Comparative evaluation of feature reduction methods for drug response prediction.

Area of Science:

• Computational biology
• Genomics
• Machine learning in medicine

Background:

• Personalized medicine requires predicting drug responses from molecular profiles.
• High-dimensional molecular data and limited samples pose challenges for machine learning models.
• Knowledge-based feature selection can improve prediction accuracy and model interpretability.

Purpose of the Study:

• To comparatively evaluate nine knowledge-based and data-driven feature reduction methods for drug response prediction.
• To assess the performance of different machine learning models using these feature selection techniques.
• To identify the most effective feature reduction strategy for predicting drug sensitivity in cancer.

Main Methods:

• Conducted a comparative evaluation of nine feature reduction methods (knowledge-based and data-driven).
• Utilized six distinct machine learning models for prediction tasks.
• Performed over 6,000 evaluation runs on cell line and tumor datasets.
• Assessed the ability to distinguish between drug-sensitive and drug-resistant tumors.

Main Results:

• Transcription factor activities emerged as the top-performing feature reduction method.
• This method demonstrated superior performance in predicting drug responses across various datasets.
• Successfully distinguished sensitive from resistant tumors for seven out of 20 evaluated drugs.
• Outperformed other tested methods in predictive accuracy and biological relevance.

Conclusions:

• Transcription factor activities offer a powerful approach for feature selection in drug response prediction.
• Leveraging biological knowledge, specifically transcription factor activities, enhances machine learning model performance in personalized oncology.
• This finding has significant implications for developing more effective, tailored cancer therapies.