Lingshu-Cell: Masked Diffusion for scRNA-seq

• Lingshu-Cell is a generative cellular world model that uses a masked discrete diffusion process to simulate full transcriptomic state distributions from scRNA-seq data.
• It discretizes sparse, count-valued gene expression data over an 18,080-gene reference without pre-filtering for highly variable genes, ensuring high-fidelity generation.
• The model supports conditional simulation to predict cellular responses under specific donor, cell identity, or perturbation contexts, enabling virtual cell experimentation.

Lingshu-Cell is a generative cellular world model for single-cell RNA sequencing introduced as a masked discrete diffusion model that learns transcriptomic state distributions and supports conditional simulation under perturbation. It is designed for sparse, discrete, orderless scRNA-seq data, operates over a fixed 18,080-gene reference without prior highly variable gene filtering or expression-rank ordering, and is presented as a route toward “virtual cells” that can be sampled either unconditionally or under specified donor, cell identity, or perturbation context (Zhang et al., 26 Mar 2026).

1. Conceptual scope

Lingshu-Cell is framed around a shift from static representation learning to explicit generative modeling of cellular state distributions. The underlying problem is not merely to encode cells into useful embeddings, but to model the full distribution of transcriptomic states well enough to generate realistic single-cell profiles, preserve heterogeneous subpopulations and subtype proportions, and predict whole-transcriptome responses under interventions such as CRISPR perturbations or cytokine stimulation. In this formulation, a “virtual cell” is not a mechanistic simulator of intracellular biochemistry; it is a probabilistic generative model of transcriptomic states and their condition-dependent variation (Zhang et al., 26 Mar 2026).

The model is motivated by four structural properties of scRNA-seq data emphasized by its authors: high dimensionality, with about 18,000 genes per cell; sparsity or zero inflation; discrete count-valued observations; and the non-sequential, orderless character of gene-expression profiles. This positioning is used to argue against direct transplantation of common generative paradigms. Autoregressive models impose an arbitrary left-to-right gene order, continuous diffusion perturbs counts with Gaussian-like noise poorly aligned to sparse integer data, and variational autoencoders are described as often prioritizing reconstruction and latent regularization rather than high-fidelity transcriptome simulation. Lingshu-Cell is therefore proposed as a masked discrete diffusion model whose corruption and denoising processes are better matched to the empirical structure of scRNA-seq (Zhang et al., 26 Mar 2026).

The paper also situates Lingshu-Cell relative to existing single-cell foundation models such as scGPT, Geneformer, scFoundation, and CellFM. Those models are treated as useful pretrained encoders, but not as explicit simulators of $p(x)$ or $p(x\mid c)$. The central claim is that a virtual-cell use case requires population-level generative fidelity rather than only downstream-transfer utility (Zhang et al., 26 Mar 2026).

2. Discrete transcriptome representation and architecture

Lingshu-Cell represents each cell over a fixed gene list of length

$G = 18{,}080.$

There is one token per gene position. Rather than using exact UMI counts as tokens, the model quantizes counts into bins through a gene-wise discretization map

$$q:\mathbb{Z}_{\ge 0}\to \{0,1,\dots,B-1\}\cup\{\mathrm{OVF}\}.$$

Counts $0$–$99$ receive exact bins, larger counts are binned with decade-dependent resolution that preserves roughly the first two significant digits, and counts above a cap $C$ map to an overflow token. The default is

$C=9999.$

This yields 100 bins for $[0,99]$, 90 bins for $[100,999]$, and 90 bins for $p(x\mid c)$0, so the non-overflow vocabulary has

$p(x\mid c)$1

and including overflow

$p(x\mid c)$2

before special tokens such as $p(x\mid c)$3. A cell is therefore encoded as

$p(x\mid c)$4

The denoiser is a bidirectional Transformer with LLaMA-style blocks. Reported backbone hyperparameters are embedding dimension $p(x\mid c)$5, 13 Transformer blocks, 10 attention heads, head dimension $p(x\mid c)$6, feed-forward dimension $p(x\mid c)$7, SwiGLU activation, RMSNorm, RoPE positional encoding, bidirectional attention, and bias-free linear layers. Because $p(x\mid c)$8 is long for direct Transformer processing, the model compresses the internal sequence representation rather than filtering genes. Let the embedded gene sequence be

$p(x\mid c)$9

A fixed random permutation $G = 18{,}080.$0 is sampled at initialization, the permuted sequence is partitioned into groups of size $G = 18{,}080.$1, flattened blockwise, and linearly down-projected to a compressed sequence of length

$G = 18{,}080.$2

The default is $G = 18{,}080.$3, giving

$G = 18{,}080.$4

whereas genetic perturbation prediction uses $G = 18{,}080.$5, giving

$G = 18{,}080.$6

After Transformer processing, the model up-projects and restores the original gene order. This design preserves gene-level input and output while reducing the sequence length seen by the backbone. Training uses AdamW with $G = 18{,}080.$7, $G = 18{,}080.$8, weight decay $G = 18{,}080.$9, cosine annealing with linear warmup, global batch size $$q:\mathbb{Z}_{\ge 0}\to \{0,1,\dots,B-1\}\cup\{\mathrm{OVF}\}.$$0, bfloat16 mixed precision, DDP, and NVIDIA A800 GPUs (Zhang et al., 26 Mar 2026).

