Heterogeneity Aware Random Forest for Drug Sensitivity Prediction.

Area of Science:

• Pharmacogenomics
• Computational Biology
• Cancer Research

Background:

• Pharmacogenomics databases contain samples from diverse cancer types, presenting challenges for drug sensitivity prediction.
• Inter-tumor heterogeneity can impact the performance of predictive models trained on mixed cancer samples.

Purpose of the Study:

• To investigate whether accounting for cancer type heterogeneity improves drug sensitivity prediction models.
• To develop a novel ensemble method that incorporates sample heterogeneity into Random Forests.

Main Methods:

• Proposed Heterogeneity Aware Random Forests (HARF), an ensemble method weighting trees based on sample category.
• Treated heterogeneity as a latent class allocation problem using a covariate-free approach based on leaf node distribution.
• Applied and evaluated HARF on the Cancer Cell Line Encyclopedia (CCLE) and Genomics of Drug Sensitivity in Cancer (GDSC) databases.

Main Results:

• Ensemble model predictions were superior when cancer type information was known, even with smaller sample sizes.
• HARF demonstrated improved performance over traditional Random Forest when average drug responses across cancer types varied.
• The covariate-free class allocation approach effectively managed heterogeneity.

Conclusions:

• Accounting for cancer type heterogeneity is crucial for building accurate drug sensitivity predictive models.
• HARF offers a robust approach to enhance predictive accuracy by addressing tumor heterogeneity.
• This method has significant implications for personalized medicine and drug development in oncology.