Cancer Drug Sensitivity Prediction Based on Deep Transfer Learning

Abstract

In recent years, many approved drugs have been discovered using phenotypic screening, which elaborates the exact mechanisms of action or molecular targets of drugs. Drug susceptibility prediction is an important type of phenotypic screening. Large-scale pharmacogenomics studies have provided us with large amounts of drug sensitivity data. By analyzing these data using computational methods, we can effectively build models to predict drug susceptibility. However, due to the differences in data distribution among databases, researchers cannot directly utilize data from multiple sources. In this study, we propose a deep transfer learning model. We integrate the genomic characterization of cancer cell lines with chemical information on compounds, combined with the Encyclopedia of Cancer Cell Lines (CCLE) and the Genomics of Cancer Drug Sensitivity (GDSC) datasets, through a domain-adapted approach and predict the half-maximal inhibitory concentrations (IC50 values). Afterward, the validity of the prediction results of our model is verified. This study effectively addresses the challenge of cross-database distribution discrepancies in drug sensitivity prediction by integrating multi-source heterogeneous data and constructing a deep transfer learning model. This model serves as a reliable computational tool for precision drug development. Its widespread application can facilitate the optimization of therapeutic strategies in personalized medicine while also providing technical support for high-throughput drug screening and the discovery of new drug targets.

Keywords: deep transfer learning; domain-adapted approach; drug sensitivity; multi-source data.

Figure 1

The 10 drugs with the lowest and highest log(IC50) values in the unknown drug response experiment.

Figure 2

Pathways enriched for key genes in ZR-75-30 cells. Multiple pathways are involved in the mechanism of action or pathogenesis of cancer.

Figure 3

Feature space before domain adaptation.

Figure 4

Feature space after domain adaptation.

Figure 5

General framework for deep transfer learning. The source domain and the target domain are mapped into a common data space.

Figure 6

Architecture of a stacked autoencoder. x and x’ denote the input and the reconstructed output, respectively, and h denotes the encoded feature representation.

Figure 7

Flowchart of our DADSP-A, which consists of two feature extractors, a regression predictor, and a domain discriminator. (A): Gene feature representation, which uses an SAE to map high-dimensional gene expression to low-dimensional representations. (B): Drug feature representation. (C): Transfer learning part, which achieves the goal of feature proximity by maximizing the domain classification error. (D): Test module.

Figure 8

Flowchart of DADSP-B, which consists of two stacked autoencoders and a regression predictor. Part (A) uses only the source domain dataset GDSC through the feature extractor and regressor. Part (B) is the module of transfer learning, which minimizes the MMD loss of the gene features from the source domain and the target domain through the hidden layer features after the feature extractor to achieve a close feature distance between the two domains. The loss of the overall training is the regression loss of the MSE and MMD losses of the common training. Part (C) is the test part, which uses only the target domain dataset CCLE to predict drug sensitivity on the multitrained model.