Research map / Image generation / Autoregressive and latent generation
A versatile informative diffusion model for single-cell ATAC-seq data generation and analysis
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ATAC-Diff, a latent diffusion model for scATAC-seq conditioned on GMM-prior auxiliary latent variables with a mutual-information regularizer, supporting generation, denoising, imputation, and clustering.
Problem
scATAC-seq data is extremely sparse and noisy, and existing methods are task-specific; there is a need for a single generative framework that can generate high-quality scATAC-seq data and also support downstream analysis such as clustering.
Approach
A latent diffusion model conditioned on semantic latent variables produced by an auxiliary encoder with a Gaussian mixture prior and auxiliary decoder, with a mutual-information maximization term (InfoVAE-style) to keep the latent variables from being ignored by the diffusion decoder.
Key results
- Clustering: ATAC-Diff reports NMI 0.740 on Forebrain and 0.647 on Hematopoiesis, with best ASW on Forebrain and Hematopoiesis.
- Unconditional generation: ATAC-Diff reaches SCC 0.925 / PCC 0.992 on Forebrain, versus SCALE 0.548 / 0.728.
- Denoising and imputation are competitive with SCALE and PeakVI rather than uniformly best (e.g., Forebrain denoising SCC 0.718 vs SCALE 0.777).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add GMM latent prior (ablation) | ATAC-Diff · Forebrain clustering | NMI: 0.558 -> 0.740 | Helped |
| Add GMM latent prior (ablation) | ATAC-Diff · Hematopoiesis clustering | NMI: 0.489 -> 0.647 | Helped |
| Add GMM latent prior (ablation) | ATAC-Diff · PBMC10k clustering | NMI: 0.586 -> 0.733 | Helped |
| Add mutual-information regularization (ablation) | ATAC-Diff · Forebrain clustering | NMI: 0.596 -> 0.740 | Helped |
| Add mutual-information regularization (ablation) | ATAC-Diff · Hematopoiesis clustering | NMI: 0.490 -> 0.647 | Helped |
| Add mutual-information regularization (ablation) | ATAC-Diff · PBMC10k clustering | NMI: 0.590 -> 0.733 | Helped |
| Add GMM latent prior (ablation) | ATAC-Diff · PBMC10k unconditional generation | SCC: 0.693 -> 0.964 | Helped |
| Add GMM latent prior (ablation) | ATAC-Diff · Forebrain unconditional generation | SCC: 0.919 -> 0.925 | Helped |
| Add mutual-information regularization (ablation) | ATAC-Diff · PBMC10k unconditional generation | SCC: 0.704 -> 0.964 | Helped |
| Add GMM latent prior (ablation) | ATAC-Diff · PBMC10k denoising | SCC: 0.843 -> 0.863 | Helped |
| Add GMM latent prior (ablation) | ATAC-Diff · Forebrain imputation | SCC: 0.710 -> 0.716 | Helped |
| Add mutual-information regularization (ablation) | ATAC-Diff · PBMC10k denoising | SCC: 0.851 -> 0.863 | Helped |
Surprising findings
- Removing the GMM or MI component reduces clustering performance and generation quality only modestly in some settings, and ATAC-Diff does not dominate denoising and imputation baselines.
Limitations
- Code is promised but not yet released at submission time.
- Evaluated on only three datasets with small cell counts (about 2000 to 9600 cells).
- Unconditional sampling relies on a separately trained vanilla VAE to sample latent variables, and the paper lacks a dedicated limitations discussion.
Experimental setup
Datasets: Forebrain Hematopoiesis PBMC10k
Benchmarks: Forebrain Hematopoiesis PBMC10k
Metrics: NMI ARI Homogeneity ASW SCC PCC
Baselines: HVP PCA cisTopic SCALE PeakVI
Builds on: diffusion autoencoder InfoVAE latent diffusion model DDPM
Closest papers by meaning
- ChromFound: Towards A Universal Foundation Model for Single-Cell Chromatin Accessibiltiy Data (NeurIPS 2025 · similarity 0.87)
- A Simple and Comprehensive Benchmark for Single-Cell Transcriptomics (AAAI 2025 · similarity 0.86)
- Semi-supervised Knowledge Transfer Across Multi-omic Single-cell Data (NeurIPS 2024 · similarity 0.86)
- Absorb & Escape: Overcoming Single Model Limitations in Generating Heterogeneous Genomic Sequences (NeurIPS 2024 · similarity 0.83)
- Cradle-VAE: Enhancing Single-Cell Gene Perturbation Modeling with Counterfactual Reasoning-based Artifact Disentanglement (AAAI 2025 · similarity 0.83)
- scPilot: Large Language Model Reasoning Toward Automated Single-Cell Analysis and Discovery (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.