Use DynamicMPNN Online
Commercially Available DynamicMPNN No-Code Web Server
DynamicMPNN: Breakthrough Inverse Folding for Conformational Ensembles and Bioswitches
DynamicMPNN is a pioneering geometric deep learning model specifically engineered for multi-state protein design. By training directly on joint conditional distributions across conformational ensembles, DynamicMPNN generates sequences uniquely optimized to satisfy multiple structural constraints simultaneously. It bypasses the limitations of single-state aggregation to efficiently engineer synthetic bioswitches, allosteric regulators, and advanced molecular machines.
Explicit Multi-State Learning: The first dedicated inverse folding model explicitly trained to capture the sequence constraints required for multiple functional conformations.
Unprecedented Self-Consistency: Achieves up to a 25% reduction in decoy-normalized RMSD and a 12% improvement in sequence recovery over ProteinMPNN Multi-State Design.
Environment-Aware Encoding: Conditions conformational changes on protein binders and oligomeric states, paving the way for controllable, stimulus-responsive switch designs.
Estimated run time: 3–5 minutes per design batch.
How DynamicMPNN Works
Traditional structural biology has long operated under the "one sequence, one structure" paradigm. When applied to multi-state engineering, previous deep learning strategies relied on post-hoc aggregation—averaging independent single-state logits during decoding. This heavily biases results toward one dominant conformation, leading to dismal experimental success rates (as low as 0.05% in silico).
DynamicMPNN overcomes this gap through a fundamentally distinct pipeline:
Joint Conditional Distribution (p(Y|X_1, ..., X_m)): Instead of averaging single-state models, DynamicMPNN auto-regressively factors a single sequence across a complete ensemble of distinct target backbones.
Conformation Order-Invariant Architecture: Built upon the SE(3)-equivariant Geometric Vector Perceptron (GVP) network. Independent encoder channels process each independent functional state and its surrounding chemical environment. Node features are pooled across states using a conformation order-invariant Deep Set mechanism.
Massive Multi-Conformational Training: Trained on a custom, ML-ready dataset of 46,033 sequence-aligned structural clusters derived from the PDB and CoDNaS databases, capturing critical conformational diversity across 75% of all CATH superfamilies.
Unmatched Performance on Challenging Benchmarks
DynamicMPNN was evaluated on a rigorous benchmark containing 96 of the largest documented 2-state conformational changes in the proteome, spanning metamorphic proteins, transporters, and hinges. Validated via a template-based AlphaFold 3 evaluation pipeline, DynamicMPNN dramatically outperforms previous paradigms:
Sequence Recovery: Reaches up to 42.7% sequence recovery compared to 38.0% for traditional ProteinMPNN Multi-State Design (MSD).
Refoldability Metrics: In comparison to single-state models, DynamicMPNN produces sequences that reliably fold into their intended alternative states rather than collapsing into a single state, achieving an average AF3 pLDDT of 79.23 (vs. 63.10 for single-state baselines).
Decoy-Normalized Control: Demonstrates a 25% superior structural specificity over alternative tools when measured against structurally dissimilar negative-control decoys.
What is Tamarind Bio?
Tamarind Bio is a pioneering no-code bioinformatics platform built to democratize access to powerful computational tools for life scientists and researchers. Recognizing that cutting-edge deep learning models in structural biology are often difficult to deploy, optimize, and scale, Tamarind provides an intuitive, web-based environment. The platform completely abstracts away the complexities of high-performance computing, GPU orchestration, command-line interfaces, and delicate software dependencies. By handling the technical heavy lifting, Tamarind Bio empowers structural biologists, chemists, and pharmaceutical researchers to focus exclusively on their scientific workflows and accelerate therapeutic discovery.
How to Use DynamicMPNN on Tamarind Bio
Tamarind Bio abstracts away complex GVP graph featurization, providing a seamless interface to design sequences for multiple states simultaneously. Follow this streamlined workflow:
Access the Platform: Log in to your secure account on the Tamarind Bio website.
Select DynamicMPNN: Choose the DynamicMPNN tool from the computational design suite.
Upload Conformation States: Upload the structural coordinate files (PDB or CIF format) corresponding to the alternative conformations of your target protein (e.g., open vs. closed transporter states, or divergent metamorphic folds).
Define Target Chains & Environments: Specify which chains are to be inverse folded, and include any structural files for binding partners or oligomeric complexes to properly condition the chemical environment.
Set Sampling Parameters: Select your preferred training variant (such as Combined Training for optimal sequence-structure balance) and specify the number of design variants you wish to generate.
Generate and Evaluate: Click Submit. Tamarind Bio manages the backend GPU resources, executes the autoregressive decoding, and returns your sequence candidates. Your outputs can be sent directly to our integrated AlphaFold 3 template pipeline to automatically compute structural self-consistency, pLDDT, and decoy-normalized RMSD scores.