Core Concepts
FOLDFLOW, a family of simulation-free generative models based on the flow-matching paradigm over the group SE(3), enables accurate and efficient modeling of protein backbones. The FOLDFLOW models offer several key advantages over previous approaches, including stability, faster training, and the ability to map any invariant source distribution to any invariant target distribution over SE(3).
Abstract
The paper introduces FOLDFLOW, a family of continuous normalizing flow (CNF) models tailored for distributions on SE(3)^N, which represents protein backbones. The authors propose three new CNF-based models that learn SE(3)^N-invariant distributions to generate protein backbones:
FOLDFLOW-BASE: A simulation-free approach to learning deterministic continuous-time dynamics and matching invariant target distributions on SE(3).
FOLDFLOW-OT: Accelerates the training of FOLDFLOW-BASE by constructing shorter and more stable flows using Riemannian Optimal Transport (OT).
FOLDFLOW-SFM: Learns a stochastic bridge on SE(3)^N by coupling Riemannian OT and simulation-free training.
The FOLDFLOW models offer several advantages over previous approaches, including stability, faster training, and the ability to map any invariant source distribution to any invariant target distribution over SE(3). Empirically, the authors validate FOLDFLOW on protein backbone generation of up to 300 amino acids, demonstrating high-quality, designable, diverse, and novel samples. They also show the utility of FOLDFLOW on equilibrium conformation generation by learning to simulate molecular dynamics trajectories.
Stats
The paper does not contain any key metrics or important figures to support the author's key logics.
Quotes
The paper does not contain any striking quotes supporting the author's key logics.