3. Diffusion formulation and sampling

The generative process is a masked discrete diffusion over transcriptome token arrays. For a discrete sequence

$$q:\mathbb{Z}_{\ge 0}\to \{0,1,\dots,B-1\}\cup\{\mathrm{OVF}\}.$$1

the forward process independently replaces each token by a mask token $$q:\mathbb{Z}_{\ge 0}\to \{0,1,\dots,B-1\}\cup\{\mathrm{OVF}\}.$$2 with probability $$q:\mathbb{Z}_{\ge 0}\to \{0,1,\dots,B-1\}\cup\{\mathrm{OVF}\}.$$3. For position $$q:\mathbb{Z}_{\ge 0}\to \{0,1,\dots,B-1\}\cup\{\mathrm{OVF}\}.$$4,

$$q:\mathbb{Z}_{\ge 0}\to \{0,1,\dots,B-1\}\cup\{\mathrm{OVF}\}.$$5

This is a masking diffusion rather than an additive-noise diffusion.

The reverse model predicts original token identities at masked positions through

$$q:\mathbb{Z}_{\ge 0}\to \{0,1,\dots,B-1\}\cup\{\mathrm{OVF}\}.$$6

Training uses masked-token cross-entropy weighted by $$q:\mathbb{Z}_{\ge 0}\to \{0,1,\dots,B-1\}\cup\{\mathrm{OVF}\}.$$7: $$q:\mathbb{Z}_{\ge 0}\to \{0,1,\dots,B-1\}\cup\{\mathrm{OVF}\}.$$8 The paper states that this objective provides a variational upper bound on negative log-likelihood.

Sampling begins from a fully masked sequence and iteratively denoises it. The reverse trajectory is discretized into $$q:\mathbb{Z}_{\ge 0}\to \{0,1,\dots,B-1\}\cup\{\mathrm{OVF}\}.$$9 steps with a cosine timestep schedule. At each step the model predicts categorical distributions over masked positions, samples token values, and then remasks a fraction so that the next state matches the intended marginal corruption level. For unconditional generation the reported setting is

$0$0

with direct categorical sampling and no temperature scaling or top-$0$1 truncation. This makes generation non-autoregressive and bidirectional at every denoising step, which is the core architectural answer to the orderless nature of transcriptomic data (Zhang et al., 26 Mar 2026).

4. Conditional simulation and perturbation modeling

For perturbation response prediction, Lingshu-Cell augments the transcriptome sequence with condition tokens $0$2 that are prepended and never masked. The conditional model is

$0$3

The condition consists of a source or background identity together with a perturbation identity. Examples given in the paper include cell line plus target gene for genetic perturbation and donor plus cytokine for PBMC stimulation. Control cells are assigned a specific neutral control condition $0$4, so perturbed and control generation are learned within one conditional model.

Inference uses classifier-free guidance. If

$0$5

is the logit for token $0$6 under condition $0$7, the guided logits are

$0$8

equivalently

$0$9

The reported settings are $99$0 for genetic perturbation prediction and $99$1 for cytokine perturbation prediction. Conditional generation uses only

$99$2

reverse steps.

For genetic perturbation prediction, the model also incorporates an inference-time biological prior. A downregulated prior set $99$3 is constructed from external CRISPR datasets, and the initial state for reverse diffusion is set by

$99$4

These prior-fixed positions remain fixed during sampling. The paper characterizes this not as a learned regularizer but as an initialization and clamping strategy.

The principal conditional benchmarks are the Virtual Cell Challenge H1 genetic perturbation task and cytokine stimulation in PARSE 10M PBMCs. For H1, the target partition contains 150 training targets, 50 validation targets, and 100 test targets; the model is trained on 183,097 H1 cells from training targets plus 323,913 external perturbed cell-line cells overlapping the 300 H1 targets, and evaluated on 60,751 validation cells and 132,670 test cells. For the cytokine setting, the data comprise 12 donors under 90 cytokine conditions plus PBS control, with 6,499,077 perturbed cells and 629,701 control cells. The paper explicitly claims prediction for novel combinations of identity and perturbation; a plausible implication is that the model is intended for compositional generalization across background identity and intervention, although the paper does not formalize this as a strict zero-shot guarantee (Zhang et al., 26 Mar 2026).

