scripts/dto_prep/io.R

# Writing the DTO input tree.
#
# The tree is split in two at the top level, by how each file is consumed on the
# cluster:
#
#   <outdir>/lists/<pert_db>/<binding_db>/
#       binding/<sample_id>.csv        target_locus_tag,rank   (no header)
#       pr/effect/<sample_id>.csv
#       pr/pvalue/<sample_id>.csv      only when the dataset reports p-values
#
#   <outdir>/index/<pert_db>/<binding_db>/
#       background.csv
#       lookup.txt                     TSV: binding, pr_effect[, pr_pvalue]
#       incomplete.csv
#   <outdir>/index/manifest.csv
#   <outdir>/index/pair_summary.csv
#
# `lists/` is the bulk -- tens of thousands of files of a few KB each, of which
# a given array task reads exactly one or two. Unpacking that onto a parallel
# filesystem is metadata-bound and slow, so it is meant to be shipped as a
# single SquashFS image and read in place:
#
#   mksquashfs <outdir>/lists dto_lists.sqsh -comp zstd -b 1M -no-xattrs -noappend
#
# `index/` is the handful of files per pair that must stay loose and readable:
# the submit script globs `index/*/*/lookup.txt` to enumerate array jobs, and
# every task reads `background.csv`. Together they are a few hundred files, not
# tens of thousands.
#
# The paths written into lookup.txt are `<scratch_path>/lists/...`, mirroring
# this layout, so the same tree works unpacked on scratch or mounted from the
# image -- see DTO_INPUT_ROOT in dto_cluster/run_dto.sh.

library(cli)
library(dplyr)
library(readr)
library(tibble)

# Top-level split: bulk ranked lists vs the per-pair files the job scripts read.
.DTO_LISTS_SUBDIR <- "lists"
.DTO_INDEX_SUBDIR <- "index"

#' Write one ranked list per sample
#'
#' Two implementations of the same contract, both consuming a SELECT that emits
#' `sample_id, target_locus_tag, rank_value, sort_key` and both writing
#' `<dir>/<sample_id>.csv` holding `target_locus_tag,rank_value` with no header,
#' ordered by `sort_key`.
#'
#' The DuckDB path materialises the ranked result once, indexes it on
#' `sample_id`, and issues one small `COPY` per sample. On the worst pair in the
#' collection (hackett x rossi_mindel: 6.6M rows, 1260 samples) that is ~43s and
#' ~150MB of R memory; pulling the same frame into R took minutes and 4.4GB.
#'
#' `COPY ... PARTITION_BY (sample_id)` looks like the obvious way to do this in
#' one statement and is *not* used: it reorders rows within a partition, so the
#' ranked lists come out shuffled.
#'
#' The R path pulls the frame across and splits it. It is slower and much
#' hungrier, and exists as a cross-check on the DuckDB path.
#'
#' @param vdb A VirtualDB handle.
#' @param select_sql SELECT producing the four columns above.
#' @param dir Directory to write into. Cleared first.
#' @param use_duckdb_copy Use the DuckDB path.
#' @return Character vector of sample_ids written.
dto_write_ranked_lists <- function(vdb, select_sql, dir, use_duckdb_copy = TRUE) {
    if (dir.exists(dir)) unlink(dir, recursive = TRUE)
    dir.create(dir, recursive = TRUE, showWarnings = FALSE)

    if (use_duckdb_copy) {
        return(.dto_write_via_duckdb(vdb, select_sql, dir))
    }

    df <- vdb$query(select_sql)
    if (!nrow(df)) {
        return(character())
    }

    df <- df[order(df$sample_id, df$sort_key), ]
    by_sample <- split(df[, c("target_locus_tag", "rank_value")], df$sample_id)
    for (sid in names(by_sample)) {
        write_csv(by_sample[[sid]], file.path(dir, paste0(sid, ".csv")), col_names = FALSE)
    }
    names(by_sample)
}

.DTO_RANKED_TBL <- "_dto_ranked"

.dto_write_via_duckdb <- function(vdb, select_sql, dir) {
    dto_execute(vdb, glue::glue(
        "CREATE OR REPLACE TEMP TABLE {.DTO_RANKED_TBL} AS {select_sql}"
    ))
    on.exit(dto_execute(vdb, glue::glue("DROP TABLE IF EXISTS {.DTO_RANKED_TBL}")), add = TRUE)

