# Structured penalized regression for drug sensitivity prediction

**Authors:** Zhi Zhao, Manuela Zucknick  
**arXiv:** [1902.04996](https://arxiv.org/abs/1902.04996)  
**Date:** 2019-02-13  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/1902.04996v3/

## Key points

- Large-scale {\it in vitro} drug sensitivity screens are crucial in personalized oncology for predicting the efficacy of potential cancer treatments.
- Zhao and Zucknick propose a novel approach using joint penalized regression with structured penalty terms to address the challenge of analyzing high-dimensional heterogeneous multi-omics data with strong correlations between outcome variables.
- The method leverages correlation structure between drugs through group-lasso-type penalties and accounts for heterogeneity between omics data sources by introducing data-source-specific penalty factors.
- The authors introduce the IPF-tree-lasso method, which combines integrative penalty factors with tree-guided group lasso, outperforming other lasso-type methods, particularly for heterogeneous data sources.
- The application of their methodology to analyze data from the Genomics of Drug Sensitivity in Cancer project contributes valuable insights into improving drug sensitivity prediction in personalized oncology.

## Summary

, , , , 

In the field of personalized oncology, large-scale {\it in vitro} drug sensitivity screens play a crucial role in predicting the efficacy of potential cancer treatments. These screens involve assessing the sensitivity of cancer cell lines to a variety of drugs, presenting a complex multivariate regression problem. The input data for these analyses consist of high-dimensional heterogeneous multi-omics data, with strong correlations between the outcome variables representing sensitivity to different drugs. To address this challenge, Zhao and Zucknick propose a novel approach using joint penalized regression with structured penalty terms. This method leverages the correlation structure between drugs through group-lasso-type penalties while also accounting for heterogeneity between omics data sources by introducing data-source-specific penalty factors. By combining integrative penalty factors (IPF) with tree-guided group lasso, they introduce the IPF-tree-lasso method. The authors present a unified framework that transforms more general IPF-type methods into the original penalized method. To optimize multiple penalty parameters efficiently, they demonstrate the use of the interval-search Efficient Parameter Selection via Global Optimization (EPSGO) algorithm. Simulation studies indicate that the IPF-tree-lasso approach outperforms other lasso-type methods, particularly for heterogeneous data sources. Finally, Zhao and Zucknick apply their new methodology to analyze data from the Genomics of Drug Sensitivity in Cancer project. Their work contributes valuable insights into improving drug sensitivity prediction in personalized oncology and showcases the potential of structured penalized regression approaches in this domain.