acceptodds
Under review as a conference paper at ICLR 2027

scMOD: Completing and Generating the Virtual Cell with Trimodal Latent Diffusion

Abstract

Virtual cells aim to simulate cellular states, but a faithful model must capture multiple molecular modalities and operate when any subset is missing. Existing translation methods are tied to predefined assay pairs and require an observed source. Shared-latent VAEs provide broader coverage, but aligning heterogeneous assays within one latent space can obscure modality-specific information, while their simple priors limit de novo generation. We introduce scMOD, to our knowledge the first latent diffusion model for trimodal RNA, chromatin accessibility (ATAC), and surface protein (ADT) data. scMOD combines assay-specific variational codecs with a shared diffusion Transformer trained by block-causal flow matching over randomized modality orders. The first block learns marginal generation, and subsequent blocks learn conditional generation from a clean multimodal prefix. This design allows a single checkpoint to perform pairwise, one-to-many, many-to-one, and unconditional generation while retaining assay-specific structure. Across TEA-seq and DOGMA-seq, scMOD achieves the highest average Pearson correlation among the evaluated multimodal VAE baselines, leads specialized translators in several directions, and improves unconditional RNA and ADT generation over prior-sampled baselines. scMOD provides a unified generative framework for completing and generating multimodal virtual cells.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Virtual Cell, Latent Diffusion, Multimodal Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_3iDAOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a framework for jointly analyzing multi-modal single-cell data, specifically focus on data sets containing three types of data: RNA-seq, ATAC-seq, and surface protein data. This is the first paper to do this kind of analysis on this particular set of three data types. Furthermore, the proposed method uses generative techniques, including that diffusion and flow matching, that have previously only been applied to RNA-seq but not multi-omics data. The paper shows empirically that the proposed model, called scMOD, outperforms state-of-the-art methods in a variety of settings.

Critical strengths

Empirically, the results are strong, and the experiments appear to be well done. I am convinced that the method delivers better performance on the stated tasks than the state-of-the-art methods that you compared against. The ablations also support this conclusion.

The paper does a very nice job of clearly laying out the proposed modeling framework, providing general background as well as the details of the proposed model. I appreciated the attention to detail in the appendix.

Critical weaknesses

I don't think this paper is written to be accessible to an ICLR audience. The ML parts are fine, but I don't think you have done a good job of explaining what the data is or motivating why these problems are interesting or important. For example, I work in this field but was not familiar with the cell surface marker data that you are working with here. A typical ICLR reader will not even know much about RNA-seq or ATAC-seq, much less ADT. (Indeed, what does "ADT" even stand for?) A background section should have been provided, or the introduction should have been expanded, to explain what this data is. Furthermore, I'm not convinced that the problems you are trying to solve are important. Yes, translating from one data type to another is potentially useful, but I can't imagine a scenario in which unconditional generation of multi-omic data is a worthwhile thing to do. Again, providing background and motivation would have helped here.

The other main weakness is that you haven't really separated out the two potential avenues of novelty here. One avenue is your particular modeling paradigm, and the other is its application to this particular trimodality of data types. If you want to argue that the paradigm is generally useful, then applying it to some other data types (e.g., NEAT-seq, scNMT-seq, or scTrio-seq2) would be more convincing.

Questions for authors

I have no questions, other than the ones implied above:

Why is unconditional generation of multi-omics data interesting or useful?

How well does this approach work with other combinations of data types?

4 comments
AuthorsOpenReview ↗

Response to Reviewer 3iDA (1/2)

Thank you for highlighting the need to better explain the biological motivation and the scope of our claims. In response, we have revised the Introduction and Background to make the RNA–ATAC–ADT setting more accessible to a general ICLR audience and to clarify the motivation for multimodal completion and generation.


Weakness 1: "I don't think this paper is written to be accessible to an ICLR audience. The ML parts are fine, but I don't think you have done a good job of explaining what the data is or motivating why these problems are interesting or important. For example, I work in this field but was not familiar with the cell surface marker data that you are working with here. A typical ICLR reader will not even know much about RNA-seq or ATAC-seq, much less ADT. (Indeed, what does ADT even stand for?) A background section should have been provided, or the introduction should have been expanded, to explain what this data is."

Response.

Thank you for this suggestion. We have added a dedicated background subsection, Single-cell multi-omics measurements (Section 2.1), that introduces the three modalities and the two benchmark assays in terms accessible to an ML audience.

The three modalities provide complementary views of the same cell. RNA sequencing (RNA-seq) [1] measures gene expression through transcript counts. The assay for transposase-accessible chromatin using sequencing (ATAC-seq) [2] measures which genomic regions are accessible, providing a view of regulatory potential. Antibody-derived tags (ADTs) [3] use DNA-barcoded antibodies to measure selected cell-surface proteins, which help characterize cell identity and phenotype. For the model, each modality is a count vector over genes, genomic regions, or antibody targets.

TEA-seq and DOGMA-seq measure all three profiles in the same cell. They therefore allow us to test both whether scMOD recovers an unmeasured profile and whether generated profiles describe a coherent cell. The revised Introduction now connects this biological setting directly to the completion and generation tasks.

References:

  • [1] Fuchou Tang, Catalin Barbacioru, Yangzhou Wang, Ellen Nordman, Clarence Lee, Nanlan Xu, Xiaohui Wang, John Bodeau, Brian B. Tuch, Asim Siddiqui, Kaiqin Lao, and M. Azim Surani. mrna-seq whole-transcriptome analysis of a single cell. Nature Methods.
  • [2] Jason D. Buenrostro, Beijing Wu, Ulrike M. Litzenburger, Dave Ruff, Michael L. Gonzales, Michael P. Snyder, Howard Y. Chang, and William J. Greenleaf. Single-cell chromatin accessibility reveals principles of regulatory variation. Nature.
  • [3] Marlon Stoeckius, Christoph Hafemeister, William Stephenson, Brian Houck-Loomis, Pratip K. Chattopadhyay, Harold Swerdlow, Rahul Satija, and Peter Smibert. Simultaneous epitope and transcriptome measurement in single cells. Nature Methods.
AuthorsOpenReview ↗

Response to Reviewer 3iDA (2/2)

Weakness 2: "Furthermore, I'm not convinced that the problems you are trying to solve are important. Yes, translating from one data type to another is potentially useful, but I can't imagine a scenario in which unconditional generation of multi-omic data is a worthwhile thing to do. Again, providing background and motivation would have helped here."

