Metabodecon Models — mdm • metabodeconplus

Fit (fit_mdm()) or cross-validate (benchmark()) a binary classification model built on a set of NMR spectra. Both run the full deconvolute -> align -> snap -> featurize -> fit pipeline with sensible defaults and expose only the parameters a typical user tunes; the classification backend is chosen via model. Power users who need to swap individual pipeline stages can call the internal engines fit_mdm_internal() / benchmark_internal(), which take pluggable decon_fun / align_fun / snap_fun / feat_fun / fit_fun / predict_fun arguments.

fit_mdm() runs the pipeline once, or iterates over the cartesian product of npmax / maxShift / maxCombine when any is a vector and returns the row with the highest acc (ties broken by auc). benchmark() wraps fit_mdm() in outer k-fold cross-validation to estimate end-to-end performance on held-out spectra.

Usage

fit_mdm(
  x,
  y,
  model = c("ranger", "lasso"),
  npmax = -1L,
  maxShift = -1L,
  maxCombine = 10L,
  nworkers = 1L,
  seed = 1L,
  verbosity = 1L,
  ...
)

benchmark(
  x,
  y,
  model = c("ranger", "lasso"),
  npmax = -1L,
  maxShift = -1L,
  maxCombine = 10L,
  nworkers = 1L,
  seed = 1L,
  verbosity = 2L,
  k = 3L,
  ...
)

Arguments

x: Spectra object.
y: Factor vector with class labels for each spectrum.
model: Classification backend. One of "ranger" (default, probability random forest) or "lasso" (L1-penalised logistic regression via glmnet).
npmax: Max peaks per spectrum. Integer in {-2, -1, 0, 1, ...}, scalar or vector. -1 (default) selects the median per-spectrum Kneedle elbow; -2 selects each spectrum's own elbow.
maxShift: Max CluPA shift in datapoints. Integer >= -1, scalar or vector. -1 (default) means auto (sweep to the alignment-correlation dip).
maxCombine: Reference-snapping window in datapoints. Integer, scalar or vector. Default 10.
nworkers: Number of workers for deconvolution, alignment and the inner fitter.
seed: Random seed. Forwarded to the fitter; also used for stratified fold assignment inside benchmark(). May be a vector for repeated CV.
verbosity: Verbosity level.
...: Further arguments passed on to the internal engine (fit_mdm_internal() / benchmark_internal()), e.g. sfr, igrs, deg, use_rust. Rarely needed.
k: Number of outer folds for benchmark().

Value

fit_mdm() returns an object of class mdm with elements model (trained backend model of the best grid row), ref (a list list(align, snap) for prediction-time replay), params (resolved pipeline parameters of the best row), the scalar performance of the best row (acc, auc, acc_se, auc_se), and mog (the augmented grid with per-row performance).

benchmark() returns a list with elements models (one fitted model per outer fold), predictions (per-spectrum out-of-fold predictions), performance (per-fold acc / auc) and overall (pooled acc / auc).

Examples

# Small, fast illustrative run. `deg` restricts deconvolution to a single
# parameter set and scalar npmax/maxShift/maxCombine give a one-row model
# grid (`mog`); benchmark() does a single 2-fold cross-validation round.
i <- c(1:3, 51:53)                       # 3 spectra per class
x <- sim2[i]
y <- attr(sim2, "group")[i]
deg <- expand.grid(nfit=3, smit=1, smws=3, delta=1.6)
m <- fit_mdm(x, y, npmax=10L, maxShift=1L, maxCombine=2L, deg=deg, verbosity=0)
bm <- benchmark(x, y, npmax=10L, maxShift=1L, maxCombine=2L, deg=deg, k=2L, verbosity=0)
# `model = "lasso"` selects L1-penalised logistic regression instead.