Abstract

Of the protein structures deposited in the Protein Data Bank, less than half have pockets suitable for the binding of drugs. Even when proteins contain pockets in their ground state structures (e.g., the nucleotide-binding active site in myosin motors), achieving specificity remains a central challenge in drug design as many protein families share common structural motifs. Cryptic pockets are cavities absent in ligand-free experimental structures that form due to protein fluctuations in solution. They provide a means to specifically target proteins currently considered undruggable. While cryptic pockets are alluring drug targets, it remains difficult to predict which proteins will form cryptic pockets. It is also unclear how certain compounds that bind at cryptic pockets discriminate between similar targets, even though those targets all have closed pockets in experimental structures. To address these problems, I develop a graph neural network called PocketMiner that predicts whether a protein is likely to form a cryptic pocket based on its ground state structure. I demonstrate that PocketMiner achieves improved performance (ROC-AUC: 0.87) compared to existing methods at >1,000-fold faster run times. To further accelerate cryptic pocket discovery, I leverage the protein structure prediction algorithm AlphaFold to generate ensembles of structures. I show that AlphaFold-generated ensembles often sample cryptic pocket opening, and that using these ensembles as starting structures for molecular dynamics simulations can enhance sampling of a rare cryptic pocket opening in an antimalarial drug target. To connect cryptic pocket opening to drug specificity, I show that differences in the probability of cryptic pocket opening underpin the specificity of a myosin inhibitor known to bind at a cryptic site. By combining Markov state models with molecular docking, we accurately predict the affinity of blebbistatin for different myosin proteins. Finally, to demonstrate the utility of these methods for drug discovery applications, I use simulations of a cancer drug target, PPM1D phosphatase, to discover a novel cryptic pocket. Docked poses of compounds bound to this cryptic pocket can be fed to a neural network that predicts affinities to accurately rank compounds by their experimental affinities. Taken together, these results represent an important advancement towards rational drug design against previously undruggable targets.

Committee Chair

Gregory Bowman

Committee Members

Michael Greenberg; Alex Holehouse; Gautam Dantas; Gregory Bowman; Leslie Leinwand; Michael Greenberg; Nathan Stitziel

Degree

Doctor of Philosophy (PhD)

Author's Department

Biology & Biomedical Sciences (Computational & Systems Biology)

Author's School

Graduate School of Arts and Sciences

Document Type

Dissertation

Date of Award

5-8-2025

Language

English (en)

Author's ORCID

https://orcid.org/0000-0002-5504-2684

Download

Included in

Biology Commons

Share

COinS