Session: COMP Poster Session:

Understanding fundamental details of natural enzymes and their chemical processes provide a prime example of efficient energy conversion that take place in the realms of alternative energy generation and storage, biofuels production, medicine, and artificial intelligence-driven discovery of high throughput molecular design. Microbial organisms and their key enzymes have evolved over millions of years to perform challenging chemical transformations with performances and efficiencies that are achieved only rarely by synthetic catalytic mimics. Computational elucidation of protein-ligand interaction can help in characterizing a wide variety of molecular and macromolecular properties from substrate binding, and therapeutic candidate design and discovery. In this context, we recently introduced the expanse of a novel, distance aware, graph based machine learning approach to analyze and predict protein-ligand interaction. The system’s graph structures are embedded with information directly from docked structures to retain and simulate a 3D chemical space and increase atomic level understanding. As a next step, we have started using deep generative models that have demonstrated their ability to explore the huge chemical space efficiently and to find out novel molecules with target activity and properties. The molecule generation while preserving critical scaffolds of the molecule is an efficient and most effective way to achieve this goal. We propose a novel graph based generative model that can take 3D structure coordinates of the molecules as an input and generates 3D coordinates of novel candidates as an output while always preserving the desired scaffolds in generated molecules. We show that our model generates predominantly valid, unique, novel molecules and have drug-like properties similar to the molecules in the training set. Moreover, our model performs well for relatively small volume of training data and generalize equally well for generating the molecules with any new scaffolds thus making our model applicable in different domains.

Andrew D. McNaughton

Computational Scientist, Pacific Nortwest National Laboratory

Appearances