SeaMoon

Structure prediction largely solved the problem of getting one plausible conformation from a sequence. What proteins actually do, however, depends on how they deform — hinge closures, domain rotations, loop excursions. Most attempts to capture that repurpose AlphaFold 2 or a related structure predictor to sample alternative conformations, which means the answer is expressed as a set of discrete structures and the pipeline still needs 3D input or heavy sampling.

SeaMoon takes a different route: it asks whether a continuous, compact representation of protein motion can be read directly out of a sequence. Developed by Valentin Lombard, Dan Timsit, Sergei Grudinin, and Elodie Laine at Sorbonne University, CNRS, and Université Grenoble Alpes, it was posted as a preprint in September 2024 and published in Structure in 2025. The model is a lightweight 1D convolutional network that consumes protein language model embeddings and emits three sets of 3D displacement vectors, one per Cα atom, defined up to global translation, rotation, and scaling. Those vectors are interpretable as linear motions and are directly comparable to the principal components of experimental conformational ensembles.

Because the network is a thin head over a frozen embedding, it inherits the language model's generalisation: SeaMoon produces useful predictions for proteins with no detectable sequence similarity to anything it was trained on.

#Key Features

• Sequence-only input: No 3D structure is required at inference, which distinguishes SeaMoon from normal mode analysis and from structure-conditioned ensemble samplers.
• Continuous motion representation: Output is a small basis of displacement vector fields rather than a set of sampled conformations, so it can be composed, compared, or used to deform a structure along a trajectory.
• Interchangeable backbones: The same head trains over embeddings from ESM-2, ESM3, or ProstT5, and the paper shows that structure-aware language models transfer motion information better than sequence-only ones.
• Invariance-aware loss: Training compares predictions to ground truth under optimal rotation and scaling, and is invariant to sign flip, reflection, and permutation of the predicted components; disabling either invariance degrades performance by 6–15%.
• Openly released: Code and model weights are MIT-licensed with versioned Zenodo-archived releases, and the development and evaluation data are separately archived.

#Technical Details

The architecture is deliberately small: a linear layer over the pLM embedding (1,280 dimensions for ESM-2, 1,536 for ESM3, 1,024 for ProstT5), two hidden 1D convolutional layers with filter widths of 15 and 31, and three parallel linear output heads producing L×3 tensors. Being convolutional, it handles sequences of arbitrary length. An ablation showed a 7-layer transformer with roughly the same parameter count performing worse.

Training and evaluation used approximately 17,000 collections of experimental conformations extracted with the DANCE method from a non-redundant subset of the PDB at 80% sequence identity, with principal components serving as ground-truth motions; a data augmentation variant supplies five alternative reference conformations per collection. On a test set of 1,121 proteins, and scoring a prediction as successful when its normalised sum-of-squares error falls below 0.6, SeaMoon succeeds for 40% of proteins with ProstT5 embeddings, 39% with ESM3, and 31% with ESM-2, against 27% for normal mode analysis. On a redundancy-reduced test set the ProstT5 figure is 37%. SeaMoon recovers 75% of the hundred proteins that normal mode analysis handles best, while also capturing motions NMA cannot; on a 21-protein elastic network benchmark it matched NMA, capturing 12 targets at high accuracy including adenylate kinase's hinge motion and the GroEL chaperone.

#Applications

The model suits screening scenarios where a structure is unavailable or where running physics-based analysis over many sequences is impractical — surveying likely conformational changes across a protein family, generating deformation trajectories to seed docking or molecular dynamics, or flagging proteins whose motions are poorly explained by rigid-body geometry. An optional torque mode orients predictions relative to a supplied 3D structure by minimising the conformation's angular velocity, which makes the output usable for visualisation and for structure-based follow-up.

#Impact

SeaMoon is best read as a proof of concept with an honest headline number: sequence embeddings carry real information about functional motion, and a very small head is enough to extract it, but a correct prediction for 40% of test proteins is a starting point rather than a solved problem. Its most transferable finding is that language models exposed to structure during pretraining encode motion better than purely sequence-trained ones — an argument for structure-aware backbones in dynamics work generally. It sits alongside other sequence-based dynamics predictors such as ESMDynamic, and the authors note the head is cheap to retrain as new or updated language models appear.