    # Turns each per-sample COPY into a point lookup rather than a full scan.
    dto_execute(vdb, glue::glue(
        "CREATE INDEX {.DTO_RANKED_TBL}_sid ON {.DTO_RANKED_TBL}(sample_id)"
    ))

    ids <- vdb$query(glue::glue(
        "SELECT DISTINCT sample_id FROM {.DTO_RANKED_TBL} ORDER BY sample_id"
    ))$sample_id
    if (!length(ids)) {
        return(character())
    }

    for (sid in ids) {
        dest <- file.path(dir, paste0(sid, ".csv"))
        dto_execute(vdb, glue::glue(
            "COPY (
                 SELECT target_locus_tag, rank_value
                 FROM {.DTO_RANKED_TBL}
                 WHERE sample_id = {.sql_str(sid)}
                 ORDER BY sort_key
             ) TO {.sql_str(dest)} (FORMAT CSV, HEADER false)"
        ))
    }

    ids
}

#' Build the lookup and incomplete-case tables for one pair
#'
#' Full-joins the two sides on regulator so that samples present on only one
#' side surface as incomplete cases rather than silently vanishing. Carried over
#' from `create_pr_lookups()` in the pre-refactor script.
#'
#' @param binding_map Tibble of `sample_id`, `regulator_locus_tag`.
#' @param pert_map Same shape, perturbation side.
#' @param scratch_pair_dir Directory the cluster job will see, used to build the
#'   paths written into `lookup.txt`.
#' @param include_pvalue Emit the third (`pr_pvalue`) column. `FALSE` for
#'   datasets with no p-value column, which get no `pr/pvalue/` directory -- a
#'   column pointing at files that were never written would break the job.
#' @param pr_sizes Tibble of `sample_id`, `n_targets` from
#'   `dto_list_sizes_sql()`. Required for `min_pr_targets` to do anything.
#' @param min_pr_targets Drop lookup rows whose perturbation list holds fewer
#'   than this many targets. A handful of targets cannot produce a meaningful
#'   overlap: DTO returns intersection 0 and an empirical p-value of 1, having
#'   spent an array task and 1000 permutations to say so. Dropped rows are
#'   recorded in `incomplete.csv` as `pr_below_min` rather than vanishing. `0`
#'   disables the floor.
#' @return List with `lookup` and `incomplete` tibbles.
dto_build_lookup <- function(binding_map, pert_map, scratch_pair_dir,
                             include_pvalue = TRUE,
                             pr_sizes = NULL,
                             min_pr_targets = 0) {
    binding_samples <- binding_map %>%
        distinct(binding_id = sample_id, regulator_locus_tag)

    pr_samples <- pert_map %>%
        distinct(pr_id = sample_id, regulator_locus_tag)

    lookup_df <- binding_samples %>%
        full_join(pr_samples, by = "regulator_locus_tag", relationship = "many-to-many") %>%
        mutate(
            binding = if_else(
                !is.na(binding_id),
                file.path(scratch_pair_dir, "binding", paste0(binding_id, ".csv")),
                NA_character_
            ),
            pr_effect = if_else(
                !is.na(pr_id),
                file.path(scratch_pair_dir, "pr", "effect", paste0(pr_id, ".csv")),
                NA_character_
            ),
            pr_pvalue = if_else(
                !is.na(pr_id),
                file.path(scratch_pair_dir, "pr", "pvalue", paste0(pr_id, ".csv")),
                NA_character_
            )
        )

    lookup_cols <- if (include_pvalue) {
        c("binding", "pr_effect", "pr_pvalue")
    } else {
        c("binding", "pr_effect")
    }

    lookup_df <- if (!is.null(pr_sizes) && min_pr_targets > 0) {
        sizes <- pr_sizes %>%
            transmute(
                pr_id = as.character(sample_id),
                pr_n_targets = as.integer(n_targets)
            )
        lookup_df %>%
            left_join(sizes, by = "pr_id") %>%
            mutate(pr_below_min = !is.na(pr_id) & pr_n_targets < min_pr_targets)
    } else {
        lookup_df %>%
            mutate(pr_n_targets = NA_integer_, pr_below_min = FALSE)
    }

    list(
        lookup = lookup_df %>%
            filter(!is.na(binding_id), !is.na(pr_id), !pr_below_min) %>%
            select(all_of(lookup_cols)),
        incomplete = lookup_df %>%
            filter(is.na(binding_id) | is.na(pr_id) | pr_below_min) %>%
            mutate(missing_type = case_when(
                is.na(binding_id) & is.na(pr_id) ~ "both",
                is.na(binding_id) ~ "binding",
                is.na(pr_id) ~ "pr",
                pr_below_min ~ "pr_below_min",
                TRUE ~ "unknown"
            )) %>%
            select(regulator_locus_tag, binding_id, pr_id, pr_n_targets, missing_type) %>%
            distinct()
    )
}

#' Write the run manifest
#'
#' Records db_name -> HuggingFace repo/config so nothing downstream has to
#' hard-code the `repo;config;sample_id` composite ID prefixes.
#'
#' @param vdb A VirtualDB handle.
#' @param binding_dbs Binding db_names in the run.
#' @param pert_dbs Perturbation db_names in the run.
#' @param outdir Root of the DTO input tree. The manifest is written under
#'   `index/`, with the other files the cluster reads directly.
#' @return The manifest tibble, invisibly.
dto_write_manifest <- function(vdb, binding_dbs, pert_dbs, outdir) {
    manifest <- bind_rows(
        dto_hf_coords(vdb, binding_dbs) %>% mutate(role = "binding"),
        dto_hf_coords(vdb, pert_dbs) %>% mutate(role = "perturbation")
    )

    index_dir <- file.path(outdir, .DTO_INDEX_SUBDIR)
    dir.create(index_dir, recursive = TRUE, showWarnings = FALSE)
    write_csv(manifest, file.path(index_dir, "manifest.csv"))
    invisible(manifest)
}

#' Prepare every DTO input file for one (binding, perturbation) pair
#'
#' The unit of work: stage the two temp tables, write the three ranked-list
#' sets, the background, the lookup and the incomplete cases, then drop the temp
#' tables.
#'
#' @param vdb A VirtualDB handle, with `dto_universe` already registered.
#' @param binding_db db_name of the binding dataset.
#' @param pert_db db_name of the perturbation dataset.
#' @param outdir Root of the DTO input tree.
#' @param scratch_path Root the cluster job will see, used for `lookup.txt`.
#' @param max_rows Cap each ranked list at this many ranks (or rows -- see
#'   `truncate_by`), applied after every filter and after ranking, so surviving
#'   rank values are the untruncated ones. `NULL` writes full lists. DTO scales
#'   with the square of the number of distinct ranks, so this is the main
#'   runtime lever. `background.csv` is not capped -- it stays the size it would
#'   be with full lists.
#' @param truncate_by `"rank"` cuts at whole rank blocks, keeping every row tied
#'   at the last surviving rank; `"row"` is a hard row cap that can split a tie.
#'   See `.limit_where()`.
#' @param min_pr_targets Omit a regulator from `lookup.txt` when its
#'   perturbation list holds fewer targets than this. See `dto_build_lookup()`.
#' @param use_duckdb_copy Passed to `dto_write_ranked_lists()`.
#' @return A one-row tibble summarising the pair, invisibly.
dto_prepare_pair <- function(vdb,
                             binding_db,
                             pert_db,
                             outdir = here::here("results/dto"),
                             scratch_path = "/scratch/mblab/chasem/dto",
                             max_rows = 300,
                             truncate_by = c("rank", "row"),
                             min_pr_targets = 10,
                             use_duckdb_copy = TRUE) {
    truncate_by <- match.arg(truncate_by)
    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}}.")
    }

    t0 <- Sys.time()

    # Two roots per pair: the ranked lists, which get packed into an image, and
    # the files the job scripts read directly off the filesystem. The scratch
    # path written into lookup.txt points at the lists root, so a mounted image
    # and an unpacked tree are interchangeable.
    lists_pair_dir <- file.path(outdir, .DTO_LISTS_SUBDIR, pert_db, binding_db)
    index_pair_dir <- file.path(outdir, .DTO_INDEX_SUBDIR, pert_db, binding_db)
    scratch_pair_dir <- file.path(scratch_path, .DTO_LISTS_SUBDIR, pert_db, binding_db)
    dir.create(lists_pair_dir, recursive = TRUE, showWarnings = FALSE)
    dir.create(index_pair_dir, recursive = TRUE, showWarnings = FALSE)

