Lingshu-Cell: A generative cellular world model for transcriptome modeling toward virtual cells

Abstract: Modeling cellular states and predicting their responses to perturbations are central challenges in computational biology and the development of virtual cells. Existing foundation models for single-cell transcriptomics provide powerful static representations, but they do not explicitly model the distribution of cellular states for generative simulation. Here, we introduce Lingshu-Cell, a masked discrete diffusion model that learns transcriptomic state distributions and supports conditional simulation under perturbation. By operating directly in a discrete token space that is compatible with the sparse, non-sequential nature of single-cell transcriptomic data, Lingshu-Cell captures complex transcriptome-wide expression dependencies across approximately 18,000 genes without relying on prior gene selection, such as filtering by high variability or ranking by expression level. Across diverse tissues and species, Lingshu-Cell accurately reproduces transcriptomic distributions, marker-gene expression patterns and cell-subtype proportions, demonstrating its ability to capture complex cellular heterogeneity. Moreover, by jointly embedding cell type or donor identity with perturbation, Lingshu-Cell can predict whole-transcriptome expression changes for novel combinations of identity and perturbation. It achieves leading performance on the Virtual Cell Challenge H1 genetic perturbation benchmark and in predicting cytokine-induced responses in human PBMCs. Together, these results establish Lingshu-Cell as a flexible cellular world model for in silico simulation of cell states and perturbation responses, laying the foundation for a new paradigm in biological discovery and perturbation screening.

• The paper introduces a masked discrete diffusion model for high-fidelity simulation of heterogeneous single-cell transcriptomes.
• It uses a two-stage diffusion process with Transformer-based token prediction to accurately recapitulate tissue- and cell-type-specific gene expression.
• The model outperforms established baselines in unconditional simulation and conditional perturbation response, validated by robust quantitative metrics.

Lingshu-Cell: A Discrete Diffusion World Model for Single-Cell Transcriptome Simulation and Perturbation Response

Introduction

Lingshu-Cell introduces a masked discrete diffusion model (MDDM), establishing a new generative framework for modeling the state space of single-cell transcriptomic profiles and simulating conditional dynamics in silico. The principal objective is to move beyond static, descriptive representations of single-cell data, enabling high-fidelity unconditional simulation of heterogeneous cell populations across tissues and species, and accurate prediction of transcriptional changes under perturbation. Lingshu-Cell distinguishes itself from prior foundation and generative models by explicitly capturing transcriptome distributions without introducing inductive biases incompatible with the discrete, sparse, and non-sequential nature of scRNA-seq data.

Methodology

Lingshu-Cell operates through a two-stage diffusion process over discrete gene expression tokens. The forward process randomly masks gene expression measurements in a continuous time framework, and in the reverse process, a Transformer-based model iteratively predicts and reconstructs the masked tokens, effectively modeling the data's orderless combinatorial structure. The model leverages sequence compression to scale to $\sim$18,000 genes per cell, with a group-based embedding downprojection and upprojection to alleviate computational burden and improve robustness to single-gene noise.

Conditional simulation is realized by prepending discrete context tokens (cell type, donor identity, perturbation) to the expression sequence, enabling the model to generate transcriptome states specific to experimental or biological conditions. At inference, classifier-free guidance is utilized to amplify perturbation-specific signal, and prior biological knowledge from external datasets is optionally introduced to guide generation for challenging perturbations.

Empirical Results

Unconditional Simulation Across Tissues and Species

Lingshu-Cell demonstrates superior fidelity in simulating transcriptomic distributions at both the population and subpopulation levels. In large PBMC datasets (PARSE 10M), the model recapitulates cell-type-specific gene expression patterns and subtype proportions with high quantitative concordance, as measured by Pearson/Spearman correlations, maximum mean discrepancy (MMD), gene-averaged 1-Wasserstein distance (1-WD), and local inverse Simpson’s index (iLISI). Across eight human tissues and four non-human species, Lingshu-Cell achieves near-perfect expression correlations (r > 0.996) and outperforms established baselines (scDiffusion, scVI) in all distributional metrics, indicating precise recovery of biological heterogeneity and inter-tissue identity.

Conditional Prediction of Genetic and Cytokine Perturbation Responses

On the Virtual Cell Challenge H1 genetic perturbation benchmark, Lingshu-Cell outperforms all competing models on average rank, achieving the lowest mean absolute error (MAE = 0.052) and highest Pearson-$\Delta$ for expression change (0.306). Ablation studies highlight the additive benefit of classifier-free guidance (CFG, optimal $w=2$), sequence compression, and prior injection: each component yields substantial improvement on perturbation response metrics. The model accurately extrapolates to cytokine-induced perturbations in a complex PBMC donor panel, outperforming both predictive and generative baselines (STATE, scGPT, scVI), and generalizes across unseen donor-perturbation pairs. Metrics such as PDS, DES, AUPRC, and Spearman-based effect size correlations confirm robust identification of distinct perturbation-induced signatures and their relative magnitudes.

Model Architecture and Training

The mask predictor is instantiated as a 13-layer bidirectional Transformer with $D=640$ embedding dimension, multi-head self-attention (n=10, RoPE), and SwiGLU activation. Input gene tokens ($G=18,080$) are compressed by group-wise random linear projection to an internal sequence of length $G_c$ (group size $S=8$ or $32$). Conditional tokens are appended and exempted from masking. Training employs cross-entropy loss over masked tokens, AdamW, cosine learning rate schedules, and distributed mixed-precision optimization. Inference proceeds over a discretized time schedule, with CFG and optional prior injection modulating the generative trajectory.

Theoretical and Practical Implications

Lingshu-Cell’s paradigm resolves key limitations of prior approaches. By modeling the full gene set without selection or ordering, it avoids the inductive biases of AR models and the noise mismatch of continuous DDPMs. Bidirectional and non-autoregressive architecture is inherently compatible with sparse, permutation-invariant transcriptomic data. Critically, Lingshu-Cell bridges the gap between static foundation model representation and generative simulation, supporting a cohesive in silico experimental framework—essential for hypothesis generation, rational perturbation screening, and mapping of dynamic cellular trajectories.

However, the model’s fidelity is measured primarily at the distributional level; single-cell biological plausibility and causal structure are not guaranteed. Predictions should be understood as probabilistic, subject to experimental validation due to the absence of explicit mechanistic constraints. The model currently addresses transcriptome data; integration of multi-omic modalities (e.g., chromatin accessibility, protein, spatial) remains to be addressed, as does the simulation of dose/time-dependent responses and combinatorial interventions.

Future Directions

A natural progression is extending the framework to:

• Jointly model chromatin/spatial/protein data within a unified MDDM,
• Simulate explicit temporal dynamics,
• Generalize to drug, multi-target, and combinatorial/intersectional perturbations,
• Integrate closed-loop experiment-guidance, where model-informed interventions iteratively improve in silico predictions and resource allocation.

These would establish an adaptive computational foundation for high-throughput biological discovery, advancing toward predictive, causal cellular world models.

Conclusion

Lingshu-Cell establishes masked discrete diffusion as a tractable, robust paradigm for comprehensive generative modeling of single-cell transcriptomes. It achieves high-fidelity unconditional simulation and highly competitive conditional prediction of perturbation responses, outperforming domain-specific methods. Its design aligns computational and biological structure, setting the stage for future integrative, multimodal, and dynamic cellular world models and in silico experimentation frameworks ["Lingshu-Cell: A generative cellular world model for transcriptome modeling toward virtual cells" (2603.25240)].