UniFlow is introduced, the first scalable generative model that unifies protein ensemble generation and machine-learned coarse-grained force fields for molecular dynamics simulation within a single framework, and paves the way for a unified class of models that bridges generative ensemble modeling with physics-based molecular simulation.
Abstract
Molecular dynamics (MD) provides a principled method for modeling equilibrium protein conformational energy landscapes, but its computational cost limits access to long timescales and larger protein systems. Recently, generative protein ensemble models and machine-learned coarse-grained force fields have emerged as complementary approaches for accelerating conformational sampling. However, they are typically developed separately despite modeling the same underlying equilibrium distribution. We introduce UniFlow, the first scalable generative model that unifies protein ensemble generation and machine-learned coarse-grained force fields for molecular dynamics simulation within a single framework. UniFlow employs an internal-coordinate normalizing flow that supports efficient i.i.d. sampling, exact likelihood evaluation, and differentiable energy and force computation. Across diverse protein systems, UniFlow generates ensembles that closely match reference MD simulations, generalizes to proteins beyond its training dataset, and samples substantially faster than diffusion-based ensemble-generation baselines. The same learned density further enables stable long-timescale molecular dynamics simulations. Together, UniFlow paves the way for a unified class of models that bridges generative ensemble modeling with physics-based molecular simulation. Code https://github.com/Harrydirk41/UniFlow.git
Proteins are critical biomolecular machines that populate ensembles of interconverting conformations. Many biological processes depend on transitions between metastable states. Although molecular dynamics (MD) simulations provide a physically grounded route to characterize these motions, routine sampling of large-scale conformational transitions remains computationally demanding. Recent advances in protein structure prediction have created new opportunities for ensemble generation, but many existing approaches require noising inputs, task-specific training, supervised fitting on extensive MD data, or experimentally-informed restraints. Here, we introduce Pi-Ensemble (Predicting Interpolated Ensemble), a sequence-guided framework for generating protein conformational ensembles interpolating between two structural anchor states. Unlike previous methods, Pi-Ensemble alternately leverages inverse-folding and structure-prediction models to propose intermediate conformations between known protein states, generating diverse ensembles without additional training. We evaluate Pi-Ensemble across diverse protein systems, including enzymes, transporters, receptors, and benchmark cases with reference MD simulations or experimental Double Electron-Electron Resonance (DEER) data. Pi-Ensemble recovers physically plausible intermediate conformations, captures transition pathways observed in large-scale MD simulations, and generates structures consistent with experimental distance distributions. Furthermore, Pi-Ensemble-generated conformations provide effective starting seeds for parallel MD simulations, improving conformational exploration and accelerating convergence relative to simulations initiated only from endpoint structures. These results establish sequence-guided structural interpolation as a practical strategy for probing protein conformational landscapes. By generating diverse and physically reasonable conformational proposals without long-timescale MD or model retraining, Pi-Ensemble provides an extensible framework for studying protein flexibility, guiding adaptive sampling, and accelerating mechanistic investigations of protein function.
Hassan Nadeem, D. Kleiman, Yuming Zhou et al.· bioRxiv· 0 citations
GeoNet is a physicochemical-principle-guided framework for modeling dual-range atomic interactions that achieves the smallest model size and the shortest training time, demonstrating both superior predictive performance and computational efficiency.
A new physics-informed representation using Fourier transforms as an inductive bias for the multiscale temporal nature of protein dynamics, DynaMode is presented, achieving strong performance across a set of ensemble-based metrics.
H. Phipps, M. Cagiada, S. Villalba et al.· 0 citations
Transition paths between metastable protein states encode both equilibrium structure statistics and dynamical connectivity yet are costly to obtain with molecular dynamics (MD) and remain challenging to emulate with machine learning. Here, we present TPS-Flow, a physics-guided flow-based generative framework for end point-conditioned conformational path sampling between predefined protein states (not equilibrium ensembles). TPS-Flow represents structures as residue-level SE(3) transforms, uses a spatiotemporal gated attention encoder to learn a flow-matching interpolation velocity field from MD trajectories, and incorporates optional energy and structure-aware constraints together with a short physics-based relaxation step. Across a mycobacterial membrane transporter, a monomeric protein, a protein-protein complex, and a soluble enzyme, TPS-Flow preserves residue-wise fluctuation patterns with damped amplitudes and occupies TICA-projected conformational corridors consistent with reference MD, provides conformational coverage complementary to finite reference MD sampling and generates intermediates with reference-comparable docking scores while reducing model size and computational cost compared to a state-of-the-art trajectory generator (MDGen). In an out-of-distribution structural generalization test using PN-subdomain mutants, TPS-Flow preserved fold continuity and wild-type-like global RMSF patterns when conditioned on AF3-derived mutant end point structures, thereby bridging atomistic simulation and deep generative modeling of protein transition paths.
Kai Xu, Likun Zhao, Yanan Tian et al.· Journal of Chemical Informat...· 0 citations
We introduce implicit machine learning force fields (I-MLFFs), which replace explicit stacks of neural network layers with self-consistent fixed-point equations. In molecular simulations, this formulation enables intermediate representations to be reused across successive timesteps, thereby warm-starting force evaluation. The resulting models effectively combine the computational footprint of a shallow, single-layer MLFF with the representational capacity and accuracy of a deep neural network. Our approach unlocks architecture-agnostic efficiency gains that are inaccessible when force prediction and trajectory integration are considered separately. We demonstrate this across three major classes of graph neural networks: invariant, equivariant Cartesian tensor, and SO(3)-equivariant spherical-tensor architectures. Each yields a two- to five-fold reduction in compute and memory footprint. Crucially, these gains are achieved while retaining full atomistic resolution and the original integration timestep, avoiding spatial or temporal coarse graining. Our contribution therefore advances the scaling frontier of quantum-mechanically faithful molecular simulation, enabling longer trajectories and larger atomistic systems within fixed GPU memory and compute budgets, and thereby opening access to new insights across biomolecular and material systems.
J. Maess, Leon Werner, J. Frank et al.· 0 citations
This work proposes a novel approach that learns hierarchical discrete representations of protein structures using vector quantization, and outperforms state-of-the-art models such as ESMDiff across challenging benchmark datasets, including BPTI MD trajectories and conformational-changing pairs.
Seokjun On, Yujin Jeong, Kanghyeon Kim et al.· Bioinformatics· 0 citations