Process simulated and bulk data for Scaden — scaden

Aligns simulated pseudobulk training data with real bulk data, filters genes, log-transforms, and applies per-sample scaling. Two gene-filtering strategies are available via `mode`.

Usage

scaden_process(
  sim_data,
  bulk_data,
  mode = c("tape", "scaden"),
  variance_threshold = 0.98,
  scaler = c("mms", "ss"),
  var_cutoff = 0.1,
  top_var_genes = NULL,
  verbose = TRUE
)

Arguments

sim_data: A list returned by [scaden_sim_pb()], with `train_x` and `train_y`.
bulk_data: Bulk expression matrix or `data.frame`. Gene names must appear in row or column names; orientation is inferred.
mode: Character scalar. `"tape"` (default) reproduces TAPE's `ProcessInputData`; `"scaden"` reproduces original Scaden's preprocessing.
variance_threshold: Numeric scalar in `[0, 1)`. `mode = "tape"` only. Fraction of genes defining the variance cutoff, applied separately to each matrix.
scaler: Character scalar. `mode = "tape"` only. `"mms"` for per-sample min-max, `"ss"` for per-sample standardization.
var_cutoff: Numeric scalar. `mode = "scaden"` only. Genes in `bulk_data` with variance at or below this value are dropped. Set to `NULL` to skip.
top_var_genes: Optional integer scalar. `mode = "scaden"` only. Keep only the most variable genes after `var_cutoff`.
verbose: Logical scalar. Print progress.

Value

A named list with `train_x`, `train_y`, `test_x`, `genename`, `celltypes`, `samplename`, `mode`, and `elapsed`.

Details

The two modes differ only in gene filtering and scaler choice:

`"tape"`: Top-quantile variance cutoff applied to **both** matrices, then intersect. `log(x + 1)`, then `mms` or `ss`.
`"scaden"`: Absolute variance cutoff applied to the **bulk matrix only**, then intersect with the training genes. `log2(x + 1)`, then per-sample min-max (`log_min_max`); `scaler` is ignored.

The log base is in fact immaterial: both scalers are invariant to multiplication by a positive constant, so `log2` and `log` give identical output up to floating-point rounding. The bases are kept literal to their respective sources.

Both scalers operate per sample (`fit_transform(x.T).T` in sklearn), use population statistics (`ddof = 0`), and map constant samples to exactly zero, matching sklearn's `_handle_zeros_in_scale`. The `"tape"` variance filter uses `ddof = 1` to match `pandas.DataFrame.var`.