Force-Guided Bridge Matching for Full-Atom Time-Coarsened Dynamics of Peptides

TL;DR

FBM introduces a force-guided bridge matching framework to learn full-atom time-coarsened peptide dynamics while directly targeting Boltzmann-consistent distributions. By injecting an intermediate force field, FBM incorporates physics priors into the generative process, enabling sampling that respects thermodynamics without extra resampling. Empirical results on Alanine-Dipeptide and PepMD demonstrate improved validity, flexibility, and distributional similarity, with strong transferability to unseen peptides. The approach advances efficient, physics-informed MD simulations with potential applicability to larger biomolecules and materials systems.

Abstract

Molecular Dynamics (MD) is crucial in various fields such as materials science, chemistry, and pharmacology to name a few. Conventional MD software struggles with the balance between time cost and prediction accuracy, which restricts its wider application. Recently, data-driven approaches based on deep generative models have been devised for time-coarsened dynamics, which aim at learning dynamics of diverse molecular systems over a long timestep, enjoying both universality and efficiency. Nevertheless, most current methods are designed solely to learn from the data distribution regardless of the underlying Boltzmann distribution, and the physics priors such as energies and forces are constantly overlooked. In this work, we propose a conditional generative model called Force-guided Bridge Matching (FBM), which learns full-atom time-coarsened dynamics and targets the Boltzmann-constrained distribution. With the guidance of our delicately-designed intermediate force field, FBM leverages favourable physics priors into the generation process, giving rise to enhanced simulations. Experiments on two datasets consisting of peptides verify our superiority in terms of comprehensive metrics and demonstrate transferability to unseen systems.

Figures (8)

• Figure 1: Illustration of how molecular conformations transfer from one state to another by MD (path in white) and time-coarsened dynamics (path in black).
• Figure 2: The overall framework of FBM- BASE and FBM. A. Firstly, FBM- BASE leverages the bridge matching framework to learn time-coarsened dynamics from the data distributions $q_0$ and $q_1$. B. With the guidance of the intermediate force field $\nabla\varepsilon_t$ at diffusion time $t$, the marginal distribution admits $p_t(\Vec{{\mathbf{X}}}_t)\propto{q_t(\Vec{{\mathbf{X}}}_t)\exp(-k\varepsilon_t(\Vec{{\mathbf{X}}}_t))}$, thereby the target distribution of FBM is debiased to the Boltzmann-constrained distribution $p_1$.
• Figure 3: Ramachandran plots of alanine dipeptide with conformation ensembles generated by models. The initial state is indicated with the red cross. Contours represent the kernel densities estimated by the MD trajectory and the generated conformations are shown in scatter.
• Figure 4: The visualization of comprehensive metrics on peptide 1e28:C. a. Plots of the slowest two TIC components analyzed by feature projections. b. The distribution of the radius-of-gyration. c. The residue contact map, where the data in the lower and upper triangle are obtained from FBM and MD, respectively. d. Cumulated valid conformations during inference over 3 independent runs.
• Figure 5: Comparisons between FBM and MD on conformation transitions over time. a. Comparison between the reference equilibrium conformations (blue) and the selected samples of FBM (yellow) of peptide 1e28:C. $C_{\alpha}$-RMSD values are reported below each cluster. b.$C_{\alpha}$-RMSD values along trajectories compared with the initial state of peptide 1e28:C over 3 independent runs. c. The effective sample size per second measured on the test set. All specific values are converted to multiples of the median value of MD, which is shown as the blue dashed line for reference.

Theorems & Definitions (4)

• Proposition 3.1
• Proposition 3.2
• Proposition 3.3
• Proposition 3.4