A single-cell simulator with a dial for batch strength: scDeepSim varies batch effects and trajectory geometry on demand
Synopsis
Benchmarking single-cell methods has been limited by costly annotation of real data and by datasets that capture only one observed condition. scDeepSim encodes expression into latent subvectors assigned to cell type, batch and residual variation, learns their distribution with latent diffusion, and at generation time modifies only the target subvector, imposing batch effects or constructing branching trajectories at user-set strengths. On scIB human pancreas data at intervention strength α=1, the structured model reached a mean cell-type silhouette width of 0.519 versus 0.432 for a plain variational autoencoder.
Interpretation
A single-cell simulator can now tune batch-effect strength continuously while keeping cell-type structure, and that strength itself becomes known ground truth for benchmarking. Earlier statistical simulators exposed interpretable parameters for batches and differentiation paths, but realism depended on their modeling assumptions; deep generative models learned complex expression patterns yet offered no transparent way to change one biological or technical factor alone. On scIB human pancreas data, restricting the intervention to the batch subspace raised within-cell-type batch silhouette width and lowered iLISI as strength α rose from 0 to 1, while cell-type silhouette width and cLISI stayed comparatively stable; a plain variational autoencoder also increased batch separation but lost cell-type silhouette width and showed a marked cLISI increase.
Generation runs in three steps: each expression profile is encoded into a latent vector split into cell-type, batch and residual blocks, the distribution of those vectors is learned, and only the target subspace is affinely transformed before decoding. The structured variational autoencoder uses supervised classifiers so each label is predictable from its designated subvector, and conditional adversarial classifiers with gradient reversal suppress leakage into the other subvectors; a latent diffusion model then samples by denoising in the compact latent space. Interventions build a whitening-recoloring map from the source and target batch subvector means and covariances and interpolate between the original subvector and its mapped counterpart by α, extrapolating when α exceeds 1; the batch experiments fix one simulator trained with seed 42 and use three generation seeds, each producing a cohort pair with 500 cells per cell type.
Perspective
The demonstrated scope is simulation realism and batch intervention on scIB human lung, pancreas and immune data, plus branching-trajectory experiments on scVelo mouse pancreatic endocrinogenesis data that vary one parameter at a time; intervened samples have no direct experimental counterpart, so realism is supported only indirectly by agreement in per-gene means and standard deviations with held-out real data.
Batch and cell-type information is still partly encoded outside the designated subspaces, and stronger interventions weaken cell-type structure; current controls do not support gene-specific interventions; extrapolation beyond the observed range may push latent representations into regions poorly covered by training data, where decoded expression profiles are harder to assess for realism.