Elucidating the Design Space of Generative Models for Single-Cell Perturbation Prediction

We introduce $\texttt{ExpressionVAE}$, the first discrete-latent perturbation model for single-cell data: a vector-quantized variational autoencoder paired with a perturbation-conditioned discrete prior. On Replogle and Parse~1M it achieves state-of-the-art on every distributional and cell-eval state metric we evaluate, with order-of-magnitude gaps on Fr\'echet distance and $\mathrm{MMD}^2$ over the strongest continuous-latent baseline. We test two prior families (autoregressive and masked discrete diffusion) and find they achieve effectively identical numbers, isolating the gain to the discrete latent. A controlled output-head ablation further reveals a single design axis governing decoder-head choice, the richness of the inference-time sampling distribution, with standard evaluation metrics partitioning into three groups whose rankings flip along it. Finally, on a held-out CRISPRi reversion benchmark of $1{,}732$ perturbations under inflammatory cytokine stress, the frozen encoder effectively matches scGPT model (trained on $10\times$ larger dataset) on biological selectivity.