acceptodds
Under review as a conference paper at ICLR 2027

Single-cell Gene Expression Generation with Hybrid Discrete-Continuous Diffusion

Abstract

Single-cell gene expression data is an important pathway to understanding the fundamentals of various cellular applications, like disease tracking and drug targeting. Obtaining high-quality single-cell data is challenging due to the high cost compared to bulk sequencing. This has led to the use of generative models to create synthetic data. The goal is to efficiently generate realistic cells, as that is vital for stronger downstream analysis for applications like cell trajectories and differentiations. Current directions either use purely continuous or purely discrete diffusion to tackle this challenge. Using a purely continuous method allows the model to learn the joint expression levels through a continuous score function. This lets the model learn any potential cellular connections between each gene expression. Using a discrete method respects the sparse nature of the data, removing any representational mismatch that occurs from working in a continuous to discrete space. We propose to leverage the best of both worlds through : single-cell continuous and discrete diffusion. We highlight the differences between continuous and discrete diffusion through a case study, pinpointing the exact differences in the generated cell profiles under discrete, continuous, and hybrid settings. We demonstrate that scCANDI achieves strong performance against current state-of-the-art masked discrete diffusion methods while requiring \textbf{5-10\times} fewer function evaluations.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
diffusion models, AI for physical sciences

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ociXOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents scCANDI, a hybrid discrete–continuous diffusion framework for generating single-cell gene expression profiles. The method combines discrete diffusion, which better matches the sparse and discrete nature of count data, with continuous updates that jointly refine multiple gene positions and aim to improve sampling efficiency. The framework is evaluated on three single-cell datasets under unconditional and conditional generation settings. Results show that scCANDI provides a trade-off between the MMD performance of discrete diffusion and the Wasserstein-distance performance of continuous diffusion, while requiring substantially fewer function evaluations than the discrete baseline.

Critical strengths

A main strength is the paper’s clear motivation for combining discrete and continuous diffusion. The authors identify complementary properties of the two approaches and provide an empirical analysis showing that they behave differently under MMD and Wasserstein distance. This makes the hybrid design intuitive and provides some insight into why the two diffusion paradigms may be useful in different regimes.

Critical weaknesses
  1. The methodological novelty appears limited. scCANDI largely adapts the existing CANDI hybrid discrete–continuous diffusion framework to single-cell gene expression using the representation and setup of DCM. The paper therefore needs to more clearly establish what methodological insight is introduced beyond combining these existing components.

  2. The evaluation is insufficient to support the claim of improved single-cell generation. The experiments rely mainly on MMD, Wasserstein distance, and cell-wise mean expression distributions. These results establish a trade-off between discrete and continuous diffusion, but do not demonstrate that the generated cells are biologically more faithful. Important properties such as marker expression, gene–gene relationships, biological programs, cell identity, differential expression, or downstream utility are not evaluated. This is particularly important because one of the main motivations for the continuous component is its ability to jointly model gene relationships, yet this property is not directly tested. The claimed efficiency improvement is also based on the number of function evaluations rather than actual runtime or computational cost.

  3. There is a noticeable mismatch between the scope suggested by the figures and the experiments actually presented. Figure 1 explicitly presents trajectory inference, gene regulatory network inference, and perturbation-effect analysis as downstream applications of the generated cells, yet none of these tasks is evaluated in the paper. The empirical study instead remains almost entirely at the level of generic distributional metrics. This makes the presentation suggest broader biological utility than is actually demonstrated.

Questions for authors
  1. What is the main methodological contribution of scCANDI beyond adapting the existing CANDI hybrid diffusion framework to the single-cell representation and pipeline used by DCM?

  2. Can the authors provide biological-fidelity evaluations showing whether scCANDI better preserves properties such as marker expression, gene–gene correlations, biological programs, cell identity, or conditional perturbation effects? Such results would materially affect whether the current improvements in MMD and Wasserstein distance translate into better synthetic cells.

