DELPHAI predicts heterogeneous perturbation responses with learned single-cell fitness

Current perturbation response modelling in single-cell transcriptomics assumes conserved cell mass and loses gene expression information to latent-space decoding. We propose DELPHAI, training a fitness network and an optimal transport network jointly without biological priors, and during inference applying a fitness-gated transport with a direct gene-space retrieval. Demonstrated across two benchmark frameworks, DELPHAI ranks first in predicting differentially expressed genes, while revealing which cell lineages a perturbation depletes.