scripts/dto_prep/spot_check.R

# Interactive spot-checking of one (binding, perturbation) pair.
#
# dto_prepare_pair() drops its temp tables when it returns, which is right for a
# batch run and useless at the console. These helpers stage the same two temp
# tables and leave them up, then query them with the *same* SQL generators the
# writers use -- so what you see here is what lands in the CSV, not a
# reimplementation that can drift from it.
#
# Typical session:
#
#   source("R/yeast_comparative_analysis/dto_prep/spot_check.R")   # after the others
#   vdb <- dto_vdb(); dto_register_universe(vdb)
#   dto_stage_pair(vdb, "rossi_mindel", "degron")
#
#   dto_regulator(vdb, "YLR451W")        # which samples does this TF have?
#   dto_ranked(vdb, "binding", "180")    # the file that would be written
#   dto_binding_rows(vdb, "180", "YAL038W")   # why that target got that rank
#   dto_source_rows(vdb, "rossi_mindel", "180", "YAL038W")  # untouched source row
#   dto_overlap(vdb, "180", "119_261", n = 100)
#
#   dto_unstage(vdb)

library(cli)
library(glue)
library(tibble)

# What dto_stage_pair() last set up, so the other helpers need only the ids.
.dto_stage <- new.env(parent = emptyenv())

# vdb$query() hands back a reticulate-converted pandas frame, which drags a
# RangeIndex and its attributes along. Strip them so a console `identical()`
# against a frame read off disk compares data and not provenance.
.as_tbl <- function(df) {
    df <- as.data.frame(df)
    attributes(df) <- attributes(df)[c("names", "class", "row.names")]
    rownames(df) <- NULL
    tibble::as_tibble(df)
}

#' Stage the two per-pair temp tables and leave them in place
#'
#' @param vdb A VirtualDB handle with `dto_universe` already registered.
#' @param binding_db,pert_db db_names, both of which must be in the registries.
#' @return `invisible(list(binding_db, pert_db, binding_spec, pert_spec))`.
dto_stage_pair <- function(vdb, binding_db, pert_db) {
    b_spec <- BINDING_DTO_REGISTRY[[binding_db]]
    p_spec <- PERTURBATION_DTO_REGISTRY[[pert_db]]
    if (is.null(b_spec)) cli_abort("No BINDING_DTO_REGISTRY entry for {.val {binding_db}}.")
    if (is.null(p_spec)) cli_abort("No PERTURBATION_DTO_REGISTRY entry for {.val {pert_db}}.")

    dto_execute(vdb, dto_pert_table_sql(pert_db, p_spec))
    dto_execute(vdb, dto_bind_table_sql(binding_db, b_spec))

    .dto_stage$binding_db <- binding_db
    .dto_stage$pert_db <- pert_db
    .dto_stage$b_spec <- b_spec
    .dto_stage$p_spec <- p_spec

    nb <- vdb$query(glue("SELECT COUNT(*) n, COUNT(DISTINCT sample_id) s FROM {.DTO_BIND_TBL}"))
    np <- vdb$query(glue("SELECT COUNT(*) n, COUNT(DISTINCT sample_id) s FROM {.DTO_PERT_TBL}"))
    cli_alert_success(
        "Staged {binding_db} ({nb$n} rows / {nb$s} samples) x \\
         {pert_db} ({np$n} rows / {np$s} samples). \\
         Tables {.code { .DTO_BIND_TBL }} and {.code { .DTO_PERT_TBL }} are live."
    )

    invisible(as.list(.dto_stage))
}

#' Drop the staged temp tables
#'
#' @param vdb A VirtualDB handle.
#' @return `invisible(NULL)`.
dto_unstage <- function(vdb) {
    dto_execute(vdb, dto_drop_pair_tables_sql())
    rm(list = ls(.dto_stage), envir = .dto_stage)
    invisible(NULL)
}

.stage_or_abort <- function() {
    if (is.null(.dto_stage$binding_db)) {
        cli_abort("Nothing staged. Call {.code dto_stage_pair(vdb, binding_db, pert_db)} first.")
    }
}

