pith. sign in

arxiv: 2506.21107 · v3 · submitted 2025-06-26 · 💻 cs.LG · q-bio.MN

Doloris: Dual Conditional Diffusion Implicit Bridges with Sparsity Masking Strategy for Unpaired Single-Cell Perturbation Estimation

Changxi Chi , Jun Xia , Yufei Huang , Zhuoli Ouyang , Cheng Tan , Yunfan Liu , Jingbo Zhou , Chang Yu
show 4 more authors
Liangyu Yuan Siyuan Li Zelin Zang Stan Z. Li
This is my paper

Pith reviewed 2026-05-19 08:00 UTC · model grok-4.3

classification 💻 cs.LG q-bio.MN
keywords single-cell perturbationunpaired datadiffusion modelssparsity maskinggenerative modelinggene expressiondrug screening
0
0 comments X

The pith

Dual conditional diffusion models implicitly align unpaired control and perturbed single-cell distributions in a shared Gaussian latent space while using sparsity masking to preserve response diversity.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces Doloris as a generative framework for estimating how individual cells respond to perturbations when the before-and-after measurements come from different cells and cannot be paired. It trains two separate conditional diffusion models, one on unperturbed cells and one on perturbed cells, then connects them by mapping both distributions to the same Gaussian noise space so that a sample from the control distribution can be transformed into a plausible perturbed sample without ever seeing the same cell twice. A separate mask model learns to identify which genes are expressed as zero, freeing the diffusion process to model only the non-zero expression patterns and thereby retain the natural variety of possible cell responses in high-dimensional sparse data. If this implicit alignment works, it removes the need for destructive paired experiments, speeds up computational drug screening, and produces more realistic perturbation predictions than methods that either ignore the pairing problem or overfit to zeros.

Core claim

Doloris defines a new paradigm for modeling unpaired, high-dimensional, and sparse single-cell perturbation data. It leverages dual conditional diffusion models for separate learning of control and perturbed distributions, complemented by a sparsity masking strategy to enhance prediction of zero-valued genes.

What carries the argument

Dual conditional diffusion implicit bridges that map control and perturbed distributions into a shared Gaussian latent space without explicit cell pairing, together with a sparsity masking strategy that predicts zero-expressed genes so the diffusion models focus on meaningful non-zero patterns.

If this is right

  • Perturbation effects can be predicted for individual cells without requiring paired pre- and post-perturbation measurements of the same cell.
  • The sparsity mask allows the diffusion process to capture varied gene-expression patterns instead of defaulting to the dominant zero values in sparse data.
  • State-of-the-art performance is achieved on public single-cell perturbation datasets by effectively modeling response diversity.
  • Key genes can be identified and drug-screening efficiency improved through computational estimation rather than exhaustive wet-lab experiments.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The shared latent-space alignment may generalize to other unpaired biological modalities such as spatial transcriptomics where direct cell pairing is infeasible.
  • Interpolating within the shared Gaussian space could enable prediction of cellular responses to novel or combined perturbations not seen in training.
  • If the implicit bridge holds, similar dual-diffusion constructions might serve as distribution-matching tools for other unpaired high-dimensional settings outside single-cell biology.

Load-bearing premise

The control and perturbed distributions can be implicitly aligned through a shared Gaussian latent space without explicit cell pairing.

What would settle it

Running the model on a held-out single-cell dataset and finding that the generated perturbed profiles have significantly lower diversity or poorer match to the true perturbed distribution than competing methods, or that disabling the shared latent space alignment collapses performance to baseline levels.

Figures

Figures reproduced from arXiv: 2506.21107 by Changxi Chi, Chang Yu, Cheng Tan, Jingbo Zhou, Jun Xia, Liangyu Yuan, Siyuan Li, Stan Z. Li, Yufei Huang, Yunfan Liu, Zelin Zang, Zhuoli Ouyang.

