Development and evaluation of a java-based deep neural network method for drug response predictions

Abstract

Accurate prediction of drug response is a crucial step in personalized medicine. Recently, deep learning techniques have been witnessed with significant breakthroughs in a variety of areas including biomedical research and chemogenomic applications. This motivated us to develop a novel deep learning platform to accurately and reliably predict the response of cancer cells to different drug treatments. In the present work, we describe a Java-based implementation of deep neural network method, termed JavaDL, to predict cancer responses to drugs solely based on their chemical features. To this end, we devised a novel cost function and added a regularization term which suppresses overfitting. We also adopted an early stopping strategy to further reduce overfit and improve the accuracy and robustness of our models. To evaluate our method, we compared with several popular machine learning and deep neural network programs and observed that JavaDL either outperformed those methods in model building or obtained comparable predictions. Finally, JavaDL was employed to predict drug responses of several aggressive breast cancer cell lines, and the results showed robust and accurate predictions with r ² as high as 0.81.

Keywords: artificial intelligence (AI); deep learning; deep neural network; drug response; multilayer neural network (MNN); quantitative structure activity relationship (QSAR); triple-negative breast cancer (TNBC).

Figure 1

Activation function based on the Tanh function tanh(x) and its derivative tanh'(x).

Figure 2

The overall framework of JavaDL including data processing and DNN implementation.

Figure 3

The overall flowchart of JavaDL Algorithm.

Figure 4

Predicted vs. actual activities obtained for test sets. (A) Prediction of Caco-2 permeability using JavaDL with five descriptors: BCUT_ PEOE_3, GCUT_SLOGP_1, mr, a_base, vsa_base. (B) Prediction of hERG activity with five descriptors: GCUT_PEOE_3, reactive, SlogP_V SA1, SlogP_V SA9, and vdw_area.

Figure 5

Predicted vs. actual biological activities for the Merck Molecular Activity Challenge big data set in Kaggle competition. The model was built using 1,569 compounds with 4,505 descriptors. Among them 250 compounds are picked out randomly for test predictions, and the best model derived (A) q² = 0.99, shown in the left panel, and (B) r² = 0.65, shown in the right panel.

Figure 6

Predicted vs. actual activities (IC₅₀) of drugs in TNBC, including (A) HCC-1937, (B) MDA-MB-231, (C) MDA-MB-453, and (D) MDA-MB-436. The red-circled compound 3-Methyladenine as an outlier was not accurately predicted. Further analysis shows that no compound in the training set is similar enough to 3-Methyladenine while there are always similar compounds to the accurately predicted ones (blue-circled Afatinib and UNC-0638).

Figure 7

From left to right is 3-Methyladenine, Afatinib and UNC-0638, respectively. 3-Methyladenine was not accurately predicted while the other two compounds obtained accurate prediction of their activity in TNBC.