Fit the DA-cVAE single-sample discriminator

Trains a domain-adversarial conditional VAE on the standardised Stage-1 structural-ILR balances (encoder \(\to\mu\), FIT-only conditional decoder \(p(\mathrm{balances}\mid z,\mathrm{platform})\), KL prior, FIT-only gradient-reversal platform adversary, supervised disease label head on \(\mu\); via reticulate-python torch, full-batch Adam), EXPORTS the encoder (trunk up to and including the \(\mu\) head) and the disease LABEL head to R as plain numeric arrays, fits a frozen kNN-distance novelty reference over the train control latents, and DISCARDS the python module. The fitted model holds no external pointer; scoring is pure base-R with no platform/cohort input.

Usage

fit_dacvae(X_train, y_train, meta_train = NULL, annotation, hp = list())

Arguments

X_train: Numeric matrix (samples x features) of non-negative abundances with unique, non-empty feature names (colnames).
y_train: 0/1 disease labels (1 = case), length nrow(X_train), with at least one case and one control.
meta_train: Optional per-sample metadata carrying a platform/assay column (the conditional-decoder + adversary domain). With \(\ge 2\) platforms the conditional decoder is always used; the adversary is engaged only when grl_lambda > 0 (so grl_lambda = 0 here is the no-GRL ablation: conditional decoder preserved, adversary off). Must have one row per row of X_train.
annotation: data.frame with columns feature (miRNA id), cluster (genomic-cluster id), gc (GC fraction) — the Stage-1 SBP priors.
hp: Optional list of hyperparameters (all frozen into the model). Fields: enc_hidden (encoder trunk widths; default c(64L,32L)), dec_hidden (FIT-only decoder widths; default c(32L,64L)), latent_dim (\(L\); default 16L), activation ("relu"/"tanh"), epochs (default 200L), lr (default 1e-3), weight_decay (default 1e-4), kl_beta (\(\beta\); default 1.0), recon_weight (\(\rho\); default 1.0), grl_lambda (adversary strength; default 1.0; 0 = no-GRL ablation), domain_hidden (FIT-only adversary widths; default c(32L)), platform_col (NULL = auto-detect), novelty_k (kNN novelty neighbours; default 5L), input_type ("abundance" (default) or "ct" for qPCR cycle-threshold input, converted to relative abundance via a frozen per-miRNA train-median reference), ct_efficiency (assay efficiency \(E\) for \(E^{-\Delta Ct}\); default 2), ct_floor (censored-Ct MZR floor abundance; NULL = fit-computed), pseudocount / aggregate / min_balance_coverage / min_part (Stage-1 balance controls), min_features (score-time overlap floor; default 3L), device ("cpu"/"cuda"/"auto"), seed (default 42L).

Value

Object of class dacvae_model: a list with feature_universe, sbp, balance_names, balance controls (pseudocount, aggregate, min_balance_coverage), center/scale (frozen balance standardisation), weights (exported encoder layers), activation, head_w/head_b (frozen disease head), latent_dim, platform_levels, n_platforms (platforms used by the conditional decoder; 1 = unconditional), adv_engaged (FALSE = no-GRL ablation), grl_lambda, novelty_ref (control-latent reference matrix), novelty_k, device, seed, hp.

Usage

Arguments

Value

See also