Fit frozen logistic-normal empirical-Bayes prior from training data

Estimates per-feature CLR mean and variance on a training cohort. At apply time (apply_logistic_normal_eb), each test-sample CLR observation is shrunk toward the per-feature prior using a precision-weighted Bayesian update: $$\hat{z}_{i,f} = w_f z_{i,f} + (1 - w_f) \mu_f,$$ where $w_f = \sigma^2_f / (\sigma^2_f + \sigma^2_{\mathrm{obs},i})$. The observation variance $\sigma^2_{\mathrm{obs},i}$ scales inversely with sample total counts so shallow-coverage samples shrink harder. Per-feature variance is truly load-bearing (not a scalar heuristic).

Usage

fit_logistic_normal_eb(x_train, pseudocount = 0.5)

Arguments

x_train: Numeric matrix (samples $\times$ features). Must have column names. All values must be non-negative.
pseudocount: Numeric $> 0$. Added before log transform. Default 0.5.

Value

Object of class logistic_normal_eb_fit with components: feature_names, pseudocount, prior_mean, prior_var, median_total_train.

References

Aitchison J, Shen SM. (1980) Biometrika 67: 261–272. Efron B. (2010) Large-Scale Inference. Cambridge UP.

Examples

if (FALSE) { # \dontrun{
set.seed(42)
x_train <- matrix(abs(rnorm(50 * 20, 100, 30)), nrow = 50,
                  dimnames = list(NULL, paste0("miR-", seq_len(20))))
fit <- fit_logistic_normal_eb(x_train)
length(fit$prior_var)
} # }