annotate mutational_patterns.R @ 2:aea952be68cb draft

"planemo upload for repository https://github.com/ARTbio/tools-artbio/tree/master/tools/mutational_patterns commit cd8f633245d53cf47eaf860a4e0ae8d806c34419"

# load packages that are provided in the conda env
options( show.error.messages=F,
       error = function () { cat( geterrmessage(), file=stderr() ); q( "no", 1, F ) } )
loc <- Sys.setlocale("LC_MESSAGES", "en_US.UTF-8")
warnings()
library(optparse)
library(rjson)
library(grid)
library(gridExtra)

# Arguments
option_list = list(
  make_option(
    "--inputs",
    default = NA,
    type = 'character',
    help = "json formatted dictionary of datasets and their paths"
  ),
  make_option(
    "--genome",
    default = NA,
    type = 'character',
    help = "genome name in the BSgenome bioconductor package"
  ),
  make_option(
    "--levels",
    default = NA,
    type = 'character',
    help = "path to the tab separated file describing the levels in function of datasets"
  ),
  make_option(
    "--signum",
    default = 2,
    type = 'integer',
    help = "selects the N most significant signatures in samples to express mutational patterns"
  ),
  make_option(
    "--nrun",
    default = 2,
    type = 'integer',
    help = "Number of runs to fit signatures"
  ),
  make_option(
    "--rank",
    default = 2,
    type = 'integer',
    help = "number of ranks to display for parameter optimization"
  ),
    make_option(
    "--newsignum",
    default = 2,
    type = 'integer',
    help = "Number of new signatures to be captured"
  ),

  make_option(
    "--output_spectrum",
    default = NA,
    type = 'character',
    help = "path to output dataset"
  ),
  make_option(
    "--output_denovo",
    default = NA,
    type = 'character',
    help = "path to output dataset"
  ),
  make_option(
    "--output_cosmic",
    default = NA,
    type = 'character',
    help = "path to output dataset"
  )
)

opt = parse_args(OptionParser(option_list = option_list),
                 args = commandArgs(trailingOnly = TRUE))

################ Manage input data ####################
json_dict <- opt$inputs
parser <- newJSONParser()
parser$addData(json_dict)
fileslist <- parser$getObject()
vcf_files <- attr(fileslist, "names")
sample_names <- unname(unlist(fileslist))
ref_genome <- opt$genome
vcf_table <- data.frame(sample_name=sample_names, path=vcf_files)

library(MutationalPatterns)
library(ref_genome, character.only = TRUE)

# Load the VCF files into a GRangesList:
vcfs <- read_vcfs_as_granges(vcf_files, sample_names, ref_genome)
if (!is.na(opt$levels)[1]) {  # manage levels if there are
    levels_table  <- read.delim(opt$levels, header=FALSE, col.names=c("sample_name","level"))
    library(dplyr)
    metadata_table <- left_join(vcf_table, levels_table, by = "sample_name")
#    detach("package:dplyr", unload=TRUE)
    tissue <- metadata_table$level
}

##### This is done for any section ######
mut_mat <- mut_matrix(vcf_list = vcfs, ref_genome = ref_genome)


###### Section 1 Mutation characteristics and spectrums #############
if (!is.na(opt$output_spectrum)[1]) {
    pdf(opt$output_spectrum, paper = "special", width = 11.69, height = 11.69)
    type_occurrences <- mut_type_occurrences(vcfs, ref_genome)
    
    # mutation spectrum, total or by sample
    
    if (is.na(opt$levels)[1]) {
        p1 <- plot_spectrum(type_occurrences, CT = TRUE, legend=TRUE)
        plot(p1)
    } else {
        p2 <- plot_spectrum(type_occurrences, by = tissue, CT=TRUE, legend=TRUE) # by sample
        p3 <- plot_spectrum(type_occurrences, CT=TRUE, legend=TRUE) # total
        grid.arrange(p2, p3, ncol=2, widths=c(4,2.3), heights=c(4,1))
    }
    plot_96_profile(mut_mat, condensed = TRUE)
    dev.off()
}

