univi.perturbation.predict_feature_perturbation

univi.perturbation.predict_feature_perturbation(model, adata, *, source_modality, target_modality, features, mode='off', value=0.0, device='cpu', batch_size=256)

Compare baseline and edited-input decoder means without retraining.

adata.X must already be in the model’s input space. features are exact source var_names, not latent coordinates or genes mapped heuristically to peaks. Supported edits are off, set, add, scale. This reports model sensitivity, not a causal perturbation effect. The input object is never modified.