Research map / Probabilistic and scientific modeling / Molecules and proteins
Multi-Modal and Multi-Attribute Generation of Single Cells with CFGen
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.
Introduces CFGen, a latent flow-matching generative model with negative binomial and Bernoulli decoders that generates discrete multi-modal single-cell data and extends classifier-free guidance to compositional multi-attribute generation.
Problem
Existing deep generative models for single-cell data operate on continuous normalized approximations, ignoring the discrete nature of counts, and offer limited controllable multi-modal and multi-attribute generation.
Approach
A conditional latent flow-matching prior conditioned on cell type and size factor, paired with count-based decoders (negative binomial for RNA, Bernoulli for ATAC), plus a compositional classifier-free guidance formulation for flow matching that composes single-attribute fields with per-attribute guidance weights.
Key results
- c-CFGen achieves the lowest MMD and WD for conditional generation on all four uni-modal datasets (e.g., MMD 0.54 vs 0.58 scVI on HLCA; WD 16.94 vs 17.66 scVI on PBMC3K).
- Multi-modal CFGen achieves the best ATAC WD (18.86 vs 20.84 PeakVI) and best RNA WD (13.90 vs 14.38 scVI).
- In batch correction, CFGen reports batch score 0.63 on NeurIPS and 0.68 on C. elegans, surpassing the second-best method by about 8% and 7%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use conditional CFGen instead of the strongest baseline | Conditional scRNA-seq generation · PBMC3K | MMD: 0.94 -> 0.85 | Helped |
| Use conditional CFGen instead of the strongest baseline | Conditional scRNA-seq generation · Dentate gyrus | WD: 22.56 -> 21.55 | Helped |
| Use conditional CFGen instead of the strongest baseline | Conditional scRNA-seq generation · HLCA | MMD: 0.58 -> 0.54 | Helped |
| Use unconditional CFGen instead of scGAN | Unconditional scRNA-seq generation · PBMC3K | MMD: 0.36 -> 0.44 | Hurt |
| Use unconditional CFGen instead of scGAN | Unconditional scRNA-seq generation · Dentate gyrus | MMD: 0.42 -> 0.42 | No clear effect |
| Use unconditional CFGen instead of scGAN | Unconditional scRNA-seq generation · Dentate gyrus | WD: 22.52 -> 21.20 | Helped |
| Use unconditional CFGen instead of scGAN | Unconditional scRNA-seq generation · Tabula Muris | MMD: 0.25 -> 0.08 | Helped |
| Use unconditional CFGen instead of scGAN | Unconditional scRNA-seq generation · HLCA | WD: 10.81 -> 10.63 | Helped |
| Generate RNA and ATAC jointly rather than RNA-only | CFGen · PBMC10K RNA | WD: 14.30 -> 13.90 | Helped |
| Generate RNA and ATAC jointly rather than RNA-only | CFGen · PBMC10K RNA | MMD: 0.86 -> 0.89 | Hurt |
| Use multi-modal CFGen instead of MultiVI | Multi-modal generation · PBMC10K ATAC | MMD: 0.96 -> 0.92 | Helped |
| Use CFGen instead of scVI for batch correction | Batch correction · NeurIPS | Batch: 0.45 -> 0.63 | Helped |
Surprising findings
- scGAN is slightly better than the unconditional CFGen on PBMC3K (MMD 0.36 vs 0.44), so CFGen does not win every unconditional comparison.
Limitations
- Relies on independence assumptions (e.g., conditional independence of attributes) that may not hold in all biological contexts.
- Autoencoder and flow are trained separately, which is inefficient and memory-intensive.
Experimental setup
Datasets: PBMC3K Dentate gyrus Tabula Muris Human Lung Cell Atlas PBMC10K multiome NeurIPS 2021 bone marrow C. elegans molecular atlas PBMC COVID
Benchmarks: PBMC3K Dentate gyrus Tabula Muris HLCA PBMC10K multiome NeurIPS 2021 C. elegans atlas PBMC COVID
Metrics: RBF-kernel MMD 2-Wasserstein distance scIB batch correction scIB biological conservation cell-type recall
Baselines: scVI scDiffusion scGAN MultiVI PeakVI scANVI scPoli
Builds on: Flow Matching classifier-free guidance scVI compositional diffusion guidance
Closest papers by meaning
- Cell2Sentence: Teaching Large Language Models the Language of Biology (ICML 2024 · similarity 0.84)
- CellFlux: Simulating Cellular Morphology Changes via Flow Matching (ICML 2025 · similarity 0.84)
- CellVerse: Do Large Language Models Really Understand Cell Biology? (NeurIPS 2025 · similarity 0.84)
- Improving Flow Matching by Aligning Flow Divergence (ICML 2025 · similarity 0.83)
- GeneFlow: Translation of Single-cell Gene Expression to Histopathological Images via Rectified Flow (NeurIPS 2025 · similarity 0.83)
- CellPLM: Pre-training of Cell Language Model Beyond Single Cells (ICLR 2024 · 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.