Improving stability of prediction models based on correlated omics data by using network approaches

Area of Science:

• Bioinformatics
• Biostatistics
• Systems Biology

Background:

• High-dimensional omics datasets (transcriptomics, proteomics, metabolomics) present challenges for prediction model development.
• Standard regularized regression methods struggle with correlated data, leading to unstable models and difficulties in selection.
• Existing approaches often fail to adequately address the inherent complexity and correlations within omics data.

Purpose of the Study:

• To propose a novel, robust strategy for selecting prediction models from high-dimensional omics data.
• To enhance the predictability and stability of models built on complex biological datasets.
• To provide practical recommendations for choosing prediction modeling strategies based on data characteristics and analytical goals.

Main Methods:

• A three-step approach: 1) weighted correlation network and Gaussian graphical modeling for network construction, 2) hierarchical clustering for module/pathway derivation, and 3) prediction model building using module information.
• Incorporation of biological pathway information via group-based variable selection or group-specific penalization.
• Comparative performance evaluation against standard regularized regression techniques using simulation studies.

Main Results:

• The proposed network-based, module-informed approaches demonstrate improved performance and stability compared to standard regularized regression methods in simulations.
• The study identifies optimal strategies for prediction model building based on dataset size and specific analytical objectives.
• Successful application to real-world datasets for predicting body mass index (DILGOM study) and drug response in breast cancer cell lines.

Conclusions:

• The novel three-step strategy offers a more reliable method for building prediction models from complex omics data.
• Integrating network and pathway information enhances model selection and predictability, addressing limitations of traditional methods.
• The findings provide valuable guidance for researchers in bioinformatics and biostatistics aiming to leverage omics data for predictive modeling.