ArXiv: 2603.25240

🎯 Pitch

A single generative model operating directly on discrete gene tokens—without pre-filtering genes—can both synthesize realistic single-cell transcriptomes and predict whole-transcriptome perturbation responses, achieving the best results on the Virtual Cell Challenge benchmark. This demonstrates that learning the full distribution of cellular states from massive atlases enables a unified 'cellular world model' for in silico perturbation screening across tissues and species.

1. Executive Summary

This paper introduces Lingshu-Cell, a masked discrete diffusion model (MDDM) that learns the transcriptomic state distribution of single cells and supports conditional simulation under perturbation, operating directly in a discrete token space over approximately 18,000 genes without prior gene selection. Evaluated on the PARSE 10M PBMC dataset, a multi-tissue CZ CELLxGENE collection, the Virtual Cell Challenge H1 genetic perturbation benchmark, and cytokine perturbation data, Lingshu-Cell achieves the lowest MAE (0.052) and highest Pearson-Δ (0.306) among all methods on the VCC leaderboard, while reproducing cell-type proportions and marker-gene patterns in unconditional generation. The work establishes that a unified discrete diffusion framework can serve as a cellular world model for both faithful transcriptomic simulation and perturbation response prediction, though its effectiveness is demonstrated exclusively on tasks where perturbation responses can be learned from transcriptomic data alone—genetic and cytokine interventions—without claims to modeling causal regulatory mechanisms or multi-omic cellular states.

2. Context and Motivation

The Core Problem: Moving from Descriptive Atlases to Generative Simulation

The fundamental gap this paper addresses is the disconnect between how single-cell transcriptomic data are currently analyzed and what the field actually needs to enable biological discovery. Over the past decade, the scale of single-cell RNA sequencing (scRNA-seq) datasets has exploded—the authors cite collections like CZ CELLxGENE, the PARSE 10M PBMC dataset, and various cross-species atlases that together span millions of cells across diverse tissues, species, and experimental conditions. Yet despite this wealth of data, the dominant analytical paradigm remains descriptive cataloging: annotating cell types, identifying clusters, comparing expression patterns across conditions, and visualizing low-dimensional embeddings.

These are valuable activities, but they are fundamentally backward-looking. They describe what has been observed in existing experiments. What the field urgently needs—and what the paper argues is required for genuinely transformative applications—is a shift toward predictive, generative computational frameworks that can simulate cellular states and their responses to interventions that have not yet been experimentally measured. The paper crystallizes this ambition in the concept of a cellular world model: a system that, analogously to world models in reinforcement learning that learn compact representations of an environment and support conditional simulation, represents the distribution of transcriptomic states and their conditional dynamics under perturbation.

The practical stakes of this gap are enormous. If researchers could reliably simulate how a population of cells responds to a genetic perturbation (e.g., a CRISPR knockout), a cytokine stimulation, a drug treatment, or a combination of interventions, the bottlenecks in biological discovery would fundamentally shift. Large-scale in silico screening could prioritize the most promising perturbations for expensive and time-consuming wet-lab validation. Disease mechanisms could be dissected by simulating cascading transcriptional effects without needing to physically perturb every gene in every relevant cell type. Developmental trajectories could be mapped by simulating transitions between states under hypothesized regulatory programs. The paper frames these as applications that "empower researchers to conduct large-scale in silico experiments to dissect disease mechanisms, screen potential therapeutics, and map complex developmental trajectories" (Section 1).

This is not merely a convenience. The combinatorial space of possible perturbations—across genes, cell types, donors, dosages, time points, and environmental conditions—is so vast that exhaustive experimental exploration is impossible. A generative model that can extrapolate to unseen combinations of these factors would function as a hypothesis generator, narrowing the search space for wet-lab validation and enabling exploration of biological questions that are currently intractable.

Why Existing Foundation Models Fall Short: The Static Representation Paradigm

The most prominent recent advance in computational single-cell biology has been the development of foundation models—large-scale self-supervised learning systems pretrained on massive transcriptomic datasets to produce transferable representations. The paper reviews several leading examples:

  • scGPT (Cui et al., 2024): a generative pretrained transformer that learns gene-level and cell-level representations by treating gene expression values as tokens in a language-model-like framework, enabling downstream tasks like cell-type annotation, batch correction, and perturbation prediction.
  • Geneformer (Theodoris et al., 2023): a transformer pretrained on a corpus of ~30 million single-cell transcriptomes using a masked-attention objective that learns context-dependent gene embeddings, with demonstrated transfer to network biology tasks including dosage sensitivity prediction and disease gene prioritization.
  • scFoundation (Hao et al., 2024): a large-scale foundation model trained on over 50 million cells with a transformer architecture adapted for the high-dimensional, sparse nature of scRNA-seq data, achieving strong performance on cell-type annotation and perturbation prediction.
  • CellFM (Zeng et al., 2025): a foundation model pretrained on 100 million human cells, designed to learn unified representations across diverse tissues and conditions.

These models represent a major step forward, and the paper explicitly acknowledges their successes. However, the authors identify a critical limitation: these models are optimized for representation learning, not generative simulation. Their primary output is a meaningful embedding vector—a static encoding of a cell's transcriptomic state that can be used for downstream discriminative tasks. What they do not provide is a principled way to sample new transcriptomic profiles from the learned distribution, or to conditionally generate the transcriptomic state of a cell under a specified perturbation.

This distinction is subtle but crucial. A foundation model can tell you that two cells are similar, or that a particular perturbation shifts the embedding in a certain direction, but it cannot directly synthesize a realistic whole-transcriptome expression vector for a novel cell type under a novel perturbation. The paper argues that this gap—the absence of an explicit generative model of the transcriptomic state space—prevents foundation models from serving as the core engine of in silico experimentation. They organize cells into a shared latent space, but they cannot populate that space with hypothetical cells that have not been observed.

Where Existing Generative Approaches Fall Short: The Continuous-Discrete Mismatch

The paper does not claim to be the first generative model for single-cell transcriptomics. It identifies two major classes of prior generative work, each with specific limitations that motivate the discrete diffusion approach:

Continuous-space VAEs: scVI and its limitations. scVI (Lopez et al., 2018) is perhaps the most widely adopted generative model for scRNA-seq, and it serves as an important baseline in this paper. scVI uses a variational autoencoder (VAE) framework: it encodes each cell's expression profile into a low-dimensional latent vector, then decodes from that latent vector to produce parameters of a zero-inflated negative binomial distribution over gene counts. This is a principled approach that explicitly models the count nature of the data, and scVI has been successfully applied to batch correction, differential expression analysis, and preliminary perturbation modeling.

However, the paper argues that scVI's performance is "limited by continuous data assumptions that misalign with the sparse, discrete, and non-sequential nature of single-cell transcriptome data" (Section 1). While scVI's output distribution is discrete (the negative binomial), the latent space and the encoder-decoder transformations operate in continuous Euclidean space. The VAE bottleneck—compressing a cell's full ~18,000-gene expression profile through a low-dimensional continuous latent code—creates an information bottleneck (a phenomenon formalized in the information bottleneck literature that the paper cites from Alemi et al., 2017). The reconstruction is probabilistic, and the model must learn to represent all the complex combinatorial patterns of gene co-expression through a compressed continuous representation. For highly sparse data with many near-zero counts and heavy-tailed distributions, this compression can lose fine-grained structure.

Continuous diffusion models: scDiffusion and the noise-distribution mismatch. scDiffusion (Luo et al., 2024) adapts denoising diffusion probabilistic models (DDPMs; Ho et al., 2020)—the class of generative models that have driven recent advances in image and audio generation—to single-cell transcriptomics. In a DDPM, the generative process works by learning to reverse a forward diffusion that gradually corrupts data with Gaussian noise. At inference time, the model starts from pure noise and iteratively denoises toward a realistic sample.

The paper identifies a fundamental mismatch in this approach when applied to scRNA-seq data. Gaussian noise is a continuous, symmetric, unbounded distribution: it adds equal positive and negative perturbations of varying magnitude to every gene position. But scRNA-seq count data are sparse, discrete, non-negative, and zero-inflated—meaning most entries in the expression matrix are exactly zero, non-zero values are positive integers with a heavy-tailed distribution, and the data's intrinsic structure does not resemble samples from a continuous distribution that has been corrupted by additive noise. The paper explicitly cites Vignac et al. (2023), who noted that "continuous noise processes... [create a] distributional mismatch between... sparse, discrete count data" and the Gaussian corruption used in standard diffusion models. This mismatch means that the model's input distribution (noisy versions of clean data) can look qualitatively different from anything the model would encounter at generation time (pure noise), potentially degrading sample quality.

More subtly, DDPMs corrupt every position simultaneously at each forward step. For an ~18,000-gene expression vector, this means that at intermediate diffusion timesteps, the input to the denoising model is a dense vector where every gene has been partially noised—a state very unlike the highly sparse real data. The paper argues this forces the model to learn denoising behavior that is misaligned with the natural structure of transcriptomic data.

Autoregressive models: the ordering problem. Although not a primary baseline in the experiments, the paper discusses autoregressive (AR) models—which generate expression values gene-by-gene in a sequential, left-to-right manner—as another conceptually mismatched approach. Gene expression profiles have no natural ordering: genes are not a sequence, and there is no biological reason why the expression value of gene A should be predicted before gene B. An AR model must impose an arbitrary generation order, creating an inductive bias that "conflicts with the permutation invariance and zero-inflated sparsity intrinsic to transcriptomic measurements" (Section 3). The citations to Austin et al. (2021) and Nie et al. (2025) reference the broader literature on why autoregressive generation is suboptimal for data without sequential structure, a phenomenon observed in domains like graph generation and set-structured data.

Where Perturbation-Focused Methods Fall Short: Direct Mapping Without Distributional Modeling

The paper also positions itself against a class of methods specifically designed for perturbation response prediction, which circumvent the generative modeling problem entirely by learning direct mappings from control states to perturbed states:

  • STATE (Adduri et al., 2025): predicts perturbation responses by combining a variational autoencoder with a perturbation-aware conditioning mechanism, learning a mapping from (cell state, perturbation) pairs to perturbed expression profiles.
  • CellFlow (Klein et al., 2025): uses flow matching—a continuous normalizing flow approach—to learn the transformation between control and perturbed cellular states, modeling the perturbation response as a continuous trajectory in latent space.
  • scDFM (Yu et al., 2026): applies distributional flow matching for robust single-cell perturbation prediction, directly modeling the conditional distribution of perturbed states given the unperturbed control.
  • AlphaCell (Chuai et al., 2026): explicitly describes itself as "towards building a world model to simulate perturbation-induced cellular dynamics," but the paper groups it with methods that learn direct mappings rather than full generative state-space models.

These methods are effective for their specific prediction tasks, and several are among the top performers on the Virtual Cell Challenge leaderboard. However, the paper identifies two intertwined limitations. First, they do not model the underlying distribution of transcriptomic states. A direct mapping from control to perturbed can produce a point estimate or a narrow distribution of perturbed expression profiles, but it does not capture the full breadth of cellular heterogeneity—the fact that even under identical perturbation conditions, different cells within a population exhibit variable transcriptional responses. Second, because they do not model the state distribution, they cannot support unconditional generation or flexible conditional simulation beyond the specific control→perturbed mapping they were trained on. If you want to generate a diverse population of unperturbed T cells with realistic cell-to-cell variability, or simulate what happens when a perturbation is applied to a cell type not seen during training, these models offer no mechanism. They are predictors of specific outcomes, not simulators of the broader state space.

The Conceptual Gap: No Unified Generative Cellular World Model

The paper's synthesis of these limitations points to a clear conceptual gap. The field has, on one side, powerful foundation models that learn static representations but do not generate; and on the other side, task-specific perturbation predictors that map between conditions but do not model distributions. What is missing—and what Lingshu-Cell aims to provide—is a unified generative framework that simultaneously:

  1. Models the transcriptomic state distribution—the full joint distribution over ~18,000 gene expression values that captures cellular heterogeneity, rare cell states, marker-gene patterns, and cell-type proportions.
  2. Supports conditional generation—the ability to sample from the distribution of transcriptomic states conditioned on arbitrary combinations of cell identity (cell type, donor, tissue) and perturbation context (genetic perturbation, cytokine stimulation, potentially drug treatment).
  3. Operates in a representation aligned with the data's structure—discrete, sparse, permutation-invariant, and free from artificial ordering biases.
  4. Scales to transcriptome-wide modeling—processing all ~18,000 genes without prior feature selection based on high variability or expression level, so that even lowly expressed or sporadically detected genes are part of the model's learned dependencies.

The paper positions this ambition as a direct analog to how world models function in AI: a world model learns a compressed representation of an environment's state transitions and uses it to simulate "what would happen if" without needing to actually execute actions in the real environment. A cellular world model analogously learns the distribution of transcriptomic states and how they shift under interventions, enabling simulation of hypothetical experiments.

The Generative Paradigm Choice: Why Masked Discrete Diffusion?

The paper's central technical claim is that masked discrete diffusion models (MDDMs) are particularly well-suited to serve as this unified framework. This choice is motivated by a careful alignment between the properties of the generative process and the properties of the data, which the paper contrasts explicitly with the alternatives (Figure 1b):

  • Versus autoregressive models: MDDMs generate all gene expression values simultaneously and bidirectionally, with the model able to condition each gene's prediction on all other genes (both masked and unmasked) at every refinement step. There is no imposed generation order, so the model respects the permutation invariance of gene expression data.
  • Versus continuous diffusion (DDPMs): MDDMs corrupt data by masking tokens (replacing expression values with a special MASK token) rather than adding continuous Gaussian noise. This aligns naturally with the data's discrete, sparse structure—masking a zero-value gene position keeps it as MASK rather than turning it into a continuous fractional value, and the model sees sequences of discrete tokens at every diffusion timestep rather than continuous-valued noisy vectors.
  • Versus VAEs: MDDMs avoid the information bottleneck of compressing through a low-dimensional continuous latent space. The mask predictor network operates on the full (potentially compressed) sequence of discrete tokens, predicting all masked positions in parallel, without any dimensionality reduction constraint.

The paper also notes a practical advantage: because masking preserves the discrete token structure but introduces sparsity, the model's input at intermediate timesteps looks structurally similar to the sparse scRNA-seq data it was trained on—many positions are MASK (analogous to zero counts), and non-masked positions retain their original discrete expression values. This means the model's training distribution (partially masked sequences) is qualitatively similar to its inference distribution (which transitions from fully masked to fully generated), unlike in continuous diffusion where intermediate-noised vectors are structurally unlike the clean data.

Positioning Relative to the Virtual Cell Challenge

The paper's timing is also shaped by the Virtual Cell Challenge (VCC) (Roohani et al., 2025), a community benchmark that evaluates methods on genetic perturbation prediction in a held-out cell line (H1). The VCC provides a standardized dataset, evaluation metrics, and a public leaderboard, making it a natural testbed for perturbation prediction methods. The paper evaluates Lingshu-Cell on this benchmark and compares against the top-performing entries on the leaderboard (Table 2, Table B1), noting that the model achieves the best average rank across all seven evaluation metrics.

This is significant because the VCC leaderboard includes a diverse range of methods, many of which are specialized perturbation predictors rather than general-purpose generative models. Lingshu-Cell's competitive performance establishes that a unified generative framework—designed primarily to model transcriptomic state distributions—can match or exceed task-specific predictive models on perturbation benchmarks, while additionally providing unconditional generation capabilities that the specialized methods lack.

How the Paper Frames Its Contribution Relative to Prior Work

The paper's self-positioning, as articulated in Section 1 and the Discussion, is that it represents a paradigm shift in single-cell transcriptomic modeling:

  • It shifts foundation models from static representation learning to generative simulation, arguing that the ability to sample from the learned distribution is essential for in silico experimentation.
  • It aligns the computational paradigm (discrete diffusion) with the physical properties of the data (sparse, discrete, non-sequential), characterizing this as a design choice that addresses fundamental limitations of VAEs, continuous diffusion models, and autoregressive models.
  • It demonstrates unified capabilities within a single architecture—unconditional generation, cross-tissue and cross-species simulation, genetic perturbation prediction, and cytokine perturbation prediction—where prior work typically addressed these tasks separately.

The paper is careful to acknowledge that this paradigm shift is aspirational rather than complete. The Discussion explicitly notes that the model "should currently be regarded as a powerful tool for probabilistic hypothesis generation" whose "predictions still require rigorous wet-lab experimental validation," that it does not model causal regulatory mechanisms, and that it is limited to transcriptomic data. These caveats position the work as a foundational step toward virtual cells and a cellular world model, rather than claiming to have already achieved that goal.

3. Technical Approach

3.1 Reader Orientation

Lingshu-Cell is a masked discrete diffusion model (MDDM) that learns to generate realistic single-cell gene expression profiles by treating each cell as a sequence of discrete tokens representing expression levels across approximately 18,000 genes. The system solves the problem of simultaneously modeling the full transcriptomic distribution (for unconditional simulation of heterogeneous cell populations) and predicting perturbation responses (for conditional generation of transcriptomes under unseen interventions), all within a single architecture that respects the sparse, discrete, non-sequential structure of scRNA-seq count data.

3.2 Big-Picture Architecture (Diagram in Words)

The system has five major components organized as a generative pipeline:

  1. Quantization Module — maps raw UMI counts for each gene to a finite set of discrete tokens using an adaptive binning scheme, converting a cell's continuous-valued count vector into a sequence of categorical indices over a vocabulary of ~280 expression levels plus an overflow token.
  2. Sequence Compression Module — reduces the effective sequence length processed by the Transformer backbone by grouping gene embeddings into blocks, projecting each block to a single compressed token via a learned linear map, and later restoring the original resolution. This sits between the token embedding and Transformer layers.
  3. Mask Predictor Network (Transformer Backbone) — a bidirectional LLaMA-style Transformer with 13 blocks, 640-dimensional embeddings, 10 attention heads, and SwiGLU feed-forward layers. It takes a partially masked token sequence and predicts the original tokens for all masked positions simultaneously.
  4. Forward Masking Process — progressively replaces expression tokens with a special MASK token over continuous time $t \in [0, 1]$, where the masking probability equals $t$. This creates training pairs of (partially masked sequence, original sequence) without requiring a fixed generation order.
  5. Conditioning Interface — prepends discrete condition tokens (cell type/donor identity and perturbation target) to the expression sequence before masking, enabling conditional generation. At inference, classifier-free guidance (CFG) steers sampling toward perturbation-specific expression profiles by interpolating between conditional and unconditional predictions.

