A comprehensive benchmarking of machine learning algorithms and dimensionality reduction methods for drug sensitivity...

Area of Science:

• Computational biology
• Genomics
• Pharmacology

Background:

• Precision oncology relies on identifying molecular biomarkers for targeted cancer treatments.
• Large cancer cell line datasets are crucial for understanding the link between cellular features and drug response.
• High-dimensional data necessitates machine learning (ML) for analysis, but algorithm and feature selection remain challenging.

Purpose of the Study:

• To comprehensively benchmark machine learning methods and dimension reduction techniques for predicting drug response metrics.
• To compare ML models based on statistical performance, runtime, and interpretability.
• To provide strategies for model performance assessment and complexity trade-offs.

Main Methods:

• Benchmarking of random forests, neural networks, boosting trees, and elastic nets.
• Application of nine dimension reduction (DR) approaches to feature sets.
• Training and evaluation using the Genomics of Drug Sensitivity in Cancer cell line panel for 179 anti-cancer compounds.

Main Results:

• Complex ML models show improved performance with optimized DR strategies.
• Standard ML models can outperform complex models even with reduced feature sets.
• Performance, runtime, and interpretability were key comparison metrics.

Conclusions:

• Dimension reduction is critical for optimizing complex machine learning models in precision oncology.
• Simpler machine learning models offer a viable and potentially superior alternative, especially when computational resources or interpretability are prioritized.
• The study provides a framework for evaluating and selecting appropriate ML and DR strategies for drug response prediction.