Background & Aims: The United States Food and Drug Administration (FDA) regulates a broad range of consumer products, which account for about 25% of the United States market. The FDA regulatory activities often involve producing and reading of a large number of documents, which is time consuming and labor intensive. To support regulatory science at FDA, we evaluated artificial intelligence (AI)-based natural language processing (NLP) of regulatory documents for text classification and compared deep learning-based models with a conventional keywords-based model.Methods: FDA drug labeling documents were used as a representative regulatory data source to classify drug-induced liver injury (DILI) risk by employing the state-of-the-art language model BERT. The resulting NLP-DILI classification model was statistically validated with both internal and external validation procedures and applied to the labeling data from the European Medicines Agency (EMA) for cross-agency application.Results: The NLP-DILI model developed using FDA labeling documents and evaluated by cross-validations in this study showed remarkable performance in DILI classification with a recall of 1 and a precision of 0.78. When cross-agency data were used to validate the model, the performance remained comparable, demonstrating that the model was portable across agencies. Results also suggested that the model was able to capture the semantic meanings of sentences in drug labeling.Conclusion: Deep learning-based NLP models performed well in DILI classification of drug labeling documents and learned the meanings of complex text in drug labeling. This proof-of-concept work demonstrated that using AI technologies to assist regulatory activities is a promising approach to modernize and advance regulatory science.

Drug-induced liver injury (DILI) is a major cause of drug development failure and drug withdrawal from the market after approval. The identification of human risk factors associated with susceptibility to DILI is of paramount importance. Increasing evidence suggests that genetic variants may lead to inter-individual differences in drug response; however, individual single-nucleotide polymorphisms (SNPs) usually have limited power to predict human phenotypes such as DILI. In this study, we aim to identify appropriate statistical methods to investigate gene–gene and/or gene–environment interactions that impact DILI susceptibility. Three machine learning approaches, including Multivariate Adaptive Regression Splines (MARS), Multifactor Dimensionality Reduction (MDR), and logistic regression, were used. The simulation study suggested that all three methods were robust and could identify the known SNP–SNP interaction when up to 4% of genotypes were randomly permutated. When applied to a real-life DILI chronicity dataset, both MARS and MDR, but not logistic regression, identified combined genetic variants having better associations with DILI chronicity in comparison to the use of individual SNPs. Furthermore, a simple decision tree model using the SNPs identified by MARS and MDR was developed to predict DILI chronicity, with fair performance. Our study suggests that machine learning approaches may help identify gene–gene interactions as potential risk factors for better assessing complicated diseases such as DILI chronicity.

Drug-induced liver injury (DILI) presentation varies biochemically and histologically. Certain drugs present quite consistent injury patterns, i.e., DILI signatures. In contrast, others are manifested as broader types of liver injury. The variety of DILI presentations by a single drug suggests that both drugs and host factors may contribute to the phenotype. However, factors determining the DILI types have not been yet elucidated. Identifying such factors may help to accurately predict the injury types based on drugs and host information and assist the clinical diagnosis of DILI. Using prospective DILI registry datasets, we sought to explore and validate the associations of biochemical injury types at the time of DILI recognition with comprehensive information on drug properties and host factors. Random forest models identified a set of drug properties and host factors that differentiate hepatocellular from cholestatic damage with reasonable accuracy (69-84%). A simplified logistic regression model developed for practical use, consisting of patient's age, drug's lipoaffinity, and hybridization ratio, achieved a fair prediction (68%-74%), but suggested potential clinical usability, computing the likelihood of liver injury type based on two properties of drugs taken by a patient and patient's age. In summary, considering both drug and host factors in evaluating DILI risk and phenotypes open an avenue for future DILI research and aid in the refinement of causality assessment.