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

    on.exit(dto_execute(vdb, dto_drop_pair_tables_sql()), add = TRUE)

    binding_ids <- dto_write_ranked_lists(
        vdb, dto_binding_list_sql(b_spec, max_rows, truncate_by),
        file.path(lists_pair_dir, "binding"), use_duckdb_copy
    )
    pr_effect_ids <- dto_write_ranked_lists(
        vdb, dto_pert_list_sql(p_spec, "effect", max_rows, truncate_by),
        file.path(lists_pair_dir, "pr", "effect"), use_duckdb_copy
    )
    # Datasets with no p-value column get no p-value-ranked list at all: their
    # `pvalue` is a constant placeholder, so the list would be one rank-1 block.
    has_pvalue <- pert_has_pvalue(p_spec)
    pvalue_dir <- file.path(lists_pair_dir, "pr", "pvalue")
    pr_pvalue_ids <- if (has_pvalue) {
        dto_write_ranked_lists(
            vdb, dto_pert_list_sql(p_spec, "pvalue", max_rows, truncate_by),
            pvalue_dir, use_duckdb_copy
        )
    } else {
        # Clear one left behind by an earlier run under different settings.
        if (dir.exists(pvalue_dir)) unlink(pvalue_dir, recursive = TRUE)
        character()
    }

    background <- vdb$query(dto_background_sql())
    write_csv(
        tibble(target_locus_tag = background$target_locus_tag),
        file.path(index_pair_dir, "background.csv"),
        col_names = FALSE
    )

    # Counted off the same SELECT that wrote pr/effect, so the floor is applied
    # to exactly the lists the job would have been handed.
    pr_sizes <- as_tibble(vdb$query(dto_list_sizes_sql(
        dto_pert_list_sql(p_spec, "effect", max_rows, truncate_by)
    )))

    lookups <- dto_build_lookup(
        binding_map = as_tibble(vdb$query(dto_sample_map_sql("binding"))),
        pert_map = as_tibble(vdb$query(dto_sample_map_sql("perturbation"))),
        scratch_pair_dir = scratch_pair_dir,
        include_pvalue = has_pvalue,
        pr_sizes = pr_sizes,
        min_pr_targets = min_pr_targets
    )

    write_tsv(lookups$lookup, file.path(index_pair_dir, "lookup.txt"), col_names = FALSE)
    if (nrow(lookups$incomplete)) {
        write_csv(lookups$incomplete, file.path(index_pair_dir, "incomplete.csv"))
    }

    elapsed <- as.numeric(difftime(Sys.time(), t0, units = "secs"))
    cap <- if (is.null(max_rows)) {
        "uncapped"
    } else {
        glue::glue("<={max_rows} {if (truncate_by == 'rank') 'ranks' else 'rows'}")
    }
    pv <- if (has_pvalue) "" else ", no pr/pvalue (no p-value column)"
    n_below_min <- sum(lookups$incomplete$missing_type == "pr_below_min")
    floor_msg <- if (n_below_min) {
        glue::glue(", {n_below_min} dropped (<{min_pr_targets} targets)")
    } else {
        ""
    }
    cli_alert_success(
        "{pert_db} x {binding_db}: {length(binding_ids)} binding, \\
         {length(pr_effect_ids)} pr, {nrow(lookups$lookup)} pairs, \\
         {nrow(background)} background, {cap}{pv}{floor_msg} ({round(elapsed, 1)}s)"
    )

    invisible(tibble(
        pert_db = pert_db,
        binding_db = binding_db,
        n_binding_samples = length(binding_ids),
        n_pr_samples = length(pr_effect_ids),
        n_pr_pvalue_samples = length(pr_pvalue_ids),
        n_lookup_rows = nrow(lookups$lookup),
        n_incomplete = nrow(lookups$incomplete),
        n_pr_below_min = n_below_min,
        n_background = nrow(background),
        min_pr_targets = min_pr_targets,
        max_rows = max_rows %||% NA_integer_,
        truncate_by = if (is.null(max_rows)) NA_character_ else truncate_by,
        elapsed_sec = elapsed
    ))
}