scDeepSim: Deep Generative Simulation for Controlled Benchmarking of Single-Cell Methods

scDeepSim generates synthetic single-cell expression profiles where users can intervene on selected biological or technical factors at specified strengths, enabling controlled benchmarking with benchmark ground truth.

• Why it matters: Real single-cell benchmarks are costly to annotate and may not reflect the diverse settings where methods are applied. scDeepSim targets the need for realistic synthetic data with transparent, factor-level control.
• What they did: The simulator combines a structured variational autoencoder with latent diffusion to produce expression profiles, and it supports user-specified interventions on chosen factors at chosen strengths. It was used to benchmark batch correction and trajectory inference.
• The result: In scDeepSim benchmarks, varying batch-effect strength changes batch mixing and cell-type preservation, and altering trajectory geometry and noise affects pseudotime recovery. This supports testing how performance shifts under controlled biological and technical variations.

AI summary of the abstract · may contain errors · not medical advice · read the paper.