DeepCarc: Deep Learning-Powered Carcinogenicity Prediction Using Model-Level Representation

Abstract

Carcinogenicity testing plays an essential role in identifying carcinogens in environmental chemistry and drug development. However, it is a time-consuming and label-intensive process to evaluate the carcinogenic potency with conventional 2-years rodent animal studies. Thus, there is an urgent need for alternative approaches to providing reliable and robust assessments on carcinogenicity. In this study, we proposed a DeepCarc model to predict carcinogenicity for small molecules using deep learning-based model-level representations. The DeepCarc Model was developed using a data set of 692 compounds and evaluated on a test set containing 171 compounds in the National Center for Toxicological Research liver cancer database (NCTRlcdb). As a result, the proposed DeepCarc model yielded a Matthews correlation coefficient (MCC) of 0.432 for the test set, outperforming four advanced deep learning (DL) powered quantitative structure-activity relationship (QSAR) models with an average improvement rate of 37%. Furthermore, the DeepCarc model was also employed to screen the carcinogenicity potential of the compounds from both DrugBank and Tox21. Altogether, the proposed DeepCarc model could serve as an early detection tool (https://github.com/TingLi2016/DeepCarc) for carcinogenicity assessment.

Keywords: NCTRlcdb; QSAR; carcinogenicity; deep learning; non-animal models.

FIGURE 1

Overall workflow for the DeepCarc model including: (1) Data preparation. 863 compounds were split into training (554 compounds), development (138 compounds), and test (171 compounds) sets based on the Kennard-stone algorithm. (2) Base classifiers development. Five algorithms were used to develop the base classifiers from three different chemical representations, including Mol2vec, Mold2, and MACCS. Two base classifiers selection strategies were employed to select the optimized classifiers for meta classifier development. (3) Meta classifier development. With three chemical representations and two selection methods, six groups of base classifiers, including Mol2vec_supervised, Mol2vec_original, Mold2_supervised, were used Mold2_original, MACCS_supervised, and MACCS_original. The probability prediction from selected base classifiers was used to train the neural network. (4) Model evaluation. The DeepCarc model was evaluated on the independent test set.

FIGURE 2

The distribution of the pairwise Tanimoto coefficients calculated from Mol2vec, Mold2, and MACCS: The pink and green indicate that the pairwise Tanimoto coefficients were calculated from the carcinogenic molecules and noncarcinogenic molecules, respectively.

FIGURE 3

The performance of the developed DeepCarc models based on the proposed supervised base classifier selection strategy with the three chemical representations: the three chemical representations included Mol2vec, Mold2, and MACCS. (A): Seven performance metrics; (B): Area under the ROC curve.

FIGURE 4

Ensemble models performance on the test set.

FIGURE 5

The probability distribution of the DeepCarc prediction of the compounds from (A) DrugBank; (B) Tox21.