pull_gains.R

#' Use bioconductor tools to extract the processed data, metadata and
#' probeset annotations for the GaINS sepsis dataset.

library(ArrayExpress)
library(illuminaHumanv4.db)
library(tidyverse)
library(here)

options(timeout = 1200)

gains_accessions <- c(
    derivation_davenport = "E-MTAB-4421",
    validation_davenport = "E-MTAB-4451",
    discovery_burnham = "E-MTAB-5273",
    validation_burnham = "E-MTAB-5274"
)

out <- map(gains_accessions, ~ {
    outdir <- file.path("data/gains", .x)
    dir.create(outdir, recursive = TRUE, showWarnings = FALSE)
    ArrayExpress::getAE(.x, path = outdir)
})

#' extract the 28 day survival column info into a uniform boolean value
standardize_survival <- function(x) {
    x <- str_squish(tolower(x))
    case_when(
        str_detect(x, "non.?surviv|dead") ~ FALSE,
        str_detect(x, "^surviv|alive") ~ TRUE,
        TRUE ~ NA
    )
}

# extract the sepsisstratifier SRS1/SRS2 labels to a standardized
# format
standardize_srs <- function(x) {
    x <- str_squish(tolower(as.character(x)))
    case_when(
        str_detect(x, "1") ~ "SRS1",
        str_detect(x, "2") ~ "SRS2",
        TRUE ~ NA_character_
    )
}

# patients are either normal (no sepsis), community acquired pneumonia
# or fecal peritonitis
derive_disease_state <- function(source_name) {
    case_when(
        str_detect(source_name, "^CON") ~ "normal",
        str_detect(source_name, "^CAP") ~ "CAP",
        str_detect(source_name, "^FP") ~ "FP",
        TRUE ~ NA_character_
    )
}

extract_metadata <- function(sdrf, accession) {
    cfg <- accession_config |> filter(accession == .env$accession)

    tibble(
        sample_id      = sdrf$source_name,
        accession      = accession,
        age            = sdrf$characteristics_age,
        sex            = sdrf$characteristics_sex,
        survived_28d   = standardize_survival(sdrf[[cfg$survival_col]]),
        srs_group      = standardize_srs(sdrf[[cfg$srs_col]]),
        disease_state  = derive_disease_state(sdrf$source_name)
    )
}

sdrf_list <- map(out, ~ {
    janitor::clean_names(read_tsv(.$sdrf))
}) |> set_names(gains_accessions)

accession_config <- tribble(
    ~accession, ~survival_col, ~srs_col,
    "E-MTAB-4421", "characteristics_28_day_survival", "characteristics_sepsis_response_signature_group",
    "E-MTAB-4451", "characteristics_28_day_survival", "factor_value_sepsis_response_signature_group",
    "E-MTAB-5273", "characteristics_clinical_information", "characteristics_unsupervised_analysis",
    "E-MTAB-5274", "characteristics_clinical_information", "characteristics_unsupervised_analysis"
)


gains_metadata <- imap_dfr(sdrf_list, extract_metadata) |>
    left_join(enframe(gains_accessions, name = "author_use", value = "accession"))

gains_expr_mats <- map(out, ~ {
    read_tsv(.$processedFiles)
})

con <- illuminaHumanv4_dbconn()

address_map <- DBI::dbGetQuery(con, "SELECT IlluminaID, ArrayAddress FROM ExtraInfo") |>
    mutate(ArrayAddress = as.numeric(ArrayAddress))

translate_probeid <- function(mat, address_map) {
    mat |>
        rename(ArrayAddress = 1) |>
        mutate(ArrayAddress = as.numeric(ArrayAddress)) |>
        left_join(address_map, by = "ArrayAddress") |>
        relocate(IlluminaID) |>
        select(-ArrayAddress)
}

gains_expr_mats$discovery_burnham <- translate_probeid(gains_expr_mats$discovery_burnham, address_map)
gains_expr_mats$validation_burnham <- translate_probeid(gains_expr_mats$validation_burnham, address_map)

get_probe_ids <- function(mat) mat[[1]]

probe_membership <- imap_dfr(gains_expr_mats, ~ tibble(
    IlluminaID = get_probe_ids(.x),
    accession  = gains_accessions[[.y]]
))

gains_expr_long <- imap_dfr(gains_expr_mats, ~ {
    id_col <- colnames(.x)[1]
    .x |>
        rename(IlluminaID = all_of(id_col)) |>
        pivot_longer(-IlluminaID, names_to = "sample_id", values_to = "value") |>
        mutate(accession = gains_accessions[[.y]])
}) |>
    arrange(accession, sample_id, IlluminaID)


extra <- DBI::dbGetQuery(con, "SELECT * FROM ExtraInfo")
extra_eset <- extra[extra$IlluminaID %in% unique(gains_expr_long$IlluminaID), ]

