BoltzNCE: Learning Likelihoods for Boltzmann Generation with Stochastic Interpolants and Noise Contrastive Estimation

Area of Science:

• Computational chemistry
• Statistical mechanics
• Machine learning

Background:

• Efficiently sampling from the Boltzmann distribution is crucial for molecular modeling.
• Continuous Normalizing Flows (CNFs) are used in Boltzmann Generators but require costly Jacobian computations.
• This limits their applicability to large molecular systems.

Purpose of the Study:

• To develop a computationally efficient method for learning Boltzmann distributions without explicit likelihood computation.
• To enable accurate free energy calculations for large molecular systems.

Main Methods:

• Proposed learning the likelihood using an energy-based model trained with noise contrastive estimation and score matching.
• Utilized stochastic interpolants for annealing between prior and generated distributions.
• Combined objective functions to efficiently learn the density function.

Main Results:

• Achieved free energy profiles and energy distributions comparable to exact likelihood methods on the alanine dipeptide system.
• Demonstrated accurate estimation of free energy differences between metastable states.
• Obtained orders-of-magnitude speedup compared to existing methods.

Conclusions:

• The proposed method offers a practical and efficient alternative for Boltzmann distribution sampling in molecular modeling.
• Enables accurate and accelerated free energy calculations for complex systems.
• Overcomes the computational bottleneck associated with Jacobian computations in CNFs.