The abstract is read at the publisher; the summary is JClub's.
GLAlign integrates heterogeneous single-cell multi-omics by learning a shared latent space that jointly uses global manifold constraints and local marker-gene overlap, enabling strong modality mixing across PBMC, mouse brain, CITE-seq, and NCI-N87 datasets.
- Why it matters: Multi-omics integration is hard when modalities live in different feature spaces and cells are unpaired, and existing alignment methods can suffer from weak cross-modality signals and geometric ambiguity.
- What they did: GLAlign optimizes a shared latent representation with a global objective using maximum mean discrepancy, distortion, and penalty terms, plus a local contrastive triplet loss guided by cross-modality marker gene overlap between independently inferred clusters.
- The result: On simulated and real datasets (PBMC scRNA-seq/scATAC-seq, mouse brain SNARE-seq, single-donor and multi-batch CITE-seq, and NCI-N87 scDNA-seq/scRNA-seq), it achieves strong modality mixing and biological conservation, supports cross-modality label transfer, and enables paired cell recovery.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.