Question 2: "Why is unconditional generation of multi-omics data interesting or useful?"

Response.

Thank you for this important question. Our motivation comes from the broader goal of building virtual cells. Fully paired RNA–ATAC–ADT measurements are expensive and relatively scarce, while many studies measure only one or two of these molecular layers. Completion therefore has an immediate use: for example, inferring a cell’s surface proteins when only its RNA has been measured. Methods such as BABEL demonstrate the value of recovering such unmeasured molecular information.

A virtual-cell model should also capture the diversity of states that real cells can occupy. Even cells of the same type differ in which genes are active and how much of each protein they produce. Modeling this variation, together with the relationships between molecular layers, is a foundation for simulating cellular states. Unconditional generation directly tests this ability: the model must produce realistic, varied cells whose RNA, chromatin, and protein profiles fit together, without being given a measured profile to guide each sample. This provides a basis for future conditional simulations of how cellular states vary with cell type, disease, or perturbation, as envisioned in virtual-cell modeling.

This motivation is shared by established single-cell generation work. scLDM [1] explicitly studies unconditional RNA generation, while scDiffusion demonstrates generation of specified cell types and developmental trajectories. Generation also has practical downstream value: CFGen shows that adding generated examples helps models recognize rare cell types. Our work scMOD extends this direction to a richer description of the cell, jointly modeling gene expression, chromatin accessibility, and surface proteins in the same framework used for completion. Our results show both realistic molecular profiles (Table R1) and stronger agreement in the cell identity inferred from different modalities (Table R5), supporting this step toward multimodal virtual cells. All the references can be found in Section 2.

References:

[1] Giovanni Palla, Sudarshan Babu, Payam Dibaeinia, James D. Pearce, Donghui Li, Aly A. Khan, Theofanis Karaletsos, and Jakub M. Tomczak. Scalable Single-Cell Gene Expression Generation with Latent Diffusion Models, 2026.


Weakness 3: "The other main weakness is that you haven't really separated out the two potential avenues of novelty here. One avenue is your particular modeling paradigm, and the other is its application to this particular trimodality of data types. If you want to argue that the paradigm is generally useful, then applying it to some other data types (e.g., NEAT-seq, scNMT-seq, or scTrio-seq2) would be more convincing."

Question 3: "How well does this approach work with other combinations of data types?"

Response.

Thank you for this insightful suggestion. Our modeling formulation separates assay-specific representations from shared multimodal generation. Modality-specific codecs map heterogeneous measurements into latent blocks, and a shared block-causal flow model learns how to generate those blocks from the available modalities. Randomized modality ordering enables the same model to handle different observed and target sets. This modular design also provides a route to other assay combinations: a new measurement type can be represented by a suitable codec while retaining the block-causal training and sampling formulation.

Your suggestions of NEAT-seq, scNMT-seq, and scTrio-seq2 are particularly interesting extensions of this design. We appreciate these concrete suggestions, which would broaden the biological scope of scMOD and provide a valuable test of its formulation beyond the current modality combination. At present, we do not have suitable datasets for these additional assays with comparable cell numbers and the paired measurements required by our evaluation protocol. Obtaining appropriate datasets and adapting the modality-specific codecs is therefore an important direction for future work.

I'm not really convinced here. The addition of a paragraph describing the dataset will help ICLR readers, but the paper really needs to be substantially revised to make it accessible. As for unconditional generation of multi-omics data, the only concrete example you gave is imputing a missing modality, but of course, that's not unconditional. I agree that imputing a missing data type is useful, but imputing all three at once unconditionally seems like an interesting but not particularly useful exercise (despite the fact that others have published papers doing this). The third point wasn't really addressed, other than to say you will do it in the future.

AuthorsOpenReview ↗

2nd Response to Reviewer 3iDA

"I'm not really convinced here. The addition of a paragraph describing the dataset will help ICLR readers, but the paper really needs to be substantially revised to make it accessible. As for unconditional generation of multi-omics data, the only concrete example you gave is imputing a missing modality, but of course, that's not unconditional. I agree that imputing a missing data type is useful, but imputing all three at once unconditionally seems like an interesting but not particularly useful exercise (despite the fact that others have published papers doing this). The third point wasn't really addressed, other than to say you will do it in the future."

Response.

We thank the reviewer for clarifying the remaining concerns. We appreciate the opportunity to further explain the biological motivation, the role of unconditional generation, and the scope of our evaluation.

Improved readability towards general audience. We carefully revised the opening of the Introduction to explain what RNA, ATAC, and ADT measure before introducing the modeling tasks. It explains why we model them together using the example of cells with similar RNA profiles but different DNA accessibility or surface protein levels. Section 2.1 now describes the three count vectors and explains that a paired sample links measurements from the same cell.

Motivation for unconditional generation. Unconditional generation tests how well a model captures the data distribution without an observed input. For example, in image diffusion, DDPM [1] uses FID to evaluate unconditional CIFAR-10 samples, indicating that unconditional generation can serve as a widely recognized metric for measuring a method’s effectiveness.

We apply the same idea to joint RNA, ATAC, and ADT profiles. Instead of FID, which is used in image generation, we follow prior work and use metrics such as W2, MMD2, and 1-NN, which are also common in this field. An unconditional scMOD sample contains three profiles, so we also test whether they resemble measurements taken together in real cells. On TEA-seq, joint generation gives a lower joint MMD$^2$ against real paired cells than independent generation with profiles re-paired within the same predicted cell type (Table R6).

References:

[1] Ho, J., Jain, A., & Abbeel, P. (2020). Denoising diffusion probabilistic models. Advances in neural information processing systems, 33, 6840-6851.

Generality across assay combinations. We do appreciate the reviewer highlighting some useful data and potential applications in other subfields. These assays provide measurements paired within cells, but the measurements differ from those in our benchmarks. NEAT-seq targets nuclear proteins, while scNMT-seq and scTrio-seq2 profile DNA methylation, and scTrio-seq2 also reports detailed alterations. Testing these combinations would require preparing matched profiles, adapting new codecs, evaluating the model on held-out paired cells, and determining which metrics capture performance in these new settings. Our work can theoretically be reused in such new settings, but learning and evaluating a joint distribution on such new tasks still requires suitable measurements paired within cells. In the final phase of the rebuttal, we acknowledge that these settings can only be left to future work and appreciate your understanding.

