Accurate and interpretable in silico clinical risk assessment for drug-induced liver injury (DILI) from molecular structure

Katherine Titterton, Daniil Boiko, Ángel Alexander Cabrera, Alex Bäuerle, Akshit Tyagi, Shahin Saneinejad, Alex Beatson, Brandon White

Drug-induced liver injury (DILI) is a leading cause of clinical trial failure and drug withdrawal, but current methods fail to provide accurate and interpretable clinical risk at human-relevant exposure. We profiled >100,000 unique molecules in primary human hepatocytes (PHHs) using a multiplexed cytotoxicity assay with high content imaging & two biochemical assays (LDH & Realtime Glo/MT-Glo). We train in silico models to predict dose-dependent responses for >10 distinct cell stress and death features from the images solely from 2D molecular structure (SMILES string). Our in silico features & thousands of clinical data points are used to train an in silico clinical risk assessment model that outperforms industry-leading in vitro assays in accuracy & interpretability.