scripts/dto_prep/audit.R

# Auditing the registry against the data.
#
# Two levels, because the questions are different:
#
#   dto_registry_table()  -- what the registry *says*. No database, instant.
#                            Answers "which column, which direction, what
#                            cutoff, any blacklist" for every dataset at once.
#   dto_registry_audit()  -- whether the data *agrees*. One scan per dataset.
#                            Catches a column that does not exist, a cutoff that
#                            keeps everything or nothing, and -- the failure that
#                            is otherwise invisible -- a rank direction that is
#                            backwards.
#
# The direction check works by correlating the rank column against the tiebreak
# column, which is always a corroborating measure of the same binding event
# (enrichment beside a p-value, peak count beside a peak score). If the spec
# calls low p-values best and high enrichment best, the two must be negatively
# correlated. A sign that comes back the other way means one of the two
# directions is wrong.

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

.DTO_NUMERIC_TYPES <- c(
    "TINYINT", "SMALLINT", "INTEGER", "BIGINT", "HUGEINT",
    "UTINYINT", "USMALLINT", "UINTEGER", "UBIGINT",
    "FLOAT", "DOUBLE", "REAL", "DECIMAL"
)

.is_numeric_type <- function(x) {
    any(vapply(.DTO_NUMERIC_TYPES, function(t) startsWith(toupper(x), t), logical(1)))
}

#' The registry as a table
#'
#' Every knob for every registered dataset, side by side. This is the fast way
#' to answer "what is dataset X ranked on and which way round" -- it reads the
#' registry only, so it needs no VirtualDB and no data.
#'
#' @param which `"binding"`, `"perturbation"` or `"both"`.
#' @return A tibble, one row per registered dataset.
dto_registry_table <- function(which = c("binding", "perturbation", "both")) {
    which <- match.arg(which)

    binding <- NULL
    if (which %in% c("binding", "both")) {
        binding <- bind_rows(lapply(names(BINDING_DTO_REGISTRY), function(db) {
            s <- BINDING_DTO_REGISTRY[[db]]
            tibble(
                side = "binding",
                db_name = db,
                rank_col = s$rank_col,
                direction = if (s$rank_asc) "asc (low = best)" else "desc (high = best)",
                tiebreak = s$tiebreak_col %||% "(target_locus_tag)",
                tiebreak_dir = if (s$tiebreak_asc) "asc" else "desc",
                sig_filter = s$sig_filter %||% "(none)",
                dedup_by = if (length(s$dedup_by)) {
                    paste(s$dedup_by, collapse = ", ")
                } else {
                    "(one row per sample/target)"
                },
                blacklist = if (length(s$target_blacklist)) {
                    paste(s$target_blacklist, collapse = " ")
                } else {
                    "(none)"
                }
            )
        }))
    }

    perturbation <- NULL
    if (which %in% c("perturbation", "both")) {
        # Two rows per dataset: one per output directory, so the table states
        # what each written file is ranked on -- and which datasets get no
        # pr/pvalue directory at all.
        perturbation <- bind_rows(lapply(names(PERTURBATION_DTO_REGISTRY), function(db) {
            s <- PERTURBATION_DTO_REGISTRY[[db]]
            has_p <- !is.null(s$pvalue_col)
            opts <- paste0(
                if (s$dedup) "dedup " else "",
                if (s$exclude_wt) "exclude_wt " else "",
                if (!is.null(s$effect_na_fill)) glue("effect_na={s$effect_na_fill} ") else "",
                if (!is.null(s$pvalue_na_fill)) glue("pvalue_na={s$pvalue_na_fill}") else ""
            )
            bind_rows(
                tibble(
                    side = "pert: pr/effect",
                    db_name = db,
                    rank_col = s$effect_col,
                    direction = "desc (high |effect| = best)",
                    tiebreak = s$pvalue_col %||% "(constant 0.0)",
                    tiebreak_dir = "asc",
                    sig_filter = s$sig_filter %||% "(none)",
                    dedup_by = "",
                    blacklist = opts
                ),
                tibble(
                    side = "pert: pr/pvalue",
                    db_name = db,
                    rank_col = if (has_p) s$pvalue_col else "(not written: no p-value column)",
                    direction = if (has_p) "asc (low = best)" else "",
                    tiebreak = if (has_p) s$effect_col else "",
                    tiebreak_dir = if (has_p) "desc" else "",
                    sig_filter = if (has_p) s$sig_filter %||% "(none)" else "",
                    dedup_by = "",
                    blacklist = if (has_p) opts else ""
                )
            )
        }))
    }

    bind_rows(binding, perturbation)
}

# One dataset's worth of data-backed checks. Single scan, aggregates only.
.audit_one <- function(vdb, db, spec, blacklist_probe) {
    schema <- vdb$query(glue("DESCRIBE {db}"))
    cols <- schema$column_name
    types <- setNames(schema$column_type, cols)