Reviewer_sfN6OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes scMOD, a latent diffusion model for trimodal single-cell data covering RNA, ATAC, and ADT. Each modality is encoded by a separate VAE into latent tokens, and a shared diffusion Transformer is trained on these latents. Modalities are randomly ordered during training, and a block-causal attention mask lets each noisy block attend only to the clean latents of the preceding modalities, so that one model learns both marginal and conditional generation. Experiments on TEA-seq and DOGMA-seq evaluate unconditional generation against MIDAS and MultiVI, and six pairwise prediction directions against multimodal VAEs and translation methods.

Critical strengths
  • The paper addresses a relevant problem: completing and generating multimodal single-cell profiles with a single model.
  • The method and evaluation protocols are clearly described, and the baselines are matched in parameter count.
  • Unconditional generation improves consistently over prior sampling in multimodal VAEs.
Critical weaknesses
  1. Limited novelty and missing related work. Latent diffusion over modality-specific autoencoders, with a shared denoiser that conditions on any subset of modalities, was proposed in MLD (Bounoua et al., 2024) and UniDiffuser (Bao et al., 2023). The block-causal layout with noisy and clean copies follows Block Diffusion (Arriola et al., 2025). Multimodal diffusion models for single-cell data also exist, such as scDiffusion-X (Luo et al., 2026) and CFGen (Palma et al., 2024). None of these works is cited.

  2. The gains in cross-modal prediction are small. The average improvements over MIDAS are reported without variance across training seeds, and MultiVI is better on both ADT-target directions.

  3. Several claims are not evaluated. The paper motivates the model by completion from any subset of modalities and by learning from incomplete assay combinations. The experiments cover only pairwise directions and unconditional generation on fully paired datasets. In unconditional generation, the metrics are computed per modality, so the coherence of jointly generated profiles is not assessed.

  • Arriola, M., Gokaslan, A., Chiu, J., Yang, Z., Qi, Z., Han, J., ... & Kuleshov, V. (2025, May). Block diffusion: Interpolating between autoregressive and diffusion language models. In International Conference on Learning Representations.
  • Bao, F., Nie, S., Xue, K., Li, C., Pu, S., Wang, Y., ... & Zhu, J. (2023, July). One transformer fits all distributions in multi-modal diffusion at scale. In International Conference on Machine Learning (pp. 1692-1717). PMLR.
  • Bounoua, M., Franzese, G., & Michiardi, P. (2024). Multi-modal latent diffusion. Entropy, 26(4), 320.
  • Luo, E., Wei, L., Hao, M., Zhang, X., & Liu, Q. (2026). A multi-modal diffusion model with dual-cross-attention for multi-omics data generation and translation. Nature Communications.
  • Palma, A., et al. (2024). Multi-modal and multi-attribute generation of single cells with CFGen. arXiv:2407.11734.
Questions for authors
  1. How does scMOD perform on many-to-one and one-to-many completion, and when trained on data where some cells miss modalities? Appendix C mentions a multi-source PBMC experiment that is not reported.
  2. How much of the gain comes from the diffusion prior and how much from the modality-specific autoencoders? For example, how does a diffusion model trained on MIDAS latents perform?
  3. Are jointly generated modalities consistent with each other, for example in the cell type assigned from each modality?
5 comments
AuthorsOpenReview ↗

Response to Reviewer sfN6 (1/3)

Thank you for your careful and detailed review. Below we respond to each of your comments in turn.


Weakness 1: "Limited novelty and missing related work. Latent diffusion over modality-specific autoencoders, with a shared denoiser that conditions on any subset of modalities, was proposed in MLD (Bounoua et al., 2024) and UniDiffuser (Bao et al., 2023). The block-causal layout with noisy and clean copies follows Block Diffusion (Arriola et al., 2025). Multimodal diffusion models for single-cell data also exist, such as scDiffusion-X (Luo et al., 2026) and CFGen (Palma et al., 2024). None of these works is cited."

Response.

Thank you for these references. After reviewing them, we have revised the related-work section to cite and discuss all five.

MLD, UniDiffuser and Block Diffusion are image and text models. Their autoencoders and tokenizers are built for pixels and words, and none of them models the count measurements of RNA, ATAC and ADT, so they cannot be applied to single-cell multiomics directly.

scDiffusion-X and CFGen are indeed single-cell models, and we added both as baselines (Tables R1 and R2). However, each covers only part of what scMOD does. scDiffusion-X models only RNA and ATAC, so it cannot generate or predict ADT. CFGen is also limited to RNA and ATAC, and it only generates cells without predicting a missing modality from an observed one.

To our knowledge, scMOD remains the first latent diffusion model to jointly model RNA, ATAC and ADT within a single generative framework. With both methods added to the comparison, scMOD is still the strongest model overall. It leads RNA generation on both datasets, translates between RNA and ATAC more accurately than scDiffusion-X, and generates more consistent RNA–ATAC pairs than either method.

Table R1. Unconditional generation. Values are means over three seeds. Details in the general response.

