DrugAppy – An end-to-end deep learning framework for computational drug discovery

Identification of druggable oncogenic vulnerabilities and the design of novel chemical entities against them is crucial in cancer research due to the limited curative options for some advanced cancers. However, the drug design process is costly and time-consuming. As a result, the use of computational tools to accelerate and optimize this process is a promising approach. We present DrugAppy, a computational tool for the identification of inhibitors, built on a hybrid model that combines Artificial Intelligence (AI) algorithms and computational and medicinal chemistry methodologies using an imbrication of models such as SMINA and GNINA for High Throughput Virtual Screening (HTVS) and GROMACS for Molecular Dynamics (MD). Additionally, the prediction of key parameters such as drug pharmacokinetics, selectivity, and potential activity was conducted using both publicly available models and proprietary artificial intelligence models trained on public datasets. We validated DrugAppy through two case studies targeting Poly(ADP-ribose) polymerase (PARP) and the transcriptional enhanced associate domain (TEAD) family of proteins. Using the methodology outlined, several molecules have been identified that either match or surpass the in vitro activity of current inhibitors. For PARP1, two molecules were found with activity comparable to olaparib. For TEAD4, a compound was identified that outperforms the activity of IK-930, the reference inhibitor for this target. In this work, we demonstrate how the workflow can be effectively used to discover novel molecular structures, using the protein families PARP and TEAD as case studies. For each target, one active compound has been identified and confirmed for target engagement that matches the reference inhibitor.