Beyond Classical Force Fields: Physics-Driven Assessment of the Grappa Machine-Learned Force Field on the FoldBind...

Area of Science:

• Computational Biology and Biophysics
• Artificial Intelligence in Molecular Modeling

Background:

• Physics-based simulations require accurate force fields and efficient sampling for mechanistic insights into biomolecular systems.
• AI is advancing molecular modeling by developing machine-learned force fields, moving beyond traditional empirical models.

Purpose of the Study:

• To introduce the FoldBind benchmark set for validating new AI-driven force fields and sampling strategies.
• To assess the accuracy of force fields and sampling methods on challenging protein-folding and binding-induced folding cases.

Main Methods:

• Developed the FoldBind benchmark, comprising 18 systems (14 protein folding, 4 peptide-protein complexes).
• Employed the Modeling Employing Limited Data (MELD) framework for accelerated conformational exploration and sampling.
• Integrated ambiguous physical restraints within a Bayesian inference formalism to balance exploration and exploitation.

Main Results:

• The FoldBind benchmark effectively probes conformational transitions and binding-induced folding.
• Force field quality critically determines the stabilization of native-like states under identical data conditions.
• Consistent stabilization of the native basin across multiple states indicates a force field's physical realism.

Conclusions:

• The FoldBind benchmark provides a rigorous test for evaluating biomolecular simulation methods and force fields.
• This benchmark, combined with MELD, can distinguish and guide future force field development efforts.
• Validating AI-driven models on diverse systems is crucial for advancing computational biophysics.