TEA-seq

ModalityMetricscMOD SinglescMOD JointMIDASMultiVIscDiffusion-XCFGen
RNAW2↓W_2 \downarrow10.1310.0813.4214.2111.4711.49
MMD$^2 \downarrow$0.00840.00860.03160.05520.01530.0104
1-NN →0.5\rightarrow 0.50.7110.7030.8910.9320.7560.795
ADTW2↓W_2 \downarrow7.467.418.357.76----
MMD$^2 \downarrow$0.01000.01020.04970.0151----
1-NN →0.5\rightarrow 0.50.6040.6020.8320.700----
ATACW2↓W_2 \downarrow16.8516.9424.9834.2318.4518.66
MMD$^2 \downarrow$0.03500.03320.07000.13640.04580.0463
1-NN →0.5\rightarrow 0.50.8740.8730.8950.9570.8290.814

DOGMA-seq

ModalityMetricscMOD SinglescMOD JointMIDASMultiVIscDiffusion-XCFGen
RNAW2↓W_2 \downarrow11.3711.2915.1216.7113.7913.06
MMD$^2 \downarrow$0.00540.00570.03860.07430.01820.0068
1-NN →0.5\rightarrow 0.50.7650.7620.9120.9660.8340.851
ADTW2↓W_2 \downarrow8.508.5214.6713.10----
MMD$^2 \downarrow$0.02650.02940.17970.1214----
1-NN →0.5\rightarrow 0.50.7000.6970.9160.845----
ATACW2↓W_2 \downarrow18.1618.2528.4635.4820.2818.26
MMD$^2 \downarrow$0.07830.08180.08440.16240.05430.0292
1-NN →0.5\rightarrow 0.50.9030.8990.8570.9300.8110.779

Table R2. RNA–ATAC translation on TEA-seq with scDiffusion-X added (PCC, ↑\uparrow). Details in the general response.

MethodRNA →\rightarrow ATACATAC →\rightarrow RNA
scMOD0.61810.5787
scDiffusion-X (new)0.56020.5403
scButterfly-B0.58600.5894
BABEL0.58440.5800
kNN0.61510.5627
AuthorsOpenReview ↗

Response to Reviewer sfN6 (2/3)

Weakness 2: "The gains in cross-modal prediction are small. ..., and MultiVI is better on both ADT-target directions."

Response.

We appreciate your close reading of the prediction results. As shown in Tables R1 and R5, scMOD generates more realistic cells than MIDAS and MultiVI, with more consistent modalities within each cell. Even on prediction alone, scMOD has the highest average PCC and is the best or close to the best in most directions, and this holds on both datasets (Table 2).

Additionaly, We would like to point out, that our goal is not simply to improve the PCC of each direction's prediction, but to build one model for both unconditional generation and cross-modal prediction in all directions.

For the translation baselines in Table 3, these methods focus on predefined source–target tasks, and extending them to different modality combinations typically requires pair-specific models. They do not offer a unified framework for arbitrary trimodal completion and unconditional generation. In contrast, scMOD natively supports all six prediction directions, multiple observed or target modalities, and unconditional trimodal generation with one checkpoint.


Weakness 3: "Several claims are not evaluated. The paper motivates the model by completion from any subset of modalities and by learning from incomplete assay combinations. The experiments cover only pairwise directions and unconditional generation on fully paired datasets."

Question 1: "How does scMOD perform on many-to-one and one-to-many completion, and when trained on data where some cells miss modalities. ..."

Response.

We value this suggestion. We have added both evaluations on TEA-seq using the same trained model, with many-to-one prediction in Table R3 and one-to-many prediction in Table R4. With two observed modalities, scMOD improves over both single-source predictions for every target and achieves the highest combined PCC among all methods. When generating two missing modalities from one observed modality, scMOD is the best method in most input–target pairs.

Multi-source conditioning (Table R3). With two observed modalities, scMOD achieves the highest PCC among all compared methods for all three targets. It also improves over both of its single-source counterparts in every case. The largest gain is for ADT prediction: RNA+ATAC reaches 0.8690, compared with 0.8570 from RNA alone and 0.8650 from ATAC alone. This shows that scMOD combines complementary information across modalities.

Multi-target generation (Table R4). We evaluate both target orders for each observed modality. As shown in the table, scMOD achieves the highest average PCC. Across all six input–target pairs, changing the generation order changes the PCC by at most 0.0006, showing that scMOD is robust to target ordering.

We have added these results to Appendix D of the revised manuscript.

Table R3. Many-to-one prediction on TEA-seq (PCC, ↑\uparrow). A and B are the observed modalities. Gains are relative percentage increases over the weaker single-source PCC. Bold marks the highest A+B PCC and gain per target.

TargetMethodA onlyB onlyA+B (PCC, gain %)
ADT (A = RNA, B = ATAC)scMOD0.85950.82660.8602 (+4.1%)
MIDAS0.85510.85140.8591 (+0.9%)
MultiVI0.86160.86060.8648 (+0.5%)
RNA (A = ADT, B = ATAC)scMOD0.58270.57870.5834 (+0.8%)
MIDAS0.58260.58560.5853 (+0.5%)
MultiVI0.52090.52460.5239 (+0.6%)
ATAC (A = RNA, B = ADT)scMOD0.61810.61790.6183 (+0.1%)
MIDAS0.58300.58270.5835 (+0.1%)
MultiVI0.54190.54570.5454 (+0.6%)

Table R4. One-to-many prediction on TEA-seq, scored per target (PCC, ↑\uparrow). First and Second indicate whether the scored target is generated before or after the other target. Bold marks the best per row.

InputTargetscMOD FirstscMOD SecondscMOD Avg.MIDASMultiVI
RNAADT0.85950.85950.85950.85510.8616
ATAC0.61810.61810.61810.58300.5419
ADTRNA0.58270.58250.58260.58260.5209
ATAC0.61790.61800.61800.58270.5457
ATACRNA0.57870.57880.57880.58560.5246
ADT0.82660.82720.82690.85140.8606
AuthorsOpenReview ↗

Response to Reviewer sfN6 (3/3)

Weakness 3: "In unconditional generation, the metrics are computed per modality, so the coherence of jointly generated profiles is not assessed."

Question 3: "Are jointly generated modalities consistent with each other, for example in the cell type assigned from each modality?"

Response.

Thanks for raising this point.

Coherence across modalities is indeed central to joint generation, and it is exactly what scMOD is designed to provide.

We have added a cross-modal consistency evaluation that directly tests whether the generated RNA, ADT, and ATAC profiles describe the same cell type (Table R5, see evaluation protocols in General Response). Three single-modality classifiers predict cell types, and Cohen's κ\kappa measures pairwise agreement beyond the random re-pairing reference (κ=0\kappa = 0). Real test cells provide the measured-data reference. scMOD achieves the highest consistency in every comparison and comes closest to the real-data reference. Its three-modality score reaches 0.668, compared with 0.343 for MIDAS and 0.250 for MultiVI. These results show that scMOD preserves a shared cell identity across jointly generated modalities. We have added this evaluation to Section 4.3 and Appendix B.4 of the revised manuscript.

