Feature selection translates drug response predictors from cell lines to patients

Area of Science:

• Oncology
• Computational Biology
• Pharmacogenomics

Background:

• Identifying predictive markers for cancer targeted therapies and chemotherapies is crucial but limited by sparse patient drug response data.
• Leveraging extensive cell line drug response data requires efficient methods to translate these findings to human tumors for clinical application.

Purpose of the Study:

• To develop versatile feature selection procedures for building accurate and interpretable predictive models of cancer drug response.
• To enable the translation of cell-line-trained predictors to human tumors, improving patient stratification for cancer therapies.

Main Methods:

• Proposed versatile feature selection procedures compatible with various classifiers.
• Demonstrated methods by combining feature selection with logit (LogitDA) and K-nearest neighbor (KNNDA) classifiers trained on cell line data.
• Derived a novel adjustment for the prediction cutoff in LogitDA to enhance prediction accuracy.

Main Results:

• LogitDA and KNNDA significantly outperformed existing methods in prediction AUC (0.70-1.00 for seven of ten drugs).
• The developed models demonstrated interpretability, offering insights into drug response mechanisms.
• Achieved prediction accuracy of 0.70-0.93 for seven drugs, including erlotinib and cetuximab, with uncovered relevant pathways.

Conclusions:

• The proposed feature selection methods efficiently translate cell-line-trained predictors to human tumors.
• These advancements can aid in stratifying cancer patients for more effective targeted therapies and chemotherapies.
• The interpretability of the models provides valuable biological insights for cancer treatment strategies.