Allosteric Hit Discovery for Phosphatases with AtomNet® Virtual Screens

At the American Chemical Society Fall 2022 National Meeting & Expo, Atomwise members were selected to present their research and work. Learn what our Atoms have been working on below and visit Atomwise at ACS Fall 2022 National Meeting for other presentation sessions. 

Saulo de Oliveira, PhD 
Atomwise

Atomwise Co-Authors:  Paweł Gniewek, Victor Kenyon, Christian Laggner, Teresa Palazzo, Srimukh Prasad Veccham, Bradley Worley, Kate A. Stafford, Brandon M. Anderson, Michael Mysinger, Henry van den Bedem

Title: Allosteric hit discovery for phosphatases with AtomNet® virtual screens

Division: Computers in Chemistry

Abstract

Phosphorylation, mediated by kinases and phosphatases, is an important protein post-translational modification that governs cellular activity.  Disruption to the phosphorylation/dephosphorylation equilibrium is often implicated in disease. While kinases comprise one of the most important classes of drug targets, with more than 70 inhibitors currently available in the market, to date there are no marketed drugs that target their counterparts, phosphatases. Targeting phosphatases is challenging because their orthosteric binding site is very conserved across the phosphatase families and it is highly charged, which often leads to binders with poor selectivity and pharmacology. One way to circumvent these hurdles is the pursuit of allosteric binders. However, despite highly successful deep learning frameworks adopted in pharmaceutical virtual screening pipelines, these approaches have struggled to identify allosteric hits, primarily owing to poorly annotated training data. Lower data availability for allosteric binders limits the effectiveness of ligand-based approaches, and correct modeling of physical interactions is critical for successful allosteric binding prediction. Here, we developed a multi-task graph neural network architecture that enforces that physical interactions are learned correctly from the training data. To complement this physical awareness, we also created an automated annotation pipeline that provides better site-specific annotations. Our multi-task models trained with site-specific data achieve higher hit rates in our benchmarks for non-orthosteric sites compared to our baseline model. To validate our approach, we carried out a campaign to identify novel allosteric binders for several clinically-relevant phosphatases. Highest-scoring compounds from our models were selected for experimental testing and validation, and show moderate allosteric binding and are more drug-like than known orthosteric binders. These findings show that better data annotation and improvements in model architecture directly translate into higher success rates in virtual screening, and showcase the strength of using deep learning in the characterization of novel hits.

Download the presentation deck. 

Join our team