Reproduce “BoltzNCE: Learning Likelihoods for Boltzmann Generation with Stochastic Interpolants and Noise Contrastive Estimation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Rishal Aggarwal, Jacky Chen, Nicholas M. Boffi, David Ryan Koes

BoltzNCE estimates the likelihoods of samples from a molecular flow model with a separately trained energy-based model, enabling reweighting without repeated Jacobian-trace integration. On alanine dipeptide, it estimates a free-energy difference of 4.08 ± 0.13 kBT, close to the 4.10 ± 0.26 kBT umbrella-sampling reference, while reducing inference time. This evidence is limited to dipeptides; unseen-dipeptide accuracy is worse than the MD baseline.

Key insights

  • The combined Denoising Score Matching and InfoNCE loss synergistically improves EBM quality, yielding significantly better density approximation and mode weighting compared to individual objectives.
  • Amortizing likelihood estimation through a trained EBM offers a path to bypass the computational bottleneck of exact Jacobian trace calculations in Boltzmann Generators without sacrificing statistical accuracy.
  • Decoupling the generation of conformers from the likelihood estimation allows for greater flexibility in generative model design and enables the application of Boltzmann Generator principles to larger molecular systems.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.