Use DynamicMPNN Online
Commercially Available DynamicMPNN No-Code Web Server
DynamicMPNN: Multi-State Inverse Folding for Designing Sequences Compatible with Multiple Protein Conformations
DynamicMPNN is an inverse folding model for multi-state protein design. Instead of designing a sequence for one backbone, it is trained jointly across conformational ensembles to produce sequences that are compatible with several conformations at once. It was introduced by Alex Abrudan, Sebastian Pujalte Ojeda, Chaitanya K. Joshi, Matthew Greenig, Felipe Engelberger, Alena Khmelinskaia, Jens Meiler, Michele Vendruscolo and Tuomas P. J. Knowles on arXiv.
Structural biology has long worked under a one sequence, one structure, one function paradigm, yet many biological processes rely on proteins that switch between states. Existing multi-state design methods typically aggregate single-state predictions after the fact, and the authors note they have had poor experimental success. Learning from conformational ensembles directly is the alternative that DynamicMPNN takes.
Attribute | Details |
|---|---|
Task | Multi-state inverse folding: backbone conformations in, amino acid sequence out |
Architecture | Built on the gRNAde framework with 8 SE(3)-equivariant GVP layers in the encoder and decoder, autoregressive decoding |
Training data | 46,033 clusters from PDB and CoDNaS covering 75% of CATH superfamilies, giving 44,243 conformer pairs, plus 27,394 single-state pairs |
Test set | 96 pairs of large two-state conformational changes |
Evaluation | Refolding with AlphaFold 3, 16 sequences sampled per pair |
How DynamicMPNN Designs Sequences for Several Conformational States
Encode each state: every input conformation is encoded independently by the geometric vector perceptron encoder.
Pool across states: encodings are combined with an order-invariant Deep Set, which adds no extra parameters. A DynamicMPNN + DSS variant adds cross-attention between conformation channels.
Decode a sequence: an autoregressive decoder models the sequence conditioned on all states.
Evaluate by refolding: designed sequences are refolded with AlphaFold 3 using each target state as a template, and scored by pLDDT, C-alpha RMSD and TM-score.
DynamicMPNN Results Versus ProteinMPNN on Multi-State Benchmarks
The abstract reports that DynamicMPNN outperforms ProteinMPNN by up to 25% on decoy-normalized RMSD and by 12% on sequence recovery. The paper tables give the detail for the comparison subset.
Metric | DynamicMPNN (combined training) | ProteinMPNN |
|---|---|---|
Decoy-normalized RMSD (lower is better) | 0.206 | 0.275 |
pLDDT | 79.23 | 74.29 |
RMSD (angstrom) | 4.30 | 5.84 |
Sequence recovery | 42.7% | 38.0% |
The authors also report that models trained only on single states did considerably worse on the multi-state benchmark, with pLDDT of 62.59 and RMSD of 13.48 angstrom, compared with 79.28 and 4.37 for combined training.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with DynamicMPNN on Tamarind Bio
Switchable proteins: design sequences for proteins that adopt more than one state, such as open and closed forms or apo and holo forms.
Transporters and metamorphic proteins: the benchmark covers metamorphic proteins, apo-holo pairs and transporters, the kinds of systems where one fold is not enough.
Better starting sequences: generate sequences that refold into several target backbones to seed experimental campaigns.
Ensemble-aware design: replace post-hoc aggregation of single-state designs with a model trained on conformational ensembles.
How to Use DynamicMPNN on Tamarind Bio
Log in and open the tool: sign in at tamarind.bio and open DynamicMPNN.
Provide backbone states: upload structures for each conformational state you want the sequence to support.
Add partner chains if needed: the model can condition on binding-partner chains, if exposed.
Set sampling options: choose the number of sequences to generate, if exposed; the authors sampled 16 per pair.
Run the job: launch the design run.
Download sequences: retrieve the designed amino acid sequences.
Validate: refold each sequence against every target state with a structure predictor, as the authors did with AlphaFold 3, before ordering.
Things to Keep in Mind
Results are computational only, and the authors state it is unclear whether the approach will work experimentally.
The model handles two-state switches; continuous conformational landscapes and intrinsically disordered proteins are left for future work.
Conditioning is currently limited to protein chains.
ProteinMPNN training data overlaps with 91 of the 96 benchmark proteins, so the comparison is not fully like for like.
Refolding with AlphaFold 3 uses templates, which can bias the evaluation.
Source: Abrudan et al., Multi-state Protein Design with DynamicMPNN, arXiv:2507.21938, https://arxiv.org/abs/2507.21938. No code repository was listed on the fetched pages.