scripts/dto_prep/registry.R

# Per-dataset DTO configuration.
#
# Everything that differs between datasets lives here: which column ranks the
# targets, which direction is "better", how ties break, and what counts as a
# significant call. Adding a promoter set to the collection means adding one
# entry to BINDING_DTO_REGISTRY -- no new code.
#
# Registration is opt-in: a dataset in brentlab_yeast_collection.yaml that has
# no entry here is not run. dto_check_registry() reports those so they are not
# forgotten.

library(cli)

# Calling cards targets that tfbpshiny's top-N analysis excludes
# (materialize/comparison/topn.py::CC_TARGET_BLACKLIST).
#
# NOT applied by default: the pre-refactor DTO prep did not exclude them, and
# turning it on here would make new DTO results incomparable with published
# ones. To adopt it, pass `target_blacklist = CC_TARGET_BLACKLIST` to the
# calling cards entries below.
CC_TARGET_BLACKLIST <- c("YOR201C", "YOR202W", "YOR203W", "YCL018W", "YEL021W")

#' Describe how a binding dataset is ranked for DTO
#'
#' @param rank_col Column the targets are ranked by.
#' @param rank_asc `TRUE` when lower values rank better (p-values), `FALSE` when
#'   higher values rank better (enrichment, peak scores).
#' @param tiebreak_col Column that orders rows sharing a rank. Reproduces the
#'   stable-sort behaviour of the original `arrange(desc(enrichment))` before
#'   `arrange(pvalue_rank)`. `NULL` falls back to `target_locus_tag`, which keeps
#'   output reproducible across runs.
#' @param tiebreak_asc Direction of `tiebreak_col`.
#' @param sig_filter SQL predicate applied before ranking. `NULL` means no
#'   pre-rank filter, which is the correct handling for peak datasets where a
#'   called peak is itself the significance criterion. Note that a `NULL`
#'   `rank_col` is always dropped regardless, when the pair's temp table is built
#'   rather than at ranking time -- see `dto_bind_table_sql()`. That is what makes
#'   `sig_filter = NULL` right for the peak sets even though they report every
#'   promoter: "a peak was called here" is expressed as a non-NULL `rank_col`.
#' @param target_blacklist Target locus tags dropped before ranking.
#' @param dedup_by SQL `ORDER BY` terms, best row first, collapsing multiple
#'   rows for one (`sample_id`, `target_locus_tag`) to one row. `NULL` (the
#'   default) means the dataset already reports one row per sample and target,
#'   which is true of every dataset here except harbison. Written against the
#'   source view's own columns, and applied before ranking and before the
#'   significance filter. `rank_col` and `tiebreak_col` are appended
#'   automatically, so only the terms that come *first* belong here.
#' @return A list carrying the above.
binding_spec <- function(rank_col,
                         rank_asc,
                         tiebreak_col = NULL,
                         tiebreak_asc = FALSE,
                         sig_filter = NULL,
                         target_blacklist = character(),
                         dedup_by = NULL) {
    stopifnot(is.character(rank_col), length(rank_col) == 1)
    stopifnot(is.logical(rank_asc), length(rank_asc) == 1)
    stopifnot(is.null(dedup_by) || is.character(dedup_by))

    list(
        rank_col = rank_col,
        rank_asc = rank_asc,
        tiebreak_col = tiebreak_col,
        tiebreak_asc = tiebreak_asc,
        sig_filter = sig_filter,
        target_blacklist = target_blacklist,
        dedup_by = dedup_by
    )
}

#' Does this perturbation dataset support a p-value-ranked list?
#'
#' The single place that question is asked. A dataset with no p-value column
#' gets `pr/effect/` and nothing else.
#'
#' @param spec A `pert_spec()`.
#' @return `TRUE` if a `pr/pvalue/` list can be built.
pert_has_pvalue <- function(spec) !is.null(spec$pvalue_col)

#' Describe how a perturbation dataset is standardised for DTO
#'
#' @param effect_col Column holding the perturbation effect size.
#' @param pvalue_col Column holding the p-value. `NULL` emits a constant `0.0`,
#'   matching the original script's `mutate(pvalue = 0)`, so every row of a
#'   dataset that reports no p-value clears the significance gate. Nothing is
#'   ever *ranked* on that constant: a dataset with `pvalue_col = NULL` gets no
#'   `pr/pvalue/` directory and no `pr_pvalue` column in `lookup.txt`, because
#'   there is no p-value to order targets by. Its targets are ranked by
#'   descending |effect| in `pr/effect/` only.
#' @param dedup Keep only the max-|effect| row per (sample_id, target_locus_tag).
#'   Needed where multiple probes map to one locus.
#' @param exclude_wt Drop regulators whose locus tag starts with `WT-`.
#' @param effect_na_fill Value substituted for a NULL effect. `NULL` leaves it.
#' @param pvalue_na_fill Value substituted for a NULL p-value. `NULL` leaves it.
#' @param sig_filter SQL predicate applied before ranking, written against the
#'   standardised `effect` / `pvalue` columns.
#' @return A list carrying the above.
pert_spec <- function(effect_col,
                      pvalue_col = NULL,
                      dedup = TRUE,
                      exclude_wt = FALSE,
                      effect_na_fill = NULL,
                      pvalue_na_fill = NULL,
                      sig_filter = "pvalue <= 0.1") {
    stopifnot(is.character(effect_col), length(effect_col) == 1)

