scGSI: a graph-guided, self-supervised framework to integrate paired single-cell multi-omics

# What scGSI does scGSI is a computational framework for integrating paired single-cell multi-omics data—measurements made on the same cell across different molecular layers (for example, scRNA-seq and scATAC-seq). The method aims to align paired cells across modalities while keeping the modality-specific topological structure and biological variation that downstream analyses need.

# Why this matters Paired multi-omics give direct within-cell correspondences across molecular layers, which is valuable for dissecting cellular heterogeneity and regulatory relationships. Existing integration methods can struggle in three ways: topology mismatch across modalities, underuse of within-cell cross-modal complementarity, or improving alignment at the cost of biological fidelity. scGSI targets all three issues.

# How scGSI works (high level) scGSI combines three main components:

• Heterogeneous graph encoders: These build and preserve modality-specific neighborhood structures rather than collapsing modality differences prematurely.

• Pull-in projection module: This module stabilizes the pre-alignment step by pulling paired cell representations closer in a controlled way, reducing the risk of forcing incompatible topologies together.

• Cross-fusion with contrastive refinement: After pre-alignment, cross-fusion uses paired-cell signals to combine modalities. Contrastive refinement then refines the joint embeddings to strengthen true cross-modal correspondences while maintaining biological variation.

# What the authors tested

# Main findings scGSI improved paired cell-state alignment across the tested datasets while retaining modality-specific topological structure. The resulting embeddings supported clearer cell-type separation and more reliable trajectory inference compared with competing approaches that either over-mix modalities or fail to exploit within-cell complementarity.

# Practical implications for researchers If you work with paired single-cell multi-omics, scGSI offers an approach that:

• Uses graph-based encodings to keep modality-appropriate neighborhoods intact before alignment.
• Applies a controlled projection step to avoid destabilizing topology when bringing paired cells together.
• Applies contrastive learning on fused representations to strengthen biologically meaningful cross-modal links.

The authors provide code on GitHub and used publicly available datasets, enabling reproducibility and benchmarking on your own data.

# Where scGSI fits compared with other methods Graph- and contrastive-learning methods have already been applied to spatial and multi-omics problems. scGSI specifically targets paired single-cell multi-omics and emphasizes preserving modality topology while using paired-cell complementarity to improve alignment. That focus makes it suited when within-cell correspondence is available and biological fidelity of structure matters for downstream analysis.

# Bottom line scGSI is a targeted integration framework for paired single-cell multi-omics that balances alignment and biological preservation through graph encoders, a pull-in projection, and contrastive cross-fusion. Across multiple public datasets, it improved alignment and downstream interpretability without erasing modality-specific structure.