Table R5. Cross-modal consistency of generated cells on TEA-seq (κ\kappa, ↑\uparrow). Values are mean ±\pm standard deviation over three sampling seeds. Details in the general response.

CellsRNA–ADTRNA–ATACADT–ATACAll three
Real test cells0.9410.9240.9280.917
scMOD0.806 ±\pm 0.0030.671 ±\pm 0.0020.671 ±\pm 0.0050.668 ±\pm 0.001
MIDAS0.479 ±\pm 0.0090.407 ±\pm 0.0080.411 ±\pm 0.0030.343 ±\pm 0.004
MultiVI0.302 ±\pm 0.0160.221 ±\pm 0.0190.382 ±\pm 0.0090.250 ±\pm 0.014
scDiffusion-X--0.509 ±\pm 0.010----
CFGen--0.639 ±\pm 0.011----

Question 2: "How much of the gain comes from the diffusion prior and how much from the modality-specific autoencoders? For example, how does a diffusion model trained on MIDAS latents perform?"

Response.

Thanks for this thoughtful comment.

Both components contribute, and they matter for different tasks. The diffusion prior mainly determines generation quality, while the modality-specific autoencoders set how accurate predictions can be. For generation, every model that learns its latent prior (scMOD, scDiffusion-X and CFGen) generates more realistic cells than the Gaussian-prior VAEs MIDAS and MultiVI on both datasets (Table R1).

For prediction, the autoencoders play the larger role. scMOD's ATAC predictions already reach the reconstruction PCC of its ATAC autoencoder (0.618), and the peak-centric ATAC codec improves all but one prediction direction (Sec. 4.4). This is also the key advantage of scMOD over a diffusion model built on MIDAS latents. MIDAS compresses all modalities into one shared latent, whereas scMOD keeps a separate latent for each modality, so its diffusion prior improves generation while its autoencoders preserve the modality-specific detail needed for accurate prediction.

Reviewer_zwmaOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces scMOD, a unified generative framework for single-cell RNA expression, chromatin accessibility (ATAC), and surface protein abundance (ADT). The method combines modality-specific VAE codecs with a shared Diffusion Transformer trained through block-causal flow matching over randomized modality orderings. This formulation enables unconditional generation and conditional completion of arbitrary missing modalities, including sequential generation of multiple targets, within a single checkpoint. The architecture further incorporates peak-centric ATAC tokenization and biological graph guidance for representation learning and cross-modal conditioning. The authors evaluate scMOD on two trimodal datasets, TEA-seq and DOGMA-seq, measuring unconditional generation quality and all six pairwise cross-modal prediction directions. They report improved marginal distribution matching over multimodal VAE baselines, the highest average pairwise Pearson correlation among those baselines, and ablation studies on the contributions of the main architectural components.

Critical strengths
  1. The paper presents an interesting approach to modeling heterogeneous single-cell modalities within a single generative framework. Combining modality-specific representations with a shared block-causal flow model creates a flexible formulation for unconditional generation and arbitrary missing-modality completion. The integration of peak-centric ATAC tokenization and biological graph guidance is also a valuable aspect of the proposed design.

  2. scMOD consistently improves Wasserstein and MMD scores over the evaluated multimodal VAE baselines across RNA, ATAC, and ADT on both TEA-seq and DOGMA-seq. Several of these improvements are substantial, suggesting that learning a flexible latent generative distribution can improve sample quality compared with standard VAE prior sampling.

Critical weaknesses
  1. Although the main contribution is a unified framework for joint generation and arbitrary missing-modality completion, the experiments primarily evaluate individual modality distributions and pairwise prediction. In Section 4.2, Single and Joint generation are assessed using marginal distributional metrics, which cannot establish whether jointly generated RNA, ATAC, and ADT profiles are mutually consistent. Furthermore, no experiments evaluate multi-source conditioning or conditional multi-target generation, despite these being central capabilities of the proposed block-causal formulation. The sensitivity to modality generation order is also not examined. Finally, the block-causal ablation in Section 4.4 is evaluated only through pairwise prediction, rather than the multi-target generation scenarios that seem to motivate this component. Consequently, the experimental evidence does not adequately establish the benefits of the paper's central architectural contribution.

  2. While scMOD achieves the highest average PCC among the multimodal VAE baselines, the improvements over MIDAS are small. Considering all reported baselines across Tables 2 and 3, scMOD achieves the best PCC in only two of six directions on TEA-seq, with both improvements over kNN being marginal. Moreover, independent training seed variability is not reported for the main comparisons, and the ablations use a single training seed, making the small differences difficult to assess. Finally, the unconditional metrics are computed in reduced 30-dimensional representations and the pairwise prediction uses the true target-cell library size and measured-feature mask, which perhaps make the reported results more difficult to interpret.

  3. In particular, evaluating Single and Joint generation using joint distributional metrics on paired multimodal representations, with randomly re-paired real cells as a baseline, would directly assess cross-modal consistency. These evaluations should also include the non-block-causal ablation to establish the benefit of the proposed formulation. Multi-source prediction and multi-target generation should also be evaluated against their respective single-source and independent-generation alternatives, together with an analysis of sensitivity to modality generation order. Finally, reporting variability across independent training runs would be important for interpreting the relatively small performance differences.

Questions for authors
  1. What is the joint diffusion over all target modalities used to replace block-causal generation in section 4.4? In particular, how are observed modalities incorporated, what attention structure and training objective are used, and how does it differ from the full model in the single-target prediction tasks reported in Figure 4?

  2. Equation 10 sums the conditional losses over subsequent modality blocks, while the accompanying text states that conditional errors are averaged within each sample. Which formulation is actually implemented?

5 comments
AuthorsOpenReview ↗

Response to Reviewer zwma (1/2)

Thank you for your detailed review. Below, we respond to each of your comments point by point.


Weakness 1: "Although the main contribution is a unified framework for joint generation and arbitrary missing-modality completion, the experiments primarily evaluate individual modality distributions and pairwise prediction. In Section 4.2, Single and Joint generation are assessed using marginal distributional metrics, which cannot establish whether jointly generated RNA, ATAC, and ADT profiles are mutually consistent."