    list(
        effect_col = effect_col,
        pvalue_col = pvalue_col,
        dedup = dedup,
        exclude_wt = exclude_wt,
        effect_na_fill = effect_na_fill,
        pvalue_na_fill = pvalue_na_fill,
        sig_filter = sig_filter
    )
}

# ---------------------------------------------------------------------------
# Binding datasets
# ---------------------------------------------------------------------------
#
# Four shapes:
#   * calling cards       -- rank on poisson_pval, tiebreak on enrichment
#   * promoter enrichment -- rank on log_poisson_pval, tiebreak on enrichment
#   * peak calls          -- rank on a score, no p-value, so no sig_filter
#   * ChIP-chip           -- rank on pvalue, tiebreak on effect, plus a dedup

.cc_spec <- function() {
    binding_spec(
        rank_col = "poisson_pval", rank_asc = TRUE,
        tiebreak_col = "callingcards_enrichment", tiebreak_asc = FALSE,
        sig_filter = "poisson_pval <= 0.1"
    )
}

.promoter_enrichment_spec <- function() {
    binding_spec(
        rank_col = "log_poisson_pval", rank_asc = TRUE,
        tiebreak_col = "enrichment", tiebreak_asc = FALSE,
        sig_filter = "log_poisson_pval <= ln(0.1)"
    )
}

# The re-called peak sets report every promoter their set defines, scoring only
# the ones a peak was called in: `max_score` is NULL elsewhere, and `n_peaks` is 0
# for a promoter with no peak at all. "A peak was called here" is therefore the
# significance criterion and needs no `sig_filter` -- a NULL `rank_col` is dropped
# when the pair's temp table is built. For the HOMER (ChEC) sets a promoter whose
# only peak appeared in a single replicate also carries a NULL `max_score` while
# keeping `n_peaks > 0`, so it is excluded too, which is the intended reading of
# the authors' two-replicate requirement.
.homer_macs_peak_spec <- function() {
    binding_spec(
        rank_col = "max_score", rank_asc = FALSE,
        tiebreak_col = "n_peaks", tiebreak_asc = FALSE,
        sig_filter = NULL
    )
}

BINDING_DTO_REGISTRY <- list(
    # Calling cards, one entry per promoter set.
    callingcards_kang = .cc_spec(),
    callingcards_mindel = .cc_spec(),
    callingcards_500bp = .cc_spec(),
    callingcards_intergenic = .cc_spec(),

    # ChIP-chip (Harbison 2004).
    #
    # No condition filter. Each of the 352 sample_ids carries exactly one of the
    # 14 conditions (204 are YPD), so ranking per sample already keeps conditions
    # apart -- a stress-condition sample becomes its own ranked list rather than
    # contaminating the YPD one. This is where tfbpshiny's top-N analysis differs:
    # it aggregates across samples, so it has to restrict to YPD first
    # (topn.py::_HARBISON_DEDUP_CTE).
    #
    # 7,744 (sample, target) pairs carry two rows, so dedup_by is needed. The
    # most significant row wins; where p-values tie -- 5,157 of those pairs --
    # the larger |effect| wins. NaN is how this dataset spells missing (2.1% of
    # rows), and DuckDB sorts NaN above every number: harmless under `pvalue ASC`
    # where it lands last on its own, but it would win an `abs(effect) DESC`, so
    # isnan() demotes it explicitly. That matches the pre-refactor script, which
    # filled a missing effect with 0 before taking the max.
    harbison = binding_spec(
        rank_col = "pvalue", rank_asc = TRUE,
        tiebreak_col = "effect", tiebreak_asc = FALSE,
        sig_filter = "pvalue <= 0.1",
        dedup_by = c(
            "pvalue ASC NULLS LAST",
            "isnan(effect) ASC NULLS LAST",
            "abs(effect) DESC NULLS LAST"
        )
    ),

    # ChIP-exo promoter enrichment (Rossi 2021).
    rossi = .promoter_enrichment_spec(),
    rossi_mindel = .promoter_enrichment_spec(),
    rossi_500bp = .promoter_enrichment_spec(),
    rossi_intergenic = .promoter_enrichment_spec(),

