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.
Figures & tables
Figure 1: Overview of the Lingshu-Cell framework. a , Lingshu-Cell employs a masked discrete diffusion model to learn and generate single-cell transcriptomic data. In the forward process, gene expression values are progressively masked (from t=0 to t=T ); in the reverse process, the model iteratively predicts masked values to generate realistic scRNA-seq expression profiles. b , Comparison of generative paradigms. Unlike autoregressive (AR) models that rely on a fixed sequential order and denoising diffusion probabilistic models (DDPMs) that corrupt all positions with continuous noise, Lingshu-Cell randomly masks and predicts gene expression values in an order-independent manner, which is inherently compatible with the orderless structure of gene expression data. c , Application scenarios of Lingshu-Cell, including unconditional generation across diverse human tissues and species, and conditional generation for genetic perturbation and cytokine perturbation response prediction.
Figure 2: Unconditional generation of cell states across diverse species and tissues by Lingshu-Cell. a , UMAP visualization of real and generated cells (10,000 each, randomly sampled) from the PARSE-PBMC dataset, colored by cell type annotation (left) and normalized expression (log1p) of canonical marker genes for each cell type. b , Comparison of cell type proportions between real and generated data. c , Quantitative benchmark comparing Lingshu-Cell, scDiffusion and scVI across five metrics (Pearson correlation, Spearman correlation, MMD, 1-WD and iLISI) on the PARSE-PBMC dataset. d , Unconditional generation results across human tissues, including neocortex, heart, lung and colon, with UMAP plots showing real (top) and generated (bottom) cells colored by cell type. e , Unconditional generation results across multiple species, including mouse, rhesus macaque, zebrafish and fly.
Tissue
Pearson ( ↑ )
Spearman ( ↑ )
MMD ( ↓ )
iLISI ( ↑ )
1-WD ( ↓ )
Human tissues (CZ CELLxGENE)
Neocortex
0.9995
0.9991
0.0128
0.9053
0.0105
Heart
0.9992
0.9987
0.0196
0.8972
0.0096
Lung
0.9967
0.9970
0.0314
0.8906
0.0159
Colon
0.9966
0.9960
0.0376
0.8815
0.0152
Non-human species - Tissues
Figure 3: Unconditional generation performance of Lingshu-Cell across human tissues and non-human species.
Figure 4: Accurate prediction of single-cell transcriptomic responses to genetic perturbations in cell lines by Lingshu-Cell. a , Schematic of CRISPR-based genetic perturbation and the resulting transcriptomic changes. b , Conditional generation framework for perturbation prediction. Cell type and perturbation target are provided as conditioning inputs, and a masked diffusion model iteratively predicts gene expression values to generate perturbation-specific expression profiles. c , Three design components of Lingshu-Cell: classifier-free guidance (CFG), sequence compression, and biological prior injection (see Methods). d , Ablation study of CFG guidance weight. Bar plots show prediction performance across eight metrics (DES, PDS, MAE, Spearman #DEG, Spearman LFC, AUPRC, Pearson- Δ , and average score) on the H1 test set ( n=100 perturbation targets). e , Ablation study of sequence compression, comparing uncompressed input with patch sizes of 8 and 32. f , Ablation study of biological prior injection, comparing prediction performance with and without prior injection. In d–f , metrics highlighted in red denote improved performance relative to the corresponding baseline in each ablation setting.
Team
Avg Rank ↓
DES ↑
PDS ↑
MAE ↓
Sp. #DEG ↑
Sp. LFC ↑
AUPRC ↑
Pearson- Δ↑
Lingshu-Cell
8.7
0.216
0.748
0.052
0.394
0.331
0.272
0.306
cleopatra
9.1
0.228
0.747
0.086
0.473
0.396
0.266
0.203
xBio
10.7
0.305
0.811
0.770
0.564
0.087
0.252
0.217
Cellock Holmes
10.9
0.356
0.679
0.239
0
0.238
0.576
0.125
Shippers
11.0
0.354
0.699
0.231
0
0.227
0.576
0.123
Mean Predictors
11.4
0.305
0.741
6.723
0.294
0.213
0.582
0.217
Figure 5: Genetic perturbation prediction on the VCC leaderboard ( Roohani et al., 2025 ) . Teams are ordered by final ranking. Avg Rank: average rank across the top 25 teams. Best per column in bold. See Table B1 for the full top-25 ranking.
