Integrating computational protein flexibility measures into generative protein design tools

Determine effective strategies to integrate computational protein flexibility quantification methods—including Molecular Dynamics-derived root mean square fluctuations, Elastic Network Models such as Gaussian Network Models and Anisotropic Network Models, and AlphaFold/ESMFold pLDDT-based flexibility proxies—into state-of-the-art generative protein design and inverse folding models such as ProteinMPNN, KWDesign, and PiFold, so that sequence generation can account for and control residue-level flexibility profiles relevant to function.

Background

The paper surveys experimental and computational approaches for quantifying protein flexibility, highlighting Molecular Dynamics simulations (RMSF), Elastic Network Models (GNM/ANM), and machine-learning-based structure predictors (AlphaFold2/ESMFold) that provide confidence scores correlated with flexibility. Although these tools estimate residue-level flexibility, their outputs are not routinely or systematically used to guide generative protein design.

The authors emphasize that modern inverse folding and de novo design models (e.g., ProteinMPNN, KWDesign, PiFold) typically ignore flexibility, limiting practical applicability. They propose Flexpert-Seq and Flexpert-3D as fast predictors and introduce Flexpert-Design to steer inverse folding toward desired flexibility, but they note that, in general, how to incorporate flexibility measurements into generative pipelines remains unresolved—motivating methodologies that directly couple flexibility signals to model inputs, objectives, and constraints.