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

Han Zhang; Guo-Hua Yuan; Chaohao Yuan; Tingyang Xu; Tian Bian; Hong Cheng; Wenbing Huang; Deli Zhao; Yu Rong

Source-linked AI summary

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

Han Zhang, Guo-Hua Yuan, Chaohao Yuan, Tingyang Xu, Tian Bian, Hong Cheng, Wenbing Huang, Deli Zhao, Yu Rong

arXiv:2603.25240v1Mar 26, 2026q-bio.QMcs.AIq-bio.GN

TL;DR

Single-cell transcriptomics needs models that move beyond static representations to simulate cellular-state distributions and perturbation responses. Lingshu-Cell addresses this need with masked discrete diffusion over transcriptome-wide discrete gene-expression tokens. It reproduces cellular heterogeneity across diverse datasets and reports leading performance for genetic and cytokine perturbation prediction.

Problem
Existing single-cell transcriptomic foundation models primarily provide static representations rather than explicitly modeling cellular-state distributions for generative simulation.
Method
Lingshu-Cell uses masked discrete diffusion over discrete gene-expression tokens to model approximately 18,000 genes without prior gene selection and condition simulations on identity and perturbation.
Results
Lingshu-Cell reproduces transcriptomic distributions, marker-gene patterns and cell-subtype proportions across diverse tissues and species, while achieving leading performance on genetic and cytokine perturbation tasks.
Takeaways & Limitations
Lingshu-Cell provides a unified generative framework for simulating heterogeneous cell states and predicting responses to genetic and cytokine perturbations.
Takeaways & Limitations
Population-level distributional metrics and pseudobulk correlations cannot fully assess single-cell biological plausibility or preservation of extremely rare states, and predictions require wet-lab validation.

Abstract

from arXiv · show

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.

1 Introduction

Lingshu-Cell is introduced to address the gap between descriptive single-cell transcriptomics and generative modeling of cellular states and perturbation responses. It uses masked discrete diffusion to model transcriptome-wide expression while matching the sparse, discrete, non-sequential structure of scRNA-seq data.

Large-scale scRNA-seq atlases enable broad characterization of cell states, but most analyses remain descriptive rather than predictive.
A cellular world model should represent transcriptomic state distributions and conditional dynamics to support simulation of cellular states and intervention responses.
Existing transcriptomic foundation models mainly learn static representations, while generative approaches face assumptions mismatched to sparse, discrete, non-sequential data.
Lingshu-Cell is a masked discrete diffusion model trained with masking and prediction over discrete gene-expression tokens for non-autoregressive, bidirectional transcriptome refinement.
Lingshu-Cell models approximately 18,000 genes without prior selection and is evaluated across datasets spanning nine tissues and five species.

2 Results

Lingshu-Cell models and generates transcriptomic states with masked discrete diffusion, reproducing cellular heterogeneity across tissues and species. Its conditional framework predicts transcriptomic responses to genetic and cytokine perturbations, achieving strong performance across both settings.

Framework: Lingshu-Cell uses forward masking and reverse iterative prediction to generate single-cell transcriptomic profiles in an order-independent discrete space.The framework supports unconditional generation and conditional perturbation-response prediction.
Unconditional generation: Generated PBMC data reproduced marker-gene patterns and cell-type proportions at both 10,000-cell and 200,000-cell scales.The agreement persisted when PBMCs were subdivided into 17 subtypes.
Unconditional generation: Lingshu-Cell generalized unconditional generation across eight human tissues and four additional species despite substantial dataset heterogeneity.The evaluated species included mouse, rhesus macaque, zebrafish and fly.
Genetic perturbations: Conditional generation incorporated cell type and perturbation-target tokens for genetic perturbation response prediction.Classifier-free guidance, sequence compression and biological prior injection improved performance in ablation studies.
Genetic perturbations: Lingshu-Cell achieved the best average rank across seven VCC H1 evaluation metrics, including MAE 0.052 and Pearson-∆correlation 0.306.It provided the best overall balance across the seven criteria, although other methods led individual metrics.
Cytokine perturbations: For cytokine perturbations in donor-derived PBMCs, Lingshu-Cell achieved the highest average score and led on PDS, Pearson-∆correlation and Spearman #DEG correlation.The model was evaluated with donor identity and cytokine condition jointly held as conditioning information.

3 Discussion

Lingshu-Cell establishes masked discrete diffusion as a unified generative framework for transcriptomic simulation and perturbation-response prediction. Its results support a cellular world model, while current evaluations and modality scope limit biological interpretation.

