Runs the streamlined full CLAMP model.

This version performs latent-variable decomposition of a gene expression matrix Y guided by prior pathway annotations priorMat, with simplified and lighter regularization compared to the original extended CLAMP variant. The algorithm alternates updates of Z, B, and U, where U captures pathway-latent variable associations inferred directly from the data without ridge-regularized projections (Chat is not used).

CLAMPfull(
  Y,
  priorMat,
  Chat = NULL,
  svdres = NULL,
  clamp.base.result = NULL,
  clamp_k = NULL,
  svd_k = NULL,
  L1 = NULL,
  L2 = NULL,
  cvn = 5,
  max.iter = 30,
  trace = TRUE,
  maxPath = 10,
  doCrossval = TRUE,
  penalty.factor = rep(1, ncol(priorMat)),
  glm_alpha = 0.9,
  minGenes = 0,
  tol = 5e-04,
  seed = 123456,
  allGenes = FALSE,
  rseed = NULL,
  max.U.updates = Inf,
  pathwaySelection = c("fast", "complete"),
  multiplier = 5,
  adaptive.p = 0.05,
  useNNLS = TRUE,
  useRaw = TRUE,
  refitEvery = 3,
  useSE = FALSE,
  var.prior = TRUE,
  Uscale = FALSE,
  robust.vp = TRUE,
  use_cpp = FALSE,
  clamp_k_method = "elbow"
)

Arguments

Y

Gene expression matrix (genes x samples). Can be dense, sparse (dgCMatrix), or FBM.

priorMat

Binary or weighted prior matrix (genes x pathways) linking genes to pathways.

Chat

Ignored in this version (kept for interface compatibility).

svdres

Optional precomputed SVD result for initialization.

clamp.base.result

Optional result from CLAMPbase() providing initial values.

clamp_k

Number of latent variables for CLAMP (final model rank). If NULL, it is chosen automatically via select_clamp_k().

svd_k

Number of singular values/components to compute in the SVD. If NULL, defaults to max(2, min(n_genes, n_samples) - 1).

L1, L2

Regularization parameters for Z and B. Defaults use values from clamp.base.result.

cvn

Number of folds for pathway-level cross-validation. Default: 5.

max.iter

Maximum number of outer iterations. Default: 30.

trace

Logical; print iteration progress. Default: TRUE.

maxPath

Maximum number of pathways per LV during U-fitting. Default: 10.

doCrossval

Whether to mask prior entries for CV evaluation. Default: TRUE.

penalty.factor

Optional vector of per-pathway penalties for glmnet. Default: 1.

glm_alpha

Elastic net mixing parameter for U estimation. Default: 0.9.

minGenes

Minimum number of genes per pathway. Default: 0.

tol

Convergence tolerance for B updates. Default: 5e-4.

seed

Random seed for CV masking. Default: 123456.

allGenes

If TRUE, adds zero-filled rows for missing genes. Default: FALSE.

rseed

Reproducibility, coordinate descent updates are done in random order.

max.U.updates

Maximum number of U updates (capped by max.iter).

pathwaySelection

Pathway selection mode for U fitting ("fast" or "complete").

multiplier

Variance-prior scaling factor. Default: 5.

adaptive.p

Quantile of negative Z values used to define adaptive thresholding. Default: 0.05.

useNNLS

Whether to use non-negative least squares for U estimation. Default: TRUE.

useRaw

If TRUE, uses unthresholded Z in U updates. Default: TRUE.

refitEvery

Frequency (in U updates) of full refits. Default: 3.

useSE

Logical; whether to use the 1-standard-error rule for internal glmnet fitting. Default is FALSE.

var.prior

Logical; if TRUE, enables adaptive variance prior updates for Z. Default: TRUE.

Uscale

Logical; whether to scale U columns. Default: FALSE.

robust.vp

Logical; winsorize prior-predicted Z2 values to reduce outlier effects. Default: TRUE.

use_cpp

Logical; if TRUE, use C++ implementation for Z updates. Default is FALSE.

clamp_k_method

Method for selecting clamp_k when not provided. One of "elbow" (default), "permutation", "gavish_donoho", or "scaleSVs". Passed to select_clamp_k().

Value

A list with elements:

B

LV loadings on samples (k × samples)

Z

Gene loadings (genes × k)

U

Pathway loadings (pathways × k)

C

Prior matrix used during training (masked if CV)

Z2

Predicted Z from pathway priors

heldOutGenes

Held-out gene lists per pathway (CV mode)

Uauc, Up, summary

CV evaluation metrics if CV is enabled

priorMat, priorMatCV

Final and masked prior matrices

call

Function call

Details

Cross-validation can be used to evaluate pathway-LV specificity, and a variance-based prior (var.prior = TRUE) introduces adaptive shrinkage of Z based on how strongly each latent component aligns with prior pathways. The scaling factor multiplier (default 5) controls the strength of this adaptive shrinkage.

This implementation omits ridge-projected priors (Chat) and uses a lighter variance prior with a lower default multiplier = 5, allowing more flexible latent representations. Setting var.prior = FALSE reproduces standard CLAMP-like updates. Cross-validation, if enabled, masks 20% of gene-pathway associations per column to estimate pathway-LV specificity (reported via AUC and p-values).

Examples

set.seed(1)
mat <- matrix(rnorm(100), nrow = 10, ncol = 10)
base <- CLAMPbase(mat, clamp_k = 5, trace = FALSE, max.iter = 5)
#> ****
#> Computing SVD
#> CLAMP k is set to 5
#> L1 is set to 0.736045659723086
#> L2 is set to 2.20813697916926
prior <- matrix(sample(0:1, 10 * 6, TRUE, prob = c(0.9, 0.1)),
    nrow = 10, ncol = 6
)
fit <- CLAMPfull(
    Y = mat,
    priorMat = prior,
    clamp.base.result = base,
    doCrossval = FALSE,
    adaptive.p = 0,
    max.U.updates = 0,
    max.iter = 1,
    trace = FALSE
)
#> ** CLAMPfull **
#> using provided CLAMPbase result
#> CLAMP k is set to 5
#> L1=0.736045659723086; L2=2.20813697916926
#> Not using cross-validation. No AUCs or p-values