Figure 1
Figure 1. Figure 1: Single-cell perturbation data are unpaired as cells cannot be measured twice. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of Unlasting. Unlasting leverages DDIB [25] to predict cellular responses under unseen perturbation conditions. The source model obtain the latent embedding x l by adding DDIM-based forward noise to unperturbed cell sample x c . Then, conditioned on the perturbation, we apply DDIM denoising to x t to generate the predicted sample. 2 Related Work and Preliminaries 2.1 Gene Regulation Network Constr… view at source ↗
Figure 3
Figure 3. Figure 3: Model architecture of the source model and target model. The source and target models [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Interpreting DDIB as Data Augmentation for Unpaired Data. (a) Discrete sample points [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Cells observed under the same experimental conditions exhibit a bimodal distribution [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Ablation study results. model demonstrates superior performance, suggesting that it better captures the effects of unseen molecules on cellular behavior. 4.4 Ablation Study To further evaluate the effectiveness of Unlasting, we compare it with the following methods through an ablation study. 1)w/o µc, σc: Excludes the mean and variance of the unperturbed group from the model input. 2)w/o latent: During sam… view at source ↗
read the original abstract

Estimating single-cell responses across various perturbations facilitates the identification of key genes and enhances drug screening, significantly boosting experimental efficiency. However, single-cell sequencing is a destructive process, making it impossible to capture the same cell's phenotype before and after perturbation. Consequently, data collected under perturbed and unperturbed conditions are inherently unpaired, creating a critical yet unresolved problem in single-cell perturbation modeling. Moreover, the high dimensionality and sparsity of single-cell expression make direct modeling prone to focusing on zeros and neglecting meaningful patterns. To address these problems, we propose a new paradigm for single-cell perturbation modeling. Specifically, we leverage dual diffusion models to learn the control and perturbed distributions separately, and implicitly align them through a shared Gaussian latent space, without requiring explicit cell pairing. Furthermore, we introduce a sparsity masking strategy in which the mask model learns to predict zero-expressed genes, allowing the diffusion model to focus on capturing meaningful patterns among expressed genes and thereby preserving diversity in high-dimensional sparse data. We introduce \textbf{Doloris}, a generative framework that defines a new paradigm for modeling unpaired, high-dimensional, and sparse single-cell perturbation data. It leverages dual conditional diffusion models for separate learning of control and perturbed distributions, complemented by a sparsity masking strategy to enhance prediction of zero-valued genes. The results on publicly available datasets show that our model effectively captures the diversity of single-cell perturbations and achieves state-of-the-art performance. To facilitate reproducibility, we include the code in the supplementary materials.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript introduces Doloris, a generative framework for unpaired single-cell perturbation estimation. It employs dual conditional diffusion models to separately learn the distributions of control and perturbed cells, implicitly aligning them through a shared Gaussian latent space without explicit cell pairing. A sparsity masking strategy is proposed where a mask model predicts zero-expressed genes to allow the diffusion model to focus on meaningful patterns. The authors report that the model captures the diversity of single-cell perturbations and achieves state-of-the-art performance on publicly available datasets, with code included for reproducibility.

Significance. This work has potential significance in advancing single-cell perturbation modeling by addressing the unpaired nature of data and the challenges of sparsity in scRNA-seq. If the implicit alignment successfully preserves biological perturbation effects, it could facilitate more efficient drug screening and key gene identification. The inclusion of reproducible code is a strength that supports the assessment.

major comments (2)
  1. [Method section on dual conditional diffusion implicit bridges] The central mechanism of implicit alignment through the shared Gaussian latent space (described in the dual conditional diffusion implicit bridges section) is not sufficiently validated for preserving cell-specific perturbation effects. The skeptic's concern lands: without explicit tests (e.g., on subsets with known pairings or using biological priors), it is possible that the generated perturbations match overall statistics but fail to reflect actual response diversity on the underlying cell population. This is load-bearing for the claim of effective modeling of unpaired data.
  2. [Results section] Table reporting main results: while SOTA performance is claimed, the manuscript should provide more detailed error analysis, variance across runs, and comparison to recent baselines in the field to substantiate the improvement, particularly regarding diversity metrics.
minor comments (2)
  1. [Abstract] The abstract could benefit from specifying the exact datasets and quantitative metrics used to claim SOTA performance.
  2. [Figure captions] Some figures illustrating the sparsity masking could have more detailed explanations of how it affects the diffusion process and diversity preservation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We have addressed the concerns regarding validation of the implicit alignment mechanism and the presentation of experimental results by adding new analyses and details in the revised version.

read point-by-point responses

  1. Referee: [Method section on dual conditional diffusion implicit bridges] The central mechanism of implicit alignment through the shared Gaussian latent space (described in the dual conditional diffusion implicit bridges section) is not sufficiently validated for preserving cell-specific perturbation effects. The skeptic's concern lands: without explicit tests (e.g., on subsets with known pairings or using biological priors), it is possible that the generated perturbations match overall statistics but fail to reflect actual response diversity on the underlying cell population. This is load-bearing for the claim of effective modeling of unpaired data.

