Atomic Trajectory Modeling with State Space Models for Biomolecular Dynamics
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: We're looking at the paper: "Atomic Trajectory Modeling with State Space Models for Biomolecular Dynamics".
Marcus: Here's the gist of the paper. Understanding the dynamic behavior of biomolecules is fundamental to elucidating biological function and facilitating drug discovery.
Ines: ATMOS is a novel generative framework based on State Space Models (SSM) designed to generate atom-level MD trajectories for biomolecular systems.
Marcus: The paper introduces ATMOS, a novel generative framework based on State Space Models (SSM) designed to generate atom-level MD trajectories for biomolecular systems.
Ines: This approach bridges the gap between computationally expensive Molecular Dynamics (MD) simulations and deep generative models by formulating biomolecular dynamics generation as a sequence modeling problem.
Title and authors: Ines: So we're looking at this paper, "Atomic Trajectory Modeling with State Space Models for Biomolecular Dynamics," and it seems like they've taken something from sequence modeling and applied it to molecular dynamics trajectories.
Marcus: Right. It’s about using a State Space Model, or SSM, to generate atom-level movements for biomolecules instead of just looking at static structures. That immediately suggests they're trying to model the actual motion, not just the shape.
Ines: Exactly. The paper points out that traditional methods are either too slow for long timescales or only work well on single proteins, so this approach aims to bridge that gap by modeling dynamics across different system sizes and time scales.
Marcus: And they claim this method uses a State Space Model to handle the temporal relationships, which they say gives them linear computational complexity with respect to the trajectory length. That sounds like a big deal compared to quadratic scaling in some attention-based models.
Yuki: From a population genetics standpoint, if you can model these atom-level dynamics accurately across different systems, it means we might finally be able to see how small changes in structure translate into observable biological functions across the whole species.
Ines: That's the big picture there. The core idea is that they are training this framework not just on static structures but on actual large-scale molecular dynamics simulation data, like mdCATH and MISATO. They’re using those real simulations to teach the model how molecules actually move in their environment.
Marcus: And they do this by operating directly on the atomic coordinates during both the encoding and decoding phases. That means they aren't just working with simplified representations like carbon atoms or backbone frames, which is a key detail for getting that high-fidelity atom-level output.
Yuki: Operating at the atomic level is crucial because it allows you to capture things like subtle interactions between specific residues, which are often what drive functional changes in a protein.
Ines: That’s right. And they structure their model using a compact latent representation, defining this state as having both a time step and some other variable to balance resolution with efficiency. They call this state transition function the Pairformer module, which allows for bidirectional information flow between different parts of the track.
Title and authors: Marcus: So, on the training side, they have two main loss terms: one for latent fidelity using a distogram loss to keep the geometry consistent in that latent space, and another reconstruction loss using a diffusion-based decoder to actually recover those target frames from noisy inputs.
Yuki: That dual-loss approach sounds like it tries to enforce both the underlying physical structure of the movement and ensure the generated sequence looks realistic. It’s about making sure the generated motion makes sense biologically, not just mathematically.
Ines: Precisely. And then for using it, they have a bi-level sampling protocol where they generate a sparse set of keyframes at a coarser time resolution and then use an interpolator to fill in the gaps between those anchors.
Marcus: That interpolation step is clever because the interpolator isn't just looking at the start point; it’s conditioned on both the starting frame and a subsequent anchor frame, plus how much time is left until the end, which keeps everything coherent.
Yuki: It sounds like a very controlled way to generate long trajectories, rather than letting the model wander off in an unpredictable way over many steps. That control is what makes it useful for studying kinetic pathways.
Ines: And when they tested it, they saw state-of-the-art performance on both monomeric proteins and complex protein ligand systems. For protein dynamics on the mdCATH dataset, they got a Pearson correlation of zero point ninety-two in pairwise root mean square deviation with the reference MD trajectories.
Marcus: That correlation number is high, but I always want to be careful about how that number is calculated. They also showed superior efficiency compared to both traditional simulations and purely attention-based approaches, cutting inference time by two point seven times over ConfRover.