#' Samples for one regulator on both sides of the staged pair
#'
#' Uses the same sample-map queries that build `lookup.txt`, so a regulator that
#' shows up here with samples on both sides is exactly a regulator the pipeline
#' will test, and one missing a side is a row in `incomplete.csv`.
#'
#' @param vdb A VirtualDB handle.
#' @param regulator A regulator locus tag.
#' @param max_rows The cap the run used, so `n_rows` matches the file.
#' @param truncate_by The truncation policy the run used.
#' @return A tibble of `side`, `sample_id`, `n_rows` (rows that pass the
#'   significance gate and the pair scope -- the length of the written file).
dto_regulator <- function(vdb, regulator, max_rows = NULL,
                          truncate_by = c("rank", "row")) {
    .stage_or_abort()
    truncate_by <- match.arg(truncate_by)

    one_side <- function(side) {
        map <- .as_tbl(vdb$query(dto_sample_map_sql(side)))
        map <- map[map$regulator_locus_tag == regulator, , drop = FALSE]
        if (!nrow(map)) return(tibble(side = character(), sample_id = character(), n_rows = integer()))

        list_sql <- if (side == "binding") {
            dto_binding_list_sql(.dto_stage$b_spec, max_rows, truncate_by)
        } else {
            dto_pert_list_sql(.dto_stage$p_spec, "effect", max_rows, truncate_by)
        }
        counts <- vdb$query(glue(
            "SELECT sample_id, COUNT(*) AS n_rows FROM ({list_sql}) _l GROUP BY sample_id"
        ))
        merged <- merge(
            data.frame(side = side, sample_id = map$sample_id, stringsAsFactors = FALSE),
            counts,
            by = "sample_id", all.x = TRUE
        )
        .as_tbl(merged[order(merged$sample_id), c("side", "sample_id", "n_rows")])
    }

    out <- rbind(one_side("binding"), one_side("perturbation"))
    if (!nrow(out)) {
        bdb <- .dto_stage$binding_db
        pdb <- .dto_stage$pert_db
        cli_alert_warning(
            "{.val {regulator}} has no in-scope samples in either \\
             {.val {bdb}} or {.val {pdb}}."
        )
    }
    out
}

#' The ranked list for one sample, exactly as written to disk
#'
#' Runs the writer's own SELECT and pulls out one sample. The two columns the
#' CSV holds are `target_locus_tag` and `rank_value`; `sort_key` is the row order
#' and is shown here because tie blocks are easiest to read with it visible.
#'
#' @param vdb A VirtualDB handle.
#' @param side `"binding"` or `"perturbation"`.
#' @param sample_id The sample to pull.
#' @param n Rows to return. `Inf` for all.
#' @param ranking Perturbation side only: `"effect"` or `"pvalue"`.
#' @param max_rows The cap the run used. Set it to match the run you are
#'   checking, or the console will show rows the file does not have.
#' @param truncate_by The truncation policy the run used.
#' @return A tibble of `target_locus_tag`, `rank_value`, `sort_key`.
dto_ranked <- function(vdb, side = c("binding", "perturbation"), sample_id,
                       n = 20, ranking = c("effect", "pvalue"),
                       max_rows = NULL, truncate_by = c("rank", "row")) {
    .stage_or_abort()
    side <- match.arg(side)
    ranking <- match.arg(ranking)
    truncate_by <- match.arg(truncate_by)

    list_sql <- if (side == "binding") {
        dto_binding_list_sql(.dto_stage$b_spec, max_rows, truncate_by)
    } else {
        dto_pert_list_sql(.dto_stage$p_spec, ranking, max_rows, truncate_by)
    }
    limit <- if (is.finite(n)) glue("LIMIT {as.integer(n)}") else ""

    .as_tbl(vdb$query(glue("
SELECT target_locus_tag, rank_value, sort_key
FROM ({list_sql}) _l
WHERE sample_id = {.sql_str(as.character(sample_id))}
ORDER BY sort_key
{limit}
")))
}

# Shared body of dto_binding_rows() / dto_pert_rows(): the staged row plus the
# two flags that decide whether it reaches a file at all.
.staged_rows <- function(vdb, tbl, extra_cols, scope, self_clause, sample_id, targets, n) {
    target_clause <- if (is.null(targets)) {
        ""
    } else {
        vals <- paste(vapply(targets, .sql_str, character(1)), collapse = ", ")
        glue("AND target_locus_tag IN ({vals})")
    }
    limit <- if (is.finite(n)) glue("LIMIT {as.integer(n)}") else ""

    .as_tbl(vdb$query(glue("
SELECT
    target_locus_tag,
    {extra_cols},
    sig_ok,
    ({scope}) AS in_scope,
    ({self_clause}) AS not_self
FROM {tbl}
WHERE sample_id = {.sql_str(as.character(sample_id))}
  {target_clause}
{limit}
")))
}

