Feature selection strategies for drug sensitivity prediction.

Area of Science:

• Genomics
• Computational Biology
• Pharmacology

Background:

• Personalized medicine aims to tailor cancer treatments, but predicting drug sensitivity is challenging.
• Cancer cell drug sensitivity relies on complex interactions within a vast biological feature landscape.
• Standard data-driven methods often struggle to identify the most relevant features for accurate prediction.

Purpose of the Study:

• To compare data-driven versus prior knowledge-driven feature selection for drug sensitivity prediction.
• To evaluate the impact of incorporating drug targets, pathways, and gene expression signatures.
• To develop more interpretable and predictive models for cancer therapy design.

Main Methods:

• Utilized the Genomics of Drug Sensitivity in Cancer (GDSC) dataset.
• Assessed 2484 unique predictive models.
• Compared standard feature selection against methods incorporating prior biological knowledge.

Main Results:

• Prior knowledge-based feature selection outperformed standard methods for 23 drugs.
• Linifanib showed the highest correlation (r=0.75) between observed and predicted response.
• Integrating gene expression signatures with drug-specific features yielded optimal models for 60 drugs, notably Dabrafenib.
• Small, knowledge-driven feature sets were highly predictive, especially for drugs targeting specific pathways.

Conclusions:

• Prior knowledge integration significantly enhances drug sensitivity prediction models.
• Feature selection strategies tailored to drug mechanisms (specific vs. general) improve model performance.
• This approach facilitates the development of interpretable models crucial for guiding cancer therapy decisions.