⚗️ biochemistry

UniFlow: Unifying protein conformational ensemble generation and machine-learned force fields with a scalable normalizing Flow

UniFlow is a scalable normalizing flow framework that unifies protein conformational ensemble generation and machine-learned coarse-grained force fields, enabling both rapid sampling of equilibrium distributions and stable long-timescale molecular dynamics simulations within a single differentiable model.

Original authors: Liu, Y., Chen, M., Lin, G.

Published 2026-07-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Liu, Y., Chen, M., Lin, G.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine the world of biology as a giant, bustling dance floor. In this dance, proteins are the stars, but they don't stand still like statues. Instead, they are constantly wiggling, twisting, and folding into different shapes to do their jobs, like catching viruses or building cells. Scientists call these different shapes a "conformational ensemble." To understand how proteins work, researchers need to map out all the possible moves they can make. Traditionally, they've used a method called Molecular Dynamics (MD), which is like running a super-accurate physics simulation to watch the dance unfold. However, this method is incredibly slow and expensive, like trying to film a whole movie by moving every single atom one by one with a stopwatch. It often takes too long to see the rare, important moves the protein makes.

Recently, scientists have tried using artificial intelligence to speed things up. Some AI models act like a "generative artist," learning from past movies to instantly draw new dance poses. Others act like "force field engineers," creating simplified rules to predict how the protein should move. But until now, these two approaches have been developed separately, like two different teams working on the same puzzle without talking to each other. One team focuses on making pretty pictures quickly, while the other focuses on making sure the physics are right, but they haven't been able to combine the speed of the artist with the accuracy of the engineer.

Enter UniFlow, a new tool that tries to be the ultimate bridge between these two worlds. Think of UniFlow as a "universal translator" for protein shapes. It uses a clever mathematical trick called a "normalizing flow" to learn the exact probability of a protein being in any specific pose. Unlike other AI models that guess or guess-and-check (like diffusion models that slowly refine a blurry image into a clear one), UniFlow can instantly generate a perfect pose in a single step. It's like having a magic camera that can snap a photo of a protein in any of its thousands of possible positions instantly, without needing to wait for the physics simulation to run.

The researchers found that UniFlow is incredibly good at this. When they tested it on a wide variety of proteins, the poses it generated looked almost exactly like the ones from the slow, expensive physics simulations. It was just as accurate at predicting how far apart parts of the protein were, how round it was, and how much it moved. But the real magic happened when they looked at speed. While a competing AI model took hundreds of seconds to generate a set of poses, UniFlow did it in just a few seconds. On a standard computer chip, it was roughly 50 to 100 times faster than the previous best methods.

Even cooler, UniFlow doesn't just make pictures; it can also act as the "engine" for the physics simulation itself. Because it knows the exact math behind the protein's shapes, scientists can use it to run long-term simulations of how proteins move over time. To make this even faster, they "distilled" the big, complex UniFlow model into a tiny, lightweight version for each specific protein. This is like taking a massive library of rules and compressing it into a single, pocket-sized cheat sheet. With this cheat sheet, they could run a 2-nanosecond simulation (which is a tiny fraction of a second in real time, but huge for a computer) in just 5.3 hours for a medium-sized protein, compared to 134.2 hours if they used the full, un-simplified model.

However, the paper is careful to note that UniFlow isn't a magic wand that works without help. It still needs a "reference" pose to start with—a starting point or a native structure to measure its movements against. It learns the changes from that starting point rather than inventing a shape from thin air. While it works amazingly well for the proteins it was trained on and even for some new ones it hasn't seen before, it relies on having that initial reference frame. The authors suggest that while it's a huge step forward, future versions might need to learn to work without that reference to handle even more chaotic and diverse protein shapes. But for now, UniFlow stands as a powerful new way to unify fast AI generation with accurate physics, potentially letting scientists watch the protein dance floor in high definition without waiting years for the movie to finish.

Technical Summary of UniFlow

Problem Statement
Molecular dynamics (MD) simulations are the principled standard for modeling equilibrium protein conformational energy landscapes. However, their utility is constrained by high computational costs, which limit access to long timescales and large protein systems. While recent machine learning approaches have emerged to address this, they have developed in two largely independent silos:

Generative Protein Ensemble Models: These learn transferable conformational energy landscapes from structural and simulation data to enable efficient sampling but often lack the physical consistency required for dynamical simulations.
Machine-Learned Coarse-Grained Force Fields (ML-CGFFs): These construct physically consistent effective potentials for long-scale dynamics but are typically developed separately from generative sampling frameworks.

Both directions aim to represent the same underlying equilibrium distribution, yet they have not been unified. This gap prevents the simultaneous realization of scalable, transferable ensemble generation and stable, physics-based molecular dynamics within a single framework.