  3. The paper reports a 5–10× reduction in function evaluations. Does this correspond to a comparable improvement in wall-clock generation time or total computational cost, especially considering the additional continuous updates and the need to tune temperature and NFE settings?

2 comments

Thank you for your comments, we have addressed your concerns in order.

  1. Our contribution is not the methodology, but utilizing scCANDI to better understand what the continuous and discrete updates mean in the context of single-cell generation. CANDI was introduced in a language modeling context, which is quite different from single cell generation. We believe that it is valuable to discuss how we adapted the model for the application and what are some important features, like inspecting the distribution fidelity and temperature tuning.

  2. We will add gene-gene correlations. We have included our current results below with some analysis on what we are observing.

  3. The wall-clock generation is slower for scCANDI compared to DCM. With the same hardware and batch size the ms/NFE is around 163 for scCANDI and 119 for DCM. This is around a 1.4x cost increase. However, since scCANDI requires less NFEs to reach their comparable results, the actual cost of wall-clock generation time is slightly faster. We also wanted to add that the temperature tuning is often straightforward to adapt with slight decreases of 0.1 in either direction will result in a low NFE requirement for good results.

Summary of new results:

Runtime measurements at equal resources: scCANDI 163 ms/nfe, DCM 119 ms/nfe

Gene-gene correlation: For each dataset we compute the gene-gene Pearson correlation matrix (top-200 most variable genes) in real and in generated cells, then report the Pearson correlation between the two matrices across all gene pairs. General correlation pools all cells while cell-type correlation computes the same within each of the largest cell types and averages (3 largest cell types for Dentate Gyrus and pbmc3k, 4 cell types for pbmc_covid). We present the result in order of scCANDI, DCM, real-data.

Dentate Gyrus (conditional 1465 cells)

General: 0.915 ± 0.001, 0.844 ± 0.002, 0.988

Cell-type: 0.242 ± 0.010, 0.002 ± 0.001, 0.778

pbmc3k (conditional 1319 cells)

General: 0.790 ± 0.003, 0.908 ± 0.000, 0.939

Cell-type: 0.073 ± 0.003, 0.001 ± 0.003, 0.272

pbmc_covid (unconditional 5000 cells)

General: 0.902 ± 0.001, 0.011 ± 0.001, 0.962

Cell-type: 0.320 ± 0.024, −0.002 ± 0.003, 0.488

scCANDI preserves cell-type co-expression better than DCM on all three datasets. On the conditional datasets, DCM achieves a higher general correlation, which reflects reproducing each cell type's average expression profile rather than co-expression (it does not exceed a random average of real cells). On the unconditional dataset, scCANDI is higher on both. Cell-type co-expression measures how genes vary together across individual cells of the same type. For example, monocyte cells with a higher S100A8 also express a higher S100A9. scCANDI captures part of this structure because genes are revealed gradually and its continuous updates use the genes already revealed, so later genes are generated consistently with them. DCM reveals nearly all genes at once, so its genes are effectively sampled independently within a cell type. These results are consistent with our motivation of using continuous updates to guide generation based on revealed genes (the discrete anchors).

Thank you for the detailed response and additional experiments. I appreciate the new gene–gene correlation analysis, which provides more biologically meaningful evidence, as well as the clarification of the computational efficiency.

However, my main concern regarding novelty remains. The authors acknowledge that the contribution is not methodological, while the additional biological analyses, although useful, do not yet establish a sufficiently strong scientific contribution beyond applying an existing framework to single-cell data. In particular, it remains unclear what new capabilities or insights scCANDI provides compared with existing generative approaches more broadly.

Overall, while the rebuttal has partially addressed my concerns, I am not sufficiently convinced to change my original assessment. I will maintain my score.

Reviewer_N6RHOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

scCANDI adapts hybrid discrete–continuous diffusion to single-cell gene-expression generation. It represents binned counts as one-hot gene tokens, combines discrete masked positions with Gaussian-corrupted positions, and alternates discrete resampling with continuous probability-flow updates. The paper evaluates unconditional and label-conditioned generation on PBMC3K, Dentate Gyrus, and PBMC COVID, comparing with DCM and discrete-only and continuous-only ablations using MMD, 2-Wasserstein distance, and sparsity summaries. It reports a fidelity–compute trade-off and comparable results to DCM with 5–10× fewer function evaluations.

