Drug clustering upon toxicity effects on liver spheroids – a deep learning approach

Abstract

Motivation for this research stems from drug-induced liver injury (DILI) being a leading cause of drug failure and market withdrawal, which calls for reliable, non-invasive hepatotoxicity assessment. The basis for this research was the work performed by Dubinsky et al., who established a deep learning (DL) model to predict ATP values (non-invasive estimation of cell viability), assess damage to liver spheroids, and predict hepatotoxicity based on spheroid images. They created a public database containing approximately 20,000 brightfield microscopy images of human liver spheroids exposed over a period of up to 7 days to 108 drug compounds for assessing cell viability. Our work focused on clustering applied compounds by the intensity of damage caused to liver spheroids (level of predicted ATP values), forming groups of compounds that have similar effects on liver cells and potentially similar mechanisms of toxicity. Then we explored whether selected drugs belong to the same class of pharmacotherapeutic classification – analgesics, antipyretics, antibiotics, etc. The model can detect if some drugs behave differently than expected, leading to detecting errors and revealing potential new phenomenological causes. Further, we analysed the temporal dependence of drugs on cell viability (toxicity), providing insights into the dynamics of effects and toxicity progression over time. Future work will investigate drug effects in combination, since co-exposure of drugs can lead to different outcomes of hepatotoxicity, in comparison to individual exposure.