CpG Pattern analysis

22 Jul 2015

Introduction

This page contains the pipeline analysis for the "CpG pattern analysis". Source code is available here. The whole analysis can be computed by running the run_all.R

Data processing

Data analysis

Part A: Analysis of average CGI+SS patterns

In the first part, we compare the average CGI+SS profiles for a given dataset (i.e Cancerous breast) across the genome to assess whether specific CGI+SS profiles exist and whether they are associated with any specific gene expression levels.

DiseaseList <- c('BRCA','COAD','LUAD')
Type <- c('Cancerous','Normal')

A.1) Filter CGI+SS with at least 20 probes:

load("../../data/processed/fData/CpGIslands_probe_size.RData")
list_big_island <- which(CpGIslands.probesize >=20)

This reduces the number of CGIs studied from 27K to 1827 CGIs.

A.2) For each type of tissue and each CGI+SS, we calculate a probewise average profile:

source('fun/calculate_Mean_PC.R')
for (DiseaseName in DiseaseList)
{
        out <- calculate_Mean_PC(Disease=DiseaseName,type="Cancerous",proc=F)
        out <- calculate_Mean_PC(Disease=DiseaseName,type="Normal",proc=F)
}

A.3) We then perform dynamic time warping to assess for a given tissue and type (normal or cancerous) the distance between two different CGI+SS profiles:

source('fun/calculate_Mean_PC.R')
for (DiseaseName in DiseaseList)
{
        out <- calculate_Mean_PC(Disease=DiseaseName,type="Cancerous",proc=F)
        out <- calculate_Mean_PC(Disease=DiseaseName,type="Normal",proc=F)
}

This outputs a 1827 x 1827 matrix that gives the DTW distance between all CGI+SS profiles

A.4) We then perform a hierarchical clustering (linkage=Ward):

source('fun/CGI_analysis.R')
for (DiseaseName in DiseaseList)
{
        out <- analyze_CGI_clusters(Disease=DiseaseName,cutoff=3,type="Cancerous")
        out <- analyze_CGI_clusters(Disease=DiseaseName,cutoff=2,type="Normal")
        ## Value for cutoff (i.e number of clusters is given by the hierarchical clustering)
}

We observe 2 and 3 clusters of CGI+SS profiles for normal and cancerous tissues respectively.

A.5) We plot the characteristic profiles in each cluster i.e the CGI+SS profiles with the lowest mean distance with other CGI+SS profiles in the cluster:

source('fun/plot_characteristic_profiles.R')
for (DiseaseName in DiseaseList)
{
        out <- plot_characteristic_profiles(Disease=DiseaseName,type="Cancerous")
        out <- plot_characteristic_profiles(Disease=DiseaseName,type="Normal")
}

A.6) We assess whether a given CGI+SS is clustered in the same cluster in normal or cancerous tissues:

source('fun/compare_clusters.R')
for (DiseaseName in DiseaseList)
{
        compare_clusters(Disease1=DiseaseName,type1="Cancerous", analysis="Mean")
}

Clusters are mostly stable between normal and cancerous tissues beside the cancerous-specific cluster that is derived from CGI+SS coming from cluster 1 and 2 in normal tissues.

A.7) We then assess whether a given CGI+SS is clustered in the same cluster between different tissues:

DiseaseListbis <- c(DiseaseList, DiseaseList[1])
for (k in 1:(length(DiseaseList)-1))
{
        compare_clusters(Disease1=DiseaseList[k],Disease2=DiseaseList[k+1],type1="Normal", analysis="Mean")
        compare_clusters(Disease1=DiseaseList[k],Disease2=DiseaseList[k+1],type1="Cancerous", analysis="Mean")
}

Clusters are less stable between tissues!

