feat: add pgx.preprocess() as single source of truth for count preprocessing

[ESCRI11]

Extracts the normalization pipeline that lived in the Omics Playground Shiny upload
module (imputedX → normalizedX → cleanX) into one exported playbase function, so a
script or the compute endpoint reproduces the app's X exactly instead of only a
plain log2(counts+prior).

• pgx.preprocess(counts, samples, contrasts, annot, options): NA handling, NPX/NPQ
  back-transform, negatives→0, log2 with data-driven/method prior, group-wise
  missingness filter, imputation, normalization (CPM/TMM/quantile/methylation/
  multi-omics), outlier-sample removal.
• pgx.createPGX() gains an opt-in preprocess= arg: when set and X is NULL, X
  is built from raw counts via pgx.preprocess before de-duplication. Existing
  bare callers are unaffected (X=NULL without preprocess → old log2 fallback).
• Parity test (test-pgx-preprocess.R, 23 assertions) vs an independent copy of the
  app reactives. Verified end-to-end that createPGX(counts, preprocess=opt) yields
  identical X/counts to createPGX(counts=pp$counts, X=pp$X).

[phisanti]

is.multiomics() has two conflicting definitions in playbase, both @exported:

• R/pgx-functions.R:7: is.multiomics(pgx) (takes a pgx object)
• R/pgx-mofa.R:1914: is.multiomics(probes) (takes a character vector of row names)

This is.multiomics(probes) is the one in use (R/pgx-preprocess.R:83) due to the collate order (that in playbase is alphabetical, we do not use [at]include). We should decide which one makes sense and delete the other to prevent confusion.

[phisanti]

Minor comment: In my last code review with Ivo, he recommended avoiding passing list as an argument, such as options = list()) without giving defaults.

[phisanti]

The deduplication logic is applied inconsistently between pgx.createPGX() (in R/pgx-compute.R) and the new pgx.preprocess() (in R/pgx-preprocess.R).

    kk <- intersect(rownames(X), rownames(counts))
    X <- X[kk, , drop = FALSE]
    counts <- counts[kk, , drop = FALSE]

This runs unconditionally on every call. It exists to handle cases where a normalization step drops rows (methylation's BMIQ against its probe-annotation manifest is the one legitimate case), but intersect() returns unique names, and matrix indexing-by-name keeps only the first match for a repeated name. So if counts has duplicate feature names going in, this step silently collapses each duplicate pair down to one arbitrary row — before pgx.createPGX()'s dedup block ever runs.

Net effect: pgx.createPGX()'s average.duplicated/make_unique logic never sees the duplicates it's designed to handle, because pgx.preprocess() has already destroyed one of each pair with no warning.

[phisanti]

Add message logging for diagnostics in case of crash. For example:

  
  info("input: ", nrow(counts), " features x ", ncol(counts), " samples")
  ...
  info("NA/Inf handling + log2 transform done; ", sum(is.na(X)), " NAs remain")
  ...
  if (isTRUE(opt$filter_missing)) {
    info("filtering by missingness (threshold=", opt$filter_threshold, "): ", sum(!sel), " feature(s) excluded")

[phisanti]

norm_method = "reference" without ref_gene crashes with an unhelpful low-level error:

if (identical(opt$norm_method, "reference") && !isTruthy(opt$ref_gene)) {
  stop("[pgx.preprocess] norm_method = 'reference' requires options$ref_gene to be set")
}

This also can be addressed at the level of the upload board blocking the compute if no reference is provided.

[phisanti]

Approved with minor comments

[ESCRI11]

@phisanti addressed all your comments except the defaults one given the conversation we had via private messages

[phisanti]

Approved now