A benchmark study of deep learning-based multi-omics data fusion methods for cancer.

Area of Science:

• Computational Biology
• Bioinformatics
• Genomics

Background:

• Multi-omics data integration is crucial for understanding complex biological systems.
• High-throughput sequencing has enabled the generation of large-scale multi-omics datasets.
• Deep learning methods offer promising approaches for fusing multi-omics data.

Purpose of the Study:

• To comprehensively evaluate representative deep learning methods for multi-omics data fusion.
• To compare the performance of these methods on classification and clustering tasks.
• To assess the utility of these methods in cancer multi-omics data analysis, including survival prediction.

Main Methods:

• Benchmarking of 16 deep learning methods on simulated, single-cell, and cancer multi-omics datasets.
• Evaluation using classification metrics (accuracy, F1 macro/weighted) and clustering metrics (Jaccard index, C-index, silhouette score, Davies Bouldin score).
• Assessment of the association between dimensionality reduction results and clinical/survival data for cancer datasets.

Main Results:

• moGAT demonstrated superior classification performance across datasets.
• efmmdVAE, efVAE, and lfmmdVAE exhibited strong and promising clustering performance.
• The study identified top-performing methods for specific multi-omics data fusion tasks.

Conclusions:

• Provides a valuable reference for selecting deep learning-based multi-omics data fusion methods.
• Offers insights into future research directions for developing more effective fusion techniques.
• Open-source deep learning frameworks are available for public use and further development.