Weakness 3: "In particular, evaluating Single and Joint generation using joint distributional metrics on paired multimodal representations, with randomly re-paired real cells as a baseline, would directly assess cross-modal consistency."

Response.

Thank you for this suggestion. We have added a cross-modal consistency evaluation that directly tests whether the generated RNA, ADT, and ATAC profiles describe the same cell type (Table R5, see evaluation protocols in General Response). Three single-modality classifiers predict cell types, and Cohen's κ\kappa measures pairwise agreement beyond the random re-pairing reference (κ=0\kappa = 0). Real test cells provide the measured-data reference. scMOD achieves the highest consistency in every comparison and comes closest to the real-data reference. Its three-modality score reaches 0.668, compared with 0.343 for MIDAS and 0.250 for MultiVI. These results show that scMOD preserves a shared cell identity across jointly generated modalities. We have added this evaluation to Section 4.3 and Appendix B.4 of the revised manuscript.

Table R5. Cross-modal consistency of generated cells on TEA-seq (κ\kappa, ↑\uparrow) under three seeds.

CellsRNA–ADTRNA–ATACADT–ATACAll three
Real test cells0.9410.9240.9280.917
scMOD0.806 ±\pm 0.0030.671 ±\pm 0.0020.671 ±\pm 0.0050.668 ±\pm 0.001
MIDAS0.479 ±\pm 0.0090.407 ±\pm 0.0080.411 ±\pm 0.0030.343 ±\pm 0.004
MultiVI0.302 ±\pm 0.0160.221 ±\pm 0.0190.382 ±\pm 0.0090.250 ±\pm 0.014
scDiffusion-X--0.509 ±\pm 0.010----
CFGen--0.639 ±\pm 0.011----

Weakness 1: "Furthermore, no experiments evaluate multi-source conditioning or conditional multi-target generation, despite these being central capabilities of the proposed block-causal formulation. The sensitivity to modality generation order is also not examined."

Weakness 3: "Multi-source prediction and multi-target generation should also be evaluated against their respective single-source and independent-generation alternatives, together with an analysis of sensitivity to modality generation order."

Response.

Thank you for this suggestion. We have added experiments on TEA-seq that evaluate both capabilities using the same trained model, without retraining (Tables R3 and R4).

Multi-source conditioning (Table R3). With two observed modalities, scMOD achieves the highest PCC among all compared methods for all three targets. It also improves over both of its single-source counterparts in every case. The largest gain is for ADT prediction: RNA+ATAC reaches 0.8690, compared with 0.8570 from RNA alone and 0.8650 from ATAC alone. This shows that scMOD combines complementary information across modalities.

Multi-target generation (Table R4). We evaluate both target orders for each observed modality. As shown in the table, scMOD achieves the highest average PCC. Across all six input–target pairs, changing the generation order changes the PCC by at most 0.0006, showing that scMOD is robust to target ordering.

We have added these results to Appendix D of the revised manuscript.

Table R3. Many-to-one prediction on TEA-seq (PCC, ↑\uparrow). A and B are the observed modalities. Gains are relative percentage increases over the weaker single-source PCC.

TargetMethodA onlyB onlyA+B (PCC, gain %)
ADT (A = RNA, B = ATAC)scMOD0.85950.82660.8602 (+4.1%)
MIDAS0.85510.85140.8591 (+0.9%)
MultiVI0.86160.86060.8648 (+0.5%)
RNA (A = ADT, B = ATAC)scMOD0.58270.57870.5834 (+0.8%)
MIDAS0.58260.58560.5853 (+0.5%)
MultiVI0.52090.52460.5239 (+0.6%)
ATAC (A = RNA, B = ADT)scMOD0.61810.61790.6183 (+0.1%)
MIDAS0.58300.58270.5835 (+0.1%)
MultiVI0.54190.54570.5454 (+0.6%)
AuthorsOpenReview ↗

Response to Reviewer zwma (2/2)

Table R4. One-to-many prediction on TEA-seq, scored per target (PCC, ↑\uparrow). First and Second indicate whether the scored target is generated before or after the other target. Bold marks the best per row.

InputTargetscMOD FirstscMOD SecondscMOD Avg.MIDASMultiVI
RNAADT0.85950.85950.85950.85510.8616
ATAC0.61810.61810.61810.58300.5419
ADTRNA0.58270.58250.58260.58260.5209
ATAC0.61790.61800.61800.58270.5457
ATACRNA0.57870.57880.57880.58560.5246
ADT0.82660.82720.82690.85140.8606

Weakness 2: "While scMOD achieves the highest average PCC among the multimodal VAE baselines, the improvements over MIDAS are small. Considering all reported baselines across Tables 2 and 3, scMOD achieves the best PCC in only two of six directions on TEA-seq, with both improvements over kNN being marginal."

Response.

Thank you for this thoughtful comment. We would like to reiterate that our goal is to build a unified model for both unconditional generation and arbitrary missing-modality completion.

Compared with the multimodal VAE baselines, scMOD significantly improves unconditional generation (Table 1 in the paper and Table R1) while also achieving the highest average prediction PCC on both datasets (Table 2). It further achieves much stronger cross-modal consistency than MIDAS and MultiVI (Table R5), showing that its jointly generated profiles preserve a coherent cell identity. Thus, scMOD’s advantages extend across generation quality, prediction accuracy, and consistency between modalities not only prediction accuracy.

For the translation baselines in Table 3, these methods focus on predefined source–target tasks, and extending them to different modality combinations typically requires pair-specific models. They do not offer a unified framework for arbitrary trimodal completion and unconditional generation. The kNN baseline predicts target profiles by averaging those of neighboring training cells, so it depends on available paired reference data and is limited to reference-based interpolation. It does not learn a joint generative distribution or support de novo cell generation. In contrast, scMOD natively supports all six prediction directions, multiple observed or target modalities, and unconditional trimodal generation with one checkpoint, while matching or exceeding kNN in all six pairwise directions.


Weakness 2: "Finally, the unconditional metrics are computed in reduced 30-dimensional representations and the pairwise prediction uses the true target-cell library size and measured-feature mask, which perhaps make the reported results more difficult to interpret."

Response.

Thank you for raising this point.

For unconditional metrics, we follow scLDM [1] using 30-dimensional representation. This is also adopted in other literatures such as CFGen to mitigate the curse of dimensionality.