    base <- tibble(
        db_name = db,
        rank_col = spec$rank_col,
        direction = if (spec$rank_asc) "asc" else "desc"
    )

    if (!spec$rank_col %in% cols) {
        return(bind_cols(base, tibble(
            status = "rank_col MISSING",
            n_rows = NA_real_, n_samples = NA_real_,
            rank_min = NA_real_, rank_med = NA_real_, rank_max = NA_real_,
            frac_sig = NA_real_, med_sig_per_sample = NA_real_,
            corr_rank_tiebreak = NA_real_, direction_ok = NA,
            n_probe_rows = NA_real_, n_dup_rows = NA_real_, dedup_declared = NA
        )))
    }

    tb <- spec$tiebreak_col
    use_corr <- !is.null(tb) && tb %in% cols &&
        .is_numeric_type(types[[tb]]) && .is_numeric_type(types[[spec$rank_col]])

    tb_select <- if (use_corr) glue(", {tb}") else ""
    corr_expr <- if (use_corr) glue("corr({spec$rank_col}, {tb})") else "CAST(NULL AS DOUBLE)"
    sig <- spec$sig_filter %||% "TRUE"
    probe <- paste(vapply(blacklist_probe, .sql_str, character(1)), collapse = ", ")

    res <- vdb$query(glue("
WITH src AS (
    SELECT CAST(sample_id AS VARCHAR) AS sample_id,
           target_locus_tag,
           {spec$rank_col}{tb_select}
    FROM {db}
    WHERE target_locus_tag IN (SELECT locus_tag FROM dto_universe)
      AND regulator_locus_tag IS NOT NULL
),
agg AS (
    SELECT COUNT(*) AS n_rows,
           MIN({spec$rank_col}) AS rank_min,
           MEDIAN({spec$rank_col}) AS rank_med,
           MAX({spec$rank_col}) AS rank_max,
           COUNT(*) FILTER (WHERE {sig}) / NULLIF(COUNT(*), 0)::DOUBLE AS frac_sig,
           COUNT(*) FILTER (WHERE target_locus_tag IN ({probe})) AS n_probe_rows,
           COUNT(*) - COUNT(DISTINCT sample_id || '|' || target_locus_tag)
               AS n_dup_rows,
           {corr_expr} AS corr_rank_tiebreak
    FROM src
),
per_sample AS (
    SELECT sample_id, COUNT(*) FILTER (WHERE {sig}) AS n_sig
    FROM src GROUP BY sample_id
),
samp AS (
    SELECT COUNT(*) AS n_samples, MEDIAN(n_sig) AS med_sig_per_sample FROM per_sample
)
SELECT * FROM agg, samp
"))

    # Better-is-low for the rank column iff rank_asc; same for the tiebreak.
    # Agreeing directions must correlate positively, opposing ones negatively.
    expect_positive <- identical(spec$rank_asc, spec$tiebreak_asc)
    corr <- res$corr_rank_tiebreak[[1]]
    direction_ok <- if (is.na(corr)) NA else (corr > 0) == expect_positive

    bind_cols(base, tibble(status = "ok"), as_tibble(res[, c(
        "n_rows", "n_samples", "rank_min", "rank_med", "rank_max",
        "frac_sig", "med_sig_per_sample", "corr_rank_tiebreak", "n_probe_rows",
        "n_dup_rows"
    )]), tibble(
        direction_ok = direction_ok,
        dedup_declared = length(spec$dedup_by) > 0
    ))
}

#' Check every registered binding dataset against its data
#'
#' `n_probe_rows` is not a blacklist that is applied -- it counts rows matching
#' `blacklist_probe` so the cost of switching one on is visible before you do.
#' `n_dup_rows` counts rows a `dedup_by` would collapse: how far the dataset is
#' from one row per (`sample_id`, `target_locus_tag`).
#' No dataset currently sets `target_blacklist`; see the comment on
#' `CC_TARGET_BLACKLIST` in registry.R for why.
#'
#' @param vdb A VirtualDB handle with `dto_universe` registered.
#' @param dbs Binding db_names to audit. Defaults to the whole registry.
#' @param blacklist_probe Locus tags to count occurrences of.
#' @return A tibble, one row per dataset.
dto_registry_audit <- function(vdb,
                               dbs = names(BINDING_DTO_REGISTRY),
                               blacklist_probe = CC_TARGET_BLACKLIST) {
    out <- bind_rows(lapply(dbs, function(db) {
        spec <- BINDING_DTO_REGISTRY[[db]]
        if (is.null(spec)) cli_abort("No BINDING_DTO_REGISTRY entry for {.val {db}}.")
        cli_alert_info("auditing {db} ...")
        .audit_one(vdb, db, spec, blacklist_probe)
    }))