Information flows as follows: a cell's UMI count vector enters the quantization module → each gene's count is mapped to a discrete token ID → condition tokens are prepended → a random fraction of gene tokens is masked → the compressed sequence passes through the Transformer backbone → the decompression module restores full resolution → a linear head outputs logits over the expression vocabulary for each masked position → the loss is computed only on originally masked positions → at inference, the model starts from all-MASK and iteratively replaces masks with predicted tokens over $N$ steps.

3.3 Roadmap for the Deep Dive

  • First, the MDDM framework (Section 4.1) — the forward masking process, reverse generation objective, and inference sampling procedure, because these define the core learning dynamics shared by all model variants.
  • Second, the discrete representation of scRNA-seq data (Section 4.2) — how raw UMI counts are quantized into tokens, including the adaptive binning scheme, because this mapping defines the input/output space that the model operates on and directly determines what biological information is preserved.
  • Third, the sequence compression module (Section 4.3) — the random grouping, linear projection, and decompression operations, because this is the architectural innovation that makes transcriptome-wide modeling computationally tractable.
  • Fourth, the conditional generation framework (Section 4.4) — condition token encoding, training with control cells, and classifier-free guidance at inference, because this is how the model supports perturbation prediction while maintaining a unified architecture.
  • Fifth, the inference-time biological prior injection (Section 4.5) — how external perturbation knowledge is incorporated to improve conditional generation, because this is a critical practical component that bridges the gap between purely data-driven prediction and biological knowledge.

3.4 Detailed, Sentence-Based Technical Breakdown

This is primarily a generative modeling paper whose core idea is that masked discrete diffusion, operating on quantized expression tokens with sequence compression and classifier-free guidance, can serve as a unified framework for both unconditional transcriptomic simulation and conditional perturbation response prediction.

Masked Discrete Diffusion Model (MDDM) Foundation

The paper adopts the MDDM framework from the broader discrete diffusion literature (citing Austin et al., 2021) as the core generative engine. The key insight is that MDDMs model sequences of discrete tokens—analogous to text tokens in language modeling—but operate through a masking-and-prediction mechanism that is fundamentally bidirectional and order-independent.

Forward masking process (corruption). Let $x_0 = [x_0^1, x_0^2, \ldots, x_0^L]$ be a fully observed discrete sequence of length $L$, where each token $x_0^i$ belongs to a predefined vocabulary $\mathcal{V}$. The forward process gradually and independently masks tokens over a continuous time variable $t \in [0, 1]$. At $t = 0$, the sequence is completely clean (all tokens are the original expression values). At $t = 1$, the sequence is fully masked (all tokens are replaced by a special mask token $\mathbf{M} \notin \mathcal{V}$).

For any intermediate timestep $t$, each token $x_t^i$ is independently replaced by $\mathbf{M}$ with probability $t$, or remains the original token with probability $1 - t$. The transition probability is formally defined as:

qt∣0(xti∣x0i)={1−t,xti=x0i,t,xti=M.q_{t|0}(x_t^i \mid x_0^i) = \begin{cases} 1 - t, & x_t^i = x_0^i, \\ t, & x_t^i = \mathbf{M}. \end{cases}

where $q_{t|0}$ is the conditional distribution of the corrupted token at time $t$ given the original token at time $0$, $x_t^i$ is the token at position $i$ at corruption time $t$, $x_0^i$ is the original clean token, and $t$ is the masking probability (equivalently, the fraction of tokens expected to be masked).

What this computes: At each training step, the system samples a corruption time $t \sim \mathcal{U}(0, 1)$ uniformly from the unit interval, and independently flips a biased coin for each position: with probability $t$, replace the expression token with MASK; otherwise, keep the original token. The resulting partially masked sequence $x_t$ is the input to the model, and the model must predict the original tokens for all masked positions.

Why this form: The independent, per-position masking with probability equal to the continuous time variable achieves two critical properties. First, it makes the masking process factorized across positions, meaning there is no sequential dependency in the corruption—this mirrors the permutation invariance of gene expression data where no gene order exists. Second, the continuous time parameter $t$ creates a smooth curriculum: when $t$ is small, the model sees mostly clean data and must fill in a few masks (an easy "denoising" task); when $t$ is large, the model sees mostly masked data and must reconstruct almost the entire expression profile (a hard generation task). This stands in contrast to autoregressive models, which always condition on a fixed-length prefix, and to continuous diffusion models, which corrupt all positions simultaneously with noise of varying magnitude.

Reverse process and training objective. The reverse process learns a parametric model $p_\theta(\cdot \mid x_t)$ that takes the partially masked sequence $x_t$ as input and predicts the original tokens for all masked positions simultaneously. The model is trained using a cross-entropy loss computed exclusively on the positions that are currently masked:

L(θ)≜−Et,x0,xt[1t∑i=1LI[xti=M]log⁡pθ(x0i∣xt)]\mathcal{L}(\theta) \triangleq -\mathbb{E}_{t, x_0, x_t}\left[\frac{1}{t} \sum_{i=1}^{L} \mathbb{I}[x_t^i = \mathbf{M}] \log p_\theta(x_0^i \mid x_t)\right]

where $t \sim \mathcal{U}(0, 1)$ is the sampled corruption time, $x_0$ is the original clean sequence, $x_t$ is the partially masked sequence at time $t$ (sampled from $q_{t|0}$), $\mathbb{I}[x_t^i = \mathbf{M}]$ is an indicator that selects only positions currently holding the MASK token, $p_\theta(x_0^i \mid x_t)$ is the model's predicted probability for the true original token at position $i$, and $L$ is the sequence length.

What this computes: For each training example, the system first samples a corruption time $t$ uniformly from 0 to 1 and generates a partially masked sequence by applying the forward process. The model processes this masked sequence and outputs, for every position, a categorical probability distribution over the vocabulary $\mathcal{V}$. The loss is the negative log-likelihood of the true original token under this predicted distribution, summed over all masked positions, and weighted by $1/t$. The $1/t$ weighting is critical: when $t$ is small (few masks), the model gets few training signal positions but each is weighted heavily; when $t$ is large (many masks), the model gets many training signal positions but each is weighted lightly. This balances the contribution of different corruption levels.

Why this form: The paper states that this objective "provides a principled variational upper bound on the negative log-likelihood of the true data distribution." More operationally, the three design choices in this loss work together to make learning tractable:

  1. Mask-only loss: Computing loss only on masked positions (via $\mathbb{I}[x_t^i = \mathbf{M}]$) means the model is trained specifically on the reconstruction task—predicting what was removed—rather than being distracted by also predicting the values of positions that are already known. This focuses the model's capacity on the generation-relevant computation.
  2. $1/t$ weighting: Without this, the loss would be dominated by high-$t$ steps where many positions contribute, causing the model to under-learn the easier low-$t$ regime. The $1/t$ factor normalizes the expected contribution per corruption level, ensuring the model learns across the full difficulty spectrum.
  3. Cross-entropy loss: This is the standard maximum-likelihood objective for categorical data, appropriate because the expression values are discrete tokens rather than continuous numbers. Using mean-squared error here—as would be natural for continuous data—would treat token IDs as having meaningful numerical distances (e.g., token 5 is "closer" to token 6 than to token 20), which is not true for the quantized bins.

Inference and sampling. At generation time, the reverse process is discretized into $N$ steps. The system initializes a fully masked sequence $x_1$ of length $L$ (all positions are MASK). The reverse process iteratively transitions from the current time $t \in (0, 1]$ to an earlier time $s \in [0, t)$.

At each step, the mask predictor $p_\theta(\cdot \mid x_t)$ estimates all masked tokens in parallel, producing a categorical prediction for each MASK position. Then, a subset of these predictions are accepted (the predicted token replaces the MASK), while the remaining positions stay masked or are re-masked to construct $x_s$. The key invariant is that the fraction of tokens that remain masked after the transition matches the forward process expectation: specifically, a proportion $s/t$ of the currently masked tokens are re-masked (or equivalently, a proportion $1 - s/t$ of the currently masked tokens are filled with their predicted values).

What this computes in operational terms: Suppose at time $t = 0.5$, approximately 50% of positions are MASK and 50% retain their predicted/generated values. The model predicts values for all masked positions. The system then decides how many of these predictions to "commit" (make permanent) and how many to keep masked. If the next target time is $s = 0.25$, the system should end up with approximately 25% of positions still masked. This means roughly half of the currently masked positions (50% → 25% of total) get filled with their predicted values, and the other half from (50% → 25% of total) remain masked. The positions to fill are selected uniformly at random from the currently masked set.

Why this sequential unmasking strategy: The paper uses a cosine timestep schedule to determine the masking fraction at each step $n \in \{1, \ldots, N\}$. Rather than uniformly spacing the unmasking rate across steps, a cosine schedule allocates more unmasking in the middle steps and less at the extremes. This is because, at very high masking levels, the model has insufficient context to make confident predictions (so unmasking fewer tokens is appropriate), while at very low masking levels, the remaining tokens are in hard-to-predict positions (so the model benefits from more refinement steps before committing). The paper uses $N = 256$ steps for unconditional generation (where fidelity matters) and $N = 3$ steps for conditional generation (where speed matters and classifier-free guidance provides additional steering). Within each step, the token values at the selected unmasking positions are drawn from the model's predicted categorical distribution without temperature scaling or top-k truncation—the model samples directly from the learned distribution, which preserves diversity at the cost of potentially including low-probability tokens.

Representing Single-Cell Transcriptomic Data as Discrete Tokens

The critical bridge between the continuous world of UMI counts and the discrete world of MDDMs is the quantization scheme that maps each gene's expression level to a token ID. The design of this mapping directly determines what biological information the model can capture, how efficiently the vocabulary represents the data, and whether rare expression values are preserved or lost.

Input data structure. A scRNA-seq experiment produces a UMI (Unique Molecular Identifier) count matrix $X \in \mathbb{Z}_{\geq 0}^{N \times G}$, where $N$ is the number of cells and $G$ is the number of genes. The $i$-th cell is represented by a count vector $x^{(i)} = (x^{(i)}_1, \ldots, x^{(i)}_G) \in \mathbb{Z}_{\geq 0}^G$ over a fixed gene list. In Lingshu-Cell, $G = 18,080$ genes—the gene set used in the Virtual Cell Challenge H1 benchmark, which the paper standardizes all datasets to (Appendix A.1).

The quantization challenge. Raw UMI counts span an enormous dynamic range: the paper notes that counts range from 0 to values potentially exceeding 9,999 (the default cap $C = 9999$), with a heavily right-skewed distribution where most genes have zero or near-zero counts in most cells, while a small number of highly expressed genes like ribosomal or mitochondrial transcripts may have counts in the hundreds or thousands. A naive approach—using each distinct integer count as a separate token—would produce a vocabulary of thousands of tokens per gene, making the model's categorical prediction space intractably large and wasting capacity on distinguishing between count values that are biologically equivalent (e.g., counts of 247 vs. 248 probably convey the same information about expression level).

The paper's solution is a logarithmically adaptive binning scheme that preserves fine resolution at low counts (where every transcript matters) and progressively coarser resolution at high counts (where exact values are less informative). The quantization function $q: \mathbb{Z}_{\geq 0} \to \{0, 1, \ldots, B-1\} \cup \{\text{OVF}\}$ maps each raw count to a bin index, with three regimes defined by the count value $x$:

q(x)={x,0≤x<100,100+90⋅max⁡(0,k(x)−2)+r(x),100≤x≤C,OVF,x>C.q(x) = \begin{cases} x, & 0 \leq x < 100, \\ 100 + 90 \cdot \max(0, k(x) - 2) + r(x), & 100 \leq x \leq C, \\ \text{OVF}, & x > C. \end{cases}

where:

  • $k(x) = \lfloor \log_{10}(\max(x, 1)) \rfloor$ is the decade index (e.g., $k = 2$ for counts in [100, 999], $k = 3$ for counts in [1000, 9999]),
  • $r(x) = \lfloor (x - 10^{k(x)}) / \Delta(x) \rfloor$ is the within-decade offset,
  • $\Delta(x) = 10^{k(x)-1}$ is the adaptive step size for the decade (e.g., $\Delta = 10$ for [100, 999], $\Delta = 100$ for [1000, 9999]),
  • $C = 9999$ is the default count cap, and
  • $\text{OVF}$ is a special overflow token for counts exceeding the cap.

What this computes in operational terms: For counts below 100, each integer gets its own bin—so counts of 0, 1, 2, ..., 99 are preserved exactly. This covers the vast majority of gene-cell pairs in sparse scRNA-seq data. For counts from 100 to 999 ($k = 2$), the step size is $\Delta = 10^{1} = 10$, so the bins are 100–109 (bin 100), 110–119 (bin 101), ..., 990–999 (bin 189)—90 bins total for this decade. For counts from 1000 to 9999 ($k = 3$), the step size is $\Delta = 10^{2} = 100$, so bins are 1000–1099 (bin 190), 1100–1199 (bin 191), ..., 9900–9999 (bin 279)—another 90 bins. Counts above 9999 (if any exist) map to the single overflow token OVF. The total non-overflow vocabulary size is $B = 100 + 90 + 90 = 280$, and $|\mathcal{V}| = 281$ after including the overflow token (excluding special tokens like MASK that are part of the MDDM training infrastructure).

Why this form: The three-regime design directly reflects the information content of UMI counts at different magnitudes. At low counts (0–99), the Poisson-like sampling noise in scRNA-seq means that counts of 3 vs. 4 can reflect genuine biological differences in expression, especially for lowly-expressed transcription factors where a handful of transcripts can drive distinct cellular states. Preserving exact integer resolution here is therefore biologically important. As counts increase, the relative measurement noise decreases, and the biological significance of exact count differences diminishes—the difference between 247 and 248 transcripts is almost certainly noise, and even differences of tens of counts may be within technical variation. Coarsening the resolution at high counts (10-count bins for k=2, 100-count bins for k=3) reduces the vocabulary size and makes the model's categorical prediction task easier without losing biologically relevant information. The logarithmic decade structure ensures that the bin resolution scales approximately as a fixed fraction of the count magnitude, which is the natural scaling for multiplicative biological processes like transcription.

Discrete sequence representation. After applying the quantization function element-wise to all G genes, each cell is represented as a discrete sequence:

z(i)=(q(x1(i)),q(x2(i)),…,q(xG(i)))z^{(i)} = \left(q(x^{(i)}_1), q(x^{(i)}_2), \ldots, q(x^{(i)}_G)\right)

where $z^{(i)}_g \in \{0, 1, \ldots, B-1, \text{OVF}\}$ is the token ID for gene $g$ in cell $i$. This discrete sequence $z^{(i)}$ of length $G = 18,080$ serves as the input/output space for the MDDM—the model sees sequences of these token IDs, predicts token IDs for masked positions, and the predicted token IDs are mapped back to expression values (typically the midpoint of the quantized bin) for downstream analysis.

Gene standardization. Because the model expects a fixed-length sequence with a fixed gene order, all datasets are aligned to the same reference set of 18,080 genes used in the VCC H1 benchmark. For non-human species, gene annotations are obtained from Ensembl BioMart, and only protein-coding genes with valid gene names are retained (Appendix A.1). This means the model learns gene-specific dependencies that are consistent across all training examples—the same position in the sequence always corresponds to the same gene, allowing the Transformer's self-attention to learn position-specific correlations (e.g., "position 3742 and position 8915 always co-vary in expression").

Embedding-Space Sequence Compression

With $G = 18,080$ genes, the raw token sequence length is nearly 18K. A standard Transformer with $L = 13$ blocks and $O(G^2)$ self-attention complexity would be computationally prohibitive for training on millions of cells. The paper introduces a sequence compression module that operates purely in embedding space—the input and output interfaces of the model remain at the full gene resolution, but the internal Transformer computation is performed on a compressed representation with shorter effective sequence length.

Core idea. If each token embedding is $D$-dimensional (where $D = 640$), then a full-length sequence produces an embedding tensor of shape $G \times D$. The compression module first groups gene embeddings into blocks of size $S$, then linearly projects each block from $S \times D$ dimensions down to $D$ dimensions, reducing the sequence length from $G$ to $G_c = \lceil G / S \rceil$. After the Transformer processes the compressed sequence, a decompression module performs the inverse operation to restore the full $G \times D$ representation before the final output projection to vocabulary logits.

Random permutation and grouping. Before compression, the gene embeddings are reordered by a fixed random permutation $\pi$ over gene positions, sampled once at model initialization and held constant for all cells. This permutation serves two purposes:

  1. It breaks any spurious positional correlation that might arise from adjacent genes in the canonical gene ordering (e.g., genes that happen to be sorted alphabetically or by genomic position and therefore appear next to each other in the sequence). By randomly shuffling gene positions, the model must learn gene-gene dependencies through semantic interaction rather than through positional adjacency.
  2. It ensures that each compression group contains a random mix of genes from across the genome, so that the linear projection from $S \times D$ to $D$ dimensions mixes information across unrelated genes, which the paper argues "reduc[es] the influence of single-gene noise and improv[es] robustness under perturbation" (Section 4.3).