    Authors: We agree that explicit validation of cell-specific preservation is important for substantiating the implicit alignment. In the revised manuscript, we have added a new subsection under Methods describing experiments on a synthetic dataset constructed from known cell populations with simulated perturbations, allowing access to ground-truth pairings. We evaluate per-cell correlation between generated and true perturbed states, as well as preservation of known biological pathways as priors. These results, now included in the revised Results section with a supporting figure, show that the shared latent space captures individual cell responses beyond aggregate statistics. revision: yes

  2. Referee: [Results section] Table reporting main results: while SOTA performance is claimed, the manuscript should provide more detailed error analysis, variance across runs, and comparison to recent baselines in the field to substantiate the improvement, particularly regarding diversity metrics.

    Authors: We appreciate this suggestion to strengthen the empirical claims. The revised manuscript now includes an expanded main results table with mean performance and standard deviations computed over five independent runs. We have added a dedicated error analysis subsection reporting per-gene and per-cell error distributions. For diversity, we include additional metrics such as response entropy and unique gene count variance, along with comparisons to two recent baselines in the single-cell perturbation literature. These updates appear in the revised Table 1 and accompanying text. revision: yes

Circularity Check

0 steps flagged

Standard diffusion machinery with domain assumptions; no load-bearing circularity

full rationale

The paper proposes Doloris as a generative framework using dual conditional diffusion models to separately learn control and perturbed distributions, implicitly aligned via a shared Gaussian latent space, plus a sparsity masking strategy for high-dimensional sparse scRNA-seq data. This is framed as a new paradigm for unpaired perturbation estimation, with empirical SOTA claims on public datasets. No equations or steps in the provided text reduce predictions to fitted inputs by construction, nor do self-citations form a load-bearing chain for the core claims. The implicit alignment is presented as a modeling choice/assumption rather than a derived necessity that loops back to itself. This qualifies as a normal non-circular proposal resting on standard diffusion techniques and domain knowledge.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard single-cell biology assumptions and diffusion-model components rather than newly invented entities. No free parameters are explicitly fitted to the target result in the abstract, and no new physical or biological entities are postulated.

axioms (2)
  • domain assumption Single-cell sequencing is destructive, so perturbed and unperturbed measurements are inherently unpaired.
    Stated directly in the abstract as the core motivation for the unpaired modeling problem.
  • domain assumption High-dimensional single-cell expression data is sparse with many zero entries that must be handled separately from expressed genes.
    Explicitly invoked to justify the sparsity masking strategy.

pith-pipeline@v0.9.0 · 5842 in / 1422 out tokens · 29768 ms · 2026-05-19T08:00:04.866075+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

29 extracted references · 29 canonical work pages · 4 internal anchors

  1. [1]

    A multiplexed single-cell crispr screening platform enables systematic dissection of the unfolded protein response

    Britt Adamson, Thomas M Norman, Marco Jost, Min Y Cho, James K Nuñez, Yuwen Chen, Jacqueline E Villalta, Luke A Gilbert, Max A Horlbeck, Marco Y Hein, et al. A multiplexed single-cell crispr screening platform enables systematic dissection of the unfolded protein response. Cell, 167(7):1867–1882, 2016

    work page 2016

  2. [2]

    Applications of crispr technologies in research and beyond

    Rodolphe Barrangou and Jennifer A Doudna. Applications of crispr technologies in research and beyond. Nature biotechnology, 34(9):933–941, 2016

    work page 2016

  3. [3]

    Modelling cellular perturbations with the sparse additive mechanism shift variational autoencoder

    Michael Bereket and Theofanis Karaletsos. Modelling cellular perturbations with the sparse additive mechanism shift variational autoencoder. Advances in Neural Information Processing Systems, 36, 2024

    work page 2024

  4. [4]

    How Attentive are Graph Attention Networks?

