For early-stage drug discovery projects, having structural information on a target of interest can be crucial for project development. Our ability as a field to obtain this information has improved significantly, with advances in instrumentation techniques and technologies to obtain protein structural data, such as cryo-EM, to advances in modeling and predicting protein structure such as with AlphaFold. With more projects becoming structure-enabled, we felt that improvements could also be made to how we incorporated structure-based drug discovery techniques into our overall drug discovery pipeline. Traditional structure-based approaches to drug discovery usually rely on filtering a large virtual library of compounds until some best in silico binders are identified. There are a number of excellent, and massive, virtual libraries available. Unfortunately, these libraries only encompass a very small portion of chemical space, finding a hit from them requires that the ideal molecule already be present in the library, and other groups are also looking through these same databases. We have incorporated into our generative models structure-based evaluations which serve as optimization criteria. Our generative models are able to propose new molecules based on predicted criteria and have been successful at solving the multi-parameter optimization problem of late stage lead optimization. By incorporating structure-based criteria into the generative models, we are able to design new molecules that have optimal predicted binding for our target of interest while at the same time optimizing the parameters necessary for promoting a compound from hit to lead and beyond. In collaboration with Oncodesign, we have successfully targeted proto-oncogene serine/threonine-protein kinase Pim-1 using a structure-based generative modeling approach. New molecules have been proposed, synthesized, and tested and serve as a proof of concept of this approach. After synthesizing and testing only 12 molecules from 5 diverse scaffolds, 4 have had activity less than 5 μM and 2 have had activity less than 1 μM. All of the synthesized compounds have shown good preliminary ADME properties and all have freedom-to-operate for intellectual property purposes. These successes validate structure-guided design using generative modeling, and we expect these technologies to continue to impact drug discovery projects.

Structure-guided de novo drug design using deep generative modeling

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.