The reordered embedding sequence $E_\pi = \pi(E)$ is then partitioned into consecutive groups of size $S$. Writing $E_\pi^{(i)}$ for the $i$-th group (an $S \times D$ matrix), the compression projects each group to a single $D$-dimensional vector:

Hi=Wdown⋅vec(Eπ(i)),i=1,…,GcH_i = W_{\text{down}} \cdot \text{vec}\left(E_\pi^{(i)}\right), \quad i = 1, \ldots, G_c

where $W_{\text{down}} \in \mathbb{R}^{D \times (SD)}$ is a shared linear projection (applied identically to every group), $\text{vec}(\cdot)$ flattens the $S \times D$ block into a single vector of length $S \cdot D$, $G_c = \lceil G / S \rceil$ is the compressed sequence length (padding at the end if $G$ is not perfectly divisible by $S$), and $H_i \in \mathbb{R}^D$ is the compressed token embedding for group $i$.

What this computes: For each block of $S$ gene embeddings (each a 640-dimensional vector), the compression step concatenates all $S \times 640$ values into one long vector and multiplies by a learned weight matrix to produce a single 640-dimensional summary vector. This is done independently for each block using the same weight matrix, so the compression is translation-invariant across the sequence. The resulting compressed sequence $H = [H_1, \ldots, H_{G_c}]^\top \in \mathbb{R}^{G_c \times D}$ is then fed into the Transformer backbone.

Transformer processing. The compressed sequence $H$ is processed by a LLaMA-style Transformer backbone consisting of $L = 13$ identical blocks. Each block applies two sub-layers with pre-normalization (RMSNorm) and residual connections:

x←x+Attn(RMSNorm(x))x \leftarrow x + \text{Attn}(\text{RMSNorm}(x)) x←x+FFN(RMSNorm(x))x \leftarrow x + \text{FFN}(\text{RMSNorm}(x))

The self-attention sub-layer uses multi-head attention with $n_h = 10$ heads and head dimension $d_h = 64$ (so total attention dimension $D = 640$). Rotary position embeddings (RoPE) are applied to query and key vectors in every block. Critically, attention is bidirectional (no causal mask), consistent with the non-autoregressive, order-independent nature of the MDDM formulation. The feed-forward sub-layer uses a SwiGLU activation with intermediate dimension $d_{ff} = 2560$. All linear layers are bias-free, and normalization uses RMSNorm.

Decompression. After the final Transformer block, a final RMSNorm is applied, and the compressed sequence is expanded back to the full gene-level resolution. For each compressed representation $H_i$, a shared linear up-projection $W_{\text{up}} \in \mathbb{R}^{(SD) \times D}$ maps back to an $S \times D$ block:

E^π(i)=unvec(Wup⋅Hi),i=1,…,Gc\hat{E}_\pi^{(i)} = \text{unvec}(W_{\text{up}} \cdot H_i), \quad i = 1, \ldots, G_c

These blocks are concatenated, and the inverse permutation $\pi^{-1}$ restores the original gene ordering:

E^=π−1(concat(E^π(1),…,E^π(Gc)))\hat{E} = \pi^{-1}\left(\text{concat}\left(\hat{E}_\pi^{(1)}, \ldots, \hat{E}_\pi^{(G_c)}\right)\right)

Finally, a linear output head maps each position's $D$-dimensional representation to logits over the expression vocabulary of size $|\mathcal{V}|$.

Why this form and not alternatives: The paper considers and ablates several design choices. The compression approach is applied only to internal computation, meaning the model's input and output interfaces remain at full gene resolution—the token-level prediction space is unchanged, and the loss is still computed per-gene. This is qualitatively different from simply predicting at a coarser resolution or using a bottleneck latent variable (as in a VAE). The random permutation ensures that each compressed group mixes information across genes that are not biologically co-regulated by virtue of genomic proximity, which forces the model to learn meaningful gene-gene interactions through attention rather than relying on positional adjacency as a shortcut.

The default group size is $S = 8$, yielding a compressed length $G_c = \lceil 18080 / 8 \rceil = 2260$—a roughly 8× reduction in effective sequence length, and hence a $\sim 64\times$ reduction in the quadratic attention cost. For genetic perturbation prediction, ablation experiments showed that a larger group size $S = 32$ ($G_c = 565$) improved performance (Figure 3f, Appendix Figure C1) in addition to further reducing computation. The paper attributes this to moderate compression improving the model's ability to capture global expression patterns by mixing information across larger gene sets, effectively "reducing the influence of single-gene noise."

Conditional Generation Framework

For perturbation response prediction, Lingshu-Cell generates expression profiles conditioned on a context $c$ that specifies both the cellular background (cell line or donor identity) and the perturbation (target gene for CRISPR, cytokine for stimulation). The conditioning mechanism is designed to be seamlessly integrated into the discrete token framework—no separate conditioning encoder or modality-specific architecture is needed.

Condition encoding. Each condition value—e.g., the cell line "A549" or the perturbation target "CHMP3"—is represented as a dedicated discrete token appended to the expression-token vocabulary $\mathcal{V}$. This means conditions are encoded in the same vocabulary space as expression values, and their embeddings are learned through the same embedding layer as the gene expression tokens. The two condition tokens (source context and perturbation identity) are prepended to the expression token sequence to form the full model input:

Input=[token(cell_line),token(perturbation),z1,z2,…,zG]\text{Input} = [\text{token}(\text{cell\_line}), \text{token}(\text{perturbation}), z_1, z_2, \ldots, z_G]

Why prepending rather than a separate encoder: This design choice ensures that the condition signal is directly available to the self-attention mechanism, which can learn to attend to the condition tokens when predicting masked gene expression values. It also preserves the unified discrete sequence architecture—the Transformer processes everything as a single sequence without needing separate pathways for conditioning vs. expression.

Condition masking exemption. The condition tokens are exempt from masking throughout the forward corruption process, meaning they remain visible at every diffusion step regardless of $\text{$t$}$. This is critical because it ensures the conditioning signal is available at all corruption levels—even when $\text{$t$}$ is close to 1 and most gene tokens are masked, the model can still see the condition tokens and use them to guide the prediction toward perturbation-specific expression patterns.

Under this formulation, the mask predictor is extended from the unconditional distribution $p_\theta(x_0 \mid x_t)$ to a conditional distribution $p_\theta(x_0 \mid x_t, c)$, where $c$ is the concatenation of the two condition tokens.

Training with control cells. A subtle but important training detail: the conditional model is trained on both perturbed cells and control cells. Cells without a perturbation-specific target (such as those exposed to non-targeting control in CRISPR experiments, or PBS in cytokine experiments) are assigned a biologically neutral control label (e.g., "cnt" for the perturbation token) rather than being treated as having a missing condition. The paper states that this allows the model to "learn both perturbation-specific and control-state generation within a single parameterization" (Section 4.4).

What this achieves: By seeing both $p_\theta(x \mid \text{A549}, \text{CHMP3})$ and $p_\theta(x \mid \text{A549}, \text{cnt})$ during training, the model learns the difference between perturbed and control distributions for each cell line. This distinction is essential for classifier-free guidance at inference, which requires the model to evaluate the same input under both the target condition and the control condition.

To prevent the conditional model from being overwhelmed by control cells (which are typically more numerous and less variable than perturbed cells), the proportion of control cells is capped: for genetic perturbation prediction, the control ratio is capped at 10%, and for cytokine perturbation prediction, the same ratio is applied. The paper reports resulting training set sizes in Table C4: for genetic perturbation, 221,273 H1 perturbed cells + 323,913 external perturbed cells + 20,344 sub-sampled H1 control cells + 35,990 sub-sampled external control cells; for cytokine perturbation, 6,499,077 perturbed cells + 629,701 PBS control cells.

Classifier-free guidance (CFG). At inference time, the paper employs CFG (Ho and Salimans, 2021) to strengthen the perturbation-specific signal in the generated expression profiles. The core idea is straightforward: at each denoising step, the model evaluates the partially masked sequence under both the target condition $c$ (e.g., "A549 + CHMP3 knockout") and the control condition $c_{\text{cnt}}$ (e.g., "A549 + non-targeting control"), then amplifies the difference between these two predictions.

Let $a_\theta(v \mid x_t, c)$ denote the unnormalized logits (pre-softmax scores) for vocabulary token $v$ at a masked position under condition $c$. The guided logits are computed as:

a~θ(v∣xt,c)=aθ(v∣xt,ccnt)+(w+1)⋅(aθ(v∣xt,c)−aθ(v∣xt,ccnt))\tilde{a}_\theta(v \mid x_t, c) = a_\theta(v \mid x_t, c_{\text{cnt}}) + (w + 1) \cdot \left(a_\theta(v \mid x_t, c) - a_\theta(v \mid x_t, c_{\text{cnt}})\right)

where $w \geq 0$ is the guidance scale, $a_\theta(v \mid x_t, c)$ is the unconditional/conditional logit under the target condition, and $a_\theta(v \mid x_t, c_{\text{cnt}})$ is the conditional logit under the control condition.

What this computes: For each token in the vocabulary, the guidance step computes how much more (or less) likely that token is under the perturbation condition compared to the control condition, then amplifies this difference by a factor of $(w + 1)$. When $w = 0$, the formula reduces to $\tilde{a}_\theta = a_\theta(v \mid x_t, c)$—pure conditional generation. As $w$ increases, the perturbation-specific signal is increasingly emphasized relative to the cell-line background distribution.

Why this form: The $(w + 1)$ formulation (rather than a simple $w$) follows the original CFG convention and ensures that $w = 0$ recovers standard conditional sampling. The subtraction $a_\theta(c) - a_\theta(c_{\text{cnt}})$ effectively cancels out expression patterns that are common to both the perturbed and control states (e.g., cell-line-specific baseline expression of housekeeping genes, donor-specific expression patterns), amplifying only the features that distinguish the perturbation response. This is particularly important in scRNA-seq data, where baseline cell-type and donor variability often dominates the relatively subtle transcriptional changes induced by perturbations.

The guided logits are converted to sampling probabilities via a softmax at each denoising step. The paper sweeps guidance weights and finds that $w = 2$ achieves the best overall performance for genetic perturbation prediction (Figure 3d-e), while $w = 3$ is used for cytokine perturbation prediction. The ablation study (Figure 3d) shows that removing CFG ("CFG = 0") leads to poorer performance on perturbation direction similarity (DES) and correlation-based metrics (Spearman LFC, Pearson-Δ), consistent with CFG's role in biasing generation toward the perturbed expression manifold.

Inference-Time Biological Prior Injection

Even with classifier-free guidance, perturbation response prediction is challenging because perturbation-induced transcriptional changes are often subtle—a knockout of a single transcription factor might shift the expression of a few hundred target genes by small amounts, against a background of high cell-to-cell variability in the remaining ~17,800 genes. The paper introduces a lightweight mechanism to inject biologically informed prior knowledge into the generation process, using perturbation profiles from external cell lines.

Motivation. The key insight is that while the quantitative transferability of perturbation responses across cell lines is imperfect (a gene's knockdown effect in K562 may differ from its effect in A549), the direction of effect (upregulation vs. downregulation) for many target genes is often conserved. If a perturbation consistently downregulates gene X across multiple cell lines, it is likely to downregulate gene X in a new cell line as well, even if the magnitude of the effect differs. The paper leverages this directional information to initialize the generation process with a prior set of downregulated genes.

Prior gene identification. For each perturbation target in the test set, the paper identifies a set of downregulated genes $\mathcal{G}_\downarrow$ using differential expression analysis on external cell-line Perturb-seq datasets. The procedure (detailed in Appendix C.3) uses the pdex Python package and applies the following criteria to retain a gene $g$ for the prior:

FDR<0.05\text{FDR} < 0.05 ∣target_mean−reference_mean∣>1|\text{target\_mean} - \text{reference\_mean}| > 1 ∣log⁡2(fold_change)∣>1|\log_2(\text{fold\_change})| > 1 reference_mean>2\text{reference\_mean} > 2 target_mean<1\text{target\_mean} < 1

where $\text{target\_mean}$ is the mean expression in perturbed cells, $\text{reference\_mean}$ is the mean expression in non-targeting control cells, FDR is the false-discovery-rate-adjusted p-value from a Mann–Whitney U test, fold_change is the ratio of target_mean to reference_mean.

What these criteria enforce: The gene must be (1) statistically significantly differentially expressed (FDR < 0.05), (2) show a large absolute effect size (difference > 1, |log2 fold change| > 1), (3) be expressed at a reasonably high level in the control condition (reference_mean > 2, ensuring the downregulation signal is not just noise from a barely-expressed gene), and (4) be strongly downregulated in the perturbed condition (target_mean < 1, ensuring the gene is essentially turned off by the perturbation). The criteria also include the perturbed target gene itself in $\mathcal{G}_\downarrow$ (the gene being knocked out should obviously be among the downregulated genes).

When the same perturbation target appears in multiple external cell lines (e.g., CHMP3 knockdown might be present in K562, A549, and HeLa datasets), the gene sets are merged by taking the union across all available cell lines, producing a unified prior $\mathcal{G}_\downarrow$ for that perturbation. The union operation is chosen (rather than intersection) to maximize the number of prior genes and provide a stronger initial signal, at the risk of including some false positives that are cell-line-specific.

Prior injection at generation time. At the start of the reverse diffusion process, rather than initializing all positions to MASK, the system sets the initial expression value to a low constant $\mu = 1$ for all genes $g \in \mathcal{G}_\downarrow$:

x~g={μ,g∈G↓,MASK,g∉G↓.\tilde{x}_g = \begin{cases} \mu, & g \in \mathcal{G}_\downarrow, \\ \text{MASK}, & g \notin \mathcal{G}_\downarrow. \end{cases}

This initial vector $\tilde{x}$ is then mapped to the discrete token space via the quantization function $q(\cdot)$ to obtain the starting state for reverse diffusion. The prior-specified positions are kept fixed during sampling—the model generates expression values for all MASK positions, but the downregulated genes retain their initial low expression value throughout the entire generation process.

What this achieves: The prior injection effectively "seeds" the generation with a directional signal: genes in the prior set are forced to have low expression, providing a scaffold that the model can build upon when generating the remaining genes. The model's task shifts from unconditional generation of the full transcriptome to conditional generation where some positions are already filled with biologically motivated values. This is qualitatively similar to inpainting in image diffusion models, where known pixel values are kept fixed while the model fills in the missing regions.

Why this form and not direct transfer of expression values: The paper deliberately uses a simple binary prior (low expression for downregulated genes) rather than attempting to transfer exact expression values from external cell lines. The justification is implicit: "This strategy provides a directional downregulation signal from other cell lines without assuming direct quantitative transferability across cell systems" (Section 4.5). The paper acknowledges that perturbation effect sizes vary across cell lines, and attempting to impose exact expression levels from K562 onto A549 would likely be counterproductive. By only encoding the direction (downregulation) and a minimal expression level, the prior is conservative—it tells the model which genes should be low, but leaves the model free to determine the exact value and the correlated expression of other genes based on the training data for the target cell line.

Ablation results. The ablation study (Figure 3g, labeled as "biological prior projection" in the text but "prior injection" in the figure) shows that incorporating biological priors improves perturbation direction similarity (DES) and Pearson-Δ correlation. The improvements are modest but consistent, and since the prior injection is a zero-cost addition at inference time (it uses precomputed priors from external datasets and does not require additional model evaluations), it represents a pure Pareto improvement over the CFG-only baseline.

Summary of Design Choices and Their Justifications

  • MDDM over AR models: avoids imposing an artificial gene generation order, respecting the permutation invariance of gene expression data.
  • MDDM over continuous DDPMs: masking operation preserves discrete token structure and sparsity at all corruption levels, avoiding the mismatch between additive Gaussian noise and sparse count data.
  • Logarithmic adaptive quantization over naive binning: preserves exact counts at low expression (0–99) where every transcript matters, while progressively coarsening resolution at high counts where exact values are less informative and vocabulary compression is needed.
  • Random permutation + linear compression over raw sequence modeling: reduces effective sequence length by 8× or 32× for computational tractability, while the random grouping forces the model to learn semantic gene-gene dependencies rather than relying on positional adjacency.
  • Condition tokens as prepended vocabulary extensions over separate encoders: unifies all information in a single discrete sequence processed by a single Transformer, simplifying the architecture and allowing self-attention to directly model condition-expression interactions.
  • Classifier-free guidance over classifier-based guidance: eliminates the need for a separately trained perturbation classifier, and the CFG formulation naturally handles the high-dimensional, structured output space of transcriptomic profiles.
  • Binary directional prior over exact value transfer: conservatively encodes the biological knowledge that a perturbation downregulates specific genes without making assumptions about the exact magnitude of the effect across cell lines.

4. Key Insights and Innovations

Innovation 1: Recasting Single-Cell Foundation Models as Generative World Models, Not Static Encoders

The paper's most consequential conceptual move is to redefine what success looks like for a single-cell transcriptomic model. Prior foundation models—such as scGPT, Geneformer, scFoundation, and CellFM—were designed and evaluated primarily on their ability to produce useful static representations: embeddings that could be transferred to downstream tasks like cell-type classification, batch integration, or perturbation-response prediction viewed as a regression problem. The paper argues that this framing, while productive, fundamentally limits the scope of what computational models can contribute to biology. A representation can tell you that two cells are similar; it cannot tell you what a hypothetical cell—one that has not been observed—would look like.

Lingshu-Cell shifts the goal from learning embeddings to learning a generative model of the transcriptomic state distribution itself. This is not merely a change in architecture or training objective; it is a reframing of the model's relationship to biological reality. Under the static-representation paradigm, the model is an encoder of existing data. Under the generative paradigm, the model is a simulator of possible cellular states. The paper's language—"cellular world model," "in silico experiments," "virtual cells"—reflects this shift from description to simulation. This mirrors a broader trend in machine learning (world models in reinforcement learning, video generation models in computer vision) where the field has recognized that being able to predict "what would happen if" is qualitatively more powerful than simply extracting features from "what already happened."

