Accelerating Cryptic Pocket Discovery Using AlphaFold

Abstract

Cryptic pockets, or pockets absent in ligand-free, experimentally determined structures, hold great potential as drug targets. However, cryptic pocket openings are often beyond the reach of conventional biomolecular simulations because certain cryptic pocket openings involve slow motions. Here, we investigate whether AlphaFold can be used to accelerate cryptic pocket discovery either by generating structures with open pockets directly or generating structures with partially open pockets that can be used as starting points for simulations. We use AlphaFold to generate ensembles for 10 known cryptic pocket examples, including five that were deposited after AlphaFold's training data were extracted from the PDB. We find that in 6 out of 10 cases AlphaFold samples the open state. For plasmepsin II, an aspartic protease from the causative agent of malaria, AlphaFold only captures a partial pocket opening. As a result, we ran simulations from an ensemble of AlphaFold-generated structures and show that this strategy samples cryptic pocket opening, even though an equivalent amount of simulations launched from a ligand-free experimental structure fails to do so. Markov state models (MSMs) constructed from the AlphaFold-seeded simulations quickly yield a free energy landscape of cryptic pocket opening that is in good agreement with the same landscape generated with well-tempered metadynamics. Taken together, our results demonstrate that AlphaFold has a useful role to play in cryptic pocket discovery but that many cryptic pockets may remain difficult to sample using AlphaFold alone.

Discovery of a cryptic pocket in the AI-predicted structure of PPM1D phosphatase explains the binding site and potency of its allosteric inhibitors

“…Using a highly flexible system, we can sample conformations and identify cryptic pockets that can be successfully used in downstream virtual screening applications. While our work was based off a single AF structure as a starting point, we are aware of efforts to use these DL protein structure prediction tools to sample multiple conformations, thus better capturing protein flexibility ( Saldanõ et al, 2022 ; Meller et al, 2023a ). To our knowledge, these methods have not been compared against MSM approaches and more research would be needed before conducting a similar analysis as described herein with a DL-generated structural ensemble.…”

AlphaFold2 and Deep Learning for Elucidating Enzyme Conformational Flexibility and Its Application for Design

“…This strategy was shown to outperform previous traditional and DL docking protocols . Meller and co-workers also developed an AF2-based strategy to find cryptic pockets . DL has also been applied for finding potential location sites of transition metals in proteins (Metal1D and Metal3D) .…”

Adaptive Sampling Methods for Molecular Dynamics in the Era of Machine Learning

“…In particular, we focus on deep-learning (DL)-based techniques. DL has revolutionized many scientific fields, but its impact in the molecular biosciences was catapulted by the high accuracy of recent protein structure prediction models. − While not directly related to adaptive sampling, structure prediction models are useful to generate initial seeds for adaptive sampling algorithms and are discussed in that context. Researchers have been concurrently working on ML models to analyze − and accelerate , MD simulations.…”

“…For brevity, we will restrict ourselves to two previous studies that applied perturbation-based methods on structure prediction models, but other recent works exist . The first study was used to gather initial structures for parallel simulations of Plasmodium falciparum plasmepsin II (PM II) with the intent to sample cryptic binding pockets . The other method was used to obtain diverse structures of the Shwachman–Bodian–Diamond syndrome protein (SBDS) and the monocarboxylate transporter 1 (MCT1) .…”

“…The result of applying this technique to PM II was the prediction of a structural ensemble that partially sampled a known cryptic pocket in the protein. By launching parallel MD simulations from these structures and building a MSM from the trajectories, it was possible to recover the free energy landscape of pocket opening . These simulations did not require any type of biasing force to find the same free energy basins that could be detected with biased methods .…”