TraceSim: A Generative Simulator and Benchmark for Joint scRNA-seq and Lineage Tracing

Single-cell lineage tracing paired with transcriptomics is fundamentally transforming our understanding of cellular differentiation. However, due to immense technical difficulties and experimental complexity, paired barcoding–scRNA-seq datasets remain scarce, fragmented across only a handful of studies, and frequently corrupted by stochastic barcode silencing and site-level dropout. Because transcriptomic sequencing is destructive, the continuous trajectories of differentiating cells cannot be observed directly, leaving accurate generative simulators as the only practical path to the standardized ground- truth benchmarks needed for rigorous method validation. Yet the few existing simulators generate gene expression as a strictly Markovian process, failing to capture the fate bias seen in real biological systems, where early progenitor cells already carry coordinated transcriptomic signatures of their distant terminal fates. Consequently, modern machine learning methods designed to predict cell fate cannot be rigorously benchmarked: the synthetic data systematically lacks the predictive signals these models are meant to learn. We introduce TraceSim, a highly controllable, continuous multi-fate simulation framework. By modeling cellular differentiation as an overdamped Langevin process on a Waddington landscape, TraceSim generates non-Markovian transcriptomic trajectories whose fate bias is tunable by construction, together with tunable parameters for asynchronous apoptosis, CRISPR barcoding errors, and transcriptomic signal-to-noise sparsity. Through a comprehensive weakness-targeted evaluation, we demonstrate that TraceSim successfully maps the predictive limits of modern architectures and shows the structural limitations of current simulators. TraceSim establishes a new standard for validating lineage reconstruction and early cell fate prediction methods.