Covalent: Interpretable and Discriminative Collective Variables Reveal Ligand-Dependent Switching in Human Cellular...

Area of Science:

• Biomolecular simulations
• Machine learning for structural biology
• Computational biophysics

Background:

• Identifying collective variables (CVs) is crucial for enhanced sampling and mechanistic analysis in biomolecular systems.
• Existing methods often struggle to find CVs that are both discriminative and interpretable.
• This challenge limits the mechanistic insights gained from molecular dynamics (MD) simulations.

Purpose of the Study:

• To present Covalent, a novel supervised machine learning pipeline for discovering interpretable collective variables (CVs).
• To apply Covalent to human cellular retinol-binding protein II (CRBP2) to understand its conformational dynamics and ligand interactions.
• To demonstrate Covalent's ability to generate physically transparent CVs for mechanistic interpretation.

Main Methods:

• Covalent employs a supervised machine learning approach combining a filter-wrapper-substitution feature funnel.
• It utilizes a Riemannian-optimized variant of harmonic linear discriminant analysis (GDHLDA) and post hoc subspace rotation.
• The pipeline was tested on MD trajectories of CRBP2 in apo, retinol-bound, and 2-lauroylglycerol (2-LaG)-bound states.

Main Results:

• Covalent identified linear CVs with clear mechanistic interpretations, outperforming principal component analysis in class separability.
• The learned CVs highlighted gating switches, ligand-induced stabilization of Ser76, and rearrangements around the CRBP2 cavity.
• Analysis revealed ligand-specific interaction switching and a decrease in internal void volume, indicating tighter packing upon ligand binding.

Conclusions:

• Covalent provides a practical method for discovering physically transparent CVs from MD data.
• The identified CVs offer valuable mechanistic insights into protein dynamics and ligand binding.
• The approach successfully demonstrates that holo states of CRBP2 are not pre-organized but require ligand-induced stabilization.