Machine learning approaches to predict Acetylcholinesterase inhibition and model applications

Summary Organophosphorus (OP) compounds are one of the most common causes of poisoning worldwide. There are nearly 3 million poisonings per year resulting in three hundred thousand deaths. OPs bind to acetylcholinesterase (AChE), rendering the enzyme incapable of hydrolyzing acetylcholine (ACh) in the cholinergic synapses and neuromuscular junctions. Subsequent accumulation of ACh leads to overstimulation of the affected neurons acting through muscarinic and nicotinic receptors. Some of the adverse effects of pesticides on non-target organisms such as fish, amphibians and humans have also occurred as a result of biomagnifications of the toxic compounds. Since AChE is a very well-studied target, the likelihood of finding further new inhibitors amongst well-known drugs and molecules is likely unlikely and therefore we need to look further afield and do this in a more efficient manner. Computational approaches such as machine learning are an approach that can be used to learn from diverse literature to enable virtual screening. Following our preliminary work generating Bayesian models for human AChE we have performed a high throughput screen of 2431 compounds and identified 66 compounds that were “active” against eel AChE. We recently completed a phase I SBIR in which we expanded the use of these machine learning models to curate additional AChE data, built and validated AChE models with many algorithms in parallel and used them to identify AChE inhibitors from various species to predict OP induced poisoning and possible threats to the environment. In addition, we have recently curated further literature data and generated models for the closely related enzyme BChE using a contrastive learning algorithm approach to score a library of 5 million molecules from various vendors in order to select molecules for testing and have identified several novel selective inhibitors. In Phase II we will develop the MegaAChE software product further for prediction of AChE inhibition and incorporate uncertainly prediction features and new algorithms such as large language models. We will also use this software to design and develop AChE reactivators that could provide a new treatment for OP pesticide and nerve agent exposure. We now propose the following aims to further develop and validate the MegaAChE software as a commercial product: Aim 1: Modeling uncertainty in AChE models Aim 2: Large language models for AChE and BChE datasets Aim 3: Develop new reactivators to treat AChE poisoning using generative design Our ultimate goal will be to provide software and models to predict AChE inhibition which will be a commercial product. In addition we will develop new molecule intellectual property which we can then license to a larger company. Computational models for AChE therefore have beneficial dual-use potential both for identifying molecules that could have environmental or human toxicity, while alternatively such machine learning models could also help us identify molecules with therapeutic utility for Alzheimer’s disease or as treatments for poisoning.