Train a feature-level texture model for the mediator panel
Source:R/feature_texture.R
train_feature_texture.RdLearns the joint distribution of a features x samples mediator matrix using a Gaussian copula with flexible marginal distributions. For each feature, the marginal is fitted as either an empirical CDF (default) or the best parametric family (normal, log-normal, gamma, beta) selected by AIC with a Kolmogorov-Smirnov goodness-of-fit check. Cross-feature dependence is captured by the Gaussian copula correlation matrix.
Arguments
- M_matrix
Mediator panel,
features x samples(e.g. gene or transcript expression, one row per feature, one column per sample).- marginal_method
Marginal fitting method:
"auto"(default, uses parametric if KS test passes at p > 0.05, otherwise empirical),"empirical"(always use empirical CDF), or"parametric"(always use best parametric fit by AIC).
Value
An iconic_feature_texture S3 object: a named list with
marginals (list of per-feature marginal specs), copula_cor
(p x p correlation matrix), n_features, n_samples,
marginal_method, and marginal_types (summary of how many
features used each method).
Details
The resulting model can be sampled via sample_feature_texture() to
produce realistic synthetic feature vectors that preserve the marginal
shapes and correlation structure of the user's mediator panel.
Examples
M <- matrix(rnorm(30 * 200), 30, 200) # 30 transcripts, 200 samples
ft <- train_feature_texture(M)
draws <- sample_feature_texture(ft, 100) # 30 x 100 synthetic draws