Enhanced co-expression extrapolation (COXEN) gene selection method for building anticancer drug response prediction models.

Yitan Zhu,Rick L Stevens,Alexander Partin,Thomas S Brettin,Hyunseung Yoo,Maulik Shukla,Fangfang Xia

doi:10.1200/jco.2020.38.15_suppl.e14073

Abstract

e14073 Background: Accurate prediction of tumor response to a drug treatment is of paramount importance for precision oncology. The co-expression extrapolation (COXEN) gene selection approach has been successfully used in multiple studies to select genes for predicting the response of tumor cells to a specific drug. Here, we enhance the original COXEN approach to select genes that are predictive of the efficacies of multiple drugs simultaneously for building general drug response prediction model. Methods: We implemented two methods to select predictive genes. The first method ranks the genes according to their prediction power for each individual drug and then takes a union of top predictive genes of all the drugs. The second method uses a linear regression model to evaluate the prediction power of a gene for all drugs while the drugs are one-hot encoded in the regression model. Among the predictive genes, we further select genes by evaluating the preservation of co-expression patterns between cancer cases with drug response data available and cancer cases for which drug response needs to be predicted, because the preservation of co-expression patterns indicates the similarity of genomic regulations between cancer cases. Results: To test the enhanced COXEN method, we used a lightGBM regression model to predict drug response based on the selected genes on two benchmark in vitro drug screening datasets. The table below compares the performance of prediction models built based on 200 genes selected by the enhanced COXEN method to that of models built on 200 genes randomly picked from the LINCS gene set, which includes 976 “landmark” genes well-representing cellular transcriptomic changes identified in the Library of Integrated Network-Based Cellular Signatures (LINCS) project. The enhanced COXEN approach selects genes better than random LINCS genes as demonstrated by the increased average coefficient of determination (R2) for predicting the area under the dose response curve through cross-validation. Pair-wise t-test indicates the improvement is statistically significant (p-value ≤ 0.05) on both datasets. Conclusions: Our result demonstrates the benefit of using an enhanced COXEN approach to select genes for building general drug response prediction model. [Table: see text]

Highlights

Cancer is a heterogeneous disease at both the histologic and genetic levels
To evaluate the proposed enhanced co-expression extrapolation (COXEN) method, we applied it on three benchmark in vitro drug screening datasets, the Cancer Cell Line Encyclopedia (CCLE) dataset [18], the Genentech Cell response
We developed an enhanced COXEN method to select predictive and generalizable genes for building general drug response prediction models

Summary

Introduction

Cancer is a heterogeneous disease at both the histologic and genetic levels. Patients with the same cancer histology can respond differently to the same treatment. Accurate prediction of a patient’s response to a drug treatment is of paramount importance to the success of precision oncology. Multiple types of tumor omics data have been used in many studies for predicting anti-cancer drug response [1,2,3,4,5], among which transcriptome data have been shown to be the most important for drug response prediction [6,7]. Because the transcriptome data usually contain the expression values of about 20,000 genes, which can be computationally expensive for training prediction models and cause model overfitting on data without a large number of samples, gene selection is frequently applied to select a group of genes most useful for the prediction of drug response [6,8,9].

Methods

Results

Conclusion