###### Section 2: De novo mutational signature extraction using NMF #######
if (!is.na(opt$output_denovo)[1]) {
    mut_mat <- mut_mat + 0.0001 # First add a small psuedocount to the mutation count matrix
    # Use the NMF package to generate an estimate rank plot
    library("NMF")
    estimate <- nmf(mut_mat, rank=1:opt$rank+1, method="brunet", nrun=opt$nrun, seed=123456)
    # And plot it
    pdf(opt$output_denovo, paper = "special", width = 11.69, height = 11.69)
    p.trans <- plot(estimate)
    grid.arrange(p.trans)
    # Extract 4 (PARAMETIZE) mutational signatures from the mutation count matrix with extract_signatures
    # (For larger datasets it is wise to perform more iterations by changing the nrun parameter
    # to achieve stability and avoid local minima)
    nmf_res <- extract_signatures(mut_mat, rank=opt$rank, nrun=opt$nrun)
    # Assign signature names
    colnames(nmf_res$signatures) <- paste0("NewSig_", 1:opt$rank)
    rownames(nmf_res$contribution) <- paste0("NewSig_", 1:opt$rank)
    # Plot the 96-profile of the signatures:
    p.trans <- plot_96_profile(nmf_res$signatures, condensed = TRUE)
    grid.arrange(p.trans)
    # Visualize the contribution of the signatures in a barplot
    pc1 <- plot_contribution(nmf_res$contribution, nmf_res$signature, mode="relative", coord_flip = TRUE)
    # Visualize the contribution of the signatures in absolute number of mutations
    pc2 <- plot_contribution(nmf_res$contribution, nmf_res$signature, mode="absolute", coord_flip = TRUE)
    # Combine the two plots:
    grid.arrange(pc1, pc2)
        
    # The relative contribution of each signature for each sample can also be plotted as a heatmap with
    # plot_contribution_heatmap, which might be easier to interpret and compare than stacked barplots.
    # The samples can be hierarchically clustered based on their euclidean dis- tance. The signatures
    # can be plotted in a user-specified order.
    # Plot signature contribution as a heatmap with sample clustering dendrogram and a specified signature order:
    pch1 <- plot_contribution_heatmap(nmf_res$contribution,
                                      sig_order = paste0("NewSig_", 1:opt$rank))
    # Plot signature contribution as a heatmap without sample clustering:
    pch2 <- plot_contribution_heatmap(nmf_res$contribution, cluster_samples=FALSE)
    #Combine the plots into one figure:
    grid.arrange(pch1, pch2, ncol = 2, widths = c(2,1.6))
    
    # Compare the reconstructed mutational profile with the original mutational profile:   
    plot_compare_profiles(mut_mat[,1],
                          nmf_res$reconstructed[,1],
                          profile_names = c("Original", "Reconstructed"),
                          condensed = TRUE)
    dev.off()
}
##### Section 3: Find optimal contribution of known signatures: COSMIC mutational signatures ####

if (!is.na(opt$output_cosmic)[1]) {
    pdf(opt$output_cosmic, paper = "special", width = 11.69, height = 11.69)
    sp_url <- paste("https://cancer.sanger.ac.uk/cancergenome/assets/", "signatures_probabilities.txt", sep = "")
    cancer_signatures = read.table(sp_url, sep = "\t", header = TRUE)
    mut_mat <- mut_mat + 0.0001 # First add a small psuedocount to the mutation count matrix
    new_order = match(row.names(mut_mat), cancer_signatures$Somatic.Mutation.Type)
    cancer_signatures = cancer_signatures[as.vector(new_order),]
    row.names(cancer_signatures) = cancer_signatures$Somatic.Mutation.Type
    cancer_signatures = as.matrix(cancer_signatures[,4:33])
    
    # Plot mutational profiles of the COSMIC signatures
    colnames(cancer_signatures) <- gsub("Signature.", "", colnames(cancer_signatures)) # shorten signature labels
    p.trans <- plot_96_profile(cancer_signatures, condensed = TRUE, ymax = 0.3)
    grid.arrange(p.trans, top = textGrob("COSMIC signature profiles",gp=gpar(fontsize=12,font=3)))

    # Hierarchically cluster the COSMIC signatures based on their similarity with average linkage
    # hclust_cosmic = cluster_signatures(cancer_signatures, method = "average")
    # store signatures in new order
    # cosmic_order = colnames(cancer_signatures)[hclust_cosmic$order]
    # plot(hclust_cosmic)
    
    # Similarity between mutational profiles and COSMIC signatures
    
    
    # The similarity between each mutational profile and each COSMIC signature, can be calculated
    # with cos_sim_matrix, and visualized with plot_cosine_heatmap. The cosine similarity reflects
    # how well each mutational profile can be explained by each signature individually. The advantage
    # of this heatmap representation is that it shows in a glance the similarity in mutational
    # profiles between samples, while at the same time providing information on which signatures
    # are most prominent. The samples can be hierarchically clustered in plot_cosine_heatmap.
    # The cosine similarity between two mutational profiles/signatures can be calculated with cos_sim :
    # cos_sim(mut_mat[,1], cancer_signatures[,1])
    
    # Calculate pairwise cosine similarity between mutational profiles and COSMIC signatures
    
    # cos_sim_samples_signatures = cos_sim_matrix(mut_mat, cancer_signatures)
    # Plot heatmap with specified signature order
    # p.trans <- plot_cosine_heatmap(cos_sim_samples_signatures, col_order = cosmic_order, cluster_rows = TRUE)
    # grid.arrange(p.trans)
    