Critical strengths

1 The method combines discrete anchors, which preserve count-valued structure, with continuous score updates that jointly refine corrupted positions. This is a clear and plausible adaptation of hybrid diffusion to sparse gene-expression data.

2 The evaluation spans three datasets, conditional and unconditional generation, discrete/continuous ablations, and NFE and temperature sweeps. These comparisons offer useful initial evidence about the trade-off between distribution metrics and inference steps.

Critical weaknesses

1 The evidence for realistic and useful synthetic cells is limited. The evaluation relies on MMD, Wasserstein distance, and marginal summaries such as sparsity and mean expression; it does not test downstream tasks or other biological properties, despite motivating applications such as trajectory and gene-regulatory-network inference. The sample count is also unclear: the setup says 100 generated cells per experiment, while Figure 3b refers to 200.

2 The central quality and efficiency comparisons need clarification. Table 2 reports Dentate Gyrus MMD of 0.36 for scCANDI versus 0.022 for DCM, which is difficult to reconcile with the text’s characterization of the results as comparable. Table 3 calls the PBMC3K conditions “8 perturbations,” while the dataset description lists 8 cell types. In addition, the 5–10× claim compares selected scCANDI NFE counts with DCM’s standard 100-step setup; the paper does not report matched runtime or the exact temperature settings used in the main tables, despite Figure 4 showing strong temperature sensitivity.

Questions for authors

1 Is the Dentate Gyrus MMD value for scCANDI in Table 2 (0.36) correct? How should it be interpreted relative to DCM’s 0.022 and the claim of comparable fidelity?

2 Are the eight PBMC3K conditioning labels cell types or perturbations? Please clarify the sample counts used for Tables 2–3, given the difference between the experimental setup and Figure 3b.

3 What temperature settings were used for the main comparisons, and does the 5–10× efficiency claim refer only to NFE counts? If runtime measurements already exist, please clarify how they compare under the same hardware and batch size.

2 comments

Thank you for your comments, we have addressed your concerns in order.

  1. The MMD values are correct. We looked at gene-gene correlation. For the Dentate Gyrus dataset, scCANDI outperformed DCM on both metrics (the specific details are attached below). Overall, our takeaway most of the results were quite close in terms of MMD and WD leading to our claim of comparable fidelity. However, for this specific dataset there is indeed a large gap for the MMD values. Based on the results from the gene-gene correlation, scCANDI is able to find much better correlations between cell-type gene relationships compared to DCM from the continuous updates.

  2. They are cell types. We will adjust the manuscript to fix any mistakes regarding this. The sample counts used to Tables 2-3 are consistent with the experimental setup, only Figure 3b did we use a higher cell count. We update the manuscript to make this clear.

  3. The temperature settings used in the main comparisons was 0.9. And the 5-10x efficiency claim was only referring to the NFE counts. However, we have run experiments with the same hardware and batch size. The results are around scCANDI 163 ms/nfe, DCM 119 ms/nfe

Summary of new results:

Runtime measurements at equal resources: scCANDI 163 ms/nfe, DCM 119 ms/nfe

Gene-gene correlation: For each dataset we compute the gene-gene Pearson correlation matrix (top-200 most variable genes) in real and in generated cells, then report the Pearson correlation between the two matrices across all gene pairs. General correlation pools all cells while cell-type correlation computes the same within each of the largest cell types and averages (3 largest cell types for Dentate Gyrus and pbmc3k, 4 cell types for pbmc_covid). We present the result in order of scCANDI, DCM, real-data.