Yuki: High accuracy on those benchmarks is good for the general case, but what about the complexity of a truly novel system? Can this model handle dynamics in environments that aren't well-represented in their training data?
Ines: That’s a fair question. The paper does mention they successfully model coupled protein ligand dynamics on the MISATO dataset, and they achieved the best scores across all four metrics there, including trajectory accuracy error.
Title and authors: Marcus: But we gotta remember that their efficiency gain is relative to those specific comparison points, so we need to see how it holds up when you push it further into entirely new biological systems. The paper notes that they successfully generalize from monomeric proteins to complex biomolecular systems, which is the main claim for broader applicability.
Yuki: Generalization across system types is important because biology isn't just about one specific molecule; it’s about interactions between many things, and being able to model those larger systems is where the real biological insight comes from.
Ines: So, to wrap up on the results, they show this unified framework can handle both protein dynamics and these more complex protein ligand scenarios with high accuracy across large datasets. It extends static structure prediction into the dynamics domain.
Marcus: They also make a point about how this approach paves the way towards foundation models for dynamics. That means we’re moving toward systems that can potentially model many different types of molecular motion, not just one specific protein or interaction type.
Yuki: It sounds like a step toward building models that understand the fundamental rules of movement in biological matter rather than just fitting data points for specific examples.
Ines: That's the implication. The paper, "Atomic Trajectory Modeling with State Space Models for Biomolecular Dynamics," shows how state space models can be adapted to efficiently generate long-horizon trajectories at an atomic level.
Marcus: They achieve that by using a specific structure—the state transition function—that captures long-range dependencies linearly, which is what makes the inference so much faster than older methods.
Yuki: It's about making the simulation of molecular motion tractable for studying biological systems in a way that’s computationally feasible for researchers.
Ines: So, this work gives us a new tool to study the movement of biomolecules at a very fine resolution, backed by strong performance on established benchmarks like mdCATH and MISATO.
Marcus: It's promising because it marries the efficiency of state space models with the detail needed for atomic-level trajectory generation.
Yuki: It’s a good piece of work for understanding how physical constraints translate into observable biological motion across different molecular scales.
The paper's summary: Ines: So, to recap, this paper is about using a State Space Model to generate actual atom-level movements for molecules instead of just looking at their static shapes and how that matters for biology.
Marcus: Right. It boils down to this framework called ATMOS which uses the SSM structure to handle long stretches of time in a way that’s computationally efficient, linear with respect to the length of the trajectory.
Yuki: And they are training it on real data from large molecular dynamics simulations, like those found in mdCATH and MISATO, so it learns how molecules actually move under different conditions.
Ines: What I find really interesting is that they model this at the full atomic level during both the encoding and decoding parts. That means they aren't just looking at simplified representations like carbon atoms or backbone frames; they are dealing with every single atom’s coordinate.
Marcus: That level of detail is what lets them capture physical realism, which I think is a big step because it allows you to see the actual forces and interactions happening on that molecular scale.
Yuki: From a population genetics viewpoint, if you can model these atomic movements accurately across different protein systems, it means we might finally be able to link those tiny structural changes directly to observable biological functions across different species.
Ines: Exactly. They show this method can predict flexibility and how accurate the distribution of movement is when testing on protein dynamics data for the mdCATH set, showing a Pearson correlation of about zero point nine two with the reference MD trajectories for a specific protein.
Marcus: That correlation is high, but we have to consider the context—it’s zero point nine two against other established methods like conformation models and previous trajectory models. It’s a good number, but it doesn't tell us if it works reliably on entirely new biological systems.
Yuki: You’re right about the context. The real test for this is its ability to handle complex protein-ligand interactions, which they tested on the MISATO dataset and said they got the best scores across all four metrics there.
Ines: They achieved that, suggesting it can model how a molecule moves when it’s bound to something else, which is crucial for drug discovery because understanding those binding motions tells you how to design better drugs.
