SurvBoard: standardized benchmarking for multi-omics cancer survival models

Area of Science:

• Computational biology
• Cancer research
• Bioinformatics

Background:

• Multi-omics data (genomic, transcriptomic, epigenetic, proteomic) are crucial for cancer patient outcome prediction.
• Existing studies highlight the need for standardized methods to compare cancer survival prediction models.

Purpose of the Study:

• Introduce SurvBoard, a benchmark framework to standardize multi-omics cancer survival model evaluation.
• Enable comparisons between single-cancer and pan-cancer models and assess the utility of incomplete patient data.

Main Methods:

• Developed SurvBoard, a framework standardizing experimental design for multi-omics survival models.
• Addressed common pitfalls in preprocessing and validation.
• Applied SurvBoard to exemplary use cases.

Main Results:

• Statistical models generally outperform deep learning methods in survival prediction, particularly in survival function calibration.
• Pan-cancer models demonstrate improved performance.
• Leveraging samples with missing omics modalities benefits model performance.

Conclusions:

• SurvBoard provides a standardized approach for evaluating multi-omics cancer survival models.
• The framework facilitates reproducible research and highlights the advantages of pan-cancer analysis and handling missing data.