Research map / Vision-language and multimodal / Cross-modal learning
Comprehensive View Embedding Learning for Single-Cell Multimodal Integration
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 CoVEL, an unsupervised deep learning method that integrates unpaired single-cell multimodal data by combining graph-linked, fine-grained cell, and contrastive cell embeddings.
Problem
Unsupervised integration of unpaired single-cell multimodal (multi-omics) data, where different modalities have different feature spaces, causing information loss and modality gaps while needing to preserve biological heterogeneity.
Approach
Learns single-cell embeddings from three complementary views at once: a graph VAE over a cross-modality regulatory guidance graph (GLUE-style), a Performer/Transformer-based fine-grained cell embedding trained with reconstruction, and a teacher-student contrastive cell embedding on the unit hypersphere.
Key results
- On Chen-2019, CoVEL reaches Batch.r 0.881 and Bio.c 0.593, versus 0.856 and 0.574 for the next best GLUE.
- On Muto-2021, CoVEL reaches Batch.r 0.998 and Bio.c 0.721, outperforming GLUE (0.969, 0.638).
- Ablation on mouse skin: removing both fine-grained and contrastive learning drops Batch.r from 0.968 to 0.881 and Bio.c from 0.562 to 0.537.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CoVEL instead of GLUE | unsupervised multimodal integration · Chen-2019 | Batch removal: 0.856 -> 0.881 | Helped |
| Use CoVEL instead of GLUE | unsupervised multimodal integration · Chen-2019 | Biology conservation: 0.574 -> 0.593 | Helped |
| Use CoVEL instead of GLUE | unsupervised multimodal integration · 10x-Multiome | Batch removal: 0.989 -> 0.981 | Hurt |
| Use CoVEL instead of Seurat v3 | unsupervised multimodal integration · 10x-Multiome | Biology conservation: 0.613 -> 0.639 | Helped |
| Use CoVEL instead of Seurat v3 | unsupervised multimodal integration · Muto-2021 | Batch removal: 0.982 -> 0.998 | Helped |
| Use CoVEL instead of Seurat v3 | unsupervised multimodal integration · Muto-2021 | Biology conservation: 0.681 -> 0.721 | Helped |
| Use CoVEL instead of GLUE | unsupervised multimodal integration · Yao-2021 | Batch removal: 0.989 -> 0.994 | Helped |
| Use CoVEL instead of LIGER | unsupervised multimodal integration · Yao-2021 | Biology conservation: 0.634 -> 0.681 | Helped |
| Add fine-grained learning by retaining the Teacher Transformer layer instead of replacing it with a Dense layer (ablation) | CoVEL · Mouse skin | Batch removal: 0.919 -> 0.968 | Helped |
| Add fine-grained learning by retaining the Teacher Transformer layer instead of replacing it with a Dense layer (ablation) | CoVEL · Mouse skin | Biology conservation: 0.560 -> 0.562 | Helped |
| Add contrastive learning and use its optimized Teacher branch embedding as the joint embedding (ablation) | CoVEL · Mouse skin | Batch removal: 0.904 -> 0.968 | Helped |
| Add contrastive learning and use its optimized Teacher branch embedding as the joint embedding (ablation) | CoVEL · Mouse skin | Biology conservation: 0.559 -> 0.562 | Helped |
Limitations
- Reliance on a guidance graph built from genome coordinates or gene names for cross-modality edges.
- Some clusters are not dense enough, which the authors attribute to possible subtypes.
- Evaluation limited to two metrics (batch removal, biology conservation) and a few public datasets.
Experimental setup
Datasets: Mouse skin (SHARE-seq, Ma et al. 2020) Mouse cortex (Saunders 2018; Luo 2017) Chen-2019 10x-Multiome Muto-2021 Yao-2021 Bastidas-Ponce 2019 mouse pancreas RNA Duvall 2022 mouse pancreas ATAC
Benchmarks: Mouse skin dataset Chen-2019 10x-Multiome Muto-2021 Yao-2021 Mouse cortex 3-modal
Metrics: Batch removal (graph connectivity) Biology conservation (cell type average silhouette width)
Baselines: iNMF LIGER bindSC Harmony Seurat v3 GLUE
Builds on: GLUE Performer scBERT Concerto Graph VAE
Closest papers by meaning
- Propensity Score Alignment of Unpaired Multimodal Data (NeurIPS 2024 · similarity 0.84)
- A Cross Modal Knowledge Distillation & Data Augmentation Recipe for Improving Transcriptomics Representations through Morphological Features (ICML 2025 · similarity 0.84)
- Multimodal Lego: Model Merging and Fine-Tuning Across Topologies and Modalities in Biomedicine (ICLR 2025 · similarity 0.84)
- ES-Merging: Biological MLLM Merging via Embedding Space Signals (NeurIPS 2026 · similarity 0.83)
- ViTally Consistent: Scaling Biological Representation Learning for Cell Microscopy (ICML 2025 · similarity 0.83)
- scSSL-Bench: Benchmarking Self-Supervised Learning for Single-Cell Data (ICML 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.