Does combining numerous data types in multi-omics data improve or hinder performance in survival prediction? Insights...

Area of Science:

• Genomics
• Bioinformatics
• Cancer Research

Background:

• Multi-omics data integration shows promise for improving predictive modeling over single-omics approaches.
• Current research often assumes more omics data types lead to better predictions, overlooking potential performance degradation from redundant or less informative data.
• Identifying optimal omics data combinations is crucial for cost-effective and accurate predictive modeling.

Purpose of the Study:

• To systematically evaluate all 31 possible combinations of five genomic data types (mRNA, miRNA, methylation, DNAseq, copy number variation) for cancer survival prediction.
• To determine the most effective combinations of omics data types for enhancing predictive performance.
• To challenge the prevailing assumption that increased data types invariably improve multi-omics predictive accuracy.

Main Methods:

• Utilized 14 cancer datasets with right-censored survival outcomes from the TCGA database.
• Assessed all 31 combinations of five omics data types, up-weighting clinical data in all models.
• Employed Harrell's C-index and integrated Brier Score for performance evaluation, with bootstrap analysis for robustness.

Main Results:

• For most cancer types, mRNA data alone or combined with miRNA data provided sufficient predictive performance.
• Inclusion of methylation data improved predictions for some cancer types.
• Adding more data types frequently led to decreased performance, varying between the two metrics used.

Conclusions:

• Findings challenge the common belief that integrating numerous omics data types enhances multi-omics survival prediction.
• The study suggests reconsidering the strategy of incorporating maximum data types to avoid suboptimal predictions and reduce costs.