Integrative Network Fusion: A Multi-Omics Approach in Molecular Profiling.

Area of Science:

• Computational Biology
• Bioinformatics
• Oncology

Background:

• Large-scale pan-cancer datasets with multi-omics and clinical information are now available through initiatives like The Cancer Genome Atlas (TCGA).
• There is a growing need for advanced computational methods to integrate these multi-modal data for improved cancer subtyping and biomarker identification.

Purpose of the Study:

• To develop and evaluate the Integrative Network Fusion (INF) pipeline for multi-modal oncogenomics data integration.
• To assess INF's performance in cancer classification tasks, including subtyping and survival prediction, compared to existing methods.

Main Methods:

• The Integrative Network Fusion (INF) pipeline combines multiple omics layers using Similarity Network Fusion (SNF) within a machine learning framework.
• INF incorporates a feature ranking scheme (rSNF) and uses classifiers like Random Forest (RF) and linear Support Vector Machine (LSVM) on juxtaposed multi-omics features (juXT).
• A compact RF model (rSNFi) is derived from top-ranked biomarkers, and all models are validated using a 10x5-fold cross-validation schema.

Main Results:

• INF demonstrated comparable or improved classification accuracy (Matthews Correlation Coefficient) with significantly smaller feature sizes (83-97% reduction) compared to juXT on TCGA datasets.
• INF improved performance for predicting survival in renal clear cell carcinoma (KIRC-OS) and achieved strong results for estrogen receptor status (BRCA-ER) and breast cancer subtypes (BRCA-subtypes).
• INF predictions generally outperformed one-dimensional omics models, with transcriptomics consistently playing a key role, and yielded more compact biomarker signatures.

Conclusions:

• The INF framework effectively integrates multiple data levels in oncogenomics classification tasks, outperforming single-layer analysis and naive data juxtaposition.
• INF provides a robust method for cancer subtyping and biomarker identification, offering improved accuracy and significantly reduced feature set sizes.
• The developed computational pipeline offers a valuable tool for leveraging complex multi-omics data in cancer research and clinical applications.