Marcus: And on the efficiency side, they claim this approach is much faster than traditional simulation methods or purely attention based models; they show a time reduction of two point seven times compared to ConfRover.
Yuki: So, we’ve got a framework that can generate detailed atomic dynamics efficiently across different system types—monomers and complexes—and the implication is that we can study molecular motion with much higher resolution than before.
Ines: It opens the door for building foundation models for dynamics, which means instead of training a model just for one specific protein or interaction, you could potentially train one that understands the fundamental rules of how different biomolecules move.
The paper's improvements: Tom: So we're going to look at how the authors suggest they could make this ATMOS framework even better, focusing on their proposed improvements for the model itself.
Ines: The paper discusses refining the latent representation, specifically how they define that state variable—the time step and whatever other piece is there—to get a better balance between detail and speed.
Marcus: They propose using a more structured state space definition, calling it the Pairformer module again, but making sure that bidirectional information flow between the single track and pair tracks is even stronger.
Yuki: I wonder if this tighter coupling helps with capturing those subtle, long-range dependencies in how a protein folds or moves over very long time scales.
Ines: It does; they suggest that by optimizing this state transition function, the model can better capture the kinetic pathways, which is key because we want to know not just where it ends up, but how it gets there.
Marcus: From a data science standpoint, they also talk about improving the loss functions. They suggest focusing more on the reconstruction loss from a diffusion perspective to ensure those generated frames look physically plausible and don't just match the overall statistics of the training data.
Yuki: That sounds like it’s trying to enforce physical constraints directly into the learning process, which is important for making sure what we generate isn't just statistically similar but biologically valid.
Ines: They also discuss improving their inference protocol, moving beyond just keyframes and interpolation to a more dynamic way of sampling that decouples the large conformational changes from the smaller local fluctuations.
Marcus: That means they’re trying to make the trajectory generation itself more intelligent, using those anchor frames not just as starting points but as anchors for generating fine detail between them.
Yuki: If they can do that, it could mean we get much cleaner and more accurate simulations of protein behavior without needing an impossibly long computation time for every single frame.
Ines: But they also flag a limitation: the model still relies heavily on the quality and diversity of the training data, so if you ask it to model a system completely outside those datasets, its performance might drop significantly.
Marcus: That makes sense; it’s still a generative model trained on specific molecular dynamics simulations, so generalization outside that learned distribution is always a concern.
Yuki: So the paper isn't claiming perfection across the board, but rather suggesting iterative improvements to the architecture and training objectives to push those limits further.
Ines: The real implication here is that this isn't just one solution; it’s a roadmap for how we can systematically enhance these generative models to handle more complex biological questions.
Conclusion: Tom: So we're wrapping up this look at "Atomic Trajectory Modeling with State Space Models for Biomolecular Dynamics" by summarizing what we’ve seen about it and where it leads next.
Ines: Basically, this paper shows how you can use a State Space Model to generate high-resolution atom-level movements for biomolecules, taking us past just looking at static structures.
Marcus: It achieves that by using the SSM structure which gives them linear complexity for the trajectory length, which is a big win over older methods that scale quadratically with time.
Yuki: From a population genetics perspective, this means we might finally be able to connect structural changes to observable biological traits across different species.
Ines: It really shows that by training on real simulations like mdCATH and MISATO, the AI can predict flexibility and distribution accuracy with very high correlation numbers against those reference MD trajectories.
Marcus: That high correlation is impressive, but we need to keep in mind it’s a comparison against other established models, so we have to be careful about how robust this performance is when applied to entirely new biological systems.
Yuki: It also demonstrates that the AI can handle complex protein-ligand dynamics, which is vital because understanding how a molecule binds tells us a lot about drug design.
Ines: So what this means for us is that we have a unified generative framework that can tackle both monomeric proteins and these more intricate systems.
Marcus: It’s paving the way toward foundation models for dynamics, which means we could eventually build AI systems that understand the fundamental rules of how biomolecules actually move.