5. Evaluation, reported performance, and stated limitations

The evaluation covers both unconditional generation and perturbation-conditioned prediction. Unconditional data include 629,701 PBS-control PBMCs from PARSE 10M, 2,602,318 human cells across eight CELLxGENE tissues, and 247,899 non-human cells spanning mouse ovary, rhesus macaque lung, zebrafish embryo, and fly brain. Human data are aligned to the 18,080-gene reference used in VCC H1. For CELLxGENE human tissues, retained cells satisfy detected genes $99$5, counts $99$6, mitochondrial fraction $99$7, genes detected in at least 3 cells, and Scrublet-based doublet removal; the model explicitly does not rely on prior HVG selection (Zhang et al., 26 Mar 2026).

For unconditional generation, the paper compares against scDiffusion and scVI using Pearson correlation of mean expression, Spearman correlation of mean expression, MMD in joint PCA space, gene-averaged 1-Wasserstein distance, and iLISI for real/generated mixing. For perturbation prediction on VCC H1, the reported metrics are DES, PDS, MAE, Spearman #DEG, Spearman LFC, AUPRC, and Pearson-$99$8, with average ranking across these seven metrics used for leaderboard comparison. Cytokine perturbation is compared against PertMean, STATE, scGPT, and scVI.

Setting	Data	Reported outcome
Unconditional PBMC generation	PARSE PBS control, 629,701 cells	Lingshu-Cell MMD $99$9; scDiffusion $C$0; scVI $C$1
Cross-tissue and cross-species generation	2,602,318 human cells; 247,899 non-human cells	Human neocortex: Pearson $C$2, Spearman $C$3, MMD $C$4, iLISI $C$5, 1-WD $C$6
Genetic perturbation prediction	VCC H1	Avg Rank $C$7; MAE $C$8 and Pearson-$C$9 $C=9999.$0 are best among compared teams
Cytokine perturbation prediction	PARSE PBMC, 12 donors	Highest average score; ranked first in PDS, Pearson-$C=9999.$1, and Spearman #DEG

Qualitatively, the paper states that generated cells reproduce major PBMC lineages, marker-gene expression patterns, major cell-type proportions, and finer subtype composition across 17 PBMC subtypes at larger generation scale. Across tissues and species, the reported takeaway is robust transcriptomic distribution fidelity rather than a claim of mechanistic biological simulation.

The paper’s limitations are explicit. Evaluation is described as largely population-level, so metrics such as MMD, iLISI, and pseudobulk correlations may not fully assess single-cell realism or very rare states. High-fidelity generation is not taken to imply causal or mechanistic understanding of regulatory biology. Predictions are positioned as probabilistic hypotheses that still require wet-lab validation. The current model is transcriptomics-only, not multimodal. Generalization is said to remain bounded by available training data, particularly for perturbation tasks. Interpretability analyses are limited to marker-level and perturbation-metric evidence rather than deep mechanistic dissection (Zhang et al., 26 Mar 2026).

6. Nomenclature, disambiguation, and relation to adjacent research

Despite the name, Lingshu-Cell is not the same object as Gilbert Ling’s “Ling’s cell” in association-induction theory. In that earlier literature, the resting cell is modeled as a low-entropy state of unfolded proteins, bound multilayer water, and selective $C=9999.$2 adsorption under ATP control, with excitation and injury interpreted as ATP depletion, water and $C=9999.$3 desorption, protein folding, and aggregation (Matveev et al., 2011). Related work formalizes “Ling’s cell” as a stationary non-equilibrium state enabled by non-ergodicity and additional first integrals (Prokhorenko et al., 2010), and proposes van der Waals and protein-conformation microstructure models in which resting protoplasm is unfolded and non-ergodically constrained whereas dead protoplasm is folded and aggregated (Prokhorenko et al., 2011). These papers belong to a thermodynamic and statistical-mechanical tradition rather than a generative modeling program.

Lingshu-Cell is also distinct from LangCell. LangCell is introduced as “the first Language-Cell pre-training framework” for aligning scRNA-seq with biomedical text in order to perform cell identity understanding, including zero-shot, few-shot, and fine-tuned annotation and retrieval. The available paper explicitly states that there is no direct mention of the name “Lingshu-Cell,” no alias, and no explicit statement that LangCell and Lingshu-Cell are the same project (Zhao et al., 2024). The two models therefore occupy different methodological niches: LangCell is a language-cell semantic alignment model, whereas Lingshu-Cell is a masked discrete diffusion world model for transcriptome generation and perturbation-conditioned simulation.

A further adjacent method is SCCAF, which addresses automated discovery and validation of putative cell types or states from scRNA-seq by self-projection, confusion-matrix-based merging, and weighted marker-gene extraction. SCCAF is not a simulator, but it provides a criterion for whether a cluster is transcriptomically self-consistent and separable enough to be treated as a distinct cell type or state (Miao et al., 2020). This suggests a complementary relationship rather than competition: SCCAF can validate or refine transcriptomic state definitions, while Lingshu-Cell models their distributions and perturbational transformations.

Within this broader landscape, Lingshu-Cell is most precisely classified as a whole-transcriptome masked discrete diffusion model for scRNA-seq, designed to move single-cell foundation modeling from static encoding toward conditional generative simulation of cellular states (Zhang et al., 26 Mar 2026).