bio.rodeo
ModelsOrganizationsProvidersLeaderboardAboutSign in
bio.rodeo

The authoritative source for evaluating biological foundation models. No hype, just honest analysis.

Categories
  • DNA & Gene
  • RNA
  • Protein
  • Small molecule
  • Single-cell
  • Spatial omics
  • Pathology
  • Imaging
  • Metabolomics
  • Biosignals
  • Language model
bio.rodeoModelsOrganizationsProvidersLeaderboardAboutFAQSubmit a modelContact
© 2026 Pulsatance. All rights reserved. ~
Built by Pulsatance
models / protein / dynamicgt
Protein
Université de LorraineCNRSInriaLORIAUniversité Grenoble AlpesReleased March 2025

DynamicGT

Protein-protein binding interface prediction from conformational ensembles, resolving interfaces in flexible and intrinsically disordered regions.

The short version

  • —Maps binding interfaces on disordered hub proteins and transient complexes
  • —Takes a conformational ensemble as input, from MD, NMR, or generated structures
  • —Cooperative GNNs let each atom decide whether to broadcast or listen before message passing
  • —A geodesic-distance loss penalizes false positives by how far they sit from the interface
  • —Works on bound and unbound inputs alike, so no complexed structure is needed
34Openness3Citations
16GitHub stars

Where to run it

No providers recorded yet. Browse all providers

Binding site predictors are trained almost exclusively on crystal structures, which fix a protein as one rigid arrangement of atoms. That assumption breaks down for intrinsically disordered regions, which have no stable fold in isolation and often acquire structure only on contact with a partner — folding upon binding, or mutual folding, where neither partner is ordered until they meet. A single coordinate set cannot represent that, and predictors built on one perform accordingly.

DynamicGT takes a conformational ensemble as its input instead. Developed at Université de Lorraine, CNRS, Inria, and LORIA in Nancy with a collaborator at Université Grenoble Alpes, it reads a set of structures — molecular dynamics trajectories, NMR models, or conformations generated by AlphaFlow — and derives per-atom and per-edge features describing how the protein moves, not merely where its atoms sit. Those features feed a cooperative graph neural network coupled to a geometric transformer. Posted as a preprint in March 2025, the work was published in Cell Systems.

Its distinguishing idea is that message passing should be regulated rather than uniform. Two preliminary GNN layers assign each atom an action — broadcasting or listening — constraining how information flows between the protein's surface and its core. The model learns that surface residues tend to listen while core residues broadcast, and that communication concentrates near binding sites, making the routing legible rather than opaque.

#Key Features

  • Ensemble input: Prediction runs over multiple conformations of one chain, with solvent accessibility averaged across them so no single snapshot dominates the surface/core assignment.
  • Time-free dynamic features: Root mean square fluctuation and directional entropy describe atom-level flexibility; communication propensity, a motion vector, and a motion scalar capture correlated movement between atom pairs — all rotationally equivariant, with no time-series modeling.
  • Cooperative message passing: Learned broadcast and listen actions mask communication between surface and core atoms, letting the network choose which parts of the protein talk to each other.
  • Spatially aware loss: GeoLoss extends focal loss with a distance decay — Euclidean for core residues, geodesic for surface ones — so a false positive on an interface rim is penalized less than one across the protein.
  • Bound and unbound generalization: Training on complexed and free conformations alike removes the reliance on bound structures that leaks interface information into crystal-derived predictions.

#Technical Details

The training set began with 1,115 non-redundant protein chains in bound form, simulated in triplicate to yield roughly 1 ms of aggregate molecular dynamics, then combined with the ATLAS and MDDB repositories to reach ensembles for more than 3,000 chains. NMR ensembles and AlphaFlow conformations extended coverage further; 92% of generated entries scored above 70 pLDDT despite containing disordered regions. Two variants were trained: a dynamic-aware model using simulation data only, and a heterogeneity-aware model adding NMR and AlphaFlow ensembles so inference needs no new simulations.

Across four disorder benchmarks with cluster-based splitting and similarity filtering against the training set, median ROC-AUC reached 0.88 on MFIB (mutual folding), 0.84 on Test42, 0.79 on IDRBind, and 0.77 on FuzDB; by secondary structure, 0.834 on structured residues against 0.784 on flexible ones. On conventional benchmarks it holds its own against static specialists: 0.868 median ROC-AUC on the 53-complex Testing Transient set for the MD-only model, 0.851 for the generalized model run on AlphaFlow ensembles built from sequence alone, and the highest PR-AUC of 0.479 on Test60, ahead of EquiPPIS, GraphPPIS, and MaSIF-site. Ablating dynamic features costs 2.6% and 2.4% ROC-AUC on those two benchmarks. The gain is larger for proteins with low disorder content than for highly disordered ones, where the conformational space is harder to sample and the surface/core distinction largely dissolves.

#Applications

The model addresses interaction problems where a static structure is the wrong abstraction: mapping interfaces on disordered hub proteins, characterizing transient complexes, and locating binding regions on unbound structures with no solved complex. Because the heterogeneity-aware variant runs on AlphaFlow ensembles generated from sequence, no molecular dynamics run is needed, which makes it practical outside groups with simulation infrastructure. Per-residue probabilities map into the B-factor column for visualization in standard viewers, and the checkpoint applies to new proteins without retraining.

#Impact

DynamicGT is an argument that conformational dynamics is usable signal rather than experimental noise, and that a model given ensembles can match static predictors on their own benchmarks while extending to disordered regions those predictors never handled. The ablations do the load-bearing work: performance drops when dynamic features are removed, tying the gain to the dynamics rather than to architecture or training scale, and the authors report reaching it on far less training data than leading static methods. Two caveats matter for adopters: the model is released under CC BY-NC-SA 4.0, a non-commercial license that rules out industry pipelines without separate permission, and the checkpoint is distributed through Google Drive rather than a versioned archive, though the training data is on Zenodo.

At a glance

Released
March 2025
Category
Protein
License
NOASSERTION
Organizations
Université de Lorraine / CNRS / Inria / LORIA / Université Grenoble Alpes

Related models

  • ESMDynamic

    University of Illinois Urbana-Champaign

  • DynaProt

    MIT CSAIL

  • GeoGAT-site

    Harbin Institute of Technology / Heilongjiang University / Northeast Forestry University / Harbin Medical University / Heilongjiang University of Chinese Medicine / King Abdullah University of Science and Technology

  • EBM-DDG

    University of Virginia

  • DynamicsPLM

    Technion – Israel Institute of Technology / Microsoft

Links

GitHub RepositoryResearch PaperResearch PaperDataset

Tags

binding_site_predictiongraph_neural_networkintrinsically_disordered_proteinsmolecular_dynamicsprotein_protein_interaction_predictiontransformer

Something wrong?

Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.