Yuki: And it also validates the use of large molecular dynamics datasets as a training ground for learning these physical laws.
Ines: So, to sum up this paper on "Atomic Trajectory Modeling with State Space Models for Biomolecular Dynamics," we have a powerful tool for simulating atomic movement efficiently across different biological scales.
Marcus: It’s promising because it marries the speed of state space models with the detail needed for atom-level trajectory generation.
Yuki: I think the real value is in seeing how these models can support deeper questions about molecular function across evolutionary timescales.
Ines: Next time, we'll look at a paper that tackles a different side of molecular modeling, something focusing more on the energy landscape rather than just the trajectory itself.
Peking University · Montreal Quebec AI Institute
q-bio.BM, cs.LG
Submitted: 2026-03-18
Updated: 2026-11-06
Comments: Accepted at NeurIPS 2026
License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
Importance score: 100/100
The gist: Understanding the dynamic behavior of biomolecules is fundamental to elucidating biological function and facilitating drug discovery.
Key concepts
- State Space Models (SSM)
- SSMs are a type of generative model excellent for predicting long sequences. They maintain a compact internal 'state' that evolves over time, allowing the model to efficiently capture complex temporal dependencies in molecular motion without the quadratic scaling issues found in traditional methods.
- Atomic-level Modeling
- Unlike models that simplify protein structures (like using just backbone atoms), ATMOS works directly with every single atom's coordinates during generation. This allows it to produce highly detailed and physically accurate trajectories that reflect the true atomic movements of biomolecules.
- Bi-level Sampling Protocol
- This is a two-step sampling technique used during inference. First, the model identifies key moments (anchor frames) at a low resolution. Second, an interpolator then fills in the details between these anchors using information from both adjacent anchor points and the remaining time to ensure smooth and consistent motion.
Terminology
Summary
Understanding the dynamic behavior of biomolecules is fundamental to elucidating biological function and facilitating drug discovery. p. 0 ATMOS is a novel generative framework based on State Space Models (SSM) designed to generate atom-level MD trajectories for biomolecular systems. p. 0
Introduction and Motivation
The paper introduces ATMOS, a novel generative framework based on State Space Models (SSM) designed to generate atom-level MD trajectories for biomolecular systems. This approach bridges the gap between computationally expensive Molecular Dynamics (MD) simulations and deep generative models by formulating biomolecular dynamics generation as a sequence modeling problem. p. 0 ATMOS leverages SSM to efficiently capture long-horizon trajectory evolution with linear computational complexity in terms of the trajectory length. p. 0 The framework is trained across crystal structures from PDB and large-scale MD simulation data including mdCATH and MISATO. p. 0
Model Architecture
ATMOS models dynamics at the fully atomic level by operating directly on atomic coordinates during both the encoding and decoding phases, rather than coarsegraining representations like Cα atoms or backbone frames. The framework utilizes a compact coupled latent representation where the SSM state ht is defined as (st, zt) to reconcile resolution with efficiency. p. 3 The input signal is the time-varying input to the model being the configuration of the system at time t, represented by atomic coordinates xt ∈ R T ×N×3. p. 2 The state transition function Tθ evolves the latent state h from t to t + 1 using a variant of the Pairformer module which enables bidirectional information flow between the single and pair tracks. p. 3
Training Objectives
The framework is trained end-to-end with two important loss terms: (1) latent fidelity loss (distogram) and (2) reconstruction loss (diffusion). The Latent Fidelity Loss enforces geometric consistency in the latent space by imposing a distogram-based latent fidelity loss. p. 4 The Reconstruction Loss optimizes the diffusion-based decoder Dθ to (conditionally) recover the target frame xt from a noisy state x˜(γ). p. 4 The final training objective is a linear combination of these two losses: L = λrecon · Lrecon + λdisto · Ldisto. p. 4
Inference and Sampling
During inference, ATMOS samples trajectories in a specific temporal granularity ∆t, employing a bi-level sampling protocol to decouple conformation changes from local fluctuations. This protocol involves generating a sparse set of keyframes at a coarse time resolution and then filling in the intermediate frames via parallel interpolation conditioned on the adjacent keyframes. p. 7 The interpolator is conditioned on not only the starting frame xt, but also the subsequent anchor frame xt' and the remaining time horizon to ensure a coherent trajectory between endpoints.