A.8) Finally we look at the link between the CGI+SS patterns and gene expression levels

source('fun/compare_GE_clusters.R')
for (DiseaseName in DiseaseList)
{
        compare_GE_clusters(DiseaseName= DiseaseName)
}

In normal tissue, CGI+SS in cluster 2 (hypermethylated CGIs) are not significantly less expressed than CGI+SS in cluster 1 (hypomethylated CGIs)
In cancerous tissues, we observe that genes associated with cluster 3 are significantly repressed compared to genes associated with cluster 1 and 2.
Using the clustering of CGI+SS associated with cancerous tissues, and looking at the gene expression distribution in normal tissues, we observe that the genes associated with cluster 3 are also repressed in normal tissues (although the CGI+SS clustering in normal tissues do not have a cluster 3 i.e the CGI+SS are either hypo/hypermethylated). This suggests that overall methylation variations might not be causal in the repression of the genes associated.

Part B: Inter-individual methylation variations to predict gene expression variations

Average methylation patterns were not associated with gene expression variations. In the second part, we look the power of inter-individual methylation variations (in the CGI+SS) in a specific dataset, to predict the gene expression variations of the associated genes.

B.1) We build a regression setting where we assess the predictive power of methylation variations to predict gene expression variations.

We assess the predictive power by performing, for each dataset and for each CGI+SS, a cross-validation procedure (nfolds=3) where we train the parameter of a Lasso on 2/3 of the dataset and we test the prediction on the remaining 1/3 of the dataset. The performance is assessed with R^2= cor(yhat, ytest)^2 which is a value between 0 and 1 with 1 being the highest. We bootstrap the prediction procedure (nboostrap=100) and we get a final average R^2 for each gene.

We assess the predictive power using only the mean CGI information or the full CGI+SS information. Supplementary analyses include all the CGIs associated with a gene or taking the full methylome to predict the gene expression or just the methylation level of the associated chromosome.

source("fun/predict_GE.R")
for (DiseaseName in DiseaseList)
{
        predict_GE(DiseaseName= DiseaseName, type="Cancerous", preprocessing="CGIs", MethylationAnalysis="Mean")
        predict_GE(DiseaseName= DiseaseName, type="Normal", preprocessing="CGIs", MethylationAnalysis="Mean")


        predict_GE(DiseaseName= DiseaseName, type="Cancerous", preprocessing="CGIs", MethylationAnalysis="Promoter")
        predict_GE(DiseaseName= DiseaseName, type="Normal", preprocessing="CGIs", MethylationAnalysis="Promoter")

}

B.2) Summary of the results:

source('fun/analyze_prediction.R')
for (DiseaseName in DiseaseList)
{
        analyze_prediction(DiseaseName)
}

B.3) We also had the CNV information in the regression model to assess whether the performance in improved (nfolds=3, nboostrap=100):

source("fun/predict_GE_CNV.R")
for (DiseaseName in DiseaseList)
{
        predict_GE_CNV(DiseaseName= DiseaseName, type="Cancerous", preprocessing="CGIs", MethylationAnalysis="Mean")
        predict_GE_CNV(DiseaseName= DiseaseName, type="Normal", preprocessing="CGIs", MethylationAnalysis="Mean")


        predict_GE_CNV(DiseaseName= DiseaseName, type="Cancerous", preprocessing="CGIs", MethylationAnalysis="Promoter")
        predict_GE_CNV(DiseaseName= DiseaseName, type="Normal", preprocessing="CGIs", MethylationAnalysis="Promoter")

}

B.4) Summary of the results:

source('fun/analyze_prediction_CNV.R')
for (DiseaseName in DiseaseList)
{
        analyze_prediction_CNV(DiseaseName)
}

B.5) We then compare the prediction performance with noCNV info:

source('fun/compare_prediction_Normal_Cancerous.R')
for (DiseaseName in DiseaseList)
{
        compare_prediction_CNV_noCNV(DiseaseName)
}

B.6) We compare the prediction performance between different tissues:

source('fun/compare_prediction_interCancer.R')
compare_prediction_interCancer()