SpaBalance: Balanced Learning for Efficient Spatial Multi-Omics Decoding.

Area of Science:

• Computational Biology
• Systems Biology
• Genomics

Background:

• Spatially resolved multi-omics enables simultaneous profiling of multiple molecular layers in tissues.
• Integrative analysis of multi-omics data faces challenges due to assay discrepancies, causing gradient conflicts in joint learning.
• These conflicts hinder the performance of multi-omics data integration.

Purpose of the Study:

• To propose SpaBalance, a computational framework for harmonizing cross-omics learning in spatial multi-omics data.
• To overcome gradient conflicts and improve the integration of multi-omics datasets.
• To enhance the delineation of spatial domains and uncover multi-omics regulatory hubs.

Main Methods:

• SpaBalance employs a gradient equilibrium mechanism to dynamically balance inter-omics contributions during backpropagation.
• It utilizes task-specific prioritization to resolve gradient conflicts without manual weighting.
• A dual-stream architecture learns shared representations while preserving omics-specific features.

Main Results:

• SpaBalance demonstrated superior performance in delineating complex spatial domains across human tumor and brain tissue datasets.
• The framework successfully uncovered previously hidden multi-omics regulatory hubs.
• Clustering accuracy and biological interpretability were significantly improved.

Conclusions:

• SpaBalance effectively harmonizes spatial multi-omics data integration by resolving gradient conflicts.
• The framework advances spatially resolved systems biology by bridging data integration with biological discovery.
• SpaBalance offers a scalable solution for integrating multiple omics, enhancing biological insights.