Methodology: UniFlow
UniFlow is introduced as the first scalable generative framework that unifies protein ensemble generation and ML-CGFF learning. It operates as a normalizing flow defined over protein internal coordinates, specifically modeling the joint distribution of angular degrees of freedom (backbone torsions $\tau$ and valence angles $\alpha$ ) relative to a reference structure.

Internal Coordinate Representation: To remove global translational and rotational degrees of freedom, UniFlow models angular displacements $\Delta y = (\Delta \tau, \Delta \alpha)$ from native backbone torsions and ideal valence angles. Bond lengths are fixed to ideal values during generation and augmented with a stiff harmonic potential only when constructing the MD force field.
Architecture: The model employs a stack of $N=12$ $N = 12$ circular coupling layers.
- Coupling Mechanism: An alternating odd–even residue masking scheme is used. Inactive residues are held constant while active residues are transformed via invertible, periodic rational-quadratic splines.
- Geometry-Aware Conditioner: The parameters for these splines are predicted by a conditioner that processes the masked input through three stages:
  1. Backbone Reconstruction: Masked displacements are converted to Cartesian coordinates using NeRF (Neural Radiance Fields).
  2. Invariant Geometric Encoding: An SE(3)-invariant geometric Transformer (based on ESM3 architecture) encodes the reconstructed structure.
  3. Two-Track Transformer: A joint sequence–structure Transformer combines amino acid sequences with structural representations to predict spline parameters. This design allows the model to capture long-range structural dependencies despite operating in local internal coordinates.
Training Objective: The model is trained via maximum likelihood on large-scale equilibrium MD trajectories. To ensure physical realism, the negative log-likelihood (NLL) objective is augmented with two geometric losses: a clash loss (penalizing nonlocal residue pairs below a distance threshold) and a contact loss (matching long-range contact probabilities).
Force Field Distillation: To enable efficient MD, the pretrained UniFlow model (which includes heavy Transformers) is distilled into a lightweight, system-specific potential. This student model replaces the geometry-aware Transformer conditioner with a simple multilayer perceptron (MLP), trained on conformations sampled directly from the pretrained UniFlow. This distilled potential includes the learned angular density, a stiff harmonic bond-length term, and a short-range repulsive potential for stability.

Key Contributions

Unified Framework: UniFlow is the first architecture to support both scalable protein ensemble generation and machine-learned force field modeling within a single model, bridging the gap between generative modeling and physics-based simulation.
Scalable Internal-Coordinate Training: The authors developed a training strategy using internal coordinates with geometry-aware supervision. This mitigates the strong locality of internal coordinates, enabling generalization across proteins of different sizes and unseen sequences.
Distillation Recipe: A method is proposed to transfer a generalizable UniFlow model into a lightweight, system-specific potential. This enables stable Langevin simulations with significantly reduced computational cost for force evaluation.

Results

Ensemble Quality: Evaluated on diverse datasets (CATH1, Octapeptides, and the held-out ATLAS benchmark), UniFlow generates ensembles that closely match reference MD simulations. It achieves Jensen–Shannon divergence (JSD) scores comparable to the state-of-the-art BioEmu model across metrics including end-to-end distance, radius of gyration, RMSD, and TICA components.
Generalization: The model successfully generalizes to unseen peptide sequences (Octapeptides) and entirely unseen protein systems (ATLAS), demonstrating transferability beyond its training distribution.
Sampling Efficiency: UniFlow generates samples in a single inverse step, resulting in substantially faster sampling times compared to diffusion-based baselines like BioEmu (e.g., 1.0s vs. 71.1s for Octapeptides on an NVIDIA L40S GPU).
Ablation Studies: Experiments confirm that operating in internal coordinates with a geometry-aware architecture is superior to Cartesian-based flows. Furthermore, the inclusion of geometric losses is critical for capturing global structural organization in larger proteins, whereas NLL-only training performs well only on small, locally dominated systems.
MD Simulation: The distilled UniFlow potential enables stable Langevin dynamics. Free-energy landscapes recovered from UniFlow-driven MD match reference trajectories, capturing major metastable basins. Distillation combined with torch.compile reduces the wall-clock time for a 2 ns simulation of a 61-residue protein from 134.2 hours (pretrained model) to 5.3 hours.

Significance
The paper claims that UniFlow paves the way for a unified class of models that bridges generative ensemble modeling with physics-based molecular simulation. By providing exact likelihoods over protein conformations, UniFlow allows the learned distribution to serve directly as a differentiable potential for MD. This capability addresses the historical trade-off between the efficiency of generative sampling and the physical consistency required for dynamical simulations, offering a pathway toward scalable, transferable, and physics-informed protein modeling.

Technical Summary of UniFlow

More like this