Dentate Gyrus (conditional 1465 cells)

General: 0.915 ± 0.001, 0.844 ± 0.002, 0.988

Cell-type: 0.242 ± 0.010, 0.002 ± 0.001, 0.778

pbmc3k (conditional 1319 cells)

General: 0.790 ± 0.003, 0.908 ± 0.000, 0.939

Cell-type: 0.073 ± 0.003, 0.001 ± 0.003, 0.272

pbmc_covid (unconditional 5000 cells)

General: 0.902 ± 0.001, 0.011 ± 0.001, 0.962

Cell-type: 0.320 ± 0.024, −0.002 ± 0.003, 0.488

scCANDI preserves cell-type co-expression better than DCM on all three datasets. On the conditional datasets, DCM achieves a higher general correlation, which reflects reproducing each cell type's average expression profile rather than co-expression (it does not exceed a random average of real cells). On the unconditional dataset, scCANDI is higher on both. Cell-type co-expression measures how genes vary together across individual cells of the same type. For example, monocyte cells with a higher S100A8 also express a higher S100A9. scCANDI captures part of this structure because genes are revealed gradually and its continuous updates use the genes already revealed, so later genes are generated consistently with them. DCM reveals nearly all genes at once, so its genes are effectively sampled independently within a cell type. These results are consistent with our motivation of using continuous updates to guide generation based on revealed genes (the discrete anchors).

Thank you for the response. After reviewing, I decide to keep my scores.

Reviewer_TMfsOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper applies a hybrid discrete-continuous diffusion framework to single-cell gene expression generation on top of the DCM pipeline, yielding scCANDI. The discrete process preserves the sparse count structure of expression data, while the continuous process enables joint updates across genes. Ablations isolating each process reveal an MMD/WD trade-off, and scCANDI is shown to balance the two. Experiments show that scCANDI achieves fidelity comparable to DCM while requiring 5–10× fewer function evaluations.

Critical strengths

1.This work applies hybrid discrete and continuous diffusion to single-cell data. The ablation study isolates the respective contributions of the discrete and continuous components and identifies a tradeoff between MMD and WD.

2.Practical efficiency gain. Comparable or better WD than DCM at 5–10× lower NFE, with consistent trends across three datasets and both conditional/unconditional settings.

Critical weaknesses

1.Only about 100 generated cells were compared with a subset of real cells. DCM was the only strong baseline (such as CFGen/scLDM were not compared)

2.Although downstream biological analysis (trajectory analysis, perturbation prediction) was listed as the original intention of the study, no related experiments were actually conducted.

3.MMD is often markedly worse than DCM, and the authors admit that a well-tuned DCM can match the real distribution without continuous refinement. NFE savings remain the main benefit, but this is weakened by the required per-dataset temperature and NFE sweeps.

Questions for authors

See weaknesses.

4 comments

Thank you for your comments, we have addressed your concerns in order.

  1. We mainly wanted to focus on discrete methodologies for this problem, but we will add more baselines, within the timeframe. We currently have unconditional generation results of scLDM on the pbmc3k dataset. Their results are close with an MMD of 0.01 and WD of 0.13, they appear to have a better MMD but weaker WD. We will run as many baselines as we can.

  2. We will revise the paper to make it more clear that our goal was to improve the overall efficiency such that downstream analysis can be faster. However, we have added some gene-gene correlation analysis below.

  3. The MMD performance for scCANDI is often worse than DCM. However, from our gene-gene correlation results, we can see that scCANDI is able to capture the cell-type correlations better compared to DCM. This case highlights the main benefit of utilizing the continuous refinement as the pure discrete methodology will unmask independently ignoring crucial context dependent information between cell types and specific gene relationships. We have also provided more concrete runtime advantages below.

Summary of new results:

Runtime measurements at equal resources: scCANDI 163 ms/nfe, DCM 119 ms/nfe

Gene-gene correlation: For each dataset we compute the gene-gene Pearson correlation matrix (top-200 most variable genes) in real and in generated cells, then report the Pearson correlation between the two matrices across all gene pairs. General correlation pools all cells while cell-type correlation computes the same within each of the largest cell types and averages (3 largest cell types for Dentate Gyrus and pbmc3k, 4 cell types for pbmc_covid). We present the result in order of scCANDI, DCM, real-data.

