Bimodal Gene Expression in Patients with Cancer Provides Interpretable Biomarkers for Drug Sensitivity

Area of Science:

• Genomics
• Computational Biology
• Pharmacology

Background:

• Precision oncology relies on identifying biomarkers for cancer drug response.
• Existing predictive models often lack interpretability, hindering clinical application.
• Large-scale pharmacogenomic datasets offer opportunities for biomarker discovery.

Purpose of the Study:

• To develop an interpretable machine learning pipeline for predicting cancer drug response.
• To leverage bimodal gene expression patterns for biomarker discovery.
• To improve the clinical translatability of predictive models.

Main Methods:

• Utilized a logic modeling approach to build a new machine learning pipeline.
• Explored bimodally expressed genes across multiple large in vitro pharmacogenomic studies.
• Developed multivariate, nonlinear, interpretable logic-based models.

Main Results:

• Generated robust and interpretable models for 101 drugs across 17 drug classes.
• Achieved high validation rates in independent datasets.
• Demonstrated support for in vivo and clinical validation of gene expression biomarkers.

Conclusions:

• The developed pipeline effectively identifies predictive biomarkers for cancer drug response.
• Bimodal gene expression analysis facilitates interpretable and clinically translatable biomarkers.
• This approach enhances the translation of biomarkers between different model systems.