Deep learning the structural determinants of protein biochemical properties by comparing structural ensembles with...

Area of Science:

• Computational biology
• Protein structure analysis
• Biophysics

Background:

• Understanding protein structure-function relationships is crucial in biology and medicine.
• Computer simulations aid in studying protein variants, but require effective dimensionality reduction.
• Existing algorithms may overemphasize large structural changes, potentially missing subtle but important features.

Purpose of the Study:

• To develop a novel dimensionality reduction method, DiffNets, for analyzing protein structural ensembles.
• To automatically identify relevant structural features that determine protein biochemical properties.
• To overcome limitations of current algorithms that rely on potentially misleading assumptions.

Main Methods:

• Implementation of DiffNets as self-supervised autoencoders.
• Training DiffNets to learn low-dimensional representations predictive of biochemical differences.
• Application to analyze structural ensembles of protein variants, including beta-lactamase and myosin.

Main Results:

• DiffNets successfully identified subtle structural signatures correlating with protein stability and function.
• Demonstrated ability to predict relative stabilities of beta-lactamase variants.
• Showcased prediction of duty ratios for myosin isoforms.

Conclusions:

• DiffNets offer a powerful, assumption-free approach to identify critical structural determinants of protein properties.
• The method enhances the analysis of protein structure-function relationships from simulation data.
• DiffNets show broad applicability to various biological perturbations, including ligand binding.