Experimental Validation
ATMOS achieves state-of-the-art performance in generating conformation trajectories on both protein monomer and complex protein-ligand systems and demonstrates superior efficiency compared to traditional simulation or purely attention-based approaches. For protein dynamics on the mdCATH dataset, ATMOS significantly outperforms both conformation models and previous trajectory models in flexibility prediction, achieving a near-perfect correlation with the reference MD trajectories in pairwise RMSD (Pearson r = 0. p. 692) and global RMSF (r = 0. p. 590). In protein-ligand dynamics on the MISATO dataset, ATMOS attains the best scores in all four metrics, including trajectory accuracy (t-RMSD Error). p. 7 The efficiency evaluation shows that ATMOS reduces inference time by 2. p. 67× compared to ConfRover and by over 60× compared to MD simulation. p. 6
Conclusion
ATMOS presents a unified generative framework for simulating biomolecular dynamics that generalizes from monomeric proteins to complex biomolecular systems. The model achieves SOTA performance on large-scale MD benchmarks, validating its ability to generalize across diverse biological systems and paving the way toward foundational dynamics models. p. 7 Responsible deployment and transparency are essential to maximize benefits and mitigate risks. p. 8 p. 8
References
Abramson, J., Adler, J. p. 8, Dunger, J. p. 8, Evans, R. p. 8 p. 8, Green, T. p. 8 p. 8, Pritzel, A. p. 8 p. 8, Ronneberger, O. p. 8 p. 8, Willmore, L. p. 8, Ballard, A. p. 8 J., Bambrick, J. p. 8, et al. Accurate structure prediction of biomolecular interactions with alphafold 3. p. 8
Chen, X., Zhang, Y. p. 8, Lu, C., Ma, W., Guan, J., Gong, C., Yang, J., Zhang, H. p. 8, Zhang, K. p. 8, Wu, S., et al. Protenix advancing structure prediction through a comprehensive alphafold3 reproduction. p. 8
Costa, A. D. S., Mitnikov, I. p. 8, Pellegrini, F. p. 8, Daigavane, A. p. 8, Geiger, M. p. 8, Cao, Z., Kreis, K. p. 8, Smidt, T. p. 8, Kucukbenli, E. p. 8, and Jacobson, J. p. 8 Equijump: Protein dynamics simulation via so (3)-equivariant stochastic interpolants. p. 8
Del Alamo, D. p. 1, Sala, D., Mchaourab, H. p. 8 S., and Meiler, J. p. 8 Sampling alternative conformational states of transporters and receptors with alphafold2. p. 8
Feng, B., Zhang, J. p. 8 p. 8, Zhang, X. p. 8, Liu, Z., and Li, Y. Biomd: All-atom generative model for biomolecular dynamics simulation. p. 8
Gu, A., Goel, K., and Re, C. Efficiently modeling long sequences with structured state spaces. p. 8
Gu, A., Johnson, I. p. 8, Goel, K., Saab, K., Dao, T., Rudra, A. p. 8, and Re, C. Combining recurrent convolutional and continuous-time models with linear state space layers. p. 8
Henzler-Wildman, K. p. 4 and Kern, D. p. 4 Dynamic personalities of proteins. p. 8
Jing, B., Berger, B. p. 8, and Jaakkola, T. p. 8 AlphaFold meets flow matching for generating protein ensembles. p. 8
Jing, B., Stark, H. p. 8, Jaakkola, T. p. 8, and Berger, B. p. 8 Generative modeling of molecular dynamics trajectories. p. 1
Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M. p. 8, Ronneberger, O., Tunyasuvunakool, K. p. 8, Bates, R. p. 8, Zˇıdek, A., Potapenko, A. p. 8, et al. Highly accurate protein structure prediction with alphafold. p. 8
Karras, T., Aittala, M. p. 8, Aila, T., and Laine, S. p. 8 Elucidating the design space of diffusion-based generative models. p. 8