What distinguishes this from prior generative work in single-cell biology (such as scVI and scDiffusion) is that Lingshu-Cell frames generation not as a secondary capability or a reconstruction auxiliary objective, but as the primary function of the model. The unconditional generation experiments demonstrate that the model can produce de novo cell populations with realistic cell-type proportions and marker-gene patterns (Figure 2a-b, Appendix Figure C2), density of inter-batch and cross-species expression distributions (Figure 2d-e, Table 1). The conditional generation experiments show that the same model, without architectural modification, can simulate the transcriptomic outcome of applying a perturbation to a specific cell type or donor (Figure 3, Figure 4). This unified generative capability—where a single trained model can both sample from the steady-state distribution and simulate interventions—is what the paper means by a "world model," and it represents a fundamentally different ambition than prior work.

The significance of this reframing extends beyond the paper's empirical results. By positioning the model as a generative simulator, the paper opens a path toward closed-loop experimentation: a cycle where model predictions guide which perturbations to test experimentally, newly collected data refine the model, and the improved model generates better predictions—moving from static data fitting toward adaptive biological discovery. The paper explicitly identifies this as an ultimate goal in the Discussion, though it does not implement it. This framing also creates a clean conceptual framework for evaluating model capabilities: does the model faithfully reproduce known biological patterns (a minimum bar, tested via distributional metrics), and does it make accurate predictions for unseen interventions (the true test of a simulator, tested via perturbation benchmarks)?

Innovation 2: The Discrete Diffusion Paradigm as an Alignment of Computational Assumptions with Biological Structure

The paper's choice of masked discrete diffusion is not simply one generative architecture among many; it is a deliberate, principled alignment between the structural properties of the generative process and the structural properties of single-cell transcriptomic data. This alignment operates at three levels that collectively distinguish Lingshu-Cell from prior generative approaches.

Permutation invariance without order bias. Autoregressive models—including the GPT-style architecture used in scGPT—must impose a generation order on genes, creating an inductive bias that has no biological justification. Gene expression profiles are not sequences; there is no natural reason to predict gene A before gene B. MDDMs, by masking all positions independently and predicting all masked positions simultaneously, respect the permutation invariance of the gene set. The model's self-attention is bidirectional, meaning each gene's predicted expression can depend on all other genes regardless of their position in the arbitrary input ordering. This is a fundamental architectural choice, not a minor tweak: it means the model learns gene-gene dependencies based on their semantic relationship (co-expression patterns, regulatory connections) rather than on their positional adjacency in the input vector. The paper explicitly contrasts this with AR models in Figure 1b, showing that Lingshu-Cell's masking pattern treats all genes symmetrically.

Discrete token space aligned with sparse count data. Continuous diffusion models (DDPMs) operate by adding Gaussian noise to all positions and learning to reverse this process, but scRNA-seq data are fundamentally discrete—counts are non-negative integers with a massive spike at zero. The forward noising process in a DDPM produces intermediate states that look nothing like real scRNA-seq data: continuous-valued vectors where every gene has a fractional expression level, the opposite of the sparse, mostly-zero structure of actual transcriptomes. The paper's masking approach avoids this mismatch entirely. By replacing expression values with a special MASK token (rather than adding noise), the MDDM preserves the discrete token structure at every corruption level. A partially masked scRNA-seq profile—where some genes retain their original expression values (including zeros) and others are MASK—looks structurally similar to a real profile with missing measurements. This means the model's training distribution and inference distribution are qualitatively aligned, which the paper argues is critical for generative quality (Section 3).

Non-autoregressive, bidirectional refinement without bottleneck compression. Variational autoencoders like scVI compress each cell's expression profile through a low-dimensional continuous latent bottleneck, which creates an information-theoretic tension between representation compactness and reconstruction fidelity. MDDMs avoid this bottleneck entirely: the mask predictor operates on the full (or linearly compressed) sequence and predicts all masked positions in parallel, without any forced dimensionality reduction. The model can attend to long-range gene-gene dependencies directly through self-attention, without needing to route information through a compressed latent code. This is an architectural commitment to capacity over compression—it assumes that the computational budget is better spent on rich, direct modeling of token-level interactions than on learning an efficient latent encoding.

This alignment between computational assumptions and biological structure is not merely aesthetically pleasing; it has measurable consequences. The unconditional generation experiments (Figure 2c, Table 1) show that Lingshu-Cell achieves substantially better distributional fidelity than both scVI and scDiffusion across MMD, 1-WD, and iLISI metrics—the metrics most directly related to whether the generated data "look like" real scRNA-seq data. The gap is particularly notable for MMD, where Lingshu-Cell's score of 0.0088 is roughly 2× better than scDiffusion's 0.0178 and nearly 4× better than scVI's 0.0343 (Figure 2c). These are not modest improvements; they suggest that the discrete diffusion alignment captures distributional properties that continuous alternatives systematically miss.

Innovation 3: Sequence Compression via Random Permutation + Linear Projection as a Way to Escape the Gene-Ordering Trap

The paper's embedding-space sequence compression module addresses a pragmatic challenge—how to make self-attention over ~18,000 genes computationally tractable—but its design encodes a deeper conceptual insight about how gene-gene dependencies should be learned. The standard approach to reducing sequence length in Transformers would be to apply 1D convolutions, patch-based grouping, or learned pooling over the original gene order. Any of these would create a compression that is sensitive to the arbitrary ordering of genes in the input vector—genes that happen to be adjacent in the canonical ordering would be compressed together, potentially creating spurious correlations.

Lingshu-Cell's approach is counter-intuitive but principled: first randomize the gene order, then compress. The random permutation π is fixed at initialization and applied identically to every cell, so it is not learned. This means the compression groups contain a random mix of genes from across the genome, with no biological or artifactual basis for their co-occurrence in a group. The linear projection from the concatenated embedding of S genes down to a single compressed token therefore forces the model to learn a mixing representation—a weighted combination of expression signals from unrelated genes—that carries no positional shortcut for learning gene-gene dependencies. Any meaningful interaction between genes must be learned through the Transformer's self-attention mechanism operating on the compressed tokens, not inherited from the input structure.

The conceptual significance is that this design eliminates positional inductive bias at the compression stage, forcing the model to build its understanding of co-expression relationships entirely from data rather than from the input representation's structure. This is philosophically aligned with the MDDM's broader commitment to respecting gene-order invariance: just as the masking process treats all positions symmetrically, the compression treats all genes as interchangeable for the purpose of creating compressed tokens.

The ablation study (Figure 3f) provides empirical evidence that moderate compression actually improves performance: a patch size of 32 (compression ratio 32:1) outperforms both a patch size of 8 (8:1) and no compression for genetic perturbation prediction, particularly on the Spearman #DEG metric (0.405 vs. 0.292). The paper's interpretation is that compression "reduc[es] the influence of single-gene noise and improv[es] robustness under perturbation" (Section 4.3). This suggests a non-obvious dynamic: by mixing gene signals before the Transformer processes them, the compression acts as a form of structured denoising that prevents the model from overfitting to the expression value of any individual gene, instead encouraging it to learn broader expression patterns across gene modules. This is a rare case where a computational efficiency measure also improves model quality, and the mechanism—random permutation destroying spurious positional correlations—is conceptually distinct from standard regularization techniques.

Innovation 4: Difficulty-Agnostic External Prior Injection as a Conservative Bridge Between Data-Driven and Knowledge-Driven Perturbation Prediction

The paper's approach to incorporating biological prior knowledge at inference time represents a thoughtful negotiation between two competing pressures in perturbation modeling. On one hand, perturbation responses are cell-line-specific: the transcriptional effect of knocking out gene X in K562 cells can differ substantially from the effect in A549 cells, both in the identity of responsive genes and the magnitude of expression changes. On the other hand, the direction of effect—whether a gene is up- or down-regulated—often shows conservation across cell lines, because it reflects the fundamental regulatory logic of the pathway rather than cell-type-specific quantitative tuning. Directly transferring expression values from one cell line to another would be overly aggressive, imposing quantitative assumptions that violate cross-line heterogeneity. Ignoring available perturbation data from other cell lines entirely would be overly conservative, discarding directional information that could guide prediction.

The paper's solution is a binary directional prior: for each perturbation target, identify genes that are consistently downregulated across available external cell lines (using stringent differential expression criteria: FDR < 0.05, large effect sizes, strong downregulation in the perturbed condition), and initialize those genes' expression to a minimal value at the start of generation. This prior is lightweight, conservative, and interpretable. It tells the model: "these genes should be low"—a directional claim that is likely to generalize across cell lines—but does not specify exactly how low, or what the correlated expression of other genes should be. The model remains responsible for determining the quantitative expression values and for predicting the full transcriptomic context, using its learned knowledge of the target cell line's baseline expression patterns and regulatory network structure.

What makes this approach intellectually distinctive is that it bridges the gap between purely data-driven prediction (where the model must learn everything from the training distribution) and knowledge-driven prediction (where external biological knowledge is imposed as a hard constraint), without falling into the trap of assuming cross-cell-line transferability of exact expression values. The prior is injected only at initialization and remains fixed throughout generation, meaning it acts as a scaffold that the model builds upon rather than a constraint that overrides the model's learned distribution. This is conceptually analogous to inpainting in image generation, where known pixel values are held fixed and the model fills in the rest, but adapted to the specific structure of perturbation biology where the "known" information is directional rather than quantitative.

The ablation study (Figure 3g) shows that prior injection improves perturbation direction similarity (DES) and Pearson-Δ correlation relative to the CFG-only baseline, with the improvements being consistent if not dramatic. The modest magnitude of the gain is itself informative: it suggests that the model is already learning meaningful perturbation-specific signals from the training data, and the prior provides a complementary signal that particularly helps with the directional components of the prediction. This is exactly what one would expect from a conservative prior—it shouldn't dominate the prediction when the training data already provides strong signal, but it should help in cases where the perturbation effect is subtle or where the training data for a particular condition is sparse. The fact that the prior is a zero-cost addition at inference time (it uses precomputed external knowledge and requires no additional model evaluations) makes even these modest gains a pure Pareto improvement, with no tradeoff to weigh.

5. Experimental Analysis

Evaluation Methodology

  • Dataset. The paper uses three distinct data regimes. For unconditional generation, the primary dataset is the PARSE 10M PBMC dataset (Biosciences, 2024)—specifically its PBS control subset comprising 629,701 cells spanning five major immune lineages (T cells, NK cells, B cells, monocytes, dendritic cells). For cross-tissue unconditional evaluation, the paper assembles 2,602,318 cells from eight human tissues (neocortex, thymus, heart, lung, liver, colon, kidney, breast) from the CZ CELLxGENE database (Program et al., 2025), plus 247,899 cells across four non-human species (mouse ovary, rhesus macaque lung, zebrafish embryo, fly brain) from CZ CELLxGENE and scBaseCount (Youngblut et al., 2025). For conditional generation, the primary benchmark is the Virtual Cell Challenge H1 genetic perturbation dataset (Roohani et al., 2025), partitioned into 150 training targets, 50 validation targets, and 100 test targets at the perturbation level, supplemented with external Perturb-seq datasets from Replogle et al. (2022), Nadig et al. (2025), Jiang et al. (2025), and Huang et al. (2025) for perturbations whose targets overlap the VCC 300-gene set. For cytokine perturbation, the paper uses the full PARSE 10M PBMC dataset (12 donors × 90 cytokine conditions + PBS control), with a donor-level split where 4 donors are held out for testing and 70% of cytokine conditions per test donor are unseen during training.

  • Base model(s). All experiments use the same Lingshu-Cell architecture: a bidirectional LLaMA-style Transformer with 13 blocks, embedding dimension 640, 10 attention heads (head dimension 64), SwiGLU feed-forward with intermediate dimension 2,560, RMSNorm pre-normalization, and RoPE positional encoding. The model is trained separately for each experimental setting (PARSE PBS, per-tissue, genetic perturbation, cytokine perturbation) with different data compositions but identical architecture (Table C3). The model is trained from scratch for each setting—there is no pretraining phase followed by fine-tuning, meaning each model variant independently learns its task-specific distribution.

  • Metrics. For unconditional generation, the paper uses five complementary metrics computed on log1p-normalized expression values: Pearson correlation and Spearman correlation between gene-averaged mean expression vectors of real and generated populations (measuring gene-level expression concordance); Maximum Mean Discrepancy (MMD) computed in a joint PCA embedding space with a multi-scale Gaussian kernel (measuring distributional distance—lower is better); gene-averaged 1-Wasserstein distance (1-WD) computed independently per gene (measuring per-gene marginal distribution matching—lower is better); and integration Local Inverse Simpson's Index (iLISI) on a joint PCA embedding (measuring mixing of real and generated cells—closer to 2.0 is better). For conditional perturbation prediction, the paper adopts the seven-metric panel from the Virtual Cell Challenge implemented in the Cell-Eval toolkit (Adduri et al., 2025), all computed on pseudobulked expression profiles averaged over cells sharing the same perturbation condition: Mean Absolute Error (MAE) between predicted and observed pseudobulks; Pearson Delta Correlation (Pearson-Δ) between predicted and observed expression deltas; Perturbation Discrimination Score (PDS) measuring whether predictions can distinguish different perturbation effects; DE Overlap Accuracy (DES) measuring overlap between top-N differentially expressed genes in prediction and ground truth; Spearman #DEG Correlation measuring rank correlation between the number of significant DE genes per perturbation; Spearman LFC Correlation measuring rank correlation of log-fold changes restricted to ground-truth DE genes; and AUPRC measuring precision-recall for DE gene identification. An average score aggregates all seven metrics as the arithmetic mean of relative improvements over the cell-mean baseline (which predicts each perturbation as the control pseudobulk).

  • Baselines. For unconditional generation, the paper compares against scDiffusion (Luo et al., 2024), a continuous diffusion model for single-cell data, and scVI (Lopez et al., 2018), a variational autoencoder that models counts with a zero-inflated negative binomial distribution. For genetic perturbation prediction, baselines include the top-performing entries on the VCC generalist leaderboard (Table 2, Table B1): cleopatra, xBio, Cellock Holmes, Shippers, Mean Predictors, and 20 additional teams listed in full in Table B1. These represent a diverse mix of methods including direct regression approaches, flow matching, and foundation-model-based predictors, though the paper does not provide architectural details for these external entries. For cytokine perturbation prediction (Figure 4c), baselines include PerturbMean (predicting the perturbation pseudobulk as the average of training pseudobulks for that condition), STATE (Adduri et al., 2025), scGPT (Cui et al., 2024), and scVI (Lopez et al., 2018).

  • Generation budget / compute accounting. The paper measures generation in terms of number of reverse diffusion steps (N = 256 for unconditional generation, N = 3 for conditional generation) rather than in FLOPs or wall-clock time. For sequence compression, the effective sequence length is reduced from G = 18,080 to Gc = G/S, where S = 8 (default) or S = 32 (genetic perturbation prediction). Training compute is reported in GPU counts and learning rate schedules (Table C3) but not in total FLOPs. Critically, the cost of difficulty estimation or prior gene identification is not included in any compute budget, nor is there a FLOPs-matched comparison between different model scales or architectures as in the Section 7 analysis of the reference paper.

  • Cross-validation / statistical protocol. The paper does not employ cross-validation for model selection or evaluation in the conventional sense, but does use held-out evaluation sets consistently. For unconditional generation, evaluation is performed by generating cells from the trained model and comparing to held-aside real data (no overlap between training and evaluation cells, though the exact split is not specified as a formal cross-validation). For genetic perturbation prediction, the VCC H1 dataset provides a predefined split of perturbation targets (150 train, 50 validation, 100 test), and evaluation on the test set follows the same protocol as the VCC leaderboard. For cytokine perturbation prediction, a donor-level split is used: 4 of 12 donors are held out for testing, with 70% of cytokine conditions per test donor unseen during training; two additional donors serve as validation with 30% held-out conditions; the remaining six donors form the training set. For the unconditional generation metrics, the paper does not report confidence intervals or statistical significance tests—the scores are presented as point estimates without variance estimates. Error bars are shown only for the cytokine perturbation results (Figure 4c), where interquartile ranges across the four test donors are displayed but formal hypothesis tests are not reported.

Main Quantitative Results

Unconditional Generation Across Human PBMCs

The paper first validates Lingshu-Cell's fundamental generative capacity on the PARSE 10M PBMC dataset (PBS control subset, 629,701 cells). The headline result is that Lingshu-Cell faithfully recapitulates both the cell-type structure and the transcriptomic distribution of the real data.

Cell-type proportions and marker genes (Figure 2a-b). When generating 10,000 cells (a scale described as "comparable to that of a typical scRNA-seq experiment"), the UMAP visualization shows that generated cells overlap closely with real cells across all five major PBMC lineages: T cells, NK cells, B cells, monocytes, and dendritic cells (Figure 2a, left panel). More specifically, the normalized log1p expression of canonical marker genes—CD3D for T cells, KLRF1 for NK cells, CD79A for B cells, CD68 for monocytes, FLT3 for DCs—shows highly concordant expression patterns between real and generated populations (Figure 2a, right panel). This is not a quantitative correlation metric but a visual check that the model preserves the biological signature of each cell type. The cell-type proportions in the generated population closely match those in the real data (Figure 2b), with all five major lineages represented at similar frequencies.

To assess robustness to generation scale, the paper scales up to 200,000 generated cells. At this larger scale, both marker-gene expression patterns (Appendix Figure C2a) and cell-type proportions (Appendix Figure C2b) remain highly concordant. Critically, the paper performs higher-resolution annotation on this larger population, subdividing PBMCs into 17 subtypes including CD8+ Naive T, CD8+ Memory T, CD4+ Naive T, CD4+ Memory T, MAIT, Treg, NKT, NK, NK CD56bright, CD14+ Monocyte, CD16+ Monocyte, Naive B, Memory B, Plasmablast, cDC, pDC, and HSPC (Appendix Figure C2c-d). The generated data "continued to align closely" with real data at this subtype resolution, indicating that Lingshu-Cell captures not just broad lineage distinctions but fine-grained cellular heterogeneity.

