A review of multi-omics integration techniques across five machine learning method families

Area of Science:

• Computational Biology
• Bioinformatics
• Cancer Research

Background:

• Multi-omics integration is widely used in cancer research but yields variable results due to design choices.
• Lack of standardization in integration methods hinders cross-study comparisons and reliability assessment.

Purpose of the Study:

• To systematically review and analyze multi-omics integration methods in cancer studies.
• To identify common practices, emerging trends, and critical design choices influencing integration outcomes.
• To provide guidance on selecting appropriate integration strategies for robust cancer data analysis.

Main Methods:

• A PRISMA-guided systematic review of 30 cancer multi-omics integration studies published between 2020 and 2025.
• Categorization of integration methods into families (e.g., graph-based, hybrid, deep learning).
• Analysis of fusion timing (early, intermediate, late) and its alignment with method families and use cases.

Main Results:

• Graph-based/hybrid pipelines and deep learning are dominant integration methods, primarily used for survival prediction.
• Fusion timing strategies vary: graph-hybrid methods favor early-intermediate fusion, while deep learning spans all stages.
• Key trade-offs identified: early-intermediate fusion stabilizes inputs but is sensitive to imbalance; shared latent spaces handle missing data; late fusion offers stable subtypes but complicates feature attribution.

Conclusions:

• Optimal multi-omics integration requires aligning fusion choices with data properties (noise, sparsity, missingness).
• Interpretability should be an integral part of the integration architecture, not an afterthought.
• Standardized reporting and methodological choices are needed to improve the reliability and comparability of cancer multi-omics studies.