Lingshu-Cell models approximately 18,000 genes without prior gene selection, shifting single-cell foundation models from static representations toward generative simulation.
The framework generates heterogeneous cellular states across tissues and species and predicts responses to genetic and cytokine perturbations within one architecture.
Discrete expression modeling aligns the architecture with permutation-invariant, zero-inflated transcriptomic data while avoiding artificial gene-ordering bias and continuous-noise mismatch.
Current evaluations use population-level distributional metrics and pseudobulk correlations, which cannot fully assess single-cell plausibility or preservation of extremely rare states.
High-fidelity expression generation does not establish biological causality, so predictions require wet-lab validation; broader virtual cells would also integrate additional molecular modalities.
Future extensions include drug-induced, multi-target, and combinatorial interventions with dose dependence and temporal dynamics, alongside joint multimodal modeling.

4 Methods

Lingshu-Cell uses masked discrete diffusion to represent single-cell expression as discrete sequences and generate profiles through iterative masking and prediction. Compression reduces internal computation, while condition tokens support perturbation- and control-state generation.

4.1 Preliminaries: Masked discrete diffusion learns a joint distribution by progressively masking discrete tokens and training a reverse model to predict masked positions simultaneously.
4.1 Preliminaries: During inference, generation starts from a fully masked sequence and iteratively predicts tokens while remasking a proportion to follow the forward-process dynamics.
4.2 Representing Single-Cell Data as Discrete Sequences: The model represents each cell as a discrete gene sequence over a fixed gene list, with the sequence serving as the MDDM input and output space.
4.2 Representing Single-Cell Data as Discrete Sequences: UMI count profiles are quantized into finite expression bins plus an overflow token, preserving approximately the first two significant digits while reducing the discrete state space.
4.3 Embedding-Space Sequence Compression: Embedding-space compression randomly groups gene embeddings, projects each group to one vector, and later expands and reorders them to preserve the original gene-level interface.
4.4 Conditional Generation: Source-context and perturbation tokens are prepended and exempt from masking, enabling conditional generation of perturbation-specific and control states under one parameterization.

A.1.1 Data processing for unconditional generation

Unconditional-generation data combine human, mouse, rhesus macaque, zebrafish, and fly single-cell datasets selected for normal, single-cell profiles and standardized quality criteria. Gene identifiers and features are harmonized across datasets before analysis.

Human unconditional data come from PARSE 10M PBMC and CZ CELLxGENE, covering eight normal human tissues profiled with 10x Genomics 3′ v3.
Cells generally required at least 200 detected genes, 500 total counts, and mitochondrial transcript fraction no greater than 20%; genes detected in fewer than three cells were removed.
Human matrices were aligned to a reference set of 18,080 genes, and non-human datasets were standardized using Ensembl annotations restricted to protein-coding genes with valid names.
Mouse and rhesus macaque data come from CZ CELLxGENE, while zebrafish and fly data come from scBaseCount.
The downstream non-human experiments use mouse ovary, rhesus macaque lung, zebrafish embryo, and fly brain data selected without explicit disease or perturbation evidence.

A.1.2 Data processing for conditional generation

Conditional-generation experiments cover genetic perturbations in H1 cells and cytokine perturbations in human PBMCs. Data were harmonized and partitioned to test generalization across unseen perturbation targets and donors.

The conditional experiments evaluate genetic perturbation generation and cytokine perturbation generation as two distinct settings.
The VCC H1 benchmark partitions perturbation targets into 150 training, 50 validation, and 100 test targets to assess generalization to unseen perturbations.
Additional CRISPR and Perturb-seq datasets were incorporated to increase perturbation diversity during genetic-perturbation training.
Genetic perturbation matrices were harmonized to a shared feature space, perturbation labels were standardized by target gene identity, and ambiguous annotations were excluded.
The PARSE PBMC cytokine dataset defines cytokine labels as perturbations and PBS as the control, with features aligned to the 18,080-gene reference set.
Cytokine data were split at the donor level, using six donors for training and two randomly selected donors for validation with held-out cytokine conditions.

B Evaluation metrics

The evaluation combines gene-level fidelity metrics with distributional and integration measures to assess whether generated cells recapitulate real transcriptomic populations.

Gene-level fidelity: Five complementary metrics assess gene-level expression fidelity and distributional similarity between real and generated cells.All metrics use log1p-normalized expression values unless otherwise noted.
Gene-level fidelity: Pearson and Spearman correlations measure concordance and monotonic agreement between real and generated mean gene-expression vectors.Pearson evaluates expression concordance, whereas Spearman captures agreement in gene-expression ranks.
Distributional similarity: MMD measures distance between real and generated distributions in a joint PCA space using a five-scale Gaussian kernel, with lower values indicating better agreement.The kernel uses bandwidth multiplier α = 2 across five scales.
Distributional similarity: Gene-averaged 1-WD computes per-gene Wasserstein distances and averages them, rather than measuring multivariate distance across the full expression space.Lower values indicate better agreement.
Integration quality: iLISI evaluates mixing of real and generated cells in a shared embedding, where 2 indicates perfect mixing and 1 indicates complete separation.The index is calculated from the effective number of real-versus-generated groups in local neighborhoods.