#' Staged binding rows for one sample, with the gate flags
#'
#' `sig_ok`, `in_scope` and `not_self` are the three conditions
#' `dto_binding_list_sql()` requires; a target present here but absent from
#' `dto_ranked()` will have a FALSE among them.
#'
#' @param vdb A VirtualDB handle.
#' @param sample_id Binding sample.
#' @param targets Target locus tags to restrict to. `NULL` for all.
#' @param n Row cap.
#' @return A tibble.
dto_binding_rows <- function(vdb, sample_id, targets = NULL, n = 50) {
    .stage_or_abort()
    # The staged columns are named generically so one query shape serves every
    # dataset; alias them back to the registry's column names so the output
    # reads like the source. A spec with no tiebreak_col falls back to
    # target_locus_tag, which is already the first column -- don't repeat it.
    b_spec <- .dto_stage$b_spec
    cols <- c("regulator_locus_tag", glue("rank_value_raw AS \"{b_spec$rank_col}\""))
    if (!is.null(b_spec$tiebreak_col)) {
        cols <- c(cols, glue("tiebreak_value AS \"{b_spec$tiebreak_col}\""))
    }

    .staged_rows(
        vdb, .DTO_BIND_TBL,
        extra_cols = paste(cols, collapse = ", "),
        scope = .bind_scope_where,
        self_clause = "regulator_locus_tag <> target_locus_tag",
        sample_id = sample_id, targets = targets, n = n
    )
}

#' Staged perturbation rows for one sample, with the gate flags
#'
#' `not_self` is always TRUE here: `_dto_pert` drops self-targets when it is
#' built, because the binding side donates its scope from unfiltered rows.
#'
#' @inheritParams dto_binding_rows
#' @return A tibble.
dto_pert_rows <- function(vdb, sample_id, targets = NULL, n = 50) {
    .stage_or_abort()
    .staged_rows(
        vdb, .DTO_PERT_TBL,
        extra_cols = "regulator_locus_tag, effect, pvalue",
        scope = .pert_scope_where,
        self_clause = "TRUE",
        sample_id = sample_id, targets = targets, n = n
    )
}

#' Untouched source rows behind a staged row
#'
#' Every column of the VirtualDB view, before the universe filter, the NA fills
#' and the dedup. This is where to look when a staged value is not what you
#' expected -- e.g. to see the several probes a dedup collapsed.
#'
#' @param vdb A VirtualDB handle.
#' @param db_name The dataset view to read.
#' @param sample_id Sample to restrict to. `NULL` for none.
#' @param targets Target locus tags to restrict to. `NULL` for all.
#' @param n Row cap.
#' @return A tibble.
dto_source_rows <- function(vdb, db_name, sample_id = NULL, targets = NULL, n = 50) {
    clauses <- "TRUE"
    if (!is.null(sample_id)) {
        clauses <- c(clauses, glue("CAST(sample_id AS VARCHAR) = {.sql_str(as.character(sample_id))}"))
    }
    if (!is.null(targets)) {
        vals <- paste(vapply(targets, .sql_str, character(1)), collapse = ", ")
        clauses <- c(clauses, glue("target_locus_tag IN ({vals})"))
    }

    .as_tbl(vdb$query(glue(
        "SELECT * FROM {db_name} WHERE {paste(clauses, collapse = ' AND ')} LIMIT {as.integer(n)}"
    )))
}

#' Top-n overlap between a binding and a perturbation sample
#'
#' The quantity DTO scores, computed here without the permutation test: how many
#' of the top `n` binding targets are in the top `n` perturbation targets, and
#' which ones. Useful for sanity-checking a surprising DTO p-value against the
#' inputs that produced it.
#'
#' @param vdb A VirtualDB handle.
#' @param binding_sample,pert_sample Sample ids from each side.
#' @param n Rank threshold, applied to `sort_key` on both sides.
#' @param ranking Perturbation ranking to use.
#' @return A list with `n_binding`, `n_pert`, `n_overlap`, `background_size` and
#'   the overlapping `targets`.
dto_overlap <- function(vdb, binding_sample, pert_sample, n = 100,
                        ranking = c("effect", "pvalue")) {
    .stage_or_abort()
    ranking <- match.arg(ranking)

    b <- dto_ranked(vdb, "binding", binding_sample, n = n)
    p <- dto_ranked(vdb, "perturbation", pert_sample, n = n, ranking = ranking)
    bg <- vdb$query(dto_background_sql())

    shared <- intersect(b$target_locus_tag, p$target_locus_tag)
    cli_alert_info(
        "top-{n}: {nrow(b)} binding, {nrow(p)} perturbation ({ranking}), \\
         {length(shared)} shared, background {nrow(bg)}."
    )

    list(
        n_binding = nrow(b),
        n_pert = nrow(p),
        n_overlap = length(shared),
        background_size = nrow(bg),
        targets = shared
    )
}