Related Experiment Video
Updated: Jul 10, 2026
06:24
Multiplexed Analysis of Retinal Gene Expression and Chromatin Accessibility Using scRNA-Seq and scATAC-Seq
Published on: March 12, 2021
scCLIP: A contrastive masked-reconstruction framework for paired single-cell multi-omics integration.
1Department of Geriatrics, Tongling People's Hospital, Tongling Academy of Medical Sciences, Tongling, Anhui, China.
Journal of Biomedical Informatics
|July 8, 2026
Summary
scCLIP, a new framework for paired single-cell multi-omics integration, aligns RNA and protein data using contrastive learning and masked reconstruction. It outperforms existing methods in cell type classification and scales effectively for large datasets.
Area of Science:
- Single-cell multi-omics data integration
- Computational biology
- Machine learning for bioinformatics
Background:
- Paired biomedical assays generate multi-modal data from the same sample, posing statistical challenges for analysis due to differing data characteristics.
- Existing methods for integrating paired single-cell data often make specific assumptions and may not fully leverage symmetric contrastive objectives for alignment while preserving modality-specific information.
Purpose of the Study:
- To introduce scCLIP, a novel contrastive masked-reconstruction framework for effective paired single-cell multi-omics integration.
- To develop a method that aligns different molecular views (e.g., RNA and protein) within the same cell while retaining modality-specific signals.
Main Methods:
- scCLIP employs a joint training approach for RNA and ADT (Abundance of Protein measurement) branches using bidirectional cross-modal contrastive loss and masked reconstruction.
- The framework utilizes shared architectural templates with separate input/output adapters, encoder-decoder parameters, and projection heads for modality-specific processing.
- Embeddings are compared in an L2-normalized space with a learnable logit scale for alignment.
Main Results:
- scCLIP achieved superior performance in terms of Adjusted Rand Index (ARI) and Fowlkes-Mallows Index (FMI) across five paired RNA-protein datasets compared to TotalVI, BREMSC, jointDIMMSC, scMM, and SCOIT.
- On a large CITE-seq benchmark dataset (90,261 cells), scCLIP demonstrated scalability without architectural modifications and effective batch mixing.
- Direct evidence of RNA-ADT alignment was confirmed through retrieval and distance-based metrics, alongside standard batch-mixing scores.
Conclusions:
- scCLIP provides a robust and reusable template for representation learning in paired-view single-cell multi-omics data.
- The framework effectively integrates RNA and protein data, demonstrating its utility and potential for broader applications in single-cell analysis.
- Publicly available code and tutorials facilitate the application of scCLIP to new CITE-seq datasets.