    bad <- out$db_name[!is.na(out$direction_ok) & !out$direction_ok]
    missing <- out$db_name[out$status != "ok"]
    # Duplicate (sample, target) rows with no dedup_by are the silent failure:
    # the target appears twice in its ranked list and twice in the scope the
    # perturbation side is filtered against.
    undeduped <- out$db_name[!is.na(out$n_dup_rows) & out$n_dup_rows > 0 &
        !is.na(out$dedup_declared) & !out$dedup_declared]

    if (length(missing)) {
        cli_alert_danger("rank_col does not exist in: {.val {missing}}")
    }
    if (length(bad)) {
        cli_alert_danger(
            "rank direction disagrees with the tiebreak column in: {.val {bad}}"
        )
    }
    if (length(undeduped)) {
        cli_alert_danger(c(
            "Repeated (sample_id, target_locus_tag) rows and no {.arg dedup_by}: ",
            "{.val {undeduped}}"
        ))
    }
    if (!length(bad) && !length(missing) && !length(undeduped)) {
        cli_alert_success(
            "All {nrow(out)} binding datasets: column exists, direction consistent, one row per sample/target."
        )
    }

    out
}

#' What truncation costs the staged pair
#'
#' Compares the two policies on the *uncapped* lists for the staged pair:
#'
#' * `"rank"` -- `rank_value <= max_rows`. Keeps whole rank blocks, so no tie is
#'   ever split, but a list can come out longer than `max_rows` rows.
#' * `"row"` -- `sort_key <= max_rows`. Never longer than `max_rows` rows, but
#'   the boundary can land inside a block of tied ranks and keep an arbitrary
#'   subset of equals; `n_samples_tie_split` and `max_tied_rows_lost` count that.
#'
#' Both bound the thing the cap is actually for: DTO scales with the square of
#' the number of *distinct* ranks, so `med_ranks_full / med_ranks_kept` squared
#' is the rough speedup, and it is the same for either policy.
#'
#' Requires a pair staged with `dto_stage_pair()`.
#'
#' @param vdb A VirtualDB handle.
#' @param max_rows The cap to evaluate.
#' @param ranking Perturbation ranking to evaluate.
#' @return A tibble, one row per side.
dto_tie_split_report <- function(vdb, max_rows = 300, ranking = c("effect", "pvalue")) {
    if (is.null(.dto_stage$binding_db)) {
        cli_abort("Nothing staged. Call {.code dto_stage_pair(vdb, binding_db, pert_db)} first.")
    }
    ranking <- match.arg(ranking)
    n <- as.integer(max_rows)

    one <- function(side, list_sql) {
        res <- vdb$query(glue("
WITH l AS ({list_sql}),
marked AS (
    SELECT sample_id, rank_value, sort_key,
           MAX(CASE WHEN sort_key = {n} THEN rank_value END)
               OVER (PARTITION BY sample_id) AS boundary
    FROM l
),
per_sample AS (
    SELECT sample_id,
           COUNT(*) AS rows_full,
           COUNT(DISTINCT rank_value) AS ranks_full,
           COUNT(*) FILTER (WHERE rank_value <= {n}) AS rows_rank,
           COUNT(DISTINCT rank_value) FILTER (WHERE rank_value <= {n}) AS ranks_rank,
           COUNT(*) FILTER (WHERE sort_key <= {n}) AS rows_row,
           COUNT(*) FILTER (WHERE sort_key > {n} AND rank_value = boundary) AS split_lost
    FROM marked GROUP BY sample_id
)
SELECT COUNT(*) AS n_samples,
       MEDIAN(rows_full) AS med_rows_full,
       MEDIAN(ranks_full) AS med_ranks_full,
       MEDIAN(rows_rank) AS med_rows_kept_rank,
       MAX(rows_rank) AS max_rows_kept_rank,
       MEDIAN(ranks_rank) AS med_ranks_kept,
       MEDIAN(rows_row) AS med_rows_kept_row,
       COUNT(*) FILTER (WHERE split_lost > 0) AS n_samples_tie_split,
       MAX(split_lost) AS max_tied_rows_lost
FROM per_sample
"))
        bind_cols(tibble(side = side), .as_tbl(res))
    }

    out <- bind_rows(
        one("binding", dto_binding_list_sql(.dto_stage$b_spec)),
        one(paste0("perturbation (", ranking, ")"), dto_pert_list_sql(.dto_stage$p_spec, ranking))
    )
    out$approx_speedup <- round((out$med_ranks_full / out$med_ranks_kept)^2, 1)
    out
}