Figure 6: Accurate prediction of single-cell transcriptomic responses to cytokine perturbations in PBMCs by Lingshu-Cell. a , Schematic of cytokine-induced transcriptomic perturbation. b , Conditional generation framework for cytokine perturbation prediction. Donor identity and cytokine condition are provided as conditioning inputs, and a masked diffusion process iteratively predicts gene expression values to generate perturbation-specific expression profiles. c , Prediction performance of Lingshu-Cell, PertMean, STATE, scGPT, and scVI on the PARSE 10M PBMC dataset, evaluated across eight metrics (as defined in Fig. 3d). Bar plots show performance on a test set comprising 4 of 12 donors, with 70% of cytokine conditions (63 of 90) held out for each donor. Bars indicate mean performance across donors; error bars, interquartile range (Q1–Q3); points, individual donor values. Metrics for which Lingshu-Cell achieves the best performance among all methods are highlighted in red.
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
DES ↑
PDS ↑
MAE ↓
Sp. #DEG ↑
Sp. LFC ↑
AUPRC ↑
Pear.- Δ↑
Team
Avg Rank ↓
val
rk
val
rk
val
rk
val
rk
val
rk
val
rk
val
rk
Lingshu-Cell
8.7
0.216
25
0.748
9
0.052
1
0.394
12
0.331
3
0.272
10
0.306
1
cleopatra
9.1
0.228
24
0.747
10
0.086
3
0.473
9
0.396
1
0.266
11
0.203
6
xBio
10.7
0.305
13
0.811
6
0.770
13
0.564
4
0.087
20
0.252
16
0.217
3
Cellock Holmes
10.9
0.356
3
0.679
26
0.239
7
0.000
19
0.238
6
0.576
4
0.125
11
Shippers
11.0
0.354
6
0.699
22
0.231
6
0.000
19
0.227
8
0.576
4
0.123
12
Appendix
Table B1 : Full VCC generalist track leaderboard. Bold values indicate the best performance for each metric. Lingshu-Cell was evaluated post-hoc and is separated by a horizontal rule. For each metric, “val” denotes the metric value and “rk” denotes the per-metric rank across all 26 entries. Abbreviated metric names: Sp. #DEG = Spearman #DEG; Sp. LFC = Spearman LFC; Pear.- Δ = Pearson- Δ . Metric definitions are given in Section B . Original leaderboard data are available at https://virtualcellchallenge.org/leaderboard .
Figure C1 : Model architecture of Lingshu-Cell. a , Overall model network architecture. Input tokens are first embedded and then passed through a compression module that reduces the gene token sequence length. The compressed sequence is processed by a stack of 13 Transformer blocks, each consisting of RMSNorm, bidirectional self-attention with rotary position embeddings (RoPE), and a SwiGLU feed-forward network, with residual connections around each sub-layer. A final RMSNorm is applied before the decompression module restores the original gene-level resolution, and a linear head produces per-token logits over the expression vocabulary. b , Compression module. Gene token embeddings are reordered with a fixed random permutation π , partitioned into consecutive groups, and linearly projected via Wdown to produce a shorter sequence of compressed tokens. Condition prefix and tail tokens bypass this module. c , Decompression module. The inverse of compression: compressed tokens are linearly projected via Wup , ungrouped, and reordered with the inverse permutation π−1 to recover gene-level representations.
Hyperparameter
Value
Embedding dimension ( D )
640
Transformer blocks ( L )
13
Attention heads ( nh )
10
Head dimension ( dh )
64
Feed-forward dimension ( dff )
2,560
Position encoding
RoPE
Appendix
Table C2 : Architectural hyperparameters of Lingshu-Cell.
Unconditional
Conditional
Setting
PARSE
Per-tissue
Genetic perturbation
Cytokine perturbation
Dataset
PARSE PBS
CZ CELLxGENE scBaseCount
VCC H1 + auxiliary
PARSE 10M
GPUs
8
2 – 4
16
8
Peak LR
1×10−4
2×10−4
2×10−4
2×10−4
Warmup steps
100
100
1,000
1,000
Appendix
Table C3 : Training configurations for different experimental settings.
Task
Dataset
#Perturbed cells
#Original control cells
#Sub-sampled control cells
Genetic perturbation
H1
221,273
38,176
20,344
External
323,913
583,377
35,990
Cytokine perturbation
PARSE 10M
6,499,077
629,701
629,701
Appendix
Table C4 : Training data composition for conditional generation tasks.
Figure C2 : Scaling unconditional generation to larger cell populations. a , UMAP visualization of real and generated cells from the PARSE 10M PBMC dataset, colored by cell type annotation (left) and normalized expression (log1p) of canonical marker genes. b , Comparison of major cell type proportions between real and generated data. c , UMAP visualization at higher resolution, showing that generated cells accurately recapitulate cell subtype structure of real data. d , Comparison of cell subtype proportions between real and generated data, demonstrating robust consistency at higher resolution.
