Research map / Probabilistic and scientific modeling / Molecules and proteins

Atomic Trajectory Modeling with State Space Models for Biomolecular Dynamics

NeurIPS 2026 · Probabilistic and scientific modeling › Molecules and proteins · approach: Generative protein structure design · 1 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ATMOS, an all-atom SSM-style generative model that autoregressively generates biomolecular MD trajectories using a Pairformer-based state transition and diffusion decoder, evaluated on mdCATH and MISATO.

Problem

Generating atom-level biomolecular MD trajectories for long timescales is expensive with MD, and existing generative models either ignore temporal correlations (ensemble models) or cannot handle complex protein-ligand systems (trajectory models).

Approach

Treats trajectory generation as SSM-style recurrent latent propagation: a Pairformer-based state transition over coupled single/pair latents (initialized from Protenix) is driven by encoded atomic coordinates, and an EDM diffusion decoder produces each all-atom frame, with a bi-level keyframe/interpolator sampling scheme.

Key results

• On mdCATH, ATMOS achieves pairwise RMSD r=0.92 and global RMSF r=0.90 vs. best baselines 0.73 and 0.67.
• On mdCATH root mean W2 distance is 2.70 Å vs 3.40 Å for AlphaFlow-MD (second best) and t-RMSD error 1.79 Å vs 2.42 Å for TEMPO.
• On MISATO ligand-only evaluation, ATMOS reports best per-target RMSF r (0.746), W2 (0.670), clashes (0.030) and t-RMSD (0.480) among the compared methods.

Surprising findings

• Removing bi-level sampling or using deterministic ODE sampling substantially degrades ensemble quality on mdCATH.
• On full protein-ligand evaluation, Boltz-2 achieves better RMSF correlation (0.534 vs 0.523) and fewer clashes, though it may be affected by data leakage.

Limitations

• Boltz-2 comparison may be affected by data leakage; the MISATO test set is only 100 sampled systems.
• Four separate models are required for the two datasets and two temporal scales; unified model left for future work.
• Stated SSM framing is loosely realized: the transition is a Pairformer with per-step recurrence, and linear-complexity claims are not formally analyzed.

Experimental setup

Datasets: mdCATH MISATO PDB

Benchmarks: mdCATH MISATO

Metrics: Pearson r (pairwise RMSD, RMSF) root mean W2 distance PCA W2 distance PC-sim Jaccard (weak/transient contacts) t-RMSD error steric clashes inference time peak memory

Baselines: AlphaFlow-MD ESMFlow-MD MDGen TEMPO ConfRover GNNMD DenoisingLD NeuralMD-SDE VerletMD DynamicBind Boltz-2

Builds on: Protenix AlphaFold3 AlphaFlow TEMPO EDM Pairformer

Closest papers by meaning

• Generative Modeling of Molecular Dynamics Trajectories (NeurIPS 2024 · similarity 0.89)
• TEMPO: Temporal Multi-scale Autoregressive Generation of Protein Conformational Ensembles (NeurIPS 2025 · similarity 0.89)
• Action-Minimization Meets Generative Modeling: Efficient Transition Path Sampling with the Onsager-Machlup Functional (ICML 2025 · similarity 0.89)
• Simultaneous Modeling of Protein Conformation and Dynamics via Autoregression (NeurIPS 2025 · similarity 0.88)
• AlphaFold Meets Flow Matching for Generating Protein Ensembles (ICML 2024 · similarity 0.87)
• Integrating Protein Dynamics into Structure-Based Drug Design via Full-Atom Stochastic Flows (ICLR 2025 · similarity 0.87)