Development of an in silico consensus model for the prediction of the phospholipigenic potential of small molecules

Phospholipidosis (PL) describes the accumulation of phospholipids in lysosomes of cells of various tissues after prolonged exposure with drug like compounds. These cellular findings can result in a delay of drug development, cause increased costs in affected projects and potentially may halt a drug development program. The early detection of compounds which potentially cause phospholipidosis therefore is desirable for risk mitigation. Here we describe an in silico consensus model for the detection of phospholipigenic potential of small molecules. The model was trained on in house in vitro data yielding an accuracy of 94%. By employing model agnostic explainability methods, we could show that the model learns reasonable molecular properties. The consensus model showed good performance on underrepresented PL-active compounds in clusters of similar molecules of the test dataset and on external in vitro and in vivo validation data of highly structural dissimilarity to the training data. Using the external in vitro data, an applicability domain of the model was deduced.