Data driven flexible backbone protein design.

Area of Science:

• Computational structural biology
• Protein engineering
• Machine learning in bioinformatics

Background:

• Current protein design methods often rely on single structures, limiting their scope.
• Existing ensemble methods treat multiple structures independently.
• Abundant multi-structure data in the Protein Data Bank (PDB) remains underutilized for holistic design.

Purpose of the Study:

• To develop a flexible backbone strategy for protein design that learns global backbone movements from multiple protein structures.
• To enhance protein interface design by exploring diverse backbone conformations.
• To improve the identification of high-affinity protein variants for specific targets.

Main Methods:

• Introduced FlexiBaL-GP, a flexible backbone strategy utilizing Gaussian Process Latent Variable Models (GP-LVM).
• GP-LVM learns a low-dimensional representation of protein coordinates to capture backbone dynamics.
• Integrated learned backbone movements into a parallel tempered Markov Chain Monte Carlo (MCMC) framework for protein engineering.

Main Results:

• Demonstrated superior performance of FlexiBaL-GP in protein interface design using the human ubiquitin-USP21 complex as a model.
• Successfully identified tighter binding ubiquitin variants for USP21 compared to existing strategies.
• Validated the ability of the method to capture and utilize global protein backbone movements.

Conclusions:

• FlexiBaL-GP offers a significant advancement in computational protein design by leveraging multiple structures for backbone flexibility.
• The strategy effectively enhances protein interface design, particularly for identifying high-affinity binders.
• This machine learning-driven approach opens new avenues for designing proteins with tailored functions.