Comparative analysis of regression algorithms for drug response prediction using GDSC dataset

Area of Science:

• Bioinformatics
• Computational Biology
• Genomics

Background:

• Drug response prediction utilizes individual genetic profiles for personalized treatment selection.
• Machine learning approaches are increasingly applied to predict drug response.
• High-throughput sequencing generates vast datasets, posing challenges for algorithm selection.

Purpose of the Study:

• To evaluate and compare the performance of various regression algorithms for drug response prediction.
• To assess the impact of feature selection, multi-omics data, and drug categories on prediction accuracy.
• To guide bioinformatics researchers in selecting appropriate methods for drug response modeling.

Main Methods:

• Compared 13 regression algorithms on the Genomics of Drug Sensitivity in Cancer (GDSC) dataset.
• Investigated the influence of feature selection methods, multi-omics data (mutation, copy number variation), and drug categories.
• Utilized LINC L1000 gene expression data for feature selection.

Main Results:

• Support Vector Regression with LINC L1000 gene features demonstrated superior accuracy and efficiency.
• Integrating mutation and copy number variation data did not enhance prediction performance.
• Drugs targeting hormone-related pathways showed higher prediction accuracy.

Conclusions:

• This study provides insights for optimizing data processing and algorithm selection in drug response prediction.
• Findings can aid in developing robust predictive models using large-scale genomic datasets.
• Recommends specific algorithms and feature sets for improved drug response prediction accuracy.