ATMOS

ATMOS ("Atomic Trajectory MOdeling with SSMs") is a pretrained generative foundation model for biomolecular dynamics, developed by researchers at Mila – Québec AI Institute in the group of Jian Tang and released as a preprint in March 2026. The model addresses a central bottleneck in computational structural biology: while molecular dynamics (MD) simulations provide a rigorous, physics-based account of how biomolecules move, they remain prohibitively expensive at the long timescales relevant to biological function.

ATMOS reframes trajectory generation as a sequence-modeling problem. Rather than predicting a single static structure or an unordered ensemble of conformations, it generates ordered, atom-level trajectories that capture the temporal evolution of a system over time. This sets it apart from earlier deep generative approaches, which typically either ignore temporal relationships between conformations or are restricted to monomeric proteins.

Crucially, ATMOS models both protein monomers and protein-ligand complexes within a single framework, making it directly relevant to drug-discovery settings where the dynamics of a bound ligand matter as much as those of the protein itself. The authors report state-of-the-art results on established MD-trajectory benchmarks.

#Key Features

• Trajectory-level generation: Produces ordered, atom-resolution conformational trajectories rather than single structures or static ensembles, preserving temporal correlations between frames.
• State space backbone: Casts dynamics as sequence modeling, using a state space model (SSM) formulation to capture long-range temporal dependencies across many trajectory frames.
• Pairformer state transitions: Adapts the AlphaFold3-style Pairformer module for state transitions, with bidirectional flow between single and pair representations.
• Diffusion decoder: Decodes each trajectory frame with an EDM-style diffusion process initialized from Protenix weights, generating coordinates autoregressively.
• Protein and protein-ligand support: Handles both monomeric proteins and protein-ligand complexes in one model, extending dynamics prediction to bound small-molecule systems.

#Technical Details

ATMOS combines a Pairformer-based state transition mechanism with a diffusion-based decoder. The Pairformer uses 4 blocks with a single-representation dimension of 384 and a pair representation of 128, adapted from AlphaFold3's architecture. The diffusion decoder is an EDM-style process initialized from Protenix weights, using 50 diffusion steps with a noise scaling schedule of γ₀ = 0.8 and step scaling η = 1.5. Training draws on crystal structures from the PDB together with large-scale MD trajectory datasets, namely mdCATH and MISATO. On mdCATH, where the model generates 400-frame trajectories at 1 ns intervals, ATMOS reports a pairwise-RMSD correlation of 0.92, a global RMSF correlation of 0.90, and a root-mean 2-Wasserstein distance of 1.89 Å (versus 2.70 Å for the next-best method). On the protein-ligand MISATO benchmark it reports a ligand per-target RMSF correlation of 0.746 and a low steric-clash rate of 0.030. A total parameter count is not stated in the preprint.

#Applications

ATMOS targets researchers who need conformational dynamics but cannot afford long classical MD simulations. Potential use cases include sampling functionally relevant conformational states, estimating per-residue and per-ligand flexibility (RMSF), and rapidly screening how bound ligands behave within a binding pocket. The dual support for monomers and protein-ligand complexes makes it especially useful in early-stage drug discovery, where understanding ligand mobility and induced-fit effects can inform candidate prioritization without committing to costly simulation campaigns.

#Impact

As a foundation model for biomolecular dynamics, ATMOS extends the generative-modeling wave that reshaped static structure prediction into the temporal domain, joining a growing body of work aimed at learning the motion of proteins rather than only their folded shapes. Its explicit handling of protein-ligand systems is notable, since most prior conformation generators focused on isolated proteins. As a recent preprint, its broader adoption and independent validation remain to be established, and at the time of writing the authors have not released code or model weights, which currently limits reproducibility and downstream use.