrieve probe signatures from the paper.

With the data and metadata available we will create one model for each signature. The code below selects the DNA methylation values for a given set of signatures (probes) and uses the classification of each sample to create a Random forest model. Each model is described in the next subsections.

2.3.1 RF to classify between IDHmut and IDHwt

Parameters Values
trainingset whole TCGA panglioma cohort
probes signature 1,300 pan-glioma tumor specific
groups to be classified LGm1, LGm2, LGm3, LGm4, LGm5, LGm6
metadata column Pan-Glioma DNA Methylation Cluster

We will start by preparing the training data. We will select the probes signatures for the group classification from the excel file (for this case the 1,300 probes) and the samples that belong to the groups we want to create our model (in this case “LGm2” “LGm5” “LGm3” “LGm4” “LGm1” “LGm6”).

The DNA methylation matrix will be subset to the DNA methylation signatures and samples with classification.

Probes that should be masked, will be removed.

We will merge the samples with their classification. In the end, we will have samples in the row, and probes and classification as columns.

After the data prepared we will start the RF training by selecting tuning the mtry argument, which defines the number of variables randomly sampled as candidates at each split. We will use the function tuneRF to optimal mtry from values going from sqrt(p) where p is number of probes in the data up to 2 * sqrt(p) (see tuneRF function for more information). RF will use all mtry values and use the one that produced the best model.

We will use the training data to create our Random forest model. First, we will set up a repeated k-fold cross-validation. We set k=10, which will split the data into 10 equal size sets. Of the 10 set, one is retained as the validation data for testing the model, and the remaining 9 sets are used as training data. The cross-validation process is then repeated 10 times (the folds), each time using a different set as the validation data. The 10 results from the folds can then be averaged (or otherwise combined) to produce a single estimation. This procedure will be repeated 10 times, with different 10 equal size sets. The final model accuracy is taken as the mean from the number of repeats.

Create our Random forest model using the train frunction from the caret package.

The IDH model are saved in glioma.idh.model object.

## Random Forest 
## 
##  880 samples
## 1205 predictors
##    6 classes: 'LGm1', 'LGm2', 'LGm3', 'LGm4', 'LGm5', 'LGm6' 
## 
## No pre-processing
## Resampling: Cross-Validated (10 fold, repeated 10 times) 
## Summary of sample sizes: 792, 793, 792, 791, 792, 794, ... 
## Resampling results across tuning parameters:
## 
##   mtry  Accuracy   Kappa    
##    34   0.9166862  0.8928240
##    68   0.9171602  0.8935333
##    87   0.9169331  0.8932650
##   104   0.9168127  0.8931906
##   122   0.9179471  0.8946451
##   156   0.9189363  0.8959692
##   260   0.9183925  0.8953602
##   295   0.9190913  0.8962952
##   382   0.9189609  0.8961085
##   451   0.9154555  0.8916298
## 
## Accuracy was used to select the optimal model using the largest value.
## The final value used for the model was mtry = 295.

2.3.2 RF to classify IDHmut specific clusters

Parameters Values
trainingset TCGA IDHmut only
probes signature 1,308 IDHmutant tumor specific
groups to be classified IDHmut-K1, IDHmut-K2, IDHmut-K3
metadata column IDH-specific DNA Methylation Cluster

To produce the model we will use the same code above but we will change the training data (probes and labels are different)

The IDHmut model are saved in glioma.idhmut.model object.

## Random Forest 
## 
##  450 samples
## 1216 predictors
##    3 classes: 'IDHmut-K1', 'IDHmut-K2', 'IDHmut-K3' 
## 
## No pre-processing
## Resampling: Cross-Validated (10 fold, repeated 10 times) 
## Summary of sample sizes: 405, 406, 405, 405, 404, 404, ... 
## Resampling results across tuning parameters:
## 
##   mtry  Accuracy   Kappa    
##    34   0.9611239  0.9328999
##    87   0.9689140  0.9465985
##   136   0.9689386  0.9466341
##   149   0.9689188  0.9465882
##   157   0.9686966  0.9462056
##   192   0.9695952  0.9477192
##   208   0.9700203  0.9484758
##   297   0.9704795  0.9492448
##   313   0.9702621  0.9489050
##   383   0.9695706  0.9477601
##   488   0.9666953  0.9429666
##   523   0.9669323  0.9433384
##   906   0.9580071  0.9283147
## 
## Accuracy was used to select the optimal model using the largest value.
## The final value used for the model was mtry = 297.

2.3.3 RF to classify between G-CIMP-low and G-CIMP-high

Parameters Values
trainingset TCGA IDHmut-K1 and IDHmut-K2 only
probes signature 163 probes that define each TC
groups to be classified G-CIMP-low, G-CIMP-high
metadata column Supervised DNA Methylation Cluster

To produce the model we will use the same code above but we will change the training data (probes and labels are different).

The G-CIMP model was saved in glioma.gcimp.model object.

## Random Forest 
## 
## 276 samples
## 145 predictors
##   3 classes: 'G-CIMP-high', 'G-CIMP-low', 'NA' 
## 
## No pre-processing
## Resampling: Cross-Validated (10 fold, repeated 10 times) 
## Summary of sample sizes: 249, 248, 249, 249, 248, 249, ... 
## Resampling results across tuning parameters:
## 
##   mtry  Accuracy   Kappa    
##     9   0.9785038  0.8640637
##    12   0.9781334  0.8605572
##    18   0.9770488  0.8521752
##    30   0.9763468  0.8482224
##    48   0.9756184  0.8416317
##   120   0.9715963  0.8219569
## 
## Accuracy was used to select the optimal model using the largest value.
## The final value used for the model was mtry = 9.

2.3.4 RF to classify IDHwt specific clusters

Parameters Values
trainingset TCGA IDHwt only
probes signature 914 IDHwildtype tumor specific
groups to be classified IDHwt-K1, IDHwt-K2, IDHwt-K3
metadata column IDH-specific DNA Methylation Cluster

Note: In this case, samples classified into IDHwt-K3 should be further subdivided by grade.

To produce the model we will use the same code above but we will change the training data (probes and labels are different)