    Shaked Brody, Uri Alon, and Eran Yahav. How attentive are graph attention networks? arXiv preprint arXiv:2105.14491, 2021

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  5. [5]

    Learning single- cell perturbation responses using neural optimal transport

    Charlotte Bunne, Stefan G Stark, Gabriele Gut, Jacobo Sarabia Del Castillo, Mitch Levesque, Kjong-Van Lehmann, Lucas Pelkmans, Andreas Krause, and Gunnar Rätsch. Learning single- cell perturbation responses using neural optimal transport. Nature methods, 20(11):1759–1768, 2023

    work page 2023

  6. [6]

    scbutterfly: a versatile single-cell cross-modality translation method via dual-aligned variational autoencoders

    Yichuan Cao, Xiamiao Zhao, Songming Tang, Qun Jiang, Sijie Li, Siyu Li, and Shengquan Chen. scbutterfly: a versatile single-cell cross-modality translation method via dual-aligned variational autoencoders. Nature Communications, 15(1):2973, 2024

    work page 2024

  7. [7]

    Grape: Hetero- geneous graph representation learning for genetic perturbation with coding and non-coding biotype

    Changxi Chi, Jun Xia, Jingbo Zhou, Jiabei Cheng, Chang Yu, and Stan Z Li. Grape: Hetero- geneous graph representation learning for genetic perturbation with coding and non-coding biotype. arXiv preprint arXiv:2505.03853, 2025

    work page arXiv 2025

  8. [8]

    scgpt: toward building a foundation model for single-cell multi-omics using generative ai

    Haotian Cui, Chloe Wang, Hassaan Maan, Kuan Pang, Fengning Luo, Nan Duan, and Bo Wang. scgpt: toward building a foundation model for single-cell multi-omics using generative ai. Nature Methods, 21(8):1470–1480, 2024

    work page 2024

  9. [9]

    AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning

    Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  10. [10]

    Large-scale foundation model on single-cell transcriptomics

    Minsheng Hao, Jing Gong, Xin Zeng, Chiming Liu, Yucheng Guo, Xingyi Cheng, Taifeng Wang, Jianzhu Ma, Xuegong Zhang, and Le Song. Large-scale foundation model on single-cell transcriptomics. Nature methods, 21(8):1481–1491, 2024

    work page 2024

  11. [11]

    Squidiff: Predicting cellular development and responses to perturbations using a diffusion model

    Siyu He, Yuefei Zhu, Daniel Naveed Tavakol, Haotian Ye, Yeh-Hsing Lao, Zixian Zhu, Cong Xu, Sharadha Chauhan, Guy Garty, Raju Tomer, et al. Squidiff: Predicting cellular development and responses to perturbations using a diffusion model. bioRxiv, pages 2024–11, 2024

    work page 2024

  12. [13]

    Predicting cellular responses to novel drug perturbations at a single-cell resolution

    Leon Hetzel, Simon Boehm, Niki Kilbertus, Stephan Günnemann, Fabian Theis, et al. Predicting cellular responses to novel drug perturbations at a single-cell resolution. Advances in Neural Information Processing Systems, 35:26711–26722, 2022

    work page 2022

  13. [14]

    Delivering crispr: a review of the challenges and approaches

    Christopher A Lino, Jason C Harper, James P Carney, and Jerilyn A Timlin. Delivering crispr: a review of the challenges and approaches. Drug delivery, 25(1):1234–1257, 2018

    work page 2018

  14. [15]

    Learning interpretable cellular responses to complex perturbations in high-throughput screens

    M Lotfollahi, AK Susmelj, and C De Donno. Learning interpretable cellular responses to complex perturbations in high-throughput screens. biorxiv. 2021. 2021.04. 14.439903. 10

    work page 2021

  15. [16]

    scgen predicts single-cell perturbation responses

    Mohammad Lotfollahi, F Alexander Wolf, and Fabian J Theis. scgen predicts single-cell perturbation responses. Nature methods, 16(8):715–721, 2019

    work page 2019

  16. [17]

    Understanding diffusion models: A unified perspective,

    Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022

    work page arXiv 2022

  17. [18]