For pairwise PCC, the true target library size provides the total count used for decoding, and the measured-feature mask specifies the features measured by the assay, excluding any too sparse to evaluate. Both are shared evaluation conditions across all the methods. Pairwise PCC is computed on normalized profiles over measured features, so it evaluates recovery of the target profile at the specified depth and assay coverage.

References:

[1] Giovanni Palla, Sudarshan Babu, Payam Dibaeinia, James D. Pearce, Donghui Li, Aly A. Khan, Theofanis Karaletsos, and Jakub M. Tomczak. Scalable Single-Cell Gene Expression Generation with Latent Diffusion Models, 2026


Question 2: "Equation 10 sums the conditional losses over subsequent modality blocks, while the accompanying text states that conditional errors are averaged within each sample. Which formulation is actually implemented?"

Response.

Thank you for spotting this inconsistency. Each sample's conditional errors are averaged before computing the batch mean. For modality order (m1,…,mn)(m_1, \dots, m_n) and block errors ℓj\ell_j, Eq. 10 is

L=λself,ℓ1+λcond,1n−1∑j=2nℓj,L = \lambda_{\mathrm{self}}\\, \ell_1 + \lambda_{\mathrm{cond}}\\, \frac{1}{n-1} \sum_{j=2}^{n} \ell_j,

The weights are λself=0.25\lambda_{\mathrm{self}} = 0.25 and λcond=0.75\lambda_{\mathrm{cond}} = 0.75 (Appendix C), giving cells with two or three modalities the same total conditional weight.

We have made the factor 1/(n−1)1/(n-1) explicit in Eq. 10 of the revised paper so that the equation matches the implementation and accompanying text.

Thank you for the detailed responses and additional experiments. I provide further comments on each response in the same order.

  1. The additional cross-modal consistency results (Table R5) are encouraging and provide meaningful evidence that jointly generated modalities preserve consistent cell-type identities. However, Cohen's κ\kappa measures agreement between broad cell-type labels rather than the fidelity of the joint molecular distribution. Modalities could agree on cell identity while failing to preserve finer-grained cross-modal dependencies. For example, randomly re-pairing real RNA, ATAC, and ADT profiles within the same cell type could retain high κ\kappa despite breaking cell-specific molecular relationships. A complementary joint distributional evaluation, such as MMD on concatenated modality embeddings, comparing joint and single generation against real paired and randomly re-paired cells, would help assess whether the model captures dependencies beyond shared cell identity.

  2. The experiments on multi-source conditioning and multi-target generation (Tables R3 and R4) help demonstrate the framework's flexibility. However, the improvements over the stronger single-source baselines are relatively modest. Regarding multi-target generation, Table R4 demonstrates that individual target prediction accuracy is largely insensitive to generation order. Nevertheless, evaluating each target separately using PCC does not establish whether sequential generation preserves dependencies between the generated modalities, which seems to be a central motivation for the block-causal formulation. Generating the targets independently conditioned on the observed modality could yield similar per-target PCCs without capturing their joint distribution. A comparison between sequential and independent target generation using a joint distributional or cross-modal consistency metric would therefore be more informative. Similarly, stable PCC across generation orders does not necessarily imply equivalent joint distributions. Including the non-block-causal ablation in these evaluations would help establish the specific contribution of the proposed architecture.

  3. Thank you for the clarification. The additional results support the broader contribution of the framework. Nevertheless, the small predictive improvements remain difficult to interpret without variability across independent training runs. As also raised by reviewer sfN6, the relative contributions of the learned diffusion prior and the modality-specific autoencoders have not been isolated empirically. Furthermore, the specific benefit of the block-causal mechanism remains unclear, and my original question about the precise formulation of the non-block-causal ablation also appears to remain unanswered. I also noticed a discrepancy in the ATAC →\rightarrow ADT prediction results on TEA-seq: the original Table 2 reports 0.8268, whereas Tables R3 and R4 report 0.8650. Could you clarify the reason for this difference?

  4. Thank you for the clarification. It addresses my questions. It is nevertheless worth noting that the true target library size may not be available in real missing-modality applications.

  5. Thank you for the clarification and correction. This resolves my question.

Overall, the additional results strengthen the evaluation and address several of my original concerns. However, the specific benefits of the block-causal formulation, particularly for modeling dependencies between generated modalities, and the fidelity of the generated joint molecular distributions remain insufficiently established. There is also still a lack of variability across independent training runs. I will therefore maintain my score.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Response on new experiments and revisions (1/2)

We thank all reviewers for their insightful comments. We add the following experiments and supplementary clarifications to clarify the contributions of our work.

New experiments

New baselines scDiffusion-X and CFGen appear in Tables R1 and R2 (sfN6). Tables R3 and R4 add two-to-one and one-to-two prediction (zwma, sfN6). Table R5 evaluates cross-modal consistency for jointly generated cells (zwma, sfN6).

Protocol for Tables R1 and R2. We use matched training budgets for all compared baselines. Their parameter counts are within $\pm$5% of scMOD (47.8M on TEA-seq), and they use 40,000 steps per training stage, the same training cells, and final weights. Results are reported on the full test splits of 3,775 TEA-seq and 3,938 DOGMA-seq cells.

Table R1. Unconditional generation. Values are means over three seeds. Bold marks the best per row, and dashes denote missing modalities.

