Target-specific machine-learning scoring functions for reliable structure-based virtual screening

点击翻译按钮获取中文摘要

Drug leads are typically small-sized chemical compounds that tightly bind to a disease-causing protein. As the right key fits into a lock, a drug lead molecule acts by binding to the right pocket of its therapeutic target protein to potently alter its function in a way that positively impacts the associated disease. Academic research excels at discovering promising therapeutic targets. However, discovering drug leads and optimising their potency for a target is an expensive, time-consuming and particularly challenging process, which has traditionally been carried out by pharmaceutical companies with vast resources. This constitutes a barrier to translating innovative biomedical research from academia into new drug candidates, as an optimised lead is required to attract funding for further preclinical and clinical studies via out-licensing or industry partnership. Tools are therefore needed to help academics to bridge this translation gap by reducing the experimental efforts required, or even making it possible, to achieve optimised drug leads for a given target.Docking is a computational technique providing relatively fast predictions of whether and how a molecule binds to an atomic-resolution structure of the target. Very recently, the exploitation of a novel technology generating billions of make-on-demand molecules by classical docking tools have directly achieved a range of diverse and potent drug leads for several targets. Therefore, no lengthy and costly optimisation was subsequently required, thereby strongly reducing the time and cost to provide these optimised drug leads. However, the modest predictive performance of these classical tools is extensively documented. This means that their application to many other targets is likely to be much worse than that in the few targets reported so far. It is now well-known that a way to boost docking performance in other targets is by enhancing it with Artificial Intelligence (AI). Unlike the classical tools, AI models can exploit fast-growing datasets to learn to discriminate between molecules with or without potent activity for the target. AI models are furthermore likely to make drug design even faster and less expensive than the classical tools on those targets where the latter work well.This methodology research project aims at improving target structure-based drug design via the innovative application of AI techniques. We will investigate optimal ways to build target-specific AI models. For the first time, these AI models will be generated in a way that not only predicts how strongly the molecule binds to the target, but also how reliable that prediction is. We will compare the predictive accuracy of these models to that of existing models for any target, whether classical or AI-based, using the most rigorous retrospective assessment practices. We will also investigate to which extent coupling these models with ultrafast, yet less accurate, models can directly provide optimised drug leads in a fraction of the time.The project focuses on those targets for which at least some binding molecules are available. Practically all currently investigated targets have never been analysed with structure-based target-specific AI models. Here we will develop and apply AI models tailored to two of such targets, which are in addition not related to those for which ultra-large library screening has been reported so far. The discovery of the first of these targets, TRPM8, has just been partly awarded the 2021 Nobel Prize in Physiology or Medicine. Drugs targeting TRPM8 should be able to mitigate cold hypersensitivity and pain. The second target, ATM, could be a way to provide new therapeutic options for those suffering from Huntington's disease and various types of cancers including brain tumours.We will provide all the data, codes and documentation to facilitate reproducibility and future research for further improvements on these and related targets.