Benchmarking computational methods for multi-omics biomarker discovery in cancer.

Area of Science:

• Computational biology
• Biomarker discovery
• Cancer research

Background:

• Multi-omics profiling is crucial for understanding cancer biology and identifying prognostic and therapeutic biomarkers.
• Existing computational methods for multi-omics biomarker identification lack systematic evaluation, raising concerns about the reliability of nominated biomarkers.

Purpose of the Study:

• To systematically benchmark 20 statistical, machine learning, and deep learning methods for multi-omics biomarker identification.
• To evaluate method performance based on accuracy and stability of identified biomarkers.
• To provide a framework for reliable biomarker discovery in cancer.

Main Methods:

• Benchmarking 20 diverse computational methods using curated gold-standard biomarkers across five real-world datasets.
• Evaluating biomarker identification accuracy and stability.
• Utilizing simulated datasets to assess sensitivity to signal and noise.

Main Results:

• DeePathNet and DeepKEGG demonstrated superior performance in identifying reliable biomarkers.
• Effective biomarker recovery correlated with biological knowledge integration, global feature interactions, multivariate feature attribution, and regularization.
• Method and omics type biases were observed, emphasizing the need for broader omics integration.

Conclusions:

• The study provides a comprehensive benchmark for multi-omics biomarker identification methods.
• Consensus biomarker panels were constructed from top-performing methods for further investigation.
• User-friendly interfaces are available for researchers to benchmark new methods or apply existing ones to custom datasets.