Large-scale benchmark study of survival prediction methods using multi-omics data.

Area of Science:

• Computational biology
• Biostatistics
• Genomics

Background:

• Multi-omics data integration is growing for disease research.
• The predictive value of multi-omics data for disease outcomes, like survival time, remains uncertain.
• Optimal methods for deriving predictive models from multi-omics data are not well-established.

Purpose of the Study:

• To benchmark the usefulness of multi-omics data for predicting cancer survival.
• To compare various machine learning and statistical methods for multi-omics survival prediction.
• To assess the impact of incorporating omics data structure into prediction models.

Main Methods:

• Applied 11 prediction methods (boosting, penalized regression, random forest) to 18 TCGA multi-omics cancer datasets.
• Compared methods with and without consideration of omics variable group structure.
• Utilized 5-fold cross-validation, Uno's C-index, and integrated Brier score for performance evaluation.
• Used Kaplan-Meier and Cox models with clinical variables as references.

Main Results:

• Methods incorporating multi-omics data structure showed slightly improved prediction performance.
• Accounting for data structure helped preserve predictive information from clinical variables.
• Only the block forest method marginally outperformed the Cox model on average.
• The utility of multi-omics data for survival prediction in TCGA studies was found to be limited.

Conclusions:

• Multi-omics data integration offers marginal benefits for cancer survival prediction compared to traditional methods.
• Careful consideration of data structure is important for maximizing predictive power.
• Further research is needed to fully leverage the potential of multi-omics data in clinical outcome prediction.