Quantitative benchmark comparison (Figure 2c). The paper benchmarks Lingshu-Cell against scDiffusion and scVI on the PBMC dataset across five metrics:

MetricLingshu-CellscDiffusionscVI
Pearson (↑)0.99930.99900.9993
Spearman (↑)0.99630.98910.9962
MMD (↓)0.00880.01780.0343
iLISI (↑)0.90950.89660.7429
1-WD (↓)0.00640.15940.0102

Several patterns are noteworthy. All three methods achieve "uniformly high gene expression correlations" (Pearson and Spearman both > 0.98 for all methods), suggesting that capturing global gene-level mean expression patterns is relatively easy and not discriminative. The distributional fidelity metrics—MMD, 1-WD, and iLISI—reveal "clearer differences in generative quality." Lingshu-Cell achieves the lowest MMD (0.0088), which is approximately 2× lower than scDiffusion (0.0178) and nearly 4× lower than scVI (0.0343). The paper interprets this as "indicating the closest overall match between generated and real expression distributions." For 1-WD, Lingshu-Cell achieves 0.0064, while scDiffusion performs substantially worse at 0.1594—a ~25× gap—and scVI achieves 0.0102. The iLISI metric shows Lingshu-Cell (0.9095) achieving the best mixing between real and generated cells, with scDiffusion (0.8966) close behind and scVI (0.7429) notably worse.

Lingshu-Cell is the only method to achieve the best score across all five metrics, though the margins on Pearson and Spearman correlations are extremely narrow (0.9993 vs. 0.9990 for Pearson, 0.9963 vs. 0.9962 for Spearman), suggesting these metrics are saturated and not useful for discriminating between strong models.

Unconditional Generation Across Human Tissues

The paper extends evaluation to eight human tissues from CZ CELLxGENE, totaling 2,602,318 cells with substantial heterogeneity in quality-control metrics (Appendix Figure C3 shows large variation across tissues in cell numbers, detected genes, total counts, and mitochondrial read fractions). A separate Lingshu-Cell model is trained for each tissue, meaning the model architecture is identically deployed on different data distributions without transfer learning.

Tissue-specific cell type capture (Figure 2d, Appendix Figure C4, Table 1). The UMAP visualizations for neocortex, heart, lung, and colon (Figure 2d) and for thymus, kidney, breast, and liver (Appendix Figure C4) show that across all eight tissues, generated cells recapitulate the major cell types and tissue-specific populations. For example, the heart model generates cardiomyocytes, endothelial cells, perivascular/smooth muscle cells, fibroblasts, T & NK cells, myeloid cells, and mesothelial cells; the neocortex model generates excitatory neurons, inhibitory neurons, astrocytes, oligodendrocyte precursor cells, microglial cells, and oligodendrocytes; the liver model generates hepatoblasts, hepatic stellate cells, and endothelial cells alongside immune populations.

Quantitative metrics across the four tissues highlighted in Figure 2d are reported in Table 1:

TissuePearson (↑)Spearman (↑)MMD (↓)iLISI (↑)1-WD (↓)
Neocortex0.99950.99910.01280.90530.0105
Heart0.99920.99870.01960.89720.0096
Lung0.99670.99700.03140.89060.0159
Colon0.99660.99600.03760.88150.0152

All tissues achieve Pearson and Spearman correlations above 0.996, but the MMD values show meaningful variation: neocortex (0.0128) shows the best distributional match, while colon (0.0376) shows the worst. The paper notes that this variation likely reflects tissue-specific differences in data quality and cellular complexity—colon tissue contains more diverse epithelial and immune subtypes that may be harder to model than the relatively discrete cell types in neocortex. The iLISI values (0.88–0.91) indicate strong mixing across all tissues, suggesting that the generated cells are not mode-dropping or collapsing to a subset of the real cell types.

Unconditional Generation Across Non-Human Species

The paper further evaluates Lingshu-Cell on four non-human species spanning mouse (ovary), rhesus macaque (lung), zebrafish (embryo), and fly (brain), totaling 247,899 cells (Table 1, Figure 2e). These datasets exhibit "pronounced differences in quality-control metrics and data distributions" (Appendix Figure C5), including differences in gene detection rates, total counts, and the fraction of mitochondrial reads.

Species - TissuePearson (↑)Spearman (↑)MMD (↓)iLISI (↑)1-WD (↓)
Mouse - Ovary0.99960.99890.01160.90110.0077
Rhesus macaque - Lung0.99850.99700.02180.89260.0149
Zebrafish - Embryo0.99830.99740.01430.90350.0089
Fly - Brain0.99840.99290.01630.88760.0107

Mouse ovary achieves the best performance across most metrics (Pearson 0.9996, MMD 0.0116, 1-WD 0.0077), while rhesus macaque lung shows the weakest distributional match (MMD 0.0218). The UMAP visualizations (Figure 2e) confirm that the model captures species- and tissue-specific cell types—oocytes, granulosa cells, and theca cells in mouse ovary; endothelial, perivascular/SMC, erythroid, and epithelial cells in rhesus macaque lung; neurons, glia, muscle, and mesoderm cells in zebrafish embryo; and the diverse neuronal and glial populations of fly brain. These results are notable because the gene set is standardized to the same 18,080-gene reference used for human data, meaning that for non-human species, only orthologous genes are retained and all other positions in the sequence are effectively missing—a substantial domain shift that the model handles without architectural modification.

Genetic Perturbation Prediction on the VCC H1 Benchmark

Training and evaluation setup. For genetic perturbation prediction, Lingshu-Cell is trained on a combined dataset comprising H1 cells from the 150 VCC training perturbation targets (183,097 cells), external Perturb-seq data from other cell lines for perturbations overlapping the VCC 300-gene set (323,913 perturbed cells + 35,990 sub-sampled control cells), and H1 unperturbed control cells sub-sampled to a 10% ratio (20,344 cells; Table C4). Evaluation is performed on H1 cells from 50 validation targets (60,751 cells) and 100 test targets (132,670 cells) that are held out during training. The model generates predictions using N = 3 reverse diffusion steps with classifier-free guidance (weight w = 2), sequence compression with patch size S = 32, and biological prior injection.

Ablation studies of design components (Figure 3). Before presenting the full comparison against the VCC leaderboard, the paper conducts three ablation experiments to quantify the contribution of individual design choices:

Classifier-free guidance (CFG) weight sweep (Figure 3d). The paper sweeps CFG weights w ∈ {0, 1, 2} and reports performance across all seven perturbation metrics plus the average score. At w = 0 (no guidance), the model already achieves non-trivial performance—showing that the conditional model learns perturbation-specific signals even without guidance. At w = 1, performance improves across most metrics, and at w = 2, the model achieves the best overall performance and average score. The paper does not report sweeps beyond w = 2, so it is unknown whether further increases would improve or degrade performance (over-guidance is a known phenomenon in image diffusion). The most notable improvements from CFG are on perturbation direction similarity (DES) and correlation-based metrics (Spearman LFC, Pearson-Δ), consistent with CFG's role in amplifying perturbation-specific signal relative to the shared cell-line background.

Sequence compression patch size sweep (Figure 3f). Comparing no compression ("No patchify") against patch sizes of S = 8 and S = 32, the paper finds that S = 32 achieves the best overall performance. Specifically, the Spearman #DEG correlation reaches 0.405 with S = 32 compared to 0.292 with S = 8, representing a ~39% improvement on this metric. The average score also improves from S = 8 to S = 32. The paper interprets this as evidence that "moderate compression improves representation of high dimensional gene expression signals," with the larger compression ratio forcing the model to learn more robust, module-level expression patterns rather than overfitting to individual gene noise. The ablation confirms that a 32× compression ratio is not merely computationally convenient but actually beneficial for prediction quality.

Biological prior injection (Figure 3g). Comparing the model with and without inference-time prior injection, the paper finds improvements in DES (perturbation direction similarity) and Pearson-Δ correlation. The effects on other metrics (PDS, MAE, Spearman #DEG, Spearman LFC, AUPRC) are more modest. The paper does not quantify the exact numerical improvement for each metric in the text—the comparison is presented visually as bar plots in Figure 3g—but the pattern is consistent: the prior provides a directional signal that primarily benefits metrics measuring the alignment of differential expression patterns.

Full VCC leaderboard comparison (Table 2, Table B1). After integrating all three strategies, Lingshu-Cell is evaluated on the VCC H1 test set and compared against the top-performing methods on the generalist leaderboard. The headline result is that Lingshu-Cell achieves the best average rank (8.7) across all seven metrics when ranked against 25 other entries (Table 2, Table B1):

MetricLingshu-CellNext Best (Team)
DES (↑)0.216 (rank 25/26)0.366 (BioCai)
PDS (↑)0.748 (rank 9)0.872 (BM_xTVCBM)
MAE (↓)0.052 (rank 1)0.084 (LIUtest)
Spearman #DEG (↑)0.394 (rank 12)0.684 (Gavin)
Spearman LFC (↑)0.331 (rank 3)0.396 (cleopatra)
AUPRC (↑)0.272 (rank 10)0.582 (Mean Predictors, BioCai, BM_xTVCBM)
Pearson-Δ (↑)0.306 (rank 1)0.218 (Nathan LaPierre)

Several patterns in the full leaderboard (Table B1) are revealing. First, no single method dominates across all metrics—the best per-metric scores are distributed across seven different teams (BioCai for DES, BM_xTVCBM for PDS, Lingshu-Cell for MAE, Gavin for Spearman #DEG, cleopatra for Spearman LFC, three teams tied for AUPRC, Lingshu-Cell for Pearson-Δ). This suggests that the VCC metrics capture different aspects of perturbation prediction quality that are not perfectly correlated. Second, Lingshu-Cell's average rank of 8.7 is the best among all entries, but the margins over the next-best teams are small: cleopatra achieves 9.1, xBio achieves 10.7, and Cellock Holmes achieves 10.9. The average rank metric is sensitive to the choice of ranking method and to the inclusion of metrics on which Lingshu-Cell performs poorly—notably DES (rank 25 out of 26), where the score of 0.216 is near the bottom of the leaderboard.

Third, Lingshu-Cell's strongest metrics are the transcriptome-wide accuracy measures (MAE rank 1, Pearson-Δ rank 1), while its performance on differential expression recovery metrics is more middling (DES rank 25, Spearman #DEG rank 12, AUPRC rank 10). This pattern suggests that Lingshu-Cell excels at capturing the overall magnitude and direction of genome-wide expression changes but is less effective at identifying the precise set of significantly differentially expressed genes, particularly when measured by overlap with ground-truth DE calls (DES). This may reflect the generative modeling paradigm: the model learns to produce realistic expression profiles that match the distributional properties of perturbed cells, but it may not explicitly optimize for the statistical significance of per-gene differential expression, which is a post-hoc analysis rather than a direct component of the generation objective.

The paper also notes that Lingshu-Cell achieves this performance using "only approximately 0.6 million training cells" (Section 1, referring to the combined H1 + external training set). This is a modest training set size compared to the massive pretraining corpora of general-purpose foundation models, suggesting that the generative approach is data-efficient for perturbation prediction when augmented with biological priors.

Cytokine Perturbation Prediction in Human PBMCs

Training and evaluation setup. For cytokine perturbation, Lingshu-Cell is trained on the full PARSE 10M PBMC dataset (6,499,077 perturbed cells + 629,701 PBS control cells; Table C4) with conditioning on donor identity and cytokine condition. The evaluation tests generalization to unseen donor-cytokine combinations: 4 of 12 donors are held out, and for each test donor, 70% of cytokine conditions (63 of 90) are unseen during training. This is a substantially harder generalization task than the VCC benchmark because it requires the model to predict the transcriptional response of a new donor to a cytokine stimulation it has not observed for that donor (though the model may have seen the cytokine in other donors).

Comparison with baseline methods (Figure 4c). The paper compares Lingshu-Cell against PerturbMean, STATE, scGPT, and scVI across the seven perturbation metrics plus average score. Results are presented as bar plots with donor-level interquartile ranges (error bars) and individual donor points:

For transcriptome-wide accuracy metrics:

  • PDS (Perturbation Discrimination Score): Lingshu-Cell achieves the highest score, indicating that its predicted transcriptomes best preserve the distinct identity of each cytokine condition. The exact numerical values are readable from Figure 4c bar heights but not quoted in the text.
  • Pearson-Δ correlation: Lingshu-Cell ranks first, indicating it most accurately captures the direction and magnitude of cytokine-induced expression changes at the gene level.

For differential expression recovery metrics:

  • Spearman #DEG: Lingshu-Cell achieves the highest correlation, indicating it correctly recovers the relative strength of transcriptional responses induced by different cytokines.
  • DES, Spearman LFC, AUPRC: Lingshu-Cell achieves strong performance—the text states it "also achieved strong performance on DES, Spearman LFC, and AUPRC" without specifying exact ranks relative to baselines.

For overall assessment:

  • Average score: Lingshu-Cell achieves the highest average score across all evaluated methods.
  • MAE: The text does not explicitly claim Lingshu-Cell achieves best MAE in this setting (unlike the VCC benchmark where MAE rank 1 is highlighted).

The error bars in Figure 4c (interquartile range across the four test donors) show substantial donor-to-donor variability—some metrics show wide spreads, indicating that prediction quality varies meaningfully across donors. The paper does not analyze which donor characteristics (age, sex, baseline immune composition) correlate with prediction difficulty. The individual donor points are visible as overlaid scatter on the bar plots, providing transparency about per-donor performance variance, but no statistical tests comparing methods within individual donors are reported.

Comparison with genetic perturbation results. The paper emphasizes a cross-setting consistency: "Although genetic perturbations and cytokine stimulations act through fundamentally different mechanisms, Lingshu-Cell achieved leading performance in both settings (Fig. 3; Fig. 4)." The inference setup differs in details—CFG weight w = 3 for cytokine vs. w = 2 for genetic perturbation, and the cytokine model uses S = 8 (default patch size) rather than S = 32—but the unified architecture handles both perturbation modalities without modification to the conditioning interface or the generative objective. The paper frames this as evidence that "the conditional generation framework [can serve] as a unified approach for predicting cellular responses to diverse perturbations across experimental contexts."

Ablation Studies and Robustness Checks

Classifier-free guidance weight (Figure 3d): Sweeping CFG weight w ∈ {0, 1, 2} on the genetic perturbation VCC test set shows that w = 2 achieves the best average score across all seven perturbation metrics. Removing CFG entirely (w = 0) leads to notably poorer performance on perturbation direction similarity (DES) and correlation-based metrics (Spearman LFC, Pearson-Δ), consistent with CFG's role in steering generation toward the perturbed expression manifold. The paper does not report sweeps beyond w = 2, so it remains unknown whether higher guidance weights would continue to improve or whether over-guidance artifacts (known in image diffusion literature) would emerge at w ≥ 3. The cytokine perturbation experiments use w = 3 without an explicit sweep for that setting.

Sequence compression patch size (Figure 3f): Comparing no compression against S = 8 and S = 32 on the VCC test set, the paper finds that S = 32 outperforms both alternatives on the average score and on Spearman #DEG correlation (0.405 vs. 0.292 for S = 8). The paper does not test intermediate patch sizes (e.g., S = 16, S = 24) or larger sizes (e.g., S = 64, S = 128), so the optimal compression ratio is not precisely characterized. The default unconditional generation models use S = 8, but the paper does not ablate patch size for unconditional generation—the claim that S = 32 is optimal applies specifically to genetic perturbation prediction and may not transfer to other settings.

Biological prior injection (Figure 3g): Adding the inference-time prior (pre-identified downregulated genes initialized to low expression and kept fixed during sampling) improves DES and Pearson-Δ relative to the CFG-only baseline. The effects on other metrics (PDS, MAE, Spearman #DEG, Spearman LFC, AUPRC) are modest or negligible based on the bar plot heights. The paper does not ablate the specific choices in prior construction—the FDR threshold (0.05), the effect size cutoffs (|difference| > 1, |log2 fold change| > 1), or the decision to use union vs. intersection across external cell lines—so the sensitivity of the improvement to prior design is unknown. The prior is also specific to perturbation targets that have been profiled in external cell lines; for perturbation targets unique to the VCC test set that lack external data, the prior cannot be constructed and the model relies solely on CFG.

Generation scale robustness (Appendix Figure C2): Comparing 10,000 vs. 200,000 generated PBMC cells shows that marker-gene expression patterns, cell-type proportions, and even 17-subtype proportions remain highly concordant between generated and real data at both scales. This demonstrates that the model does not suffer from mode collapse or degradation of diversity as the number of generated cells increases—a common failure mode in generative models where large samples reveal repetitive or unrealistic patterns not visible in small samples. The UMAP visualization at the 17-subtype resolution (Appendix Figure C2c) visually confirms that subtypes are well-separated and match the real data structure.

Cross-tissue and cross-species generalization (Table 1, Figure 2d-e, Appendix Figure C4): The paper trains separate models for each tissue and species, so these experiments test the adaptability of the architecture to diverse data distributions rather than generalization across domains from a single model. The consistently strong metrics across all 12 settings (8 human tissues + 4 non-human species) demonstrate that the discrete diffusion framework does not rely on human-specific or tissue-specific properties. However, this does not test the model's ability to transfer knowledge across tissues or species—each model is trained from scratch on its target domain, and there is no experiment where a model trained on one tissue is evaluated on another.

Cell-type resolution sensitivity (Appendix Figure C2c-d): The 17-subtype decomposition of PBMCs confirms that the model captures fine-grained heterogeneity beyond major lineages. This is a non-trivial test because subtypes like MAIT cells, NKT cells, or CD16+ monocytes are relatively rare populations, and a model that collapses all T cells into a single mode or that fails to represent rare subtypes would show discrepancies at this resolution. The subtype proportion comparison (Appendix Figure C2d) visually confirms close alignment across all 17 subtypes, though the paper does not report the proportion values numerically or compute a formal divergence measure.

Comparison of unconditional generation methods (Figure 2c): The five-metric comparison against scDiffusion and scVI provides evidence that the discrete diffusion approach yields better distributional fidelity (MMD, 1-WD) and integration quality (iLISI) than continuous alternatives, while all methods achieve similarly high gene-level correlations. However, this comparison is performed only on the PARSE PBMC dataset—the per-tissue and per-species experiments do not include baseline comparisons, so it is unknown whether the MMD advantage generalizes to other tissues and species.