Dentate Gyrus (conditional 1465 cells)

General: 0.915 ± 0.001, 0.844 ± 0.002, 0.988

Cell-type: 0.242 ± 0.010, 0.002 ± 0.001, 0.778

pbmc3k (conditional 1319 cells)

General: 0.790 ± 0.003, 0.908 ± 0.000, 0.939

Cell-type: 0.073 ± 0.003, 0.001 ± 0.003, 0.272

pbmc_covid (unconditional 5000 cells)

General: 0.902 ± 0.001, 0.011 ± 0.001, 0.962

Cell-type: 0.320 ± 0.024, −0.002 ± 0.003, 0.488

scCANDI preserves cell-type co-expression better than DCM on all three datasets. On the conditional datasets, DCM achieves a higher general correlation, which reflects reproducing each cell type's average expression profile rather than co-expression (it does not exceed a random average of real cells). On the unconditional dataset, scCANDI is higher on both. Cell-type co-expression measures how genes vary together across individual cells of the same type. For example, monocyte cells with a higher S100A8 also express a higher S100A9. scCANDI captures part of this structure because genes are revealed gradually and its continuous updates use the genes already revealed, so later genes are generated consistently with them. DCM reveals nearly all genes at once, so its genes are effectively sampled independently within a cell type. These results are consistent with our motivation of using continuous updates to guide generation based on revealed genes (the discrete anchors).

Thank you for the authors' responses.

They have addressed some of my concerns. However, most of these points are listed by the authors as future plans. Without corresponding results, the statement still cannot support the related claims. I will keep my score. Best wishes!

We have run the additional continuous baselines, scLDM and scDiffusion. In addition to the original datasets, we look at a larger dataset HLCA (584,944 cells) compared with 200 generated cells. We note the following:

scLDM has a slight edge on smaller datasets (Dentate Gyrus and PBMC3K) while scCANDI is comparable and slightly better on larger datasets (PBMC Covid and HLCA).

scLDM and scDiffusion both use an autoencoder with continuous diffusion setup. With the variable performance in the Dentate Gyrus and PBMC3K, it appears like the main benefit comes from the autoencoder. From CCDD, using a contextual autoencoder is compatible with hybrid diffusion, and we will look into this in the future.

Table 1: Unconditional Generation

DatasetMethodWDMMD
Dentate GyrusscCANDI1.42 ± 0.050.36 ± 0.01
DCM1.48 ± 0.000.022 ± 0.01
scLDM0.73 ± 0.270.018 ± 0.01
scDiffusion4.31 ± 0.520.271 ± 0.004
PBMC3KscCANDI0.08 ± 0.010.02 ± 0.00
DCM0.11 ± 0.010.03 ± 0.01
scLDM0.04 ± 0.0010.01 ± 0.000
scDiffusion0.09 ± 0.0010.06 ± 0.002
PBMC COVIDscCANDI0.07 ± 0.010.02 ± 0.00
DCM0.13 ± 0.010.05 ± 0.01
scLDM0.12 ± 0.010.04 ± 0.00
scDiffusion0.13 ± 0.010.06 ± 0.01
HLCAscCANDI0.12 ± 0.0030.02 ± 0.000
DCM0.17 ± 0.0020.05 ± 0.001
scLDM0.13 ± 0.0110.02 ± 0.004
scDiffusion0.15 ± 0.0090.03 ± 0.003

Table 2: Conditional Generation

DatasetMethodWDMMD
PBMC3KscCANDI0.34 ± 0.050.07 ± 0.02
DCM0.39 ± 0.140.04 ± 0.00
scLDM0.30 ± 0.010.01 ± 0.000
scDiffusion0.45 ± 0.010.01 ± 0.000
PBMC COVIDscCANDI0.092 ± 0.020.07 ± 0.03
DCM0.094 ± 0.050.09 ± 0.09
scLDM0.144 ± 0.0270.03 ± 0.002
scDiffusion0.112 ± 0.0060.03 ± 0.003

Thank you for the detailed response. After reviewing, I will maintain my score.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.