    Mapping and quantifying mammalian transcriptomes by rna-seq

    Ali Mortazavi, Brian A Williams, Kenneth McCue, Lorian Schaeffer, and Barbara Wold. Mapping and quantifying mammalian transcriptomes by rna-seq. Nature methods, 5(7):621– 628, 2008

    work page 2008

  18. [19]

    Exploring genetic interaction manifolds constructed from rich single-cell phenotypes

    Thomas M Norman, Max A Horlbeck, Joseph M Replogle, Alex Y Ge, Albert Xu, Marco Jost, Luke A Gilbert, and Jonathan S Weissman. Exploring genetic interaction manifolds constructed from rich single-cell phenotypes. Science, 365(6455):786–793, 2019

    work page 2019

  19. [20]

    scperturb: harmonized single-cell perturbation data

    Stefan Peidli, Tessa D Green, Ciyue Shen, Torsten Gross, Joseph Min, Samuele Garda, Bo Yuan, Linus J Schumacher, Jake P Taylor-King, Debora S Marks, et al. scperturb: harmonized single-cell perturbation data. Nature Methods, 21(3):531–540, 2024

    work page 2024

  20. [21]

    Gears: Predicting transcriptional outcomes of novel multi-gene perturbations

    Yusuf Roohani, Kexin Huang, and Jure Leskovec. Gears: Predicting transcriptional outcomes of novel multi-gene perturbations. BioRxiv, pages 2022–07, 2022

    work page 2022

  21. [22]

    Denoising Diffusion Implicit Models

    Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  22. [24]

    Massively multiplex chemical transcriptomics at single-cell resolution

    Sanjay R Srivatsan, José L McFaline-Figueroa, Vijay Ramani, Lauren Saunders, Junyue Cao, Jonathan Packer, Hannah A Pliner, Dana L Jackson, Riza M Daza, Lena Christiansen, et al. Massively multiplex chemical transcriptomics at single-cell resolution. Science, 367(6473): 45–51, 2020

    work page 2020

  23. [25]

    Dual diffusion implicit bridges for image-to-image translation

    Xuan Su, Jiaming Song, Chenlin Meng, and Stefano Ermon. Dual diffusion implicit bridges for image-to-image translation. arXiv preprint arXiv:2203.08382, 2022

    work page arXiv 2022

  24. [26]

    Graph Attention Networks

    Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  25. [27]

    Scanpy: large-scale single-cell gene expression data analysis

    F Alexander Wolf, Philipp Angerer, and Fabian J Theis. Scanpy: large-scale single-cell gene expression data analysis. Genome biology, 19:1–5, 2018

    work page 2018

  26. [28]

    Predicting cellular responses with variational causal inference and refined relational information

    Yulun Wu, Robert A Barton, Zichen Wang, Vassilis N Ioannidis, Carlo De Donno, Layne C Price, Luis F V oloch, and George Karypis. Predicting cellular responses with variational causal inference and refined relational information. arXiv preprint arXiv:2210.00116, 2022

    work page arXiv 2022

  27. [29]

    Mole-bert: Rethinking pre-training graph neural networks for molecules

    Jun Xia, Chengshuai Zhao, Bozhen Hu, Zhangyang Gao, Cheng Tan, Yue Liu, Siyuan Li, and Stan Z Li. Mole-bert: Rethinking pre-training graph neural networks for molecules. 2023

    work page 2023

  28. [30]

    Genecompass: deciphering universal gene regulatory mechanisms with a knowledge-informed cross-species foundation model

    Xiaodong Yang, Guole Liu, Guihai Feng, Dechao Bu, Pengfei Wang, Jie Jiang, Shubai Chen, Qinmeng Yang, Hefan Miao, Yiyang Zhang, et al. Genecompass: deciphering universal gene regulatory mechanisms with a knowledge-informed cross-species foundation model. Cell Research, 34(12):830–845, 2024

    work page 2024

  29. [31]

    Uni-mol: A universal 3d molecular representation learning framework

    Gengmo Zhou, Zhifeng Gao, Qiankun Ding, Hang Zheng, Hongteng Xu, Zhewei Wei, Linfeng Zhang, and Guolin Ke. Uni-mol: A universal 3d molecular representation learning framework. 2023. 11 A Mask Model In this section, we present the design rationale and architecture of the Mask Model. Given the high-dimensional and sparse nature of gene expression data, dire...

    work page 2023