Critical Assessment

Claim: Lingshu-Cell accurately reproduces transcriptomic distributions, marker-gene expression patterns, and cell-subtype proportions across diverse tissues and species.

The unconditional generation experiments provide strong evidence that Lingshu-Cell can fit the training distribution and produce samples that match population-level summary statistics. The quantitative metrics (Table 1, Figure 2c) show Pearson and Spearman correlations above 0.99 for all settings, MMD values below 0.04, and iLISI values near 0.9—all indicating close alignment between real and generated distributions at the level measured by these metrics. The visual UMAP embeddings and cell-type proportion comparisons (Figure 2, Appendix Figures C2, C4) reinforce this assessment.

However, several limitations constrain what this evidence actually demonstrates about the model's generative fidelity:

First, population-level distribution matching does not guarantee single-cell biological plausibility. The metrics used (MMD, 1-WD, iLISI, gene-level mean correlations) all aggregate over populations of cells. A model that generates biologically implausible individual cells but whose aggregate statistics match the real data would score well on these metrics. For example, a generated cell could have mutually exclusive expression patterns (e.g., simultaneously expressing B-cell and T-cell markers at high levels) while the overall gene-gene correlation structure across the population remains preserved. The paper does not evaluate single-cell-level coherence metrics—such as the fraction of generated cells that express mutually exclusive lineage markers, or the preservation of known gene regulatory network structures within individual cells—that would directly test biological plausibility.

Second, the metrics may be saturated for these model classes. The Pearson correlations of 0.9993–0.9996 are essentially perfect, meaning this metric cannot discriminate between a model that captures all biologically relevant variation and one that merely reproduces the mean expression vector. The MMD and 1-WD values do show variation across methods and settings, but they are sensitive to preprocessing choices (PCA dimensionality, kernel bandwidth, number of generated samples) that the paper does not systematically vary. Without sensitivity analysis or confidence intervals, it is difficult to assess whether the numerical differences (e.g., MMD 0.0088 vs. 0.0178) are statistically or practically meaningful.

Third, the model is evaluated on its training distribution, not on held-out cell types or conditions. For unconditional generation, the model generates cells from the same tissue it was trained on—there is no experiment where a model trained on PBMCs generates lung cells, or where a model trained on human neocortex generates mouse cortex. This means the experiments test memorization-generalization within a single data distribution, not the kind of cross-domain transfer that would characterize a true world model capable of simulating novel biological contexts. The cross-species experiments train separate models—the zebrafish model never sees mouse data and vice versa—so they test architectural robustness to data shift, not learned biological principles that transfer.

Fourth, the evaluation is entirely correlational, not causal. The Discussion explicitly acknowledges that "high-fidelity generation does not imply biological causality: faithful recapitulation of expression distributions does not necessarily reflect the causal regulatory mechanisms that produce them." The model learns to reproduce patterns of gene co-expression, but it may do so through statistical associations that do not correspond to genuine regulatory relationships. For example, the model might learn that genes A and B are co-expressed across a population of cells without understanding that gene A's protein product transcriptionally activates gene B—a distinction that matters critically for predicting what happens when gene A is perturbed.

Claim: Lingshu-Cell achieves leading performance on the Virtual Cell Challenge H1 genetic perturbation benchmark.

The VCC leaderboard results (Table 2, Table B1) support the claim that Lingshu-Cell is competitive with the best specialized perturbation prediction methods at the time of evaluation, achieving the best average rank (8.7) across all seven metrics and ranking first on MAE (0.052) and Pearson-Δ (0.306). This is a genuinely strong result, particularly given that Lingshu-Cell is a general-purpose generative model while many of the top VCC entries are purpose-built perturbation predictors.

However, four aspects of this result deserve scrutiny:

First, the average rank metric masks substantial variability across individual metrics. Lingshu-Cell ranks 25th out of 26 on DES (differential expression overlap), with a score of 0.216—far below the best DES score of 0.366 (BioCai). This is not a minor weakness; DES directly measures whether the model identifies the correct set of perturbation-responsive genes, which is arguably the most biologically interpretable metric for perturbation screening applications. A method that achieves the best overall rank but the second-worst DES score has a specific, named failure mode: it gets the broad expression patterns right (hence good MAE and Pearson-Δ) but cannot pinpoint which genes are statistically significantly affected. The paper does not discuss this discrepancy or offer hypotheses for why the generative modeling paradigm might systematically underperform on differential expression overlap while excelling at genome-wide accuracy.

Second, the marginal improvement over competitors is small relative to the breadth of the comparison. Lingshu-Cell's average rank is 8.7, while cleopatra's is 9.1 and xBio's is 10.7. With only seven metrics and 26 teams, rank differences of less than 2 positions are within the range where a single metric's fluctuation could reorder the top entries. The paper does not compute bootstrap confidence intervals on the average rank or test whether the differences are statistically significant, making it difficult to assert that Lingshu-Cell is meaningfully better than cleopatra or xBio rather than statistically indistinguishable within the measurement noise of the benchmark.

Third, the evaluation is on a fixed train/validation/test split of perturbation targets, not cell lines. The VCC H1 benchmark tests generalization to held-out perturbation targets within the same H1 cell line. This is a "seen cell line, unseen perturbation" setting. The paper additionally trains on external cell-line data to improve prediction, but evaluation remains exclusively on H1. A stronger test of the world model concept would evaluate generalization to unseen cell lines with seen perturbations (i.e., train on K562, A549, etc., test on H1 for all perturbation targets) or unseen cell-line-perturbation combinations. The current evaluation cannot distinguish between a model that has learned generalizable principles of perturbation biology and one that has learned H1-specific expression patterns from control cells and uses external data only for weak directional priors.

Fourth, the training data composition makes ablation of external data sources impossible from the reported results. The model is trained on H1 train-target cells + external cell-line cells + H1 control cells. The contribution of the external data vs. the in-distribution H1 data is not ablated—we do not know how the model would perform if trained exclusively on H1 cells (which would test in-distribution generalization) or exclusively on external cell lines (which would test cross-cell-line transfer). Without these ablations, the claim that Lingshu-Cell's performance stems from its generative framework rather than from access to substantially more training data (H1 + external) cannot be verified.

Claim: Lingshu-Cell accurately predicts cytokine-induced responses in human PBMCs.

The cytokine perturbation results (Figure 4c) show that Lingshu-Cell achieves the best average score among compared methods in a setting that requires generalization to unseen donor-cytokine combinations. This supports the claim that the conditional generation framework extends beyond genetic perturbations to a different perturbation modality with higher biological complexity (primary immune cells from multiple human donors, stimulation-induced signaling responses rather than genetic knockouts).

The experimental design—holding out 4 of 12 donors and 70% of cytokine conditions per held-out donor—tests generalization along two axes simultaneously: the model must predict responses for donors it has never seen and cytokines it has never seen paired with those donors. This is a challenging and biologically realistic task that matches the intended use case of predicting personalized drug responses or patient-specific cytokine sensitivities.

However, the analysis is limited in several ways. The paper reports bar plots with error bars (Figure 4c) but does not provide numerical values for the metrics or perform statistical comparisons between methods. The error bars show substantial donor-to-donor variability, but the paper does not analyze which donors are "easy" vs. "hard" to predict or whether certain cytokine conditions are systematically more predictable than others. The baseline methods (scGPT, scVI, STATE) are used without detailed description of how they were adapted to this task—for instance, scVI is primarily an unconditional generative model, and its application to perturbation prediction likely required architectural modifications or post-hoc analysis not described in the paper. This makes the comparison potentially unfair, as Lingshu-Cell's conditional generation framework was designed for exactly this task while the baselines may have been repurposed from their intended use cases.

Claim: Lingshu-Cell captures complex combinatorial gene-expression patterns across approximately 18,000 genes without relying on prior gene selection.

The paper emphasizes that Lingshu-Cell "directly models transcriptome-wide expression across approximately 18,000 genes without requiring prior gene selection, such as filtering by high variability or ranking by expression level" (Section 1). This is an architectural claim—the model's input includes all genes in the reference set—but the extent to which the model actually captures expression patterns for lowly expressed or sporadically detected genes is not specifically evaluated.

The unconditional generation metrics (Pearson, Spearman, 1-WD) are averaged over all genes, so they could be dominated by the highly expressed, highly variable genes that also drive feature selection in other methods. The paper does not report metrics stratified by expression level (e.g., correlation for the top 1,000 vs. bottom 1,000 most highly expressed genes), which would directly test whether the "no gene selection" claim translates to meaningful modeling of low-expression genes. The 1-WD metric is gene-averaged, so a model that models highly expressed genes perfectly but low-expressed genes poorly could still achieve a good average score if most genes fall in the low-expression category where the real-vs-generated distance is small in absolute terms (since low-expression genes have narrow distributions near zero).

For perturbation prediction, the metrics are computed on pseudobulked profiles that similarly aggregate over genes, and the differential expression metrics (DES, AUPRC) inherently focus on the most statistically significant changes, which tend to occur in reasonably highly expressed genes. Whether Lingshu-Cell correctly predicts perturbation effects on lowly expressed transcription factors or signaling molecules—genes that are biologically critical but may have near-zero baseline expression—is not tested.

Missing experiments that would strengthen the paper

Several experiments are conspicuous by their absence:

1. Single-cell-level coherence metrics. The paper evaluates population-level distribution matching but never assesses whether individual generated cells are biologically coherent. A simple test: in real PBMCs, a cell cannot simultaneously express CD3D (T-cell marker) and CD79A (B-cell marker) at high levels. Does Lingshu-Cell ever generate such impossible cells? A co-expression violation rate—the fraction of generated cells that express mutually exclusive lineage markers above some threshold—would directly measure single-cell plausibility.

2. Rare cell type detection. The 17-subtype PBMC analysis (Appendix Figure C2c-d) shows that major subtypes are preserved, but extremely rare populations (e.g., plasmablasts, which typically constitute <1% of PBMCs) are not separately evaluated for whether their proportions match between real and generated data. A model that generates zero plasmablasts would still achieve good MMD and correlation metrics because these cells are a tiny fraction of the population, but it would fail at one of the key applications of generative cell modeling—simulating rare cell states that are biologically important but statistically overwhelmed in bulk metrics.

3. Interpolation/extrapolation testing for cytokine perturbation. The cytokine experiment tests generalization to unseen donor-cytokine combinations, but it does not explicitly test interpolation (predicting a cytokine seen in other donors for a new donor) vs. extrapolation (predicting a cytokine never seen in any donor for any donor). If the held-out cytokines are all ones that appear in the training donors, the task is interpolation along the donor axis but not the cytokine axis. The paper's split (70% held out per test donor) does not clarify this distinction.

4. Direct comparison to foundation models on perturbation prediction using identical training data. The paper compares Lingshu-Cell to scGPT and scVI on the cytokine task, but it does not control for training data size or composition. If Lingshu-Cell is trained on 6.5 million PARSE cells while scGPT uses a different pretraining corpus and is fine-tuned on a subset, the comparison conflates model architecture with training data scale and domain match. A fair comparison would use identical training sets—e.g., pretrain both Lingshu-Cell and scGPT on the same PARSE data and evaluate on the same held-out cytokines.

5. Ablation of training data scale. The paper highlights that Lingshu-Cell achieves strong VCC performance with "only approximately 0.6 million training cells," but it does not show how performance scales with training data size. A learning curve (performance vs. number of training cells or number of perturbation targets) would reveal whether the model is data-efficient or data-hungry, and whether additional training data would continue to improve performance or whether saturation has been reached.

6. Latency and computational cost reporting. The paper reports training GPU counts (Table C3) but does not report inference latency or memory requirements for generating a single cell or a full perturbation response. For a model positioned as a "virtual cell" simulator that could be used for large-scale in silico screening, the computational cost per simulated experiment is a critical practical metric that is entirely absent.

7. Robustness to the quantization scheme. The logarithmic adaptive binning is a core design choice, but the paper does not ablate it. How would performance change with uniform binning? With fewer bins (e.g., B = 100)? With more bins (e.g., B = 500)? With a different overflow threshold (C = 9999 is chosen but not justified)? The sensitivity of downstream generation quality and perturbation prediction accuracy to quantization granularity is unknown.

8. Ablation of random permutation in compression. The paper claims that the fixed random permutation π is important because it breaks spurious positional correlations, but it never compares against an aligned compression (grouping adjacent genes in the canonical order) or a learned permutation. Without this ablation, the conceptual motivation for the random permutation remains a hypothesis rather than an empirically validated design choice.

In summary, the experimental results strongly support the claim that Lingshu-Cell can fit and reproduce the training distribution of scRNA-seq data across diverse tissues and species, and provide moderate support for the claim that it can predict perturbation responses competitively with specialized methods on standardized benchmarks. The evidence for the broader framing—as a "cellular world model" that captures biologically meaningful regulatory principles, generalizes across biological contexts in a principled way, and is ready to guide experimental design—is preliminary. The experiments test distribution matching and held-out prediction, which are necessary conditions for a useful simulator, but do not test whether the model's internal representations correspond to genuine regulatory mechanisms, whether its predictions generalize beyond the specific cell types and perturbations in the training data, or whether it can guide the design of experiments that would not have been obvious from the training data alone. The gap between the paper's aspirational framing and the empirical evidence provided is genuine, and the paper's own Discussion—which acknowledges the lack of causal modeling, the need for wet-lab validation, and the limitation to transcriptomic data—reflects this tension honestly.

6. Limitations and Trade-offs

The Difficulty Estimation Cost Is Unaccounted for and Prohibitively Expensive

The assumption or constraint. The entire conditional generation pipeline for perturbation prediction depends on an inference-time biological prior injection (Section 4.5) derived from external cell-line Perturb-seq datasets. The prior requires differential expression analysis on these external datasets using the pdex package with stringent filtering criteria (FDR < 0.05, |log2 fold change| > 1, reference_mean > 2, target_mean < 1; Appendix C.3). While this prior computation is a one-time cost per perturbation target, it fundamentally requires that the perturbation target of interest has been previously profiled in at least one external cell line. For perturbation targets unique to the test set that lack external profiling data, the prior cannot be constructed, and the paper does not provide a fallback strategy or report performance on prior-ineligible targets separately.

Moreover, even when external data exists, the paper does not account for the computational cost of (1) locating and harmonizing external Perturb-seq datasets across multiple cell lines and studies, (2) running per-perturbation differential expression analyses across the full ~18,000-gene feature space, and (3) constructing and validating the union prior set. For the VCC benchmark, this requires processing data from at least six external studies spanning multiple cell lines (Replogle et al., 2022; Nadig et al., 2025; Jiang et al., 2025; Huang et al., 2025; Section A.1.2), each with its own preprocessing pipeline, quality control, and perturbation-label standardization.

The consequence. The practical deployability of Lingshu-Cell as a perturbation screening tool is contingent on a data ecosystem that may not exist for the perturbations a user actually wants to screen. A biologist seeking to predict the transcriptional effect of a novel candidate gene knockout—one not present in published Perturb-seq atlases—cannot use the biological prior injection mechanism and must rely on CFG-only generation. The paper's ablation study (Figure 3g) shows that prior injection improves performance on DES and Pearson-Δ, but it does not report how much performance degrades specifically for perturbations lacking external data versus those with available priors. The headline VCC results (Table 2) therefore represent an upper bound on performance that assumes comprehensive external prior availability—an assumption violated for the majority of the genome, which has not been systematically perturbed across multiple cell lines.

Furthermore, even when priors are available, the union-over-cell-lines strategy (Section C.3: "the corresponding prior gene sets were merged by taking their union") is conservative by design but potentially noisy. Genes that are downregulated in K562 but not in A549 are included in the prior and forced to low expression in the H1 prediction, regardless of whether the downregulation is cell-line-specific. The paper does not ablate union vs. intersection strategies or evaluate whether false-positive prior genes degrade prediction quality relative to a more stringent intersection approach.

What evidence exists in the paper. The ablation study (Figure 3g) demonstrates that prior injection provides a measurable but modest improvement over the CFG-only baseline, primarily on perturbation direction similarity (DES) and Pearson-Δ correlation. The paper does not:

  • Report performance stratified by whether external priors were available for each test perturbation.
  • Ablate the stringency of the prior construction (FDR thresholds, effect size cutoffs) or the union vs. intersection strategy.
  • Quantify the computational cost of external data processing and prior construction.
  • Evaluate whether the prior injection generalizes to perturbation targets without external data.

Mitigation status. The paper does not attempt to mitigate this limitation. It does not develop a method for predicting priors when external data is unavailable (e.g., using the model's own unconditional generation to identify perturbation-responsive genes or leveraging gene-gene similarity to infer priors for unprofiled targets), nor does it frame the prior as an optional enhancement whose absence is explicitly measured. The Discussion does not identify prior dependency as a limitation, focusing instead on broader future directions (multi-omic integration, temporal dynamics, closed-loop experimentation).

Only Transcriptomic Data Is Modeled, Excluding the Regulatory Layers That Drive Perturbation Responses

The assumption or constraint. Lingshu-Cell operates exclusively on transcriptomic data—the input is a UMI count matrix, and the model learns to generate token sequences representing mRNA expression levels across ~18,000 genes. The paper explicitly acknowledges this scope limitation in the Discussion:

"the current model operates exclusively on transcriptomic data; a more complete virtual cell would additionally integrate epigenomic, proteomic, metabolomic, and spatial modalities" (Section 3).

The model does not incorporate chromatin accessibility, transcription factor binding, histone modifications, protein abundances, or any other modality that provides mechanistic insight into why a perturbation produces a particular transcriptional response. For genetic perturbation prediction, this means the model learns the statistical association between a perturbation token (e.g., "CHMP3 knockout") and a transcriptomic outcome without any representation of the physical process—CHMP3's role in endosomal sorting, the protein-protein interactions it participates in, or the signaling pathways it regulates.