    # Find optimal contribution of COSMIC signatures to reconstruct 96 mutational profiles
    fit_res <- fit_to_signatures(mut_mat, cancer_signatures)

    # Select signatures with some contribution (above a threshold)
    threshold <- tail(sort(unlist(rowSums(fit_res$contribution), use.names = FALSE)), opt$signum)[1]
    select <- which(rowSums(fit_res$contribution) >= threshold) # ensure opt$signum best signatures in samples are retained, the others discarded
    # Plot contribution barplots
    pc1 <- plot_contribution(fit_res$contribution[select,], cancer_signatures[,select], coord_flip = T, mode = "absolute")
    pc2 <- plot_contribution(fit_res$contribution[select,], cancer_signatures[,select], coord_flip = T, mode = "relative")
    # Combine the two plots:
    grid.arrange(pc1, pc2)
    ## pie charts of comic signatures contributions in samples
    
    sig_data_pie <- as.data.frame(t(head(fit_res$contribution[select,])))
    colnames(sig_data_pie) <- gsub("nature", "", colnames(sig_data_pie))
    sig_data_pie_percents <- sig_data_pie / (apply(sig_data_pie,1,sum)) * 100
    sig_data_pie_percents$sample <- rownames(sig_data_pie_percents)
    library(reshape2)
    melted_sig_data_pie_percents <-melt(data=sig_data_pie_percents)
    melted_sig_data_pie_percents$label <- sub("Sig.", "", melted_sig_data_pie_percents$variable)
    melted_sig_data_pie_percents$pos <- cumsum(melted_sig_data_pie_percents$value) - melted_sig_data_pie_percents$value/2
    library(ggplot2)
    p.trans <-  ggplot(melted_sig_data_pie_percents, aes(x="", y=value, group=variable, fill=variable)) +
                       geom_bar(width = 1, stat = "identity") +
                       geom_text(aes(label = label), position = position_stack(vjust = 0.5), color="black", size=3) +
                       coord_polar("y", start=0) + facet_wrap(~ sample) +
                       labs(x="", y="Samples", fill = "Signatures (Cosmic_v2,March 2015)") +
                       theme(axis.text = element_blank(),
                             axis.ticks = element_blank(),
                             panel.grid  = element_blank())
    grid.arrange(p.trans)

    # Plot relative contribution of the cancer signatures in each sample as a heatmap with sample clustering
    p.trans <- plot_contribution_heatmap(fit_res$contribution, cluster_samples = TRUE, method = "complete")
    grid.arrange(p.trans)

    # Compare the reconstructed mutational profile of sample 1 with its original mutational profile
    # plot_compare_profiles(mut_mat[,1], fit_res$reconstructed[,1],
    #                       profile_names = c("Original", "Reconstructed"),
    #                       condensed = TRUE)
    
    # Calculate the cosine similarity between all original and reconstructed mutational profiles with
    # `cos_sim_matrix`
    
    # calculate all pairwise cosine similarities
    cos_sim_ori_rec <- cos_sim_matrix(mut_mat, fit_res$reconstructed)
    # extract cosine similarities per sample between original and reconstructed
    cos_sim_ori_rec <- as.data.frame(diag(cos_sim_ori_rec))
    
    # We can use ggplot to make a barplot of the cosine similarities between the original and
    # reconstructed mutational profile of each sample. This clearly shows how well each mutational
    # profile can be reconstructed with the COSMIC mutational signatures. Two identical profiles
    # have a cosine similarity of 1. The lower the cosine similarity between original and
    # reconstructed, the less well the original mutational profile can be reconstructed with
    # the COSMIC signatures. You could use, for example, cosine similarity of 0.95 as a cutoff.
    
    # Adjust data frame for plotting with gpplot
    colnames(cos_sim_ori_rec) = "cos_sim"
    cos_sim_ori_rec$sample = row.names(cos_sim_ori_rec)
    # Make barplot
    p.trans <- ggplot(cos_sim_ori_rec, aes(y=cos_sim, x=sample)) +
                      geom_bar(stat="identity", fill = "skyblue4") +
                      coord_cartesian(ylim=c(0.8, 1)) +
                      # coord_flip(ylim=c(0.8,1)) +
                      ylab("Cosine similarity\n original VS reconstructed") +
                      xlab("") +
                      # Reverse order of the samples such that first is up
                      # xlim(rev(levels(factor(cos_sim_ori_rec$sample)))) +
                      theme_bw() +
                      theme(panel.grid.minor.y=element_blank(),
                      panel.grid.major.y=element_blank()) +
                      # Add cut.off line
                      geom_hline(aes(yintercept=.95))
    grid.arrange(p.trans, top = textGrob("Similarity between true and reconstructed profiles (with all Cosmic sig.)",gp=gpar(fontsize=12,font=3)))    
    
    dev.off()
}