Inductive Transfer of Knowledge: Development of Predictive Multitask Learning Models to Find New Hit Compounds Against Trypanosomatids

Despite the outstanding capability of artificial intelligence in drug discovery for Neglected Tropical Diseases, learning from low data regimes has often resulted in low predictive performance due to insufficient structure-activity relationship examples. Here, we developed a deep learning strategy based on the inductive transfer of knowledge (multitask learning, MTL) to simultaneously predict pIC50 values of untested compounds against six parasites of the genus Trypanosoma spp. and Leishmania spp. Compared with other conventional single-task learning (STL) models focused only on a single property, our MTL models for regression and classification developed using Feedforward Neural Networks (FNNs) and graph-based Message Passing Neural Networks (MPNNs) significantly outperform past methods. Using these models, we prioritized 57 putative hits from the ChemBridge database for further in vitro testing against Trypanosomatids. So far, 17 compounds showed potent anti-T.b. brucei and anti-L. infantum activities at low nanomolar to micromolar concentrations (IC50 = 0.005–14.6 µM).