Klein, L., Foong, A. p. 8 p. 8, Fjelde, T. p. 8, Mlodozeniec, B. p. 8, Brockschmidt, M. p. 8, Nowozin, S. p. 8, Noe, F., and Tomioka R. p. 8 Timewarp: Transferable acceleration of molecular dynamics by learning time-coarsened dynamics. p. 8
Lewis, S., Hempel, T. p. 8, Jimenez-Luna, J. p. 8, Gastegger, M. p. 8, Xie, Y., Foong, A. Y., Satorras, V. p. 8 G., Abdin, O. p. 8, Veeling, B. p. 8 S., Zaporozhetsi I., et al. Scalable emulation of protein equilibrium ensembles with generative deep learning. p. 8
Lu, J., Zhong, B. p. 8, Zhang, Z. p. 8 p. 8, and Tang, J. p. 0 p. 0 Str2str: A score-based framework for zero-shot protein conformation sampling. p. 8
Lu, J., Chen, X., Lu, S. Z., Lozano, A. p. 8, Chenthamarakshan, V. p. 8, Das, P., and Tang, J. Aligning protein conformation ensemble generation with physical feedback. p. 8
Lu, W., Zhang, J., Huang, W. p. 9, Zhang, Z., Jia, X., Wang, Z., Shi, L., Li, C., Wolynes P. p. 9 G., and Zheng S. p. 1 Dynamicbind: predicting ligand-specific protein-ligand complex structure with a deep equivariant generative model. p. 9
McGibbon, R. p. 9 T., Beauchamp, K. p. 9 A., Harrigan, M. p. 9 P., Klein, C. p. 1, Swails, J. p. 9 M., Hernandez, C. p. 9 X., Schwantes, C. p. 9 R., Wang L.-P., Lane T. J., and Pande V. p. 9 S. Mdtraj: a modern open library for the analysis of molecular dynamics trajectories. p. 9
Mirarchi, A. p. 0, Giorgino, T. p. 9, and De Fabritiis G. p. 9 mdcath: A large-scale md dataset for data-driven computational biophysics. p. 9
Noe, F., Olsson, S. p. 9, K´ohler J. p. 9, and Wu H. Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning. p. 9
Passaro, S. p. 5, Corso, G. p. 9, Wohlwend J. p. 9, Reveiz M. p. 9, Thaler S. p. 9, Somnath V. p. 9 R., Getz N., Portnoi T. p. 9, Roy J., Stark H. et al. Boltz-2: Towards accurate and efficient binding affinity prediction. p. 9
Salomon-Ferrer, R. p. 9, Gotz A. W., Poole D., Le Grand S. p. 9, and Walker R. p. 9 C. Routine microsecond molecular dynamics simulations with amber on gpus. p. 9
Schreiner, M. p. 1, Winther O. p. 9, and Olsson S. p. 9 Implicit transfer operator learning: Multiple time-resolution models for molecular dynamics. p. 9
Shen, Y., Wang L., Yuan H., Wang Y., Yang B., and Gu Q. Simultaneous modeling of protein conformation and dynamics via autoregression. p. 9
Siebenmorgen, T. p. 0, Menezes F. p. 9, Benassou S. p. 9, Merdivan E. p. 9, Didi K., Mourao A. p. 9 S. D., Kitel R., Li˜o P., Kesselheim S. p. 9, Piraud M. p. 9 et al. Misato: machine learning dataset of protein–ligand complexes for structure-based drug discovery. p. 9
Vander Meersche, Y. p. 5, Cretin G. p. 9, Gheeraert A. p. 9, Gelly J.-C., and Galochkina T. p. 9 Atlas: protein flexibility description from atomistic molecular dynamics simulations. p. 9
Wang, Y., Wang L., Shen Y., Wang Y., Yuan H., Wu Y., and Gu Q. Protein conformation generation via force-guided se (3) diffusion models. p. 9
Wayment-Steele, H. p. 1 K., Ojoawo A. p. 9, Otten R., Apitz J. M., Pitsawong W. p. 9, Homberger M. p. 9, Ovchinnikov S. p. 9, Colwell L. p. 9, and Kern D. p. 4 Predicting multiple conformations via sequence clustering and alphafold2. p. 9
Williams, R. p. 4 J. and Zipser D. p. 4 A learning algorithm for continually running fully recurrent neural networks. p. 9
Wu, F. and Li S. Z. Diffmd: a geometric diffusion model for molecular dynamics simulations. p. 9
Xu, K., Wang J., Liu M., Zhou K., Lin S., Li W., Shi L., Zhou P., Liu H., and Yao X. p. 9 Efficient generation of protein and protein–protein complex dynamics via se(3)-parameterized diffusion models. p. 9