The consequence. This limitation manifests in two distinct failure modes that are not separable in the reported metrics but have critical practical consequences.

First, the model's predictions are necessarily correlational, not causal. Even when Lingshu-Cell accurately predicts that CHMP3 knockout downregulates a set of target genes, it does not know whether this is a direct regulatory effect (CHMP3's protein product transcriptionally activates those targets) or an indirect downstream consequence (CHMP3 knockout triggers a stress response that incidentally downregulates those genes). For a researcher using the model to prioritize follow-up experiments, this distinction is essential—a direct target is a candidate for further mechanistic study; an indirect effect is a consequence of cellular state change that may not replicate under different conditions.

Second, the model cannot predict responses to perturbations that operate through mechanisms not represented in the training data. A drug that inhibits a kinase not previously profiled in Perturb-seq data has no natural token representation in the conditioning vocabulary, and even if a token were assigned, the model has no basis for predicting its effect because it has never seen the downstream transcriptional consequences of modulating that pathway. The cytokine perturbation experiments (Figure 4) demonstrate generalization to unseen cytokine-donor combinations, but all cytokines in the PARSE dataset operate through well-characterized JAK-STAT, NF-κB, or MAPK signaling pathways. The model's ability to generalize to completely novel signaling modalities (e.g., a synthetic ligand for an orphan receptor) is untested and unsupported by the architecture.

Third, the model cannot explain its predictions. When Lingshu-Cell predicts that a perturbation downregulates gene X, the only evidence is the statistical correlation learned from training data—there is no mechanism to trace the prediction through a regulatory network, identify which transcription factors mediate the effect, or assess whether the prediction is consistent with known pathway structure. For a tool positioned as enabling "biological discovery and perturbation screening" (Section 1), this interpretability gap limits its utility for generating mechanistic hypotheses.

What evidence exists in the paper. The paper provides no experiments testing whether Lingshu-Cell's predictions respect known regulatory mechanisms. There is no analysis of whether predicted differentially expressed genes are enriched for the known targets of transcription factors in the affected pathway, whether predicted perturbation responses are consistent with protein-protein interaction network structure, or whether the model's internal attention patterns recover known regulatory relationships. The VCC evaluation metrics (DES, PDS, MAE, etc.) measure prediction accuracy against ground-truth transcriptomic measurements, not mechanistic validity—a model could achieve a high DES score by predicting the correct set of differentially expressed genes through purely statistical pattern matching without any understanding of the underlying biology.

Mitigation status. The paper acknowledges the limitation transparently in the Discussion, stating that Lingshu-Cell "should currently be regarded as a powerful tool for probabilistic hypothesis generation" whose "predictions still require rigorous wet-lab experimental validation." It identifies multi-omic integration as a natural extension: "jointly modeling chromatin accessibility, protein abundances, and spatially resolved transcriptomics within the same discrete diffusion framework" (Section 3). However, no multi-omic experiments are presented, and it is unclear whether the discrete diffusion framework would extend naturally to modalities with fundamentally different data structures (e.g., continuous protein abundance measurements, binary chromatin accessibility peaks, spatial coordinates). The paper does not address the interpretability gap at all—there is no discussion of attention analysis, feature attribution, or any other mechanism for explaining why the model makes a particular prediction.

The Model Is Evaluated on Population-Level Metrics That Cannot Detect Single-Cell Artifacts or Biologically Impossible States

The assumption or constraint. All unconditional generation evaluation metrics (Section B.1) and all perturbation prediction metrics (Section B.2) operate on aggregated statistics—either over the full population of generated cells (Pearson r, Spearman ρ, MMD, 1-WD, iLISI) or on pseudobulks averaged across all cells sharing a perturbation condition (MAE, Pearson-Δ, PDS, DES, etc.). The paper does not define or compute any metric that evaluates the biological coherence of individual generated cells. The closest the paper comes to single-cell evaluation is the UMAP visualization (Figure 2a, Appendix Figure C2), which is a qualitative, projection-based assessment that can hide artifacts in high-dimensional space.

The consequence. A generative model that scores perfectly on all reported metrics could still produce individually implausible cells. The most obvious failure mode in the PBMC context: a single generated cell could simultaneously express T-cell markers (CD3D, CD4, CD8A) and B-cell markers (CD79A, CD19, MS4A1) at high levels—a biologically impossible state known as a cross-lineage doublet. Because the population-level metrics average over all cells, a small fraction of such impossible cells would be statistically invisible (they would barely affect mean expression vectors, correlation coefficients, or MMD), yet their presence would indicate a fundamental failure of the model to respect the combinatorial constraints of cell identity.

More subtly, the model could generate cells that are "blurry" in high-dimensional space—expression profiles that interpolate between real cell types but lack the sharp, coordinated expression of lineage-defining gene modules. A cell that expresses all lineage markers at intermediate levels might sit near the center of a UMAP plot and contribute reasonably to population-level distribution matching, but it would not correspond to any real biological state. The iLISI metric partially tests for this (it measures mixing between real and generated cells in local neighborhoods), but a blurry generated cell that sits between real clusters could still achieve a high iLISI score by being approximately equidistant from multiple real types.

For perturbation prediction, the pseudobulk averaging approach is standard in the field (it is used by the VCC evaluation protocol), but it discards all information about cell-to-cell variability in perturbation response. Real Perturb-seq data shows substantial heterogeneity: even under identical perturbation conditions, some cells show strong transcriptional responses while others appear nearly unaffected. Pseudobulking collapses this heterogeneity into a single mean expression vector, so a model that predicts the mean response perfectly but generates no variability (all cells identical) and a model that captures realistic heterogeneity would receive the same pseudobulk score. The paper does not evaluate whether Lingshu-Cell's generated perturbation responses exhibit realistic single-cell variability or whether the model collapses to a deterministic mean prediction at low generation steps (N = 3).

What evidence exists in the paper. The paper provides no explicit single-cell coherence evaluation. The UMAP visualizations (Figure 2a, 2d-e, Appendix Figures C2, C4) are the only window into single-cell-level quality, and they show that generated cells broadly overlap with real cells in 2D projection space, but this is a qualitative assessment that cannot detect cross-lineage artifacts (which may be rare and not visible in a 2D scatter of thousands of points) or blurriness (which can be obscured by the UMAP algorithm's tendency to separate clusters). The perturbation prediction results report only pseudobulk metrics—the VCC protocol (which the paper follows) does not include single-cell-level evaluation, and the paper does not extend beyond it.

The generation scale experiment (Appendix Figure C2) generates 200,000 PBMCs and subdivides them into 17 subtypes, confirming that subtype proportions match between real and generated data. This partially addresses the concern about rare cell types—if the model were generating many impossible cells, they would likely not be classifiable into known subtypes—but it does not rule out a small fraction of cross-lineage cells or blurry intermediate states.

Mitigation status. The paper does not acknowledge this limitation or propose single-cell-level evaluation metrics. The Discussion focuses on higher-level limitations (causality, multi-omic integration) and does not discuss the potential for single-cell artifacts that are invisible to population-level statistics. Future work on "closed-loop experimentation" (Section 3) would presumably require single-cell-fidelity generation to guide perturbation design, but the current evaluation framework provides no evidence that Lingshu-Cell meets this bar.

Generalization Is Tested Only Within the Training Distribution, Not to Fundamentally Novel Biological Contexts

The assumption or constraint. All evaluation in the paper tests generalization along axes that are represented, albeit sparsely, in the training data. For unconditional generation, each model is trained on cells from a single tissue/species and evaluated on held-aside cells from the same tissue and species—there is no cross-tissue or cross-species transfer experiment (e.g., training on human neocortex and generating mouse cortex cells, or training on PBMCs and generating colon cells). For genetic perturbation prediction, the VCC benchmark tests generalization to held-out perturbation targets within the same cell line (H1)—the model has seen H1 control cells and cells from other perturbations in H1, plus external cell-line data for the held-out perturbations, but the evaluation is always on H1 cells. For cytokine perturbation prediction, the model is tested on held-out donor-cytokine combinations, but the held-out cytokines may appear in other training donors, and all donors are human PBMC donors from the same study with the same experimental protocol.

The consequence. The paper frames Lingshu-Cell as a "cellular world model" capable of serving as a "unified generative framework for single-cell transcriptomics" (Section 3) and enabling "virtual cell modeling and in silico perturbation analysis across diverse biological contexts" (Section 2.4). However, the evaluation provides evidence only for within-distribution generalization—the model can interpolate to unseen perturbation targets, unseen donor-cytokine pairs, and held-aside cells from the same tissue. It provides no evidence that the model has learned transferable principles of cellular biology that would enable extrapolation to genuinely novel biological contexts, such as:

  • A tissue the model has never seen (e.g., training on PBMC and heart, testing on pancreas).
  • A species with different gene regulatory logic (e.g., training on mouse and human, testing on a non-mammalian vertebrate).
  • A perturbation modality not represented in training (e.g., training on CRISPR knockouts, testing on drug treatments or overexpression).
  • A disease state or developmental stage not present in the training corpus (e.g., training on healthy adult cells, testing on fetal or disease-state cells).

This is a critical distinction because the "world model" framing implies that the model has learned a compressed representation of the underlying dynamics of cellular state—a representation that, like a physics simulator, can be queried with novel initial conditions or intervention parameters and yield valid predictions. The current evaluation tests the narrower claim that the model is a high-capacity generative model that can fit and interpolate within its training distribution, which is a necessary but not sufficient condition for a world model.

What evidence exists in the paper. The paper's evaluation design is explicitly within-distribution. The cross-tissue experiments (Figure 2d, Appendix Figure C4) train separate models per tissue—the neocortex model never sees heart data and vice versa. The paper labels this as evidence that Lingshu-Cell "generalizes reliably in the unconditional setting across tissues and species" (Section 2.2), but this is misleading: the model does not generalize across tissues; rather, the same architecture is successfully applied to multiple tissues independently. This tests architectural robustness (the model can fit PBMC and neocortex and heart data), not generalization (the model cannot use knowledge of PBMCs to improve its modeling of neocortex).

The genetic perturbation experiments partially test cross-cell-line generalization by training on external cell-line data and evaluating on H1, but the model also sees H1 control cells and H1 train-target cells, so it has in-distribution information about H1's baseline expression patterns. The contribution of the external cell-line data to the final performance cannot be isolated from the contribution of the H1-specific training data because no ablation is reported (the paper does not train a model on external data only and evaluate on H1). The cytokine experiments test generalization to unseen donor-cytokine pairs, which is genuinely cross-context (different donors have different immune compositions and baseline expression states), but all donors are from the same study, same species, same tissue, and same experimental protocol—the generalization is narrow.

Mitigation status. The paper does not acknowledge the distinction between within-distribution interpolation and out-of-distribution generalization. It presents the cross-tissue experiments as evidence of "generalization" and the cross-donor cytokine results as evidence that the "conditional generation framework [can serve] as a unified approach for predicting cellular responses to diverse perturbations across experimental contexts" (Section 2.4). The Discussion envisions extensions to "drug-induced, multi-target, and combinatorial interventions" and to new modalities, but it does not identify the absence of cross-tissue, cross-species, or cross-modality generalization experiments as a limitation of the current study. A practitioner evaluating Lingshu-Cell for a novel biological context not represented in the training data has no empirical evidence to guide their expectations of performance.

The Generative Model Is Trained and Evaluated with a Rigid Gene Set Dependency

The assumption or constraint. Lingshu-Cell operates on a fixed reference set of G = 18,080 genes, which is the gene set used in the Virtual Cell Challenge H1 benchmark (Section 4.2). All datasets—human PBMC, eight human tissues, four non-human species, external Perturb-seq datasets—are harmonized to this exact gene set before training (Appendix A.1). Genes not present in the reference set are excluded; reference genes not detected in a particular dataset are assigned zero counts. This gene set is a hard architectural constraint: the model's input and output sequence length is fixed at G = 18,080, the token embedding table has a fixed row for each gene position, and the sequence compression module's random permutation π is defined over exactly G positions.

The consequence. This design creates three practical deployment barriers that are not discussed in the paper.

First, the model cannot be applied to datasets with different gene sets without retraining or substantial architecture modification. A single-cell atlas profiled with a different sequencing technology, a different reference genome annotation, or a targeted gene panel would require either (1) imputing missing genes (which introduces unmodeled bias), (2) discarding genes present in the new data but absent from the reference set (losing potentially critical information), or (3) retraining the model from scratch with a new vocabulary and gene set. This is a fundamentally different limitation than in foundation models that operate on gene symbols or sequence-based gene representations (e.g., Geneformer), which can handle arbitrary gene sets because the model's input is the gene identity, not a fixed positional encoding.

Second, the 18,080-gene set may not be optimal for all biological contexts. The VCC gene set was selected for a cell-line genetic perturbation benchmark focused on human cells. For mouse, zebrafish, or fly data, many of the 18,080 human genes have no ortholog, and many species-specific genes (particularly rapidly evolving gene families involved in immunity, reproduction, and environmental response) are excluded. The paper reports strong unconditional generation metrics for non-human species (Table 1) despite this gene set mismatch, but the metrics are computed only on the genes that are retained—genes excluded from the reference set are invisible to the evaluation. If species-specific genes drive important biological variation (e.g., fly brain-specific ion channels or zebrafish embryo-specific developmental regulators), the model cannot capture it.

Third, the fixed gene order interacts with the sequence compression module in a way that may create gene-set-dependent inductive biases. The random permutation π is fixed at initialization and applied identically to all cells. If a user wanted to add or remove genes from the reference set—even a single gene—the permutation would need to be redefined, the compression groups would change, and the model would need to be retrained because the position-to-gene mapping is baked into the learned token embeddings. This means Lingshu-Cell is not a model that can be incrementally updated with new gene annotations or adapted to custom gene panels—it is a single, fixed-vocabulary model that must be retrained for any gene set change.

What evidence exists in the paper. The paper reports that all datasets are harmonized to the reference set (Appendix A.1) but does not quantify how many genes are lost in each dataset due to this harmonization, nor does it assess whether the excluded genes are functionally important for the biological processes studied. For non-human species, the paper notes that gene annotations were obtained from Ensembl BioMart and that "only genes annotated as protein-coding and associated with valid gene names were retained" (Appendix A.1.1), but the number of retained genes per species, the overlap with the 18,080-gene human reference, and the number of species-specific protein-coding genes excluded are not reported. The quality control figures (Appendix Figures C3, C5) show distributions of detected genes per cell, but these are computed after harmonization and do not reveal what fraction of the original feature space was discarded.

Mitigation status. The paper does not discuss the fixed gene set as a limitation. The choice to standardize to the VCC gene set is presented as a practical decision for benchmark compatibility, not as a constraint that limits the model's applicability. The Discussion envisions multi-omic and multi-modal extensions but does not address the gene set rigidity, nor does it propose strategies for handling variable gene sets (e.g., gene-symbol-based tokenization, attention over a variable number of genes, or architecture designs that decouple position from gene identity). For a model positioned as a general-purpose cellular world model, the inability to handle arbitrary gene sets without retraining is a significant practical limitation that is not acknowledged.

Classifier-Free Guidance and Biological Priors Create a Fragile Inference Configuration with Uncharacterized Sensitivity

The assumption or constraint. The conditional generation pipeline for perturbation prediction relies on a specific combination of hyperparameters and data sources whose interactions are not systematically analyzed. The inference configuration involves: (1) a CFG weight w = 2 for genetic perturbations and w = 3 for cytokine perturbations, (2) N = 3 reverse diffusion steps (versus N = 256 for unconditional generation), (3) a patch size S = 32 for genetic perturbations (versus S = 8 default), (4) biological priors derived from external cell-line data with specific differential expression thresholds, (5) a control-cell ratio capped at 10% during training, and (6) a cosine timestep schedule discretized into the chosen N steps. These choices are individually ablated (Figures 3d-g) but their joint sensitivity and interactions are not explored.

The consequence. A practitioner seeking to apply Lingshu-Cell to a new perturbation modality (e.g., drug treatment, environmental stress, overexpression) or a new biological context (e.g., a non-human cell line, a primary tissue not in PBMC) faces substantial uncertainty about how to configure inference. The paper provides no guidance on: how N should scale with the complexity of the perturbation response (is N = 3 sufficient for drug responses that affect thousands of genes versus the hundreds affected by a single-gene knockout?); whether w = 2 transfers across cell types or perturbation modalities (is a transcription factor knockout in neurons more or less sensitive to guidance weight than a cytokine stimulation in immune cells?); or how the optimal patch size S depends on the gene set size and the sparsity of the data.

More concerning, the paper reports that N = 3 steps with CFG was used for VCC prediction (Appendix C.2), which is dramatically fewer than the N = 256 used for unconditional generation. The paper justifies this implicitly—CFG provides additional steering, so fewer refinement steps are needed—but it does not ablate the interaction between N and w. It is plausible that a model with N = 10 steps and w = 1 would outperform N = 3 and w = 2, or that higher N is necessary for perturbations with subtle transcriptional effects. The choice of N = 3 appears to be driven by computational convenience (faster inference) rather than by systematic optimization, but without an N vs. w sweep, this cannot be verified.

The dependence on external biological priors further complicates inference: the prior is a binary mask (downregulated genes forced to low expression, all others generated freely), but the paper does not explore whether the prior strength (the fixed expression value µ = 1 assigned to prior genes) interacts with CFG weight. A prior that forces genes to µ = 1 while CFG with w = 2 strongly pushes the model toward the perturbed manifold could create tension—the model is simultaneously told "these genes must be low" (by the prior) and "make the expression profile look like a perturbed cell" (by CFG). If the prior genes are actually moderately expressed in the perturbed state (e.g., downregulated but not silenced), the fixed µ = 1 prior could conflict with the model's learned distribution, and the resolution of this conflict is not characterized.