    # ChEC-seq promoter enrichment (Mahendrawada 2025).
    chec_m2025 = .promoter_enrichment_spec(),
    chec_m2025_mindel = .promoter_enrichment_spec(),
    chec_m2025_500bp = .promoter_enrichment_spec(),
    chec_m2025_intergenic = .promoter_enrichment_spec(),

    # Peaks as published by the original authors.
    rossi_peaks = binding_spec(
        rank_col = "peak_score", rank_asc = FALSE,
        tiebreak_col = "n_peaks", tiebreak_asc = FALSE,
        sig_filter = NULL
    ),
    # chec_m2025_peaks carries peak_score only -- no peak count to break ties on.
    chec_m2025_peaks = binding_spec(
        rank_col = "peak_score", rank_asc = FALSE,
        sig_filter = NULL
    ),

    # Peaks re-called against each promoter set (MACS for Rossi, HOMER for ChEC).
    rossi_peaks_kang = .homer_macs_peak_spec(),
    rossi_peaks_mindel = .homer_macs_peak_spec(),
    rossi_peaks_500bp = .homer_macs_peak_spec(),
    rossi_peaks_intergenic = .homer_macs_peak_spec(),
    chec_m2025_peaks_kang = .homer_macs_peak_spec(),
    chec_m2025_peaks_mindel = .homer_macs_peak_spec(),
    chec_m2025_peaks_500bp = .homer_macs_peak_spec(),
    chec_m2025_peaks_intergenic = .homer_macs_peak_spec()
)

# ---------------------------------------------------------------------------
# Perturbation datasets
# ---------------------------------------------------------------------------
# These reproduce the handling in the pre-refactor dto_preparation2.R exactly.

PERTURBATION_DTO_REGISTRY <- list(
    kemmeren = pert_spec(
        effect_col = "Madj", pvalue_col = "pval",
        dedup = TRUE, exclude_wt = TRUE
    ),
    hackett = pert_spec(
        effect_col = "log2_shrunken_timecourses", pvalue_col = NULL,
        dedup = TRUE, exclude_wt = TRUE
    ),
    degron = pert_spec(
        effect_col = "log2FoldChange", pvalue_col = "padj",
        dedup = FALSE,
        effect_na_fill = 0, pvalue_na_fill = 1
    ),
    hu_reimand = pert_spec(
        effect_col = "effect", pvalue_col = "pval",
        dedup = TRUE
    ),
    hughes_knockout = pert_spec(
        effect_col = "mean_norm_log2fc", pvalue_col = NULL,
        dedup = TRUE
    ),
    hughes_overexpression = pert_spec(
        effect_col = "mean_norm_log2fc", pvalue_col = NULL,
        dedup = TRUE
    )
)

#' Reconcile the registries against the collection config
#'
#' Errors when a registry entry names a dataset the collection does not have.
#' Informs -- but does not error -- when the collection has a binding or
#' perturbation dataset with no registry entry, which is the nudge to add one
#' after a new promoter set lands in the YAML.
#'
#' @param vdb A VirtualDB handle.
#' @param quiet Suppress the "not registered" report.
#' @return `invisible(TRUE)`.
dto_check_registry <- function(vdb, quiet = FALSE) {
    available <- vdb$get_datasets()
    registered <- c(names(BINDING_DTO_REGISTRY), names(PERTURBATION_DTO_REGISTRY))

    unknown <- setdiff(registered, available)
    if (length(unknown)) {
        cli_abort(c(
            "Registry names {length(unknown)} dataset{?s} absent from the collection.",
            "x" = "{.val {unknown}}",
            "i" = "Check the {.field db_name} values in {.file brentlab_yeast_collection.yaml}."
        ))
    }

    if (quiet) {
        return(invisible(TRUE))
    }

    tagged <- function(role) {
        keep <- vapply(available, function(db) {
            tags <- reticulate::py_to_r(vdb$get_tags(db))
            identical(tags[["data_type"]], role)
        }, logical(1))
        available[keep]
    }

    unregistered_binding <- setdiff(tagged("binding"), names(BINDING_DTO_REGISTRY))
    unregistered_pert <- setdiff(tagged("perturbation"), names(PERTURBATION_DTO_REGISTRY))

    if (length(unregistered_binding)) {
        cli_alert_warning(
            "Binding dataset{?s} in the collection with no registry entry: {.val {unregistered_binding}}"
        )
    }
    if (length(unregistered_pert)) {
        cli_alert_warning(
            "Perturbation dataset{?s} in the collection with no registry entry: {.val {unregistered_pert}}"
        )
    }
    if (!length(unregistered_binding) && !length(unregistered_pert)) {
        cli_alert_success("Registry covers every binding and perturbation dataset in the collection.")
    }

    invisible(TRUE)
}