Protein Ensemble Generation Through Variational Autoencoder Latent Space Sampling.

Area of Science:

• Computational Biology
• Structural Biology
• Drug Discovery

Background:

• Mapping protein conformational ensembles is vital for understanding function and developing targeted therapeutics.
• High-dimensional protein structural data presents a significant challenge for computational modeling and ensemble generation.

Purpose of the Study:

• To develop and validate a novel approach for generating protein structural ensembles using variational autoencoders (VAEs).
• To assess the capability of the VAE-based method in accurately sampling protein conformations relevant to drug discovery, specifically for the K-Ras protein.

Main Methods:

• Utilized variational autoencoders (VAEs) to reduce the dimensionality of protein structural data, creating a continuous, low-dimensional representation.
• Employed a structure quality metric to guide a search within the VAE's latent space.
• Generated 3D protein structures using RoseTTAFold, informed by the sampled information from the latent space.
• Trained the VAE on K-Ras crystal structures and molecular dynamics (MD) simulation snapshots, evaluating performance against withheld crystal structures.

Main Results:

• The VAE-based latent space sampling rapidly generated protein ensembles with high structural quality.
• The method achieved sampling within 1 Å of held-out K-Ras crystal structures, demonstrating superior consistency compared to MD simulations and AlphaFold2.
• The generated ensembles successfully recapitulated cryptic pockets in K-Ras structures, enabling effective small molecule docking.

Conclusions:

• The developed VAE-driven approach offers an efficient and accurate method for generating protein conformational ensembles.
• This technique significantly advances the ability to model protein structures for drug target identification and the design of small molecule inhibitors.
• The improved sampling consistency and pocket recapitulation hold promise for accelerating drug discovery pipelines.