TL;DR
This paper introduces STATegra, a novel, step-wise bioinformatics framework designed to integrate multi-omics data for enhanced biological discovery. By combining component analysis, non-parametric data combination, and exploratory analysis, STATegra offers a systematic approach to uncover insights that single-omics analyses miss, as demonstrated in two cancer case studies.
What Problem Does This Paper Address?
The explosion of high-throughput technologies allows us to measure various biological layers (genomics, transcriptomics, epigenomics, etc.) simultaneously. This "multi-omics" data holds immense potential for understanding complex biological systems and diseases. However, a significant challenge lies in effectively integrating and analyzing these diverse datasets. Currently, researchers often build custom pipelines from scratch, making arbitrary choices about which tools to use and how to combine them. This ad-hoc approach can lead to inconsistent results and hinders reproducible discovery. STATegra aims to address this by providing a generic, conceptual framework and a corresponding bioinformatics pipeline for systematic multi-omics data integration.
Methods
STATegra is a four-step framework:
- Independent Data-Type Exploration and Characterization: Each omics dataset is analyzed individually using quality control and exploratory methods like Principal Component Analysis (PCA). This step helps identify potential batch effects and understand the inherent variability within each data type.
- Joint Exploration and Characterization: This step focuses on identifying "coordination" or shared variability between different omics datasets. Techniques like Joint and Individual Variation Explained (JIVE) are used to find components that are informative across multiple omics layers. This helps determine which omics combinations are most promising for further analysis.
- Integrative Differential Analysis (omicsNPC): Leveraging the coordination identified in Step 2, this step uses Non-Parametric Combination (NPC) to increase statistical power. NPC non-parametrically combines p-values from associated features across different omics datasets to identify significant features that explain a biological outcome (e.g., survival, disease subtype). The STATegra package includes an
omicsNPCfunction for this purpose. - Exploratory Analysis and Determination of the Framework's Added Value: The final step involves summarizing and exploring the biological insights gained from the integrated analysis. This is achieved using tools like GeneSetCluster, which groups gene sets and facilitates the identification of novel pathways and functional annotations.
The framework is implemented as an open-source Bioconductor package, making it accessible to researchers.
Datasets
The STATegra framework was validated using two publicly available datasets from The Cancer Genome Atlas (TCGA):
- Glioblastoma Multiforme (GBM): Data included gene expression (mRNA), miRNA, and DNA methylation.
- Skin Cutaneous Melanoma (SKCM): Data also included gene expression (mRNA), miRNA, and DNA methylation.
The paper notes that for some analyses, they considered scenarios where not all samples were profiled for all omics, demonstrating the framework's flexibility.
Key Insights and Ideas
- Systematic Integration is Crucial: The paper strongly advocates for a structured, step-wise approach to multi-omics analysis rather than ad-hoc pipeline construction.
- Coordination is Key to Integration: Identifying shared variability (coordination) between omics datasets in Step 2 is essential for effectively leveraging them in downstream analyses like NPC.
- Enhanced Statistical Power: Non-parametric combination (NPC) significantly boosts the ability to detect novel features and associations compared to single-omics analyses.
- Discovery of New Biology: The integrated framework facilitates the discovery of novel genes, miRNAs, and pathways that are not apparent when analyzing omics data in isolation.
- Framework Flexibility: STATegra can accommodate datasets with missing omics profiles for certain samples and offers both parametric and non-parametric options for analysis.
Key Results
- GBM Case Study: STATegra identified 23 new genes and 4 new miRNAs when integrating mRNA and miRNA data, and 106 new genes and 150 new CpG sites when integrating mRNA and DNA methylation data, for survival outcome analysis.
- SKCM Case Study: For primary site of disease analysis, STATegra identified 48 new genes and 14 new miRNAs (mRNA + miRNA integration), and 116 new genes and 428 new CpG sites (mRNA + DNA methylation integration).
- Gene Set Enrichment: The GeneSetCluster analysis revealed novel biological functions and canonical pathways, some of which were previously unassociated with the specific cancer types or were identified only through multi-omics integration. For example, the TREM1 signaling pathway was identified as newly associated with GBM.
Limitations and Future Work
- Model Selection in Component Analysis: The paper acknowledges that determining the optimal model for component analysis (i.e., the number of shared and distinctive components) remains an open question, and no definitive function is yet included in the STATegra package. They recommend using multiple methodologies for robustness.
- Parametric vs. Non-Parametric NPC: While the parametric version of NPC is faster for preliminary explorations, the non-parametric version is suggested for more robust results, especially when dealing with complex inter-data-set correlations.
- Future Development: The authors plan to continuously update the STATegra Bioconductor package by incorporating additional tools, such as PESCA for multi-omic component analysis and GeneSetCluster for multi-omic exploratory analysis. They also highlight the need for further developments in multi-omics visualization, simulated data generation, and advanced component analysis techniques. The framework is seen as a starting point for future development.
Summary updated