Xu, Y., Wang D., Zhou Z., Yu T., and Chen M. p. 5 TEMPO: Temporal multi-scale autoregressive generation of protein conformational ensembles. p. 9
Zheng, S., He J., Liu C., Shi Y., Lu Z., Feng W., Ju F., Wang J., Zhu J., Min Y. Predicting equilibrium distributions for molecular systems with deep learning. p. 9
Impact Statement
ATMOS has significant broader impacts for computational biology and drug discovery by providing a general-purpose framework for modeling biomolecular dynamics that could accelerate therapeutic discovery and enhance biological understanding. Responsible deployment and transparency are essential to maximize benefits and mitigate risks.
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
State Space Models for long-range temporal dependencies are used to efficiently capture long-horizon trajectory evolution with linear computational complexity in terms of trajectory length, addressing the bottleneck of existing autoregressive methods. This allows ATMOS to generate trajectories of arbitrary length T with linear scaling compared to quadratic scaling in traditional attention-based approaches.
-
The framework can model dynamics at the fully atomic level by operating directly on atomic coordinates during both encoding and decoding phases, rather than coarsegraining representations like C alpha atoms or backbone frames. This enables ATMOS to generate high-quality, physically realistic protein dynamics that capture both the equilibrium distribution and kinetic pathways of protein dynamics.
-
The bi-level sampling protocol decouples conformation changes from local fluctuations by using a keyframe generator for distinct anchor conformations and an interpolator conditioned on adjacent keyframes to fill in the dynamics between anchors. This hierarchical strategy enables the parallel generation of fine-grained frames between anchor frames and significantly enhances sampling efficiency while maintaining trajectory consistency.
-
The model can predict flexibility, distributional accuracy, and ensemble observables across large-scale MD benchmarks like mdCATH and MISATO. For instance, ATMOS achieves
near-perfect correlation with the reference MD trajectories in pairwise RMSD (Pearson r = 0.92)
on protein dynamics for the 1s79a00 protein. -
The system can model coupled protein-ligand dynamics by generating physically realistic ligand motions while closely matching the reference distribution, achieving
the best scores in all four metrics
for ligand dynamics on the MISATO dataset.
Sources
- EquiJump: Protein Dynamics Simulation via SO(3)-Equivariant Stochastic Interpolants
- BioMD: All-atom Generative Model for Biomolecular Dynamics Simulation
- Efficiently Modeling Long Sequences with Structured State Spaces
- Protein Conformation Generation via Force-Guided SE(3) Diffusion Models
Related papers
- Agent-guided epitope excision enables economical de novo binder design
- CupOFMoCA: Coupled Objective-Guided Discrete Flows for Molecular Conjugate Assembly
- BioBlobs: Unsupervised Discovery of Functional Substructures
- CPTCs Drive Somatic-Visceral Communication via the Wnt Axis in Somatic Mechanotherapy: A Single-Cell Deep Learning Study
- Speak to a Protein: An Interactive Multimodal Co-Scientist
- Discrete Backbone Elasticity and Curvature-Dependent Torsion Statistics in Proteins