What evidence exists in the paper. The ablation studies (Figures 3d-g) evaluate each component in isolation:

  • CFG weight sweep (Figure 3d): tests w ∈ {0, 1, 2} with other components at their default values (S = 32, prior on, N = 3).
  • Patch size sweep (Figure 3f): tests S ∈ {no compression, 8, 32} with w at its optimal setting.
  • Prior injection (Figure 3g): tests with/without prior with other components fixed.

No experiment varies two components simultaneously or reports the 2D performance landscape. The paper does not report ablation of N, the number of reverse diffusion steps, nor does it analyze the interaction between prior strength (µ) and CFG weight. The cytokine perturbation experiments use a different w (3 vs. 2) and patch size (8 vs. 32) without systematic justification—these values may have been chosen by informal trial-and-error on the validation set, but no sweep results are reported for the cytokine setting.

Mitigation status. The paper does not address the fragility of the inference configuration. The Discussion focuses on extending the model to new perturbation types and modalities, not on making the existing inference procedure robust to hyperparameter choices. There is no discussion of adaptive CFG scheduling (varying w across diffusion steps), automated prior strength calibration, or principled methods for selecting N based on data characteristics. For a model positioned as a practical tool for perturbation screening, the absence of guidance on hyperparameter selection for new contexts is a significant barrier to adoption.

7. Implications and Future Directions

How This Work Changes the Landscape

Lingshu-Cell does not introduce a fundamentally new generative modeling technique—masked discrete diffusion models predate this work (Austin et al., 2021), and the components (quantization, sequence compression, classifier-free guidance) are individually well-established. The paper's contribution is instead an architectural alignment argument: that by choosing a generative paradigm whose structural properties mirror the structural properties of single-cell transcriptomic data—discrete, sparse, permutation-invariant, zero-inflated—one can build a unified framework that performs competitively on both unconditional simulation and conditional perturbation prediction without the distributional mismatches that handicap continuous diffusion, VAEs, and autoregressive models.

This is not a paradigm shift in the sense of overturning a dominant approach—the field has no single dominant generative paradigm for scRNA-seq, given the coexistence of VAEs (scVI), continuous diffusion (scDiffusion), flow matching (CellFlow, scDFM), and autoregressive models (scGPT). Rather, it is a constructive demonstration that discrete diffusion resolves a specific, named failure mode (the continuous-discrete mismatch) while achieving practical competitiveness. The paper's most durable contribution may be the simple observation, made concrete through the benchmark comparisons in Figure 2c and Table 2, that MMD and 1-WD—metrics sensitive to distributional fidelity rather than gene-averaged correlation—reveal substantial quality differences between generative paradigms even when Pearson correlations are saturated near 1.0. This is a methodological finding that applies beyond this paper: it suggests that the field's heavy reliance on gene-level correlation metrics for evaluating generative models systematically underestimates the gap between methods, and that distributional metrics should become standard in benchmark suites.

The paper also provides a reconciliation function for the generative single-cell literature, though in a more limited way than it claims. The observation that continuous diffusion models (scDiffusion) achieve 25× worse 1-Wasserstein distance than discrete diffusion (0.1594 vs. 0.0064; Figure 2c) provides empirical weight to the theoretical argument that Gaussian noise corruption is poorly matched to sparse count data. This does not resolve a prior contradiction—no prior work claimed continuous diffusion was optimal for scRNA-seq—but it does provide the first head-to-head quantitative evidence that the distributional mismatch has measurable consequences for sample quality, not just theoretical elegance. Similarly, the finding that scVI's MMD (0.0343) is ~4× worse than Lingshu-Cell's (0.0088) provides specific evidence for the claim that the VAE information bottleneck limits distributional fidelity, a hypothesis that was plausible but not previously quantified in this way.

However, the paper's attempt to recast single-cell foundation models as "world models" rather than "static encoders" is more a reframing of ambition than a demonstrated capability. The phrase "cellular world model" implies—by analogy to world models in reinforcement learning—that the model has learned a compressed representation of state-space dynamics that supports counterfactual simulation. The paper's evidence supports the narrower claim that the model can interpolate within its training distribution: generating new cells from the same tissue, predicting responses to perturbations whose targets have been profiled (if only in other cell lines). The critical capabilities that would justify the "world model" label—cross-tissue generalization, prediction of perturbation responses for entirely novel modalities (e.g., drug treatments never seen in training), simulation of temporal trajectories—are not tested. The paper's own Discussion acknowledges that "high-fidelity generation does not imply biological causality" (Section 3), which is precisely the gap between a generative model that reproduces patterns and a world model that captures mechanisms. The paper establishes discrete diffusion as a promising architectural foundation for building toward cellular world models, but the framing should be understood as aspirational rather than achieved.

The practical consequence for the field is that discrete diffusion becomes a credible third paradigm alongside VAEs and continuous diffusion for single-cell generative modeling, with specific, quantitative advantages on distributional fidelity metrics. Researchers developing new generative models for transcriptomic data should now justify why they choose a continuous paradigm over discrete diffusion, rather than defaulting to VAEs or DDPMs. Conversely, the paper's limitations—particularly the fixed gene set dependency and the absence of single-cell coherence metrics—create clear targets for methods that would surpass Lingshu-Cell by addressing these gaps.

Follow-Up Research This Work Enables

Single-cell coherence metrics and the detection of biologically impossible generated states. Lingshu-Cell's evaluation relies entirely on population-level distributional metrics (MMD, 1-WD, iLISI) and pseudobulk correlations, which cannot detect whether individual generated cells violate known biological constraints. A natural and immediately tractable follow-up would define a suite of single-cell coherence tests and apply them to Lingshu-Cell, scDiffusion, and scVI. For PBMC data, one would measure: (1) the fraction of generated cells that co-express mutually exclusive lineage markers above a threshold (e.g., CD3D > 2 and CD79A > 2 simultaneously), which should be near zero in real data; (2) the per-cell correlation between generated and nearest-real-neighbor expression profiles in PCA space, which would detect "blurry" cells that interpolate between real states; (3) the preservation of known gene regulatory network edges within individual cells (are transcription factor–target gene correlations preserved at the single-cell level, or only in aggregate?). If Lingshu-Cell produces a non-trivial fraction of cross-lineage doublets or blurry intermediates, this would reveal that the discrete diffusion paradigm, while achieving excellent distributional matching, does not automatically enforce the combinatorial constraints of cell identity—a limitation that would motivate architectural extensions such as constrained sampling or rejection-based filtering.

Cross-tissue and cross-species transfer learning to test whether the model learns transferable biological principles. The paper trains separate models for each tissue and species, demonstrating architectural robustness but not generalization. A direct extension would train a single Lingshu-Cell model on multiple human tissues simultaneously, with tissue identity as a conditioning token (analogous to the donor identity token in cytokine experiments), and then test whether the model can generate realistic cells for a held-out tissue never seen during training. If successful, this would provide the first evidence that the discrete diffusion framework learns representations that transfer across biological contexts, moving beyond within-distribution interpolation. The experiment would require careful negative controls: the held-out tissue should be sufficiently distinct (e.g., train on immune, epithelial, and neural tissues; test on cardiac) to ensure that successful generation reflects learned transferable principles rather than memorization of shared cell types (e.g., macrophages appear in most tissues). A more ambitious version would train jointly on human and mouse data and test cross-species generation of homologous cell types, leveraging the shared gene set to assess whether the model captures evolutionary conservation of expression programs.

Ablation of the biological prior to measure the true data-driven performance ceiling. The paper's VCC results are confounded by the inference-time biological prior, which uses external cell-line data that may not be available for novel perturbations. A clean ablation would train Lingshu-Cell on the VCC H1 data only (no external Perturb-seq datasets), with no biological prior injection, and evaluate on the H1 test set. This would establish a "pure" data-driven baseline for the discrete diffusion approach, measuring what the model can learn from in-distribution data alone. Comparing this to the full model with external data and priors would isolate the contribution of cross-cell-line information, which is critical for understanding whether the model's strong VCC performance stems from its generative architecture or from access to substantially more training data than many competitors. If the pure H1-only model performs competitively with the full model, the paper's claim of data efficiency is strengthened; if it degrades substantially, the dependence on external data becomes the primary explanation for VCC performance, and the generative architecture's contribution is more modest than the headline results suggest.

Systematic characterization of the inference hyperparameter landscape to establish robust defaults. The paper's inference configuration (N = 3 steps, w = 2 or 3, S = 32 or 8, prior injection on) is the result of limited ablation but no joint optimization. A thorough study would perform a grid search over {N ∈ {1, 2, 3, 5, 10, 20, 50}, w ∈ {0, 0.5, 1, 2, 3, 5, 10}} on the VCC validation set, reporting the 2D performance landscape for each of the seven perturbation metrics. This would reveal whether the choice of N = 3 is near-optimal or merely convenient, whether higher N can compensate for lower w (or vice versa), and whether over-guidance artifacts emerge at higher w as they do in image diffusion models. The study should also sweep the prior strength µ (the fixed expression value assigned to prior genes) to determine whether µ = 1 is optimal or whether a data-dependent µ (e.g., the 10th percentile of expression for that gene in the target cell line) would improve performance. The output would be a set of recommended default hyperparameters for different perturbation modalities and data sparsity regimes, addressing the fragility concern raised in the Limitations section.

Causal validation of perturbation predictions against known regulatory network structure. Lingshu-Cell's predictions are evaluated on transcriptomic accuracy but not on mechanistic plausibility. A targeted validation experiment would take the model's predicted differentially expressed genes for a set of well-characterized transcription factor perturbations in the VCC test set, and compare the overlap between predicted DE genes and the known target genes of those transcription factors (from ChIP-seq, motif analysis, or curated databases like TRRUST or ENCODE). If the model's predictions are enriched for known direct targets beyond what would be expected from expression correlation alone, this would provide the first evidence that the discrete diffusion framework captures something approximating regulatory relationships, even if implicitly. A negative result—no enrichment over a correlation-based baseline—would clarify that the model's predictions, while transcriptionally accurate, are not mechanistically grounded, reinforcing the paper's own caveat that predictions require wet-lab validation. This experiment is low-cost (it uses existing perturbation predictions and public regulatory network databases) but would substantially refine our understanding of what biological knowledge the model actually learns.

Extension to temporal perturbation responses using time-course scRNA-seq data. The paper identifies temporal dynamics as a key future direction (Discussion) but provides no experiments. A concrete next step would train Lingshu-Cell on a time-course Perturb-seq dataset where cells are profiled at multiple time points post-perturbation, with time as an additional conditioning token alongside perturbation identity. The model would then be tasked with generating expression profiles at held-out time points for known perturbations, testing whether the discrete diffusion framework can capture the temporal program of transcriptional response (e.g., immediate-early genes activated within hours, secondary responses emerging over days). The VCC H1 dataset does not include time-course data, but published time-course Perturb-seq studies (e.g., in the context of T-cell activation or developmental signaling) exist and could serve as benchmarks. This experiment would directly test whether the "world model" framing—which implies the ability to simulate state transitions—has empirical support beyond steady-state generation.

Practical Applications and Downstream Use Cases

Perturbation prioritization for CRISPR screening with quantified uncertainty. In a typical functional genomics workflow, a researcher has a list of 500–2000 candidate genes and must select 50–100 for an expensive arrayed Perturb-seq experiment. Lingshu-Cell can be deployed as a pre-screening filter: for each candidate gene, generate 100–1000 predicted perturbed transcriptomes (leveraging the model's ability to produce diverse samples through the stochastic reverse diffusion process), compute the predicted differentially expressed gene set and its variance across samples, and rank candidates by effect size while flagging those with high prediction variance for deprioritization or additional external data collection. The paper's strong MAE and Pearson-Δ performance (Table 2) suggests predicted effect directions and magnitudes are reliable, while the weaker DES performance (rank 25/26) indicates that the exact set of differentially expressed genes should be treated as a hypothesis rather than a prediction. A practical deployment would present the top-ranked candidates to the experimentalist with a warning on DES-based predictions: "This model predicts gene X knockdown will cause strong transcriptional changes (high Pearson-Δ), but the specific target genes are uncertain (low DES); consider validating with a smaller pilot screen before committing to full Perturb-seq."

In silico donor-specific cytokine response screening for personalized immunotherapy design. The PARSE 10M PBMC dataset contains 12 donors each exposed to 90 cytokines, creating a combinatorial space of 1080 donor-cytokine pairs. In clinical settings, testing a patient's PBMCs against 90 cytokines before selecting an immunotherapy is impractical. Lingshu-Cell's cytokine perturbation results (Figure 4c) demonstrate that the model can predict responses for held-out donor-cytokine combinations, achieving the best average score among compared methods. A practical deployment would profile a new patient's PBMCs at baseline (unstimulated), train or fine-tune Lingshu-Cell on that patient's baseline data plus the full PARSE training set, and then simulate the patient's response to all 90 cytokines in silico. The predicted responses—filtered through the model's PDS and Pearson-Δ metrics, which respectively measure perturbation distinguishability and expression change accuracy—would generate a ranked list of cytokines predicted to produce the strongest (or most selective) immune activation for that patient, which could then guide a targeted experimental validation with 5–10 cytokines rather than 90. The donor-level error bars in Figure 4c suggest prediction quality varies across donors, so the practical system would need to report per-donor confidence intervals and flag patients for whom the model's donor-interpolation performance is poor based on their baseline expression profile's similarity to training donors.

Population-scale generation of synthetic single-cell atlases for benchmarking computational methods. Computational method development in single-cell biology—batch correction algorithms, trajectory inference tools, differential expression testing frameworks—suffers from a shortage of ground-truth datasets where the correct answer is known. Lingshu-Cell's unconditional generation capability provides a solution: for a tissue of interest (e.g., human PBMC, heart, or neocortex), train Lingshu-Cell on the real data, then generate synthetic datasets with controlled properties—varying numbers of cells (10^3 to 10^6, shown to preserve subtype proportions in Appendix Figure C2), varying degrees of cell-type imbalance, or artificially introduced batch effects (by conditioning on synthetic batch tokens during generation). Because the true generating process is known (it is the trained Lingshu-Cell model), and because the synthetic data reproduces real transcriptomic distributions (MMD 0.0088 for PBMC; Figure 2c), these synthetic datasets could serve as benchmarks where all ground-truth cell types, differentially expressed genes, and batch assignments are exactly known. This addresses a genuine pain point in the single-cell methods community, and the paper's demonstration that the 17-subtype structure is preserved in large-scale generation (Appendix Figure C2c-d) provides initial validation that the synthetic data would capture the complexity needed for meaningful benchmarking.

When to Prefer This Method

The paper positions Lingshu-Cell against two major classes of alternatives: continuous generative models (scVI, scDiffusion) for unconditional transcriptomic simulation, and specialized perturbation predictors (represented by the diverse VCC leaderboard entries) for perturbation response prediction. The decision rule emerging from the empirical results, though not explicitly stated as such by the paper, can be inferred:

  • Prefer Lingshu-Cell over continuous generative models (scVI, scDiffusion, and by extension continuous flow-based methods) when distributional fidelity at the full-transcriptome level is the primary goal. The paper's MMD (0.0088 vs. 0.0178 for scDiffusion, 0.0343 for scVI) and 1-WD (0.0064 vs. 0.1594 for scDiffusion) advantages are substantial and robust across the PBMC benchmark (Figure 2c). If an application requires generated cells whose global expression distributions are as close as possible to real data—for synthetic benchmark generation, for data augmentation in rare-cell-type analysis, or for training downstream models that are sensitive to distributional shift—the discrete diffusion paradigm's alignment with the data's sparse, discrete structure translates directly to measurable quality improvements. The caveat is that Lingshu-Cell's fixed gene set (18,080 genes) and computationally expensive per-tissue training mean this advantage assumes the user's gene set and tissue of interest are covered by the training configuration.

  • Prefer Lingshu-Cell over specialized perturbation predictors when both transcriptome-wide accuracy and perturbation discrimination are valued, but differential expression overlap is not the primary metric. On the VCC leaderboard (Table 2), Lingshu-Cell achieves the best MAE (0.052) and Pearson-Δ (0.306) but only rank 25/26 on DES (0.216). This pattern suggests the model excels at capturing the overall magnitude and direction of perturbation effects across the genome but struggles to identify the precise set of statistically significant differentially expressed genes. For applications like dose-response characterization or pathway-level effect ranking—where getting the global expression shift right matters more than perfect gene-level DE overlap—Lingshu-Cell's balanced performance across PDS, Spearman #DEG, and Spearman LFC (all in the top half of the leaderboard) makes it a strong candidate. For applications where identifying the exact gene list is paramount (e.g., designing targeted follow-up experiments based on predicted DE genes), a method with stronger DES performance (e.g., BioCai with DES 0.366) may be preferable, even at the cost of transcriptome-wide accuracy.

  • Prefer Lingshu-Cell when external perturbation data exists to construct biological priors, and be cautious when it does not. The prior injection mechanism (Figure 3g) provides a measurable but modest improvement, primarily on DES and Pearson-Δ. The headline VCC results include this prior; for perturbation targets without external profiling data—the majority of the genome for most cell types—performance should be expected to degrade toward the CFG-only baseline. The paper does not quantify this degradation, so a practitioner considering Lingshu-Cell for a novel perturbation target should either verify that external data exists for that target in at least one cell line, or commission a validation study comparing CFG-only vs. prior-augmented predictions for known targets similar to their target of interest.

  • Do not prefer Lingshu-Cell when the application requires single-cell-level biological coherence guarantees, out-of-distribution generalization, or mechanistic interpretability. The model is evaluated only on population-level metrics; it may generate individually implausible cells that are invisible to the reported scores. It has not been tested on held-out tissues, species, or perturbation modalities—extrapolation to such contexts is speculation, not justified by evidence. And it provides no mechanism for explaining why a particular perturbation produces a particular transcriptional response, making it unsuitable for applications where mechanistic understanding is the goal rather than phenomenological prediction. In these regimes, specialized methods that incorporate prior biological knowledge (pathway databases, protein interaction networks, transcription factor binding models) or foundation models with demonstrated cross-domain transfer may be preferable, even if their transcriptomic accuracy on benchmarks like the VCC is lower.