# A cross-study analysis of drug response prediction in cancer cell lines

**Authors:** Fangfang Xia, Jonathan Allen, Prasanna Balaprakash, Thomas Brettin, Cristina Garcia-Cardona, Austin Clyde, Judith Cohn, James Doroshow, Xiaotian Duan, Veronika Dubinkina, Yvonne Evrard, Ya Ju Fan, Jason Gans, Stewart He, Pinyi Lu, Sergei Maslov, Alexander Partin, Maulik Shukla, Eric Stahlberg, Justin M. Wozniak  
**arXiv:** [2104.08961](https://arxiv.org/abs/2104.08961)  
**Date:** 2021-04-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2104.08961v2/

## Key points

- Machine learning models can predict drug response for personalized cancer treatment
- Cross-validation within a single study may provide an overly optimistic estimate of prediction performance on independent test sets
- Researchers used machine learning to analyze five publicly available cell line-based data sets to assess model generalizability between different studies
- Multitasking deep neural network achieved the best cross-study generalizability
- Differences in viability assays can limit model generalizability across studies
- Drug diversity is crucial for raising model generalizability in preclinical screening more than tumor diversity
- Using diverse drugs in preclinical screening could improve model generalizability across different studies, providing an effective approach for personalized cancer treatment.

## Summary

Machine learning models have been developed to predict drug response as a function of tumor and drug features, enabling personalized cancer treatment. However, most algorithm development efforts rely on cross-validation within a single study to assess model accuracy, which typically provides an overly optimistic estimate of prediction performance on independent test sets. To provide a more rigorous assessment of model generalizability between different studies, researchers used machine learning to analyze five publicly available cell line-based data sets: NCI60, CTRP, GDSC, CCLE, and gCSI. Based on observed experimental variability across studies, the researchers explored estimates of prediction upper bounds. They reported performance results of various machine learning models and found that a multitasking deep neural network achieved the best cross-study generalizability. By multiple measures, models trained on CTRP yielded the most accurate predictions on the remaining testing data, and gCSI was the most predictable among the cell line data sets included in this study. Through experiments and further simulations on partial data, two lessons emerged: differences in viability assays can limit model generalizability across studies; and drug diversity is crucial for raising model generalizability in preclinical screening more than tumor diversity. The researchers also noted that drug sensitivity prediction is challenging due to biological complexity and experimental variability. This study provides valuable insights into improving machine learning models' ability to predict drug response for personalized cancer treatment by analyzing multiple publicly available datasets. The findings suggest that using diverse drugs in preclinical screening could improve model generalizability across different studies, thus providing an effective approach for personalized cancer treatment.