[Paper Review] Boltzmann Generators -- Sampling Equilibrium States of Many-Body Systems with Deep Learning

Frank Noé, Simon Olsson+2 more|arXiv (Cornell University)|Dec 4, 2018

Protein Structure and Dynamics2 references67 citations

TL;DR

The paper introduces Boltzmann Generators, invertible neural networks that map complex configurational spaces to a latent space for efficient one-shot sampling of Boltzmann-like distributions, enabling free energy calculations and discovery of new states without predefined reaction coordinates.

ABSTRACT

Computing equilibrium states in condensed-matter many-body systems, such as solvated proteins, is a long-standing challenge. Lacking methods for generating statistically independent equilibrium samples in "one shot", vast computational effort is invested for simulating these system in small steps, e.g., using Molecular Dynamics. Combining deep learning and statistical mechanics, we here develop Boltzmann Generators, that are shown to generate unbiased one-shot equilibrium samples of representative condensed matter systems and proteins. Boltzmann Generators use neural networks to learn a coordinate transformation of the complex configurational equilibrium distribution to a distribution that can be easily sampled. Accurate computation of free energy differences and discovery of new configurations are demonstrated, providing a statistical mechanics tool that can avoid rare events during sampling without prior knowledge of reaction coordinates.

Motivation & Objective

• Motivate the challenge of sampling equilibrium states in dense many-body systems and proteins.
• Propose a neural network based coordinate transformation to map configurations to a latent space that is easy to sample.
• Train the transformation by energy (KL loss) and by example (maximum likelihood) to avoid mode collapse.
• Demonstrate unbiased sampling and free energy calculations across model systems and a real protein, including disconnected states.

Proposed method

• Train an invertible neural network F_zx to transform latent samples z from a Gaussian prior into configurations x with Boltzmann weight.
• Minimize KL divergence between generated and Boltzmann distributions via J_KL = E_z[u(F_zx(z)) - log R_zx(z)], with R_zx the Jacobian determinant.
• Use training by energy (KL) and training by example (ML) to balance sampling low-energy states and maintaining entropy.
• Optionally include a reaction-coordinate loss to enhance sampling along chosen coordinates.
• Reweight generated samples to the Boltzmann distribution to compute thermodynamic observables.

Experimental results

Research questions

• RQ1 Can a neural network learn a coordinate transformation that yields unbiased one-shot samples from Boltzmann distributions without predefined reaction coordinates?
• RQ2 How effectively can Boltzmann Generators reproduce free energy differences and sample transition states in metastable, high-dimensional systems?
• RQ3 Can the latent-space representation facilitate exploration of configuration space and enable sampling of new metastable states?
• RQ4 Is it possible to extend to complex molecules and disconnected states, and compute temperature-dependent free energies without extensive MD?

Key findings

• Boltzmann Generators produce unbiased one-shot samples that reproduce Boltzmann statistics after reweighting.
• Training by energy plus training by example prevents mode collapse and yields accurate free energy differences in model systems.
• Latent-space interpolation between states generates physically meaningful low-energy pathways.
• For complex molecules like BPTI, the method can sample all-atom structures and reproduce realistic bond/angle distributions with reweighting support.
• Using multiple Boltzmann Generators for disconnected states enables direct free energy differences without predefined reaction coordinates, achieving substantial speedups over brute-force MD and umbrella sampling.
• The latent-space exploration approach combined with Metropolis Monte Carlo in latent space accelerates discovery of new states and transitions.