LoHi-SSL: A Multi-Level Synergistic Learning Model for Integrating Single-Cell Multi-Omics Data via Low- and...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Single-cell sequencing reveals cellular heterogeneity but integrating multi-omics data is challenging due to data variability.
• Existing methods struggle with cross-omics and intra-omics heterogeneity, limiting comprehensive cellular analysis.

Purpose of the Study:

• To develop an efficient model, LoHi-SSL, for synergistic multi-level learning to integrate single-cell multi-omics data.
• To address challenges in fusing data from multiple molecular layers by accounting for cellular heterogeneity.

Main Methods:

• LoHi-SSL employs three modules: low-order learning (Graph Autoencoder for intra-omics similarity), high-order learning (multi-omics hypergraph for cross-omics alignment), and feature integration (contrastive learning for discriminative representations).
• The model learns a unified latent space, aligning features from different omics for the same cell and separating representations for different cell types.

Main Results:

• LoHi-SSL outperformed existing methods on six datasets for clustering tasks, showing significant improvements in NMI, ARI, AMI, and ACC.
• Robustness analysis confirmed LoHi-SSL's resistance to noise.
• Latent representations accurately predicted biological evolutionary pathways through cell trajectory analysis.

Conclusions:

• LoHi-SSL offers an efficient and robust solution for single-cell multi-omics data integration.
• The model provides a powerful tool for studying cellular heterogeneity, state transitions, and regulatory mechanisms.