🧬 biology

Multiscale autoregressive modeling generates ordered protein conformational trajectories

The paper introduces TEMPO, a hierarchical autoregressive model that efficiently generates long, ordered protein conformational trajectories with significantly reduced error accumulation and high fidelity to molecular dynamics data, outperforming existing methods like AlphaFlow.

Original authors: Mingchen Chen, Yaoyao Xu, Yu Liu, Tianshu Yu

Published 2026-10-19

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mingchen Chen, Yaoyao Xu, Yu Liu, Tianshu Yu

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 ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed.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

Proteins are the working machines of life, tiny molecular structures that fold into specific shapes to perform tasks like carrying oxygen, fighting infection, or sending signals between cells. However, these machines are not static statues; they are constantly moving, breathing, and shifting their shapes to get the job done. A protein might start in a closed position, open up to grab a molecule, and then snap shut to change it. Scientists have long known that understanding these movements is just as important as knowing the final shape, because the path a protein takes to get from one state to another often holds the key to how it works or how it might be stopped by a drug. For decades, researchers have tried to simulate these movements using powerful computers that calculate the physics of every atom, but these simulations are incredibly slow and expensive, often taking weeks to generate just a few seconds of movement for a single protein. This limitation has made it difficult to study how different proteins move or to compare the pathways of many different molecules side by side.

To solve this problem, a team of researchers has developed a new artificial intelligence tool called TEMPO that can generate long, ordered sequences of protein movements in a fraction of the time it used to take. Instead of trying to calculate every single physical force between atoms, TEMPO learns from thousands of existing computer simulations to recognize the patterns of how proteins naturally evolve over time. The system works like a two-speed camera: it first predicts the major, slow shifts in the protein's shape, which act as anchor points, and then fills in the rapid, smaller wiggles and fluctuations that happen between those anchors. This approach allows the model to create a smooth, continuous movie of a protein's life cycle without the errors that usually pile up when computers try to predict step-by-step over long periods. The researchers tested this method on a wide variety of proteins, including enzymes that help cells produce energy and a large family of proteins called kinases that regulate many cellular processes.

The results show that TEMPO can recreate the statistical properties of real protein movements with remarkable accuracy. When the team compared the generated sequences to the gold standard of physics-based simulations, the new model matched the expected patterns of movement almost perfectly. For instance, the model correctly predicted how much different parts of a protein would wiggle and how the overall shape would change, achieving a level of agreement that far exceeded previous methods. In one specific test involving a protein that opens and closes like a clam, the model successfully generated a continuous path from the closed state to the open state, showing a logical progression of contact changes that mirrored what scientists see in real experiments. The speed of this new method is also a major breakthrough; generating a sequence of 400 frames took less than one minute, a task that would have taken hours or days with older techniques.

Beyond just speed and accuracy, the study demonstrated that TEMPO can be guided to produce specific types of movements. The researchers showed they could steer the model to generate proteins in specific local shapes, such as a particular configuration of a chemical motif known as DFG, which is a critical target for many drugs. By applying simple rules during the generation process, they could direct the model to create structures that looked like the "on" or "off" states of these proteins, successfully doing so for dozens of different kinases. This suggests that the model has learned the underlying rules of protein flexibility well enough to be directed toward specific outcomes without needing to be retrained for each new protein. The team also applied the tool to a group of proteins called G protein-coupled receptors, which are embedded in cell membranes and are involved in sensing the environment, showing that the model could transfer its knowledge to proteins it had never seen before.

The study confirms that proteins move in a hierarchical way, with slow, large-scale changes providing the framework for faster, local jitters. By separating these two scales of motion, the researchers were able to prevent the computer from drifting off course, a common problem where small errors accumulate over time to create a nonsensical final shape. While the current version of the tool focuses on the backbone structure of the protein and does not yet include the surrounding water, ions, or chemical ligands, it provides a powerful new framework for comparing how different proteins move. This capability opens the door to rapidly screening how mutations or drugs might alter the dynamic behavior of proteins, offering a new lens through which to view the molecular machinery of life. The work does not replace the need for detailed physical simulations but offers a rapid, reliable complement that allows scientists to explore the vast landscape of protein motion with a clarity and speed that was previously out of reach.

Technical Summary: Multiscale Autoregressive Modeling Generates Ordered Protein Conformational Trajectories

Problem Statement

Protein function relies on transitions between conformational states, necessitating an understanding of ordered structural sequences rather than just static endpoints or unstructured ensembles. While Molecular Dynamics (MD) simulations provide atomically resolved trajectories, they are computationally expensive, limiting systematic comparisons across large protein panels, mutants, or ligand states. Existing deep generative models (e.g., AlphaFlow, BioEmu) excel at sampling diverse conformational ensembles but often fail to preserve the temporal dependencies required for meaningful sequential trajectories. Conversely, sequential generators face the challenge of error accumulation during long rollouts, where geometric errors compound as each prediction becomes the context for the next. Furthermore, protein motions are inherently multiscale, involving fast local fluctuations embedded within slower collective rearrangements; a single-scale update mechanism struggles to balance short-range variation with stable long-range evolution.

Methodology: TEMPO

The authors introduce TEMPO, a hierarchical autoregressive model trained directly on MD trajectories to generate ordered protein-backbone conformations at two distinct temporal resolutions.

