Summary of Findings:

This was an exploration of how TMB is affected by using nonsynonymous filters. It addresses part 1 of this OpenPBTA issue As far as the TMB comparisons go, it doesn’t matter if we use Variant_Classification filters or not. They are highly correlated no matter what and the same general TCGA to PBTA differences seem to persist.

Usage

If both run_caller_consensus_analysis-tcga.sh and run_caller_consensus_analysis-pbta.sh have been run, you can run this command. This will prepare the files needed for this notebook and run this notebook:

# bash run_explorations.sh

Table of Contents generated with DocToc

Setup

# Magrittr pipe
`%>%` <- dplyr::`%>%`
source(file.path("..", "..", "tmb-compare", "util", "cdf-plot-function.R"))
dir.create("plots", showWarnings = FALSE)

Read in the TMB files

Read in PBTA data

Import the tmb calculations that are not filtered by nonsynonymous classifications.

tmb_pbta_no_filter <- data.table::fread(file.path(
  "..",
  "results",
  "no_filter",
  "pbta-snv-mutation-tmb-coding.tsv"
)) %>%
  dplyr::filter(experimental_strategy != "Panel") %>%
  dplyr::mutate(filter = "no filter")

Repeat same steps for original tmb calculations (with filter).

tmb_pbta_with_filter <- data.table::fread(file.path(
  "..",
  "results",
  "consensus",
  "pbta-snv-mutation-tmb-coding.tsv"
)) %>%
  dplyr::filter(experimental_strategy != "Panel") %>%
  dplyr::mutate(filter = "filter")

Combine the unfilter and filtered data sets into one.

tmb_pbta <- dplyr::inner_join(tmb_pbta_no_filter,
  dplyr::select(tmb_pbta_with_filter, Tumor_Sample_Barcode, tmb),
  by = "Tumor_Sample_Barcode",
  suffix = c("_no_filter", "_filter")
)

Read in TCGA data

Repeat the same steps for the TCGA data.

tmb_tcga_no_filter <- data.table::fread(file.path(
  "..",
  "results",
  "no_filter",
  "tcga-snv-mutation-tmb-coding.tsv"
)) %>%
  dplyr::mutate(filter = "no filter")
tmb_tcga_with_filter <- data.table::fread(file.path(
  "..",
  "results",
  "consensus",
  "tcga-snv-mutation-tmb-coding.tsv"
)) %>%
  dplyr::mutate(filter = "filter")

Join these together.

tmb_tcga <- dplyr::inner_join(tmb_tcga_no_filter,
  dplyr::select(tmb_tcga_with_filter, Tumor_Sample_Barcode, tmb),
  by = "Tumor_Sample_Barcode",
  suffix = c("_no_filter", "_filter")
)

Does the filter change a participant’s TMB?

Plot PBTA data

Correlate first with Pearson’s.

cor.test(tmb_pbta$tmb_filter, tmb_pbta$tmb_no_filter)

    Pearson's product-moment correlation

data:  tmb_pbta$tmb_filter and tmb_pbta$tmb_no_filter
t = 2488.9, df = 934, p-value < 2.2e-16
alternative hypothesis: true correlation is not equal to 0
95 percent confidence interval:
 0.9999143 0.9999337
sample estimates:
      cor 
0.9999246 
pbta_cor_plot <- ggplot2::ggplot(
  tmb_pbta,
  ggplot2::aes(x = tmb_no_filter, y = tmb_filter, color = short_histology)
) +
  ggplot2::geom_point() +
  ggplot2::theme_classic() +
  ggplot2::theme(legend.position = "none")

pbta_cor_plot

This data is very skewed, let’s try a nonparametric correlation.

cor.test(tmb_pbta$tmb_filter, tmb_pbta$tmb_no_filter, method = "spearman")
Cannot compute exact p-value with ties

    Spearman's rank correlation rho

data:  tmb_pbta$tmb_filter and tmb_pbta$tmb_no_filter
S = 3031423, p-value < 2.2e-16
alternative hypothesis: true rho is not equal to 0
sample estimates:
      rho 
0.9778195

Still pretty good.

Plot TCGA data

cor.test(tmb_tcga$tmb_filter, tmb_tcga$tmb_no_filter)

    Pearson's product-moment correlation

data:  tmb_tcga$tmb_filter and tmb_tcga$tmb_no_filter
t = 4296.2, df = 316, p-value < 2.2e-16
alternative hypothesis: true correlation is not equal to 0
95 percent confidence interval:
 0.9999893 0.9999931
sample estimates:
      cor 
0.9999914 
tcga_cor_plot <- ggplot2::ggplot(
  tmb_tcga,
  ggplot2::aes(x = tmb_no_filter, y = tmb_filter, color = short_histology)
) +
  ggplot2::geom_point() +
  ggplot2::xlim(0, 5) +
  ggplot2::ylim(0, 5) +
  ggplot2::theme_classic() +
  ggplot2::theme(legend.position = "none")

tcga_cor_plot

Correlate with nonparametric correlation since we did that for PBTA.

cor.test(tmb_tcga$tmb_filter, tmb_tcga$tmb_no_filter, method = "spearman")
Cannot compute exact p-value with ties

    Spearman's rank correlation rho

data:  tmb_tcga$tmb_filter and tmb_tcga$tmb_no_filter
S = 60671, p-value < 2.2e-16
alternative hypothesis: true rho is not equal to 0
sample estimates:
      rho 
0.9886798

Are some histologies affected more than others?

Let’s make the same plots but by short_histology.

Plotting PBTA data, but setting the scales to “free” since the histologies have very different tmb ranges.

pbta_cor_plot + 
  ggplot2::facet_wrap(~short_histology, scales = "free")