# --- extra table subset ---
extra_selected <- extra_eset |>
    as_tibble() |>
    distinct(IlluminaID, .keep_all = TRUE) |>
    select(
        IlluminaID, ProbeQuality, CodingZone, ProbeSequence,
        SecondMatches, OtherGenomicMatches, RepeatMask,
        OverlappingSNP, GenomicLocation
    )

# --- standard accessors ---
std_anno <- AnnotationDbi::select(
    illuminaHumanv4.db,
    keys = unique(gains_expr_long$IlluminaID),
    columns = c("ENSEMBL", "ENTREZID", "SYMBOL", "GENENAME", "UNIPROT"),
    keytype = "PROBEID"
)

# collapse many-to-one columns (ENSEMBL, UNIPROT) to one row per probe
std_anno_collapsed <- std_anno |>
    group_by(PROBEID) |>
    summarise(
        ENSEMBL = paste(unique(na.omit(ENSEMBL)), collapse = ";"),
        ENTREZID = first(ENTREZID),
        SYMBOL = first(SYMBOL),
        GENENAME = first(GENENAME),
        UNIPROT = paste(unique(na.omit(UNIPROT)), collapse = ";"),
        .groups = "drop"
    ) |>
    mutate(across(c(ENSEMBL, UNIPROT), ~ na_if(.x, ""))) # empty string -> NA

probe_annotation <- extra_selected |>
    left_join(std_anno_collapsed, by = c("IlluminaID" = "PROBEID"))

# Checks for accuracy

# Expected row count for gains_expr_long
expected_long_rows <- map_dbl(gains_expr_mats, ~ nrow(.x) * (ncol(.x) - 1)) |> sum()
stopifnot(nrow(gains_expr_long) == expected_long_rows)

# Every sample in the expression data has a metadata row, and vice versa
expr_samples <- gains_expr_long |> distinct(accession, sample_id)
meta_samples <- gains_metadata |> distinct(accession, sample_id)

setdiff_expr_not_meta <- anti_join(expr_samples, meta_samples, by = c("accession", "sample_id"))
setdiff_meta_not_expr <- anti_join(meta_samples, expr_samples, by = c("accession", "sample_id"))

nrow(setdiff_expr_not_meta) # samples with expression data but no metadata row
nrow(setdiff_meta_not_expr) # samples with metadata but no expression data (expected -- these are your QC/outlier-excluded SDRF rows)

# No duplicate sample rows in metadata
stopifnot(!any(duplicated(gains_metadata |> select(accession, sample_id))))

# probe_annotation is one row per unique probe, no fan-out
stopifnot(nrow(probe_annotation) == length(unique(gains_expr_long$IlluminaID)))
stopifnot(!any(duplicated(probe_annotation$IlluminaID)))

# Every probe in expression data has an annotation row (even if mostly NA)
stopifnot(all(unique(gains_expr_long$IlluminaID) %in% probe_annotation$IlluminaID))

# No unexpected NAs in join keys themselves
stopifnot(
    !anyNA(gains_expr_long$IlluminaID),
    !anyNA(gains_expr_long$sample_id),
    !anyNA(gains_expr_long$accession)
)
stopifnot(
    !anyNA(gains_metadata$sample_id),
    !anyNA(gains_metadata$accession)
)

# write out
dir.create("data/gains/parquet", recursive = TRUE, showWarnings = FALSE)

arrow::write_parquet(gains_metadata, "data/gains/parquet/sample_metadata.parquet")
arrow::write_parquet(probe_annotation, "data/gains/parquet/feature_metadata.parquet")

arrow::write_dataset(gains_expr_long, "data/gains/parquet/expression",
    partitioning = "accession"
)