SBIR project - revisions

Requested revisions: results using different imputation algorithms

We imputed chromosome 21 using GLIMPSE for sequence data and IMPUTE5 for array data. We recompute concordance on chromosome 21 from the original data and the new results.

# load in requisite packages
library(tidyverse)
library(ggthemes)
library(patchwork)
library(rms)
library(glue)
library(knitr)
library(kableExtra)

# populations
pops = read_delim("../data_revisions/all_samples_with_pops.list", delim = " ")

# load in precomputed data
concordance = read_delim("../data_revisions/all-gc-counts.txt",
                          col_names = c("cell_line", "sample", "stat_type", "set_id", "af_midpoint", "rr_hom_matches", "ra_het_matches", "aa_hom_matches",
                   "rr_hom_mismatches", "ra_het_mismatches", "aa_hom_mismatches", "dosage_rsquared", "n_genotypes", "site_type", "tech", "imputation_program"), 
                   delim = " ") %>% 
   mutate(
    # characterize into ilmn 0.5x, 1x, BGI, array
    Experiment = case_when(
      grepl("-0-2$", sample) ~ "Experiment D",
      grepl("-0-[0-9]-[01]-0$", sample) ~ "Experiment A",
      grepl("-1-[0-9]-[01]-0$", sample) ~ "Experiment B",
      grepl("_[1-3]$", sample) ~ "Experiment E (array)",
      grepl("-1$", sample) ~ "ilmn_1x_repl",
    )
  ) %>% 
  inner_join(pops, by = "cell_line") %>% 
  mutate(
    nrc_numerator = (ra_het_matches + aa_hom_matches),
    nrc_denominator =  (ra_het_matches + aa_hom_matches + rr_hom_mismatches + ra_het_mismatches + aa_hom_mismatches),
    nrc = nrc_numerator / nrc_denominator
  )

We can summarize results and plot, like so:

concordance_summary = concordance %>%  
  group_by(af_midpoint, Experiment, stat_type, site_type, tech, imputation_program, super_pop) %>%
  summarise(mean_nrc = mean(nrc)) %>% 
  rename(
    # rename to human-readable
    `Super population` = super_pop,
  )

## `summarise()` regrouping output by 'af_midpoint', 'Experiment', 'stat_type', 'site_type', 'tech', 'imputation_program' (override with `.groups` argument)

concordance_summary %>% 
  filter(stat_type == "GCsAF", site_type == "allsites") %>% 
  ggplot(aes(x = af_midpoint, y = mean_nrc, color = Experiment)) + 
  facet_grid(~`Super population`) + geom_point() + 
  geom_line(aes(linetype = factor(imputation_program))) + scale_color_tableau() + theme_few() + 
  labs(
    title = "Average NRC by non-reference allele frequency, unfiltered on chr21",
    x = "Non-Reference Allele Frequency",
    y = "Non-Reference Concordance"
  ) + scale_x_log10()

ggsave("../paper/src/figs/revisions-chr21-nrc.pdf", width = 10, height = 7)

from which we see that the qualitative trend of the sequence imputation being more accurate than the array imputation holding, with a fixed effect of the imputation program being used given an assay type/Experiment.

Interestingly, if one looks at filtered SNPs, the gap between concordance for the two array imputation algorithms (IMPUTE5 vs minimac4) widens significantly:

concordance_summary %>% 
  filter(stat_type == "GCsAF", site_type == "passing") %>% 
  ggplot(aes(x = af_midpoint, y = mean_nrc, color = Experiment)) + 
  facet_grid(~`Super population`) + geom_point() + 
  geom_line(aes(linetype = factor(imputation_program))) + scale_color_tableau() + theme_few() + 
  labs(
    title = "Average NRC by non-reference allele frequency, filtered for high confidence on chr21",
    x = "Non-Reference Allele Frequency",
    y = "Non-Reference Concordance"
  ) + scale_x_log10()

particularly at higher allele frequencies.