B.2 Metrics for perturbation prediction

Perturbation prediction is evaluated with seven complementary metrics spanning transcriptome-wide accuracy and recovery of differentially expressed genes, summarized against a cell-mean baseline.

Metric framework: Seven metrics evaluate perturbation prediction using pseudobulk expression profiles averaged over cells sharing each perturbation condition.Expression deltas compare perturbed profiles with the unperturbed control pseudobulk.
Metric framework: Transcriptome-wide accuracy measures overall expression fidelity, while differential-expression recovery assesses identification and ranking of perturbation-responsive genes.The two categories evaluate complementary aspects of predicted perturbation responses.
Transcriptome-wide accuracy: MAE measures average absolute error, Pearson-∆ measures agreement in expression deltas, and PDS measures whether distinct perturbation effects are discriminated.PDS equals 1 for perfect discrimination and 0 for random performance.
Differential expression recovery: DES, Spearman #DEG, Spearman LFC, and AUPRC evaluate overlap, effect-size relationships, log-fold-change ranking, and precision-recall recovery of significant DE genes.AUPRC uses predicted −log10(padj) values as confidence scores for ground-truth DE labels.
Aggregate evaluation: The average score summarizes relative improvement over the cell-mean baseline across all seven metrics, with zero indicating no improvement and one indicating perfect performance.MAE is smaller-is-better; the other six metrics are larger-is-better and upper-bounded by 1.
Aggregate evaluation: Lingshu-Cell achieved the best average rank, 8.7, across 26 VCC generalist leaderboard entries and ranked first on MAE and Pearson-∆.The model was evaluated post-hoc and ranked first on two of seven metrics.

C.1 Model architecture

Lingshu-Cell uses a bidirectional Transformer over compressed gene-token sequences, then decompresses representations to predict expression-vocabulary logits.

Input compression: The model embeds 18,080 gene tokens, compresses their sequence, and concatenates condition-prefix tokens before Transformer processing.The default group size S = 8 yields Gc = 2,260 compressed tokens; genetic perturbation prediction uses S = 32 and Gc = 565.
Transformer backbone: The Transformer backbone contains 13 identical pre-normalized blocks with residual connections.Each block contains self-attention and a feed-forward sub-layer.
Transformer backbone: Bidirectional multi-head attention uses 10 heads, 64-dimensional heads, and rotary position embeddings without a causal mask.The non-causal design matches the order-independent nature of masked diffusion models.
Transformer backbone: The feed-forward sub-layer uses SwiGLU with intermediate dimension dff = 2,560, and all linear layers are bias-free.These architectural choices are part of the LLaMA-style backbone.
Output reconstruction: After final RMSNorm, decompression restores gene-level sequence length and a linear head maps each position to expression-vocabulary logits.This completes the pathway from compressed representations to gene-level predictions.

C.2 Training and inference details

Training and inference combine distributed optimization, task-specific data composition, scheduled masked-token sampling, and perturbation-informed initialization for conditional generation.

Training: Training uses AdamW with gradient clipping, cosine annealing with warmup, bfloat16 mixed precision, distributed data parallelism, and NVIDIA A800 GPUs.The global batch size is 256, and the learning rate decays to 2 × 10^-5.
Training data: Unconditional models are trained independently for each tissue or species dataset, while conditional datasets cap control cells at 10%.The same control ratio is used for genetic and cytokine perturbation prediction.
Inference: Inference precomputes tokens to unmask using a cosine timestep schedule and samples token values from predicted categorical distributions.Positions are selected uniformly from the remaining masked tokens.
Inference: Unconditional generation uses 256 diffusion steps without CFG, whereas conditional generation uses 3 steps with task-specific CFG weights.The guidance weight is w = 2 for genetic perturbation prediction and w = 3 for cytokine prediction.
Perturbation priors: Perturbation-specific prior genes are selected from external CRISPR datasets using differential-expression evidence and preferentially initialize masked positions during conditional generation.The selection threshold includes |log2(fold_change)| > 1, and priors are merged across cell lines by union.
Supplementary evaluation: The supplementary figures document the architecture, scaling of unconditional generation, dataset quality control, and results across additional human tissues and non-human species.These figures cover both real-versus-generated population comparisons and dataset characteristics.