Predicting DILI (drug-induced liver injury) is challenging compared to other organ-specific toxicities. Translation from animals to humans is poor and, mechanistically, DILI can be complex. As a consequence, DILI continues to be one of the leading causes of attrition during drug development. Better human relevant models are required to improve early stage DILI prediction. Cyprotex is committed to researching and developing approaches to improve the prediction of DILI using human cell-based models in combination with novel techniques such as toxicogenomics and artificial intelligence (AI).
At the Society of Toxicology (SOT) conference on March 10-14, 2024, Cyprotex presented a poster titled, ‘An AI Approach to Drug-Induced Liver Injury Risk: Prediction of Safe Maximum Doses from Toxicogenomics Profiles’. The research evaluated 128 compounds from the FDA Liver Toxicity Knowledge Base – 68 of these compounds were associated with DILI and 60 of these compounds were not associated with DILI. Transcriptomics profiles were generated after dosing primary human hepatocytes in triplicate at 8 concentrations over 24 hr.
Machine learning is a subset of artificial intelligence which is used to find patterns, make decisions and optimise outcomes. In this study, the high throughput transcriptomics profiles of a set of known DILI-positive and DILI-negative compounds were used to train a supervised machine learning model to predict a safe maximum Cmax for novel compounds. When interpreting the results, a compound was predicted as DILI-positive if the true Cmax was above the predicted safe Cmax, and a compound was predicted as DILI-negative if the true Cmax was below the predicted safe Cmax. The model achieved the following metrics on the test set (assuming 40x Cmax level and 90% DILI score threshold):
The poster provides a detailed insight into two DILI-positive (TAK-875 and bosentan) and two DILI-negative compounds (dopamine and caffeine) to demonstrate the power of the transcriptomics and AI in predicting DILI as well as identifying specific mechanisms of toxicity. The model was able to capture the importance of cholestasis-associated genes in DILI.
To learn more about the use of transcriptomics and AI in DILI prediction: