Swapnil Chavan got his masters in Pharmaco-informatics from NIPER, India, and received his PhD in Computational Chemistry from Linnaeus University, Sweden. Then he went to NCCT (National Center for Computational Toxicology) at US – Environmental Protection Agency for his postdoctoral research where he served as a Unilever postdoc. Swapnil currently works as a senior data scientist in computational toxicology at RISE AB, Sweden.
INTERESTS: He leads innovative research towards developing novel chemical risk-assessment methods using cheminformatics, bioinformatics, image-informatics and machine learning techniques. Currently, he is focusing on integrating explainable AI and uncertainty estimation approaches towards building a comprehensive toxicity prediction tool.
Placental Barrier Penetration Prediction Using Multi-Task Learning to screen for Potential Developmental Toxicants
Swapnil Chavan a*, Mark Cronin b
a Department of Chemical and Pharmaceutical Toxicology, Research Institutes of Sweden RISE, Södertalje 151 36, Sweden.
b School of Pharmacy and Biomolecular Sciences, Liverpool John Moores University, Byrom Street, Liverpool L3 3AF, England.
An adverse outcome pathway (AOP) is a conceptual framework for describing the scientifically credible basis for relating a chemically induced molecular initiating event (MIE) to an adverse outcome (AO) via a set of key events (KEs). The MIE is the first event in an AOP, that is the initial interaction of a chemical with the biological target. Therefore, this is a very important event in any given AOP. To quantify an AOP, a quantitative understanding between two events can be a valuable predictor to ascertain whether quantitative estimate of first event predict possibility of happening of subsequent event. For example, in the case of developmental toxicity, predicting whether a drug permeates the placental barrier and, if yes, is the permeated drug concentration sufficient to be toxic to the foetus? Therefore, building a model that can predict whether a drug can permeate the placental barrier and up to what extent, will be a valuable tool towards developing a qAOP model for developmental toxicity. The goal of this study was to build a model for placental barrier permeation that will predict not only whether a drug passes through the placental barrier but also provide the fraction that will reach the foetus. To achieve this goal, we have utilized cutting-edge deep learning approach of multi-task learning to build an advanced in silico placental barrier penetration model using 2-dimensional RdKit descriptors. For the classification task, our model showed twice-repeated-10-fold cross validation AUC ROC score equal to 0.95, 0.83 and 0.95 for a training, validation and test set, respectively. For the regression task, this model showed twice-repeated-10-fold cross validation R-square of 0.65, 0.46 and 0.21 for the training, validation, and test set, respectively. We believe the multi-task placental barrier penetration model described here could be a useful tool towards construction of a qAOP model for the developmental toxicity.
