Date of Award

12-22-2021

Document Type

Dissertation

Degree Name

Doctor of Philosophy (PhD)

Department

Information Science

First Advisor

Shraddha Thakkar

Abstract

Drug-induced liver injury (DILI) is an undesirable side effect to the liver caused by the common use of drugs, which is of great concern to the patients, physicians, the pharmaceutical industry, and government regulators. The current animal testing models do not present reliable results on the human DILI risk prediction during the preclinical phase. Therefore, there is an urgent need to find alternative methods to improve the human DILI risk prediction during the early stage of development. Utilizing the accumulated high-throughput transcriptomics profiles from L1000, we constructed an 8-layer deep neural network for human DILI prediction based on the largest binary DILI classification dataset, DILIst. The developed model outperformed the other three machine learning classifiers and was specifically effectively predicted the DILI potential of agents in the anatomical therapeutic chemical (ATC) class of antineoplastic and immunomodulating with an AUC of 0.943. To screen the DILI risk with the chemical structure only, we proposed the DeepDILI model, which integrates the model-level representation from five groups of conventional ML classifiers into a deep learning framework. The DeepDILI model showed superior results than other models developed from the molecular descriptors directly and was especially effective in the ATC class of alimentary tract and metabolism with an MCC of 0.581. To evaluate the transferability of the DeepDILI’s architecture, we constructed the DeepCarc model for carcinogenic potency prediction. The DeepCarc model made better performance (such as MCC = 0.432) than the conventional ML models and four deep neural network models developed from the molecular descriptors directly, which indicated the architecture of the DeepDILI model could be transferred to other safety endpoints, such as carcinogenicity. In addition, DeepCarc was used to rank the carcinogenic potency of the compounds from DrugBank and Tox21. Our project proposed two deep learning models, which could serve as promising tools for DILI risk evaluation in the early stage of drug development with either transcriptomic profiles or chemical structures. In addition, we also validate the transferability of the DeepDILI architecture to develop the DeepCarc model for the carcinogenic potency evaluation.

Download

Share

COinS