DynamicMPNN

Protein design has largely been built on the assumption that one sequence maps to one structure, yet many essential biological processes — enzyme catalysis, membrane transport, allosteric signalling — depend on proteins that interconvert between distinct conformational states. Designing one sequence that satisfies two folds has been handled by aggregating single-state models: ProteinMPNN Multi-state Design averages the decoding logits of two independently encoded backbones. Because such averaging rewards logits biased toward one state, success rates lag far behind single-state design; ProteinGenerator reported an in-silico multi-state success rate of 0.05% against 2–10% for its single-state objectives.

DynamicMPNN is trained directly on conformational ensembles instead. It encodes each state of a protein, along with its binding partners, into a shared latent space, pools those embeddings across states, and autoregressively decodes one sequence conditioned on the whole ensemble. It was developed by researchers at the University of Cambridge, Leipzig University, and LMU Munich, first posted as a preprint in July 2025 and published at ICLR 2026.

The authors also contribute the two ingredients the problem was missing: a machine-learning-ready multi-conformational dataset built from sequence redundancy in the PDB, and a refoldability benchmark using template-conditioned AlphaFold 3 predictions with decoy normalization, which tests whether a design encodes both target states or just one dominant fold.

#Key Features

• Joint multi-state conditioning: The model learns the joint conditional distribution over sequences given all supplied conformations, so each residue is chosen against the full ensemble rather than an average of per-state predictions.
• Order-invariant ensemble pooling: Deep Set pooling aggregates per-conformation embeddings without adding parameters or imposing an ordering on states; an optional Deep Symmetric Set variant adds cross-attention between encoder channels.
• Alignment-aware featurization: Structure pairs are pooled over pairwise sequence alignments, so conformers with non-identical sequences, tags, or missing residues can be paired rather than discarded.
• Chemical environment encoding: Each state is encoded with its interaction partners, letting the model condition conformational change on binders and oligomeric context.
• Multi-state self-consistency benchmark: Designs are refolded with AlphaFold 3 using each target state as a template, then normalized against a dissimilar decoy template to separate genuine two-state specificity from template-induced bias.

#Technical Details

The architecture builds on gRNAde, a multi-state geometric GNN for RNA inverse folding, using SE(3)-equivariant Geometric Vector Perceptron layers over sparse k-nearest-neighbour graphs (k = 32), with 8 GVP layers each in encoder and decoder. Training data came from clustering PDB chains at ≥95% sequence similarity following the CoDNaS approach, yielding 46,033 conformational clusters covering 75% of CATH superfamilies — versus 21% from the roughly 12,000 available NMR ensembles — with the highest-RMSD pair retained from each. The 96-protein test benchmark draws the largest inter-state RMSDs from published sets of metamorphic, apo-holo, open-closed, and transporter proteins; 100 further pairs form the validation set. Removing training clusters within TM-score 0.4 or 30% sequence identity of held-out proteins left 44,243 conformer pairs, augmented with 27,394 single-state pairs.

On the 96-protein benchmark, the best variant recovers 42.7% of native sequence versus 38.0% for ProteinMPNN multi-state design, a 12% relative improvement. On the 61-protein subset where both methods could be refolded, it reaches a decoy-normalized RMSD of 0.206 against 0.275 for that baseline — a 25% reduction — with mean pLDDT of 79.2 versus 74.3. The same architecture trained only on single-state data collapses on this task (27.1% sequence recovery, 13.21 Å raw RMSD versus 4.30 Å), showing the gain comes from explicit multi-state training.

#Applications

The model targets design problems where function depends on switching rather than a single stable fold: bioswitches and biosensors that change conformation on binding, allosteric regulators, hinge proteins with binder-controlled state equilibria, transporters cycling between open and closed states, and synthetic molecular machines. Because each state is encoded with its binding partners, designers can condition a target conformational change on a specific interaction partner. The decoy-normalized refoldability protocol is reusable on its own as an in-silico filter for multi-state design campaigns, replacing single-conformation self-consistency checks that cannot tell a two-state design from a one-state one.

#Impact

DynamicMPNN is the first inverse folding model trained explicitly for multi-state design, and it supplies both a training set and an evaluation protocol later work can build on — an area where LigandMPNN and MoMPNN remain single-state. Its limitations are clearly bounded: it targets two-state proteins, not continuous conformational landscapes or disordered proteins; chemical-environment encoding covers proteins only, not nucleic acids or ligands; and results are entirely in silico, with no experimental characterization of designs reported. ProteinMPNN's training set also overlaps 91 of the 96 benchmark proteins, which the authors note makes the comparison conservative in its favour. Code is MIT-licensed with five pretrained checkpoints spanning multi-chain and single-chain settings at 2, 3, and 5 states; the processed training data is on Zenodo under CC BY 4.0.