RocketSHP

Static structure prediction has advanced dramatically, yet proteins are dynamic molecules whose biological function depends on conformational flexibility that a single predicted structure cannot capture. Measuring these motions traditionally requires molecular dynamics (MD) simulations, which are accurate but so computationally expensive that running them across an entire proteome is impractical. RocketSHP addresses this gap by predicting dynamic protein properties directly from an amino acid sequence or a static structure, returning results in seconds rather than the hours or days a comparable simulation demands.

RocketSHP was developed by Samuel Sledzieski and Sonya M. Hanson at the Center for Computational Biology of the Flatiron Institute. It is distributed as the open-source PLANET-MD software package, and the two names refer to the same model. Rather than predicting a full trajectory, RocketSHP outputs compact summaries of a protein's dynamic behavior, making it feasible to run dynamics-aware analysis at the scale of complete proteomes.

The model sits at the interface between static structural biology and dynamic functional understanding. By approximating quantities normally obtained from simulation, it lets researchers add a dynamics dimension to workflows that previously relied on fixed structures alone.

#Key Features

• Simultaneous multi-property prediction: A single forward pass yields per-residue root-mean-square fluctuations (RMSF), pairwise generalized correlation coefficients (GCC-LMI), and a structural heterogeneity profile (SHP) describing conformational variability.
• Sequence or structure input: Predictions can be made from sequence alone or from a static structure such as an AlphaFold or experimental model, fitting different data-availability scenarios.
• Frozen protein language model backbone: Prediction heads operate on top of frozen ESM3 embeddings, transferring learned protein representations to the dynamics task without retraining the base model.
• Proteome-scale speed: Inference in seconds per protein makes exhaustive analysis of the human proteome tractable.
• Multiple checkpoints: A full sequence-plus-structure model, a sequence-only variant, and a 1.5M-parameter lightweight version trade accuracy against speed and resource requirements.
• Open tooling: Released under the MIT license with a command-line interface and a Python API, plus built-in utilities for network and allosteric-pathway analysis.

#Technical Details

RocketSHP is a transformer encoder with specialized prediction heads mounted on frozen ESM3 embeddings; structure embeddings are supplied as an optional additional input for the full model. The heads are trained in a supervised fashion against labels derived from thousands of molecular dynamics trajectories spanning diverse protein families. The structural heterogeneity profile is formulated as a categorical distribution over structure tokens, drawing on structure-quantization methods to represent the range of conformational states a residue may adopt. Three checkpoints are provided: the full model combining sequence and structure, a sequence-only variant, and a 1.5M-parameter mini model for lightweight deployment. Pretrained weights are fetched automatically and applied to new sequences or structures without per-protein retraining.

#Applications

RocketSHP supports dynamics-aware structural analysis for researchers who need flexibility and correlation information but cannot run MD at scale. Its outputs feed variant effect prediction, allosteric and network analysis of communication pathways within a protein, and screening of conformational heterogeneity across large collections of structures. Because it accepts predicted structures as input, it integrates naturally with AlphaFold-style pipelines, and its speed enables the first comprehensive dynamics survey of the entire human proteome, benefiting structural biologists, protein engineers, and drug-discovery teams.

#Impact

RocketSHP extends the reach of structural biology from static snapshots toward the motions that underlie protein function, offering a fast approximation to expensive simulation that can be applied at proteome scale. As a preprint awaiting peer review, its claims are still being evaluated by the community, and it predicts summary dynamics statistics rather than full atomistic trajectories, so it complements rather than replaces molecular dynamics. Because its labels are derived from simulation, prediction quality is ultimately bounded by the accuracy of the underlying MD data. Distributed as free, open-source software, it lowers the barrier to incorporating dynamics into everyday protein analysis.