TEA-seq

ModalityMetricscMOD SinglescMOD JointMIDASMultiVIscDiffusion-XCFGen
RNAW2↓W_2 \downarrow10.1310.0813.4214.2111.4711.49
MMD$^2 \downarrow$0.00840.00860.03160.05520.01530.0104
1-NN →0.5\rightarrow 0.50.7110.7030.8910.9320.7560.795
ADTW2↓W_2 \downarrow7.467.418.357.76----
MMD$^2 \downarrow$0.01000.01020.04970.0151----
1-NN →0.5\rightarrow 0.50.6040.6020.8320.700----
ATACW2↓W_2 \downarrow16.8516.9424.9834.2318.4518.66
MMD$^2 \downarrow$0.03500.03320.07000.13640.04580.0463
1-NN →0.5\rightarrow 0.50.8740.8730.8950.9570.8290.814

DOGMA-seq

ModalityMetricscMOD SinglescMOD JointMIDASMultiVIscDiffusion-XCFGen
RNAW2↓W_2 \downarrow11.3711.2915.1216.7113.7913.06
MMD$^2 \downarrow$0.00540.00570.03860.07430.01820.0068
1-NN →0.5\rightarrow 0.50.7650.7620.9120.9660.8340.851
ADTW2↓W_2 \downarrow8.508.5214.6713.10----
MMD$^2 \downarrow$0.02650.02940.17970.1214----
1-NN →0.5\rightarrow 0.50.7000.6970.9160.845----
ATACW2↓W_2 \downarrow18.1618.2528.4635.4820.2818.26
MMD$^2 \downarrow$0.07830.08180.08440.16240.05430.0292
1-NN →0.5\rightarrow 0.50.9030.8990.8570.9300.8110.779

Across both test sets, scMOD leads every RNA metric and has the lowest ATAC W2W_2 on both. ATAC rankings differ by criterion. scMOD has the lowest TEA-seq MMD$^2$, while the new baselines are closer to the 1-NN golden target on both datasets. The RNA results, together with the ATAC W2W_2 advantage, are consistent with learned latent distributions.

Table R2. RNA–ATAC translation on TEA-seq with scDiffusion-X added (PCC, ↑\uparrow). Only the two directions scDiffusion-X supports are shown. Bold marks the best per column.

MethodRNA →\rightarrow ATACATAC →\rightarrow RNA
scMOD0.61810.5787
scDiffusion-X (new)0.56020.5403
scButterfly-B0.58600.5894
BABEL0.58440.5800
kNN0.61510.5627

CFGen does not report observed-modality RNA–ATAC translation, so it is excluded from this comparison.

scMOD is evaluated on the full 3,775-cell TEA-seq test split with 64 sampling seeds, from 41 through 104, and 20 Heun sampling steps in BF16. PCC is computed on log1p(CP10K) after averaging decoded predictions at each cell and feature position. Table R4 uses 32 draws per target order.

AuthorsOpenReview ↗

General Response on new experiments and revisions (2/2)

Table R3. Many-to-one prediction on TEA-seq (PCC, ↑\uparrow). A and B are the observed modalities. Gains are relative percentage increases over the weaker single-source PCC. Bold marks the highest A+B PCC and gain per target.

TargetMethodA onlyB onlyA+B (PCC, gain %)
ADT (A = RNA, B = ATAC)scMOD0.85950.82660.8602 (+4.1%)
MIDAS0.85510.85140.8591 (+0.9%)
MultiVI0.86160.86060.8648 (+0.5%)
RNA (A = ADT, B = ATAC)scMOD0.58270.57870.5834 (+0.8%)
MIDAS0.58260.58560.5853 (+0.5%)
MultiVI0.52090.52460.5239 (+0.6%)
ATAC (A = RNA, B = ADT)scMOD0.61810.61790.6183 (+0.1%)
MIDAS0.58300.58270.5835 (+0.1%)
MultiVI0.54190.54570.5454 (+0.6%)

Table R4. One-to-many prediction on TEA-seq, scored per target (PCC, ↑\uparrow). First and Second indicate whether the scored target is generated before or after the other target. Bold marks the best per row.

InputTargetscMOD FirstscMOD SecondscMOD Avg.MIDASMultiVI
RNAADT0.85950.85950.85950.85510.8616
ATAC0.61810.61810.61810.58300.5419
ADTRNA0.58270.58250.58260.58260.5209
ATAC0.61790.61800.61800.58270.5457
ATACRNA0.57870.57880.57880.58560.5246
ADT0.86470.82660.82690.85140.8606

Table R5. Cross-modal consistency of generated cells on TEA-seq (κ\kappa, ↑\uparrow). Values are mean ±\pm standard deviation over the three sampling seeds of Table R1. Bold marks the best method per column, and dashes denote modalities a baseline does not model.

CellsRNA–ADTRNA–ATACADT–ATACAll three
Real test cells0.9410.9240.9280.917
scMOD0.806 ±\pm 0.0030.671 ±\pm 0.0020.671 ±\pm 0.0050.668 ±\pm 0.001
MIDAS0.479 ±\pm 0.0090.407 ±\pm 0.0080.411 ±\pm 0.0030.343 ±\pm 0.004
MultiVI0.302 ±\pm 0.0160.221 ±\pm 0.0190.382 ±\pm 0.0090.250 ±\pm 0.014
scDiffusion-X--0.509 ±\pm 0.010----
CFGen--0.639 ±\pm 0.011----

Implementations

For each modality, we train a logistic-regression classifier that assigns cells to one of seven types, CD4⁺ T, CD8⁺ T, NK, B, monocyte, DC, or other cells. The classifiers are trained on the training cells and reach test accuracy of 0.967 for RNA, 0.977 for ADT, and 0.953 for ATAC. Cell-type labels come from Leiden clusters of real-cell profiles across all three modalities, merged using ADT and RNA markers.

To assess whether conditional generation actually uses the preceding condition modality for a pair of modalities, we use Cohen's kappa coefficient κ=(a−e)/(1−e)\kappa = (a - e) / (1 - e), where aa is the fraction of cells assigned the same type and ee is the expected performance under random re-pairing. κ=0\kappa = 0 represents pure randomness and collapses to unconditional generation.


Revisions in the PDF

We have uploaded a revised version of the paper for the rebuttal, with all edits highlighted in yellow for clarity.

  1. Background on the three measurements and two assays, with clearer motivation (3iDA).
  2. Related work covers MLD, UniDiffuser, Block Diffusion, scDiffusion-X and CFGen, and the novelty statement specifies joint RNA–ATAC–ADT modeling within a single latent diffusion framework (sfN6).
  3. The main text summarizes the new R1/R2 baselines, with detailed results in Appendix tables (sfN6).
  4. R3/R4 paper-checkpoint completion results are in Appendix tables. The R5 consistency table and available protocol are now in Section 4.3 and the Appendix (zwma, sfN6).
  5. Eq. 10 explicitly averages conditional losses. The evaluation protocol states the 30-dimensional representation used for unconditional metrics and the target library sizes and measured-feature masks used for prediction (zwma).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.