# Lingshu-Cell: A Generative Cellular World Model for Transcriptome Modeling Toward Virtual Cells

**Authors:** Han Zhang, Guo-Hua Yuan, Chaohao Yuan, Tingyang Xu, Tian Bian
**Year:** 2026
**Venue:** arXiv preprint
**arXiv:** 2603.25240

## One-sentence summary

Lingshu-Cell is a masked discrete diffusion model that learns the full distribution of transcriptomic states (not just embeddings) and achieves SOTA on the Virtual Cell Challenge H1 genetic perturbation benchmark by treating the cell as a generative world model.

## Key contribution

Unlike prior foundation models that learn point embeddings, Lingshu-Cell models the *distribution* of cellular states — enabling genuine sampling, simulation, and conditional generation. This is the key shift from "representation model" to "world model."

## Methods

- **Architecture:** Masked discrete diffusion transformer
- **Tokenization:** Operates in discrete token space — compatible with sparse, non-sequential scRNA-seq data
- **Gene coverage:** ~18,000 genes (no HVG filtering — processes the full transcriptome)
- **Conditioning:** Jointly embeds cell type / donor identity + perturbation to predict response
- **Key advantage over masked reconstruction models:** Models the full generative distribution, not a denoised point estimate

## Key findings

- Accurately reproduces transcriptomic distributions, marker-gene patterns, and cell subtype proportions across diverse tissues and species
- Achieves **SOTA on Virtual Cell Challenge H1** genetic perturbation benchmark
- Strong performance predicting cytokine-induced responses in human PBMCs
- No gene pre-selection needed — operates on all ~18k genes

## Why this matters

This is arguably the closest thing to a true Virtual Cell published as of April 2026. The shift to generative modeling of state distributions (rather than regression to mean expression) addresses a fundamental limitation of prior work. You can now *sample* from the model, not just query it.

## Limitations

- arXiv preprint — not yet peer-reviewed
- Training data details not fully specified
- Comparison baseline set not exhaustive (doesn't benchmark against all GEARS variants)
- Scalability to combinatorial perturbations (3+ gene interactions) not demonstrated

## Connections

- [../concepts/virtual-cell.md](../concepts/virtual-cell.md) — this paper advances the Virtual Cell agenda directly
- [../concepts/single-cell-foundation-models.md](../concepts/single-cell-foundation-models.md) — generative evolution of the field
- [../concepts/perturbation-biology.md](../concepts/perturbation-biology.md) — tested on perturbation benchmarks
- [../papers/scgpt.md](scgpt.md) — prior SOTA foundation model it builds beyond

## Bo's notes

This paper is directly relevant to Xaira's mission. The "cellular world model" framing aligns with Bo's view that Virtual Cell ≠ perturbation predictor — a world model can generalize causally, not just interpolate. Worth tracking the authors (Tian Bian at Tencent AI Lab) and watching for follow-up work.