Research map / Label-efficient and robust learning / Semi-supervised learning
Semi-supervised Knowledge Transfer Across Multi-omic Single-cell Data
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 DANCE, a semi-supervised framework for transferring cell types from labeled scRNA-seq to unannotated scATAC-seq under label scarcity in both omics, using OT-based pseudo-labeling, divide-and-conquer learning, and cross-omic multi-sample Mixup.
Problem
Transferring cell-type labels from scRNA-seq to scATAC-seq when both modalities have very few annotations, while reducing the cross-omic heterogeneity gap and preserving cell-type discriminability.
Approach
Combines optimal-transport pseudo-labeling to expand the scarce labeled scRNA-seq set, splits scATAC-seq into source-like samples (consistency regularization) and target-specific samples (progressive elimination of incorrect candidate labels), and aligns modalities via class-similarity-weighted multi-sample Mixup that needs no pairing or scATAC labels.
Key results
- On snRNA_10X_v3_A-ATAC at high label ratio, DANCE reaches 51.42% vs 41.86% for the best baseline scBridge; on snRNA_SMARTer-snmC, 77.94% vs 73.39% for scNCL.
- Ablation: removing OT pseudo-labeling drops snRNA_10X_v3_A-ATAC low-label accuracy from 50.23% to 31.87%.
- Reported computation: DANCE uses 1.4GB memory and 30s per epoch, comparable to scBridge and scJoint.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add OT-based pseudo-labeling for source scRNA-seq data (ablation) | DANCE · snRNA_10X_v3_A-ATAC, low label ratio | accuracy: 31.87 -> 50.23 | Helped |
| add source-like consistency regularization (ablation) | DANCE · snRNA_10X_v3_A-ATAC, low label ratio | accuracy: 48.10 -> 50.23 | Helped |
| add target-specific ambiguous-set learning (ablation) | DANCE · snRNA_10X_v3_A-ATAC, low label ratio | accuracy: 48.62 -> 50.23 | Helped |
| add cross-omic multi-sample Mixup (ablation) | DANCE · snRNA_10X_v3_A-ATAC, low label ratio | accuracy: 44.80 -> 50.23 | Helped |
| apply OT-based pseudo-labeling to target scATAC-seq data (ablation) | DANCE · CITE-ASAP, low label ratio | accuracy: 45.36 -> 34.90 | Hurt |
Surprising findings
- Increasing label count does not always help baselines: scBridge scores 36.15% at high label ratio versus 39.75% at mid on snRNA_SMARTer-snmC.
- Applying OT pseudo-labeling directly to scATAC-seq hurts performance (39.34% vs 47.45% at high label ratio on CITE-ASAP).
Limitations
- Authors note the method may not perform well under open-set transfer or label noise.
- Only static scenarios are studied; dynamic applications such as RNA velocity are not addressed.
- Closed-set setting assumes scRNA-seq and scATAC-seq share the same C cell types.
Experimental setup
Datasets: Mouse Atlas (Tabula Muris multi-omics) CITE-ASAP PBMC
Benchmarks: Mouse Atlas snRNA/scRNA to ATAC and snmC transfer CITE-ASAP PBMC
Metrics: accuracy (%) A-distance t-SNE visualization
Baselines: DAN CDAN MCC FixMatch scJoint scNCL scBridge SLA COT Seurat Harmony
Builds on: FixMatch Sinkhorn algorithm scNCL scBridge MCC
Closest papers by meaning
- Cell ontology guided transcriptome foundation model (NeurIPS 2024 · similarity 0.87)
- scPilot: Large Language Model Reasoning Toward Automated Single-Cell Analysis and Discovery (NeurIPS 2025 · similarity 0.87)
- scMRDR: A scalable and flexible framework for unpaired single-cell multi-omics data integration (NeurIPS 2025 · similarity 0.86)
- A Simple and Comprehensive Benchmark for Single-Cell Transcriptomics (AAAI 2025 · similarity 0.86)
- Unsupervised Gene-Cell Collective Representation Learning with Optimal Transport (AAAI 2024 · similarity 0.86)
- A versatile informative diffusion model for single-cell ATAC-seq data generation and analysis (NeurIPS 2024 · similarity 0.86)
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.