Figure C3 : Quality control of human unconditional generation datasets. a , Summary of datasets used for unconditional generation, including data source, species, tissue type and cell number. b , c , Distribution of the number of detected genes ( b ) and total counts ( c ) across tissues. d , Density scatter plots of mitochondrial gene percentage versus total counts for each tissue. e , Density scatter plots of the number of detected genes versus total counts for each tissue.
Figure C4 : Unconditional generation results for additional human tissues. UMAP visualization of real (top) and generated (bottom) cells for thymus, kidney, breast and liver, colored by cell type annotation.
Figure C5 : Quality control of non-human species unconditional generation datasets. a , Summary of datasets used for unconditional generation across non-human species, including data source, species, tissue type and cell number. b , c , Distribution of the number of detected genes ( b ) and total counts ( c ) across species. d , Density scatter plots of mitochondrial gene percentage versus total counts for each species. The rhesus macaque lung dataset is not shown owing to the absence of annotated mitochondrial genes. e , Density scatter plots of the number of detected genes versus total counts for each species.
Predicting cellular responses to genetic perturbations is a central capability for virtual cells and a key step toward computational modeling of biological interventions. Most existing models directly map an unperturbed molecular profile and perturba- tion to the resulting observation without explicitly representing the latent cellular transition induced by the intervention. We introduce a mechanism-centric virtual cell world model that represents cellular state as a set of Latent Mechanism Units (LMUs) and treats genetic perturbations as actions on these latent states. Each LMU combines a reusable identity grounded in multimodal biological evidence with an observation-specific state, allowing a perturbation to induce mechanism- specific stochastic transitions before decoding the resulting transcriptional response. We train VCLMU through two-stage pretraining, first on around 200K pseudo-bulk perturbation profiles and then on gene-aligned single-cell perturbation data. Across six perturbation-disjoint benchmarks, VCLMU consistently improves perturbation- specific response recovery over strong baselines while maintaining competitive global response accuracy. We further analyze learned LMUs through enrichment between perturbation responses and LMU gene sets and show that they capture structured biological response programs. These results support mechanism-level latent state transition as a useful formulation for virtual cell models that aim to predict and interpret cellular responses to biological interventions.
Yuwei Miao, Azim Dehghani Amirabad, Scott Oloff +3
Johnson & Johnson Innovative Medicine · University of Texas at Arlington
Building Virtual Cells that can accurately simulate cellular responses to perturbations is a long-standing goal in systems biology. A fundamental challenge is that high-throughput single-cell sequencing is destructive: the same cell cannot be observed both before and after a perturbation. Thus, perturbation prediction requires mapping unpaired control and perturbed populations. Existing models address this by learning maps between distributions, but typically assume a single fixed response distribution when conditioned on observed cellular context (e.g., cell type) and the perturbation type. In reality, responses vary systematically due to unobservable latent factors such as microenvironmental fluctuations and complex batch effects, forming a manifold of possible distributions for the same observed conditions. To account for this variability, we introduce PerturbDiff, which shifts modeling from individual cells to entire distributions. By embedding distributions as points in a Hilbert space, we define a diffusion-based generative process operating directly over probability distributions. This allows PerturbDiff to capture population-level response shifts across hidden factors. Benchmarks on established datasets show that PerturbDiff achieves state-of-the-art performance in single-cell response prediction and generalizes substantially better to unseen perturbations. See our project page (https://katarinayuan.github.io/PerturbDiff-ProjectPage/), where code and data will be made publicly available (https://github.com/DeepGraphLearning/PerturbDiff).
Xinyu Yuan, Xixian Liu, Ya Shi Zhang +3
Mila - Québec AI Institute · University of Montréal · University of Ottawa +3
Predicting cellular transcriptional responses to genetic perturbations is a central problem in single-cell biology, especially in the zero-shot setting where the perturbed gene or gene combination is unseen during training. A major difficulty is that perturbation effects are not determined by expression state alone: they depend on how the perturbed gene product influences other genes and proteins, how those downstream factors act on cis-regulatory elements, and which regulatory programs are active in the current cell state. To better capture this biological complexity, we propose CisTransCell, a cell-conditioned multi-modal framework for single-cell perturbation prediction that augments each gene with two complementary priors: a regulatory-sequence prior that captures how the gene is controlled, and a coding-sequence prior that captures what the gene product does. By integrating these priors with cellular expression state, CisTransCell models perturbation response as a cascade from gene function to regulatory control to downstream transcriptional change. Experiments on benchmark single-cell perturbation datasets show that CisTransCell achieves strong performance in zero-shot perturbation prediction.
Wei Zhang, Xun Jiang, Yuesi Xi +1
1L3S Research Center, Leibniz Universität Hannover, Germany · School of Clinical Medicine & Laboratory Medicine, Jiangsu University, China · Institute for Information Processing (tnt), Leibniz Universität Hannover, Germany.