InDyVAE: An interaction and dependency-aware variational autoencoder for single-cell perturbation response generation

Single-cell perturbation profiling provides a powerful means of studying how genetic interventions reshape cellular transcription, but only a small portion of the possible perturbation space can be measured experimentally. Predicting the remaining conditions requires more than transferring an average treatment effect: jointly applied perturbations may alter one another, cellular and technical states modulate the observed response, and the resulting latent variation must be faithfully expressed at the gene level. We address these coupled requirements with InDyVAE, an interaction- and dependency-aware variational autoencoder that models response generation as a connected sequence of transformations. Perturbation Interaction Learning (PIL) represents the effects of active interventions in a shared context, allowing each treatment's contribution to reflect its co-occurring perturbations while preserving its individual signal. Latent State Dependency Modeling (LSDM) then conditions the perturbation response on basal cellular and quality-associated states, retaining the distinct roles of these sources of variation. Nonlinear Expression Mapping (NEM) translates the integrated state into gene expression through a direct pathway complemented by localized functional responses in latent space. We evaluate InDyVAE on four Perturb-seq datasets using treatment-effect agreement, top-gene overlap, and quality-control compliance under three thresholds. Comparisons with recent reference models, including CRADLE-VAE and AttentionPert, show strong response recovery on Norman and metric-dependent trade-offs between response accuracy and rule-based QC pass rates on Dixit. For single-gene interventions, it achieves higher quality-control pass rates on Replogle and stronger treatment-effect estimates on Adamson, while the remaining comparisons are dataset- and metric-dependent. Ablation experiments further show that the three components contribute complementary capabilities, supporting the proposed organization of interaction, context, and expression modeling. Source code is available at https://github.com/yuqi0373-rgb/InDyVAE.