Using biological constraints to improve prediction in precision oncology.

Area of Science:

• Computational biology
• Genomics
• Cancer research

Background:

• Machine learning (ML) applied to omics data generates gene signatures for cancer.
• Clinical utility of these signatures is limited by poor interpretability and unstable performance.

Purpose of the Study:

• To investigate the impact of integrating prior biological knowledge into ML decision rules for developing robust cancer gene signatures.
• To compare the performance and interpretability of ML models trained with and without biological constraints.

Main Methods:

• Applied various ML algorithms to gene expression data for predicting cancer phenotypes.
• Developed two sets of classifiers: 'mechanistic' (using biological mechanisms) and 'agnostic' (no prior biological information).
• Tested classifiers on bladder cancer progression, triple-negative breast cancer chemotherapy response, and prostate cancer metastasis.

Main Results:

• Mechanistic models demonstrated similar or superior predictive performance compared to agnostic models.
• Mechanistic models offered enhanced interpretability of gene signatures.
• Biological constraints improved the robustness and translational potential of ML-derived gene signatures.

Conclusions:

• Embedding prior biological knowledge into ML decision rules is crucial for building robust and interpretable gene signatures.
• Mechanistic ML approaches hold significant promise for advancing cancer diagnostics and treatment strategies.
• This study advocates for the use of biological constraints in developing clinically relevant gene signatures.