Synthetic benchmarks to evaluate attribution for structure-based graph convolutional networks in drug discovery

February 24, 2022
Events, ACS2022

At the American Chemical Society Spring 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 Spring 2022 National Meeting for other presentation sessions. 

andreana-rosnik-256x256Andreana Rosnik, PhD 

Atomwise Co-Authors: Kate A. Stafford, Brandon M. Anderson, and Henry van den Bedem

Title:Synthetic benchmarks to evaluate attribution for structure-based graph convolutional
networks in drug discovery

Division: Computers in Chemistry



In drug discovery, deep learning (DL) methods can quickly identify potent and active compounds from vast on-demand chemical libraries. Understanding how structure-based DL models attribute features of molecular recognition to activity can cast light on sensitivity to receptor-ligand interactions, biases towards particular functional groups or scaffolds, target selectivity, and, importantly, help medicinal chemists advance compounds in lead optimization. However, despite a rapidly increasing use and need for deep learning in drug discovery, structure-based model interpretability remains a formidable challenge, in part owing to a dearth of reliable, ‘ground-truth’ benchmark sets. 

Here, we introduce a set of synthetic chemical benchmarks consisting of computable features such as intermolecular hydrogen bonds in docked protein-ligand complexes. We trained AtomNet® structure-based graph convolutional networks for classification and regression on the synthetic benchmarks. We visualize edge weights for models with and without attention on representative molecules. An attention-based information entropy attribution measure correctly identifies hydrogen-bonding heavy atoms. Our models very accurately predict the number of canonical hydrogen-bonds in a protein-ligand complex. In addition, we explore other chemical benchmarks of physically relevant protein-ligand interactions, and we demonstrate applications to protein selectivity prediction in targets such as kinases and G-protein-coupled receptors. The benchmarks will help evaluate and improve our models, and guide structural optimization in drug discovery.


View Presentation 


Join our team

Our team is comprised of over 80 PhD scientists who contribute to a high-performance academic-like culture that fosters robust scientific and technical excellence. We strongly believe that data wins over opinions, and aim for as little dogma as possible in our decision making. Learn more about our team and opportunities at Atomwise.