Architecture and Representation

Backbone Representation: Each residue is represented by an SE(3) backbone frame (rotation via unit quaternion, translation) and torsion angles ( $\phi, \psi, \omega$ ) encoded as sine-cosine pairs to avoid discontinuities.
Neural Architecture: The model utilizes a shared architecture for both scales, employing Invariant Point Attention (IPA) for spatial encoding of SE(3) frames and multi-head self-attention for inter-residue interactions. Temporal dependencies are modeled using Gated Recurrent Unit (GRU) layers.
Dynamics Formulation: Conformational evolution is parameterized using an SDE-inspired discrete update: $X_{t+\Delta t} = X_t + f_\theta(ctxt)\Delta t + \sigma\sqrt{\Delta t}\epsilon_t$ . The prediction is conditioned on a two-frame context, making the process non-first-order Markovian.

Hierarchical Framework

TEMPO decomposes generation into two scales to mitigate long-horizon drift:

Slow-Scale Model: Advances "anchor" conformations at a 20-ns sampling stride (matching the training data stride). These anchors organize the long-range evolution of the trajectory.
Fast-Scale Model: Generates intermediate frames at a 1-ns stride. This model is conditioned on the slow-scale context (the anchor conformations) to resolve structural variation between anchors.
Inference Process: During inference, the slow-scale model generates the sequence of anchors first. Subsequently, the fast-scale model generates the 19 intermediate frames for each 20-ns segment, assembling a continuous, ordered trajectory.

Training and Conditioning

Training Data: Trained on 320,000 trajectories from 4,000 randomly sampled mdCATH proteins.
Loss Function: Combines mean-squared error reconstruction with a steric-clash penalty (ReLU penalty for non-adjacent atoms closer than 1.2 Å).
Conditioning Mechanisms:
- Endpoint-Conditioned: A learnable encoder maps a target structure ( $X^*$ ) to a per-residue embedding, which is added to the latent state to bias the rollout toward a specific endpoint.
- Inference-Time Steering: For local motif control (e.g., DFG geometry), a steering procedure refines specific residues after each autoregressive step using a multi-term objective without updating model parameters.

Key Results

Ensemble Fidelity and Efficiency

On held-out mdCATH and ATLAS proteins, TEMPO reproduced MD-derived ensemble observables:

Correlations: Achieved Pearson correlations of 0.91 for ensemble-averaged pairwise RMSD and 0.89 for residue-level RMSF on ATLAS.
Principal Component Analysis (PCA): The leading principal-component direction agreed with MD in 76% of ATLAS cases, significantly outperforming AlphaFlow (44%).
Efficiency: A 400-frame rollout required less than 1 minute, approximately 750-fold faster than AlphaFlow under reported settings.
Drift Reduction: Compared to a single-scale autoregressive model, the multiscale hierarchy reduced long-horizon RMSD error from 8.62 Å to 1.78 Å for a 400-frame rollout.

Application 1: Adenylate Kinase (ADK) Opening

In 10 endpoint-conditioned rollouts from closed (1AKE) to open (4AKE) states:

Generated conformations formed a continuous corridor in RMSD space.
Kinetic-map tICA identified a shared closed-to-open coordinate (tIC1) showing consistent progression with frame order.
Analysis revealed progressive contact exchange: closed-state-specific contacts decreased while open-state-specific contacts accumulated, with mobility localized to the NMP and LID domains.

Application 2: Kinase DFG Steering

Inference-time steering was applied to 100 kinases to generate alternative local DFG backbone geometries:

For 68 kinases, the method successfully yielded both DFG-in-like and DFG-out-like frames.
Accepted frames occupied distinct regions of the DFG-Phe Ramachandran plot, demonstrating controllability over local backbone geometry without retraining.

Application 3: Kinase Flexibility and GPCR Transfer

Kinase Flexibility: Across 252 kinases, generated fluctuation profiles showed variable agreement with crystallographic B-factors (mean Spearman $\rho = 0.358$ ; 22% with $\rho \ge 0.5$ ).
GPCR Transfer: The model, trained solely on soluble proteins (mdCATH), was applied to six GPCRs (Classes A–C) without specific updates. Generated conformations overlapped prominent regions of receptor-only MD projections, with variation concentrated in flexible regions (N-terminus, loops) while the transmembrane core remained stable.

Significance and Claims

The paper claims that TEMPO provides a framework for generating and comparing ordered protein conformational sequences, addressing the specific challenge of error accumulation in long-rollout generation. Its primary significance lies in:

Hierarchical Stability: Demonstrating that separating generation into slow (collective) and fast (local) scales significantly reduces drift while preserving MD-like conformational statistics.
Controllability: Showing that the model supports both endpoint-conditioned generation (for transition pathways) and inference-time steering (for local motif manipulation).
Efficiency: Offering a rapid alternative to physics-based simulation for comparing ordered conformational changes across large protein panels.

The authors explicitly note limitations: the model is backbone-centric (omitting side chains, ligands, and solvent), the time labels reflect sampling strides rather than calibrated kinetic time, and conditioned/steered rollouts are not unbiased samples. The work positions TEMPO as a complement to, rather than a replacement for, physics-based simulation and experimental validation.