DynamicsPLM

DynamicsPLM is a protein language model that conditions on conformational dynamics rather than on a single static structure, developed by Dan Kalifa and Kira Radinsky at the Technion — Israel Institute of Technology with Eric Horvitz at Microsoft. Standard structure-aware protein language models pair a sequence with one structural snapshot. Yet many proteins adopt multiple, functionally relevant shapes, and it is often the exposure of a binding site or the relay of a signal in a particular state that governs interaction, localization, and catalysis. Representing a protein by a single conformation can therefore miss the state-dependent features that matter most.

DynamicsPLM instead derives state-aware representations by conditioning on ensembles of computationally generated conformations. Building on a structure-aware protein language model backbone, it incorporates the variability across an ensemble so that the resulting embeddings reflect how a protein moves, not just how it looks in one frame. The authors frame this as a shift from static to dynamics-aware modeling, treating conformational variability as informative signal rather than noise.

Posted to bioRxiv in October 2025 and published in the journal Bioinformatics, DynamicsPLM is released with open-source code and demonstrates consistent gains across four functional benchmarks.

#Key Features

• Ensemble conditioning: The model conditions on ensembles of computationally generated conformations rather than a single structure, yielding representations that encode functionally relevant conformational states.
• State-aware embeddings: By capturing conformational variability, DynamicsPLM produces features tuned to state-dependent phenomena such as binding-site exposure, signal relay, and catalytic competence.
• Broad functional gains: It improves predictive performance across protein-protein interaction, subcellular localization, enzyme classification, and metal-ion binding.
• Strong on multi-state proteins: On a curated test set enriched for proteins with multiple conformational states, its advantage over the strongest baseline widens to eleven accuracy points, versus a four-point gain on a widely used protein-protein interaction benchmark.
• Modular conformation source: Conformational ensembles are supplied by an external generator, allowing the dynamics input to be produced by different structure-ensemble methods.

#Technical Details

DynamicsPLM is built on the structure-aware SaProt-650M protein language model, whose structural vocabulary lets sequence and local structure be encoded jointly. Where a conventional pipeline would feed one predicted structure per protein, DynamicsPLM feeds a set of conformations sampled to represent the protein's accessible states, and learns representations conditioned on that ensemble. The conformational ensembles are generated by an external emulator of protein structure dynamics, and the framework can substitute alternative ensemble generators. Evaluated across four tasks — protein-protein interaction, subcellular localization, enzyme classification, and metal-ion binding — the model outperforms strong single-structure baselines, achieving a four-point accuracy improvement on a widely used protein-protein interaction benchmark and an eleven-point improvement on a curated set enriched for proteins known to adopt multiple conformational states. Code and instructions for assembling the pretraining and downstream datasets are provided in the DynamicsPLM GitHub repository under the MIT license.

#Applications

DynamicsPLM is useful for researchers predicting protein function and interactions where conformational state is mechanistically important — for example distinguishing interaction partners that depend on an exposed binding surface, classifying enzymes whose activity hinges on catalytic-state geometry, or annotating subcellular localization and metal-ion binding. Because it consumes conformational ensembles from an external generator, it can be paired with a range of structure-dynamics emulators, letting groups trade off ensemble quality against compute when producing state-aware embeddings for downstream predictors.

#Impact

DynamicsPLM provides evidence that elevating conformational variability to a first-class input, rather than collapsing each protein to a single snapshot, improves representation quality for functional prediction — most markedly for proteins that genuinely populate multiple states. The larger eleven-point margin on multi-state proteins directly supports the motivation that dynamics-aware modeling captures signal that static models miss. Peer-reviewed publication in Bioinformatics and an open-source release lower the barrier to reuse. Performance depends on the fidelity of the computationally generated conformational ensembles, so the quality of the upstream structure-dynamics generator is an important consideration when applying the model.