CrossDomainAdjust

CrossDomainAdjust is a small R package for cross-domain feature correction. The current version implements lambda-controlled projection correction; a centering-based method can be added later under the same package.

• lambda = 0: keep the original features.
• lambda = 1: fully remove the learned cross-domain projection component.
• 0 < lambda < 1: partial correction.

The method works for two, three, four, or more domains. With K domains, the domain-shift subspace has rank at most K - 1 for either built-in anchor. An arbitrary custom anchor outside the affine hull of the centroids can increase this to K, subject to the number of features. It learns an orthogonal projection matrix P = B %*% t(B) from training-domain centroid differences and corrects features by:

x_corrected = x - lambda * ((x - anchor) %*% P)

The new sample’s domain label is checked for consistency, while the correction itself is projection-based and does not apply a domain-wise mean shift.

The projection anchor is user-controlled:

fit_domain_projector(x, domain, anchor = "domain_mean") # equal-weight domain center
fit_domain_projector(x, domain, anchor = "sample_mean") # sample-size-weighted center
fit_domain_projector(x, domain, anchor = custom_center_vector)

Example

library(CrossDomainAdjust)

set.seed(1)
x <- rbind(
  matrix(rnorm(40), nrow = 10),
  matrix(rnorm(40) + 1.2, nrow = 10),
  matrix(rnorm(40) - 0.8, nrow = 10)
)
colnames(x) <- paste0("gene", 1:4)
domain <- rep(c("TCGA", "GEO_A", "GEO_B"), each = 10)

projector <- fit_domain_projector(x, domain, anchor = "domain_mean")
projector

new_sample <- c(gene1 = 2.1, gene2 = 0.4, gene3 = 1.3, gene4 = -0.2)

project_new_sample(projector, new_sample, domain = "GEO_A", lambda = 0)
project_new_sample(projector, new_sample, domain = "GEO_A", lambda = 0.5)
project_new_sample(projector, new_sample, domain = "GEO_A", lambda = 1)

Included Examples

After installation, run:

example_dir <- system.file("examples", package = "CrossDomainAdjust")
source(file.path(example_dir, "01_pca_lambda_overlap.R"))
source(file.path(example_dir, "02_two_domain_new_sample_projection.R"))
source(file.path(example_dir, "03_three_domain_new_sample_projection.R"))

The first example simulates two cohorts with a batch effect and draws PCA panels for multiple lambda values. The second example takes two cohort matrices, a lambda value, one new sample, and its source cohort, then outputs the corrected feature vector and a PCA plot showing cohort centers plus the sample before and after correction. The third example repeats the same idea for three domains, where the centers form a triangle and the projection can remove up to two independent domain directions.

Typical GI Prognosis Use

Use your training expression matrix after feature construction, for example a 15-gene within-sample rank matrix:

projector <- fit_domain_projector(rank_matrix, domain = cohort_domain, anchor = "domain_mean")
corrected_training <- project_features(projector, rank_matrix, lambda = 0.55)
corrected_one_patient <- project_new_sample(
  projector,
  sample = named_15_gene_rank_vector,
  domain = "GEO",
  lambda = 0.55
)