What is HyperMPNN?

Many thermostability design projects start with a solved or predicted protein structure and a specific question: which amino acid sequence is more likely to keep this backbone folded at high temperature? HyperMPNN answers that inverse-folding question by using ProteinMPNN weights retrained on proteins from hyperthermophilic organisms.

The native HyperMPNN repository is not a separate inference engine. It supplies retrained model weights that run through the original ProteinMPNN code path with --path_to_model_weights and --model_name. ProteinIQ follows that standard behavior and supports the HyperMPNN checkpoints as protein-only sequence design models.

HyperMPNN is most useful when thermal resilience is the design objective and the backbone is already chosen. Typical inputs include enzyme structures for high-temperature biocatalysis, vaccine nanoparticle components that need better storage stability, and scaffold proteins where functionally important residues can be fixed while the rest of the sequence is redesigned.

How to use HyperMPNN online

HyperMPNN runs online from one protein backbone, supplied as a PDB file or RCSB PDB ID, with a chosen number of thermostability-biased sequence variants. ProteinIQ returns designed sequences, mutation lists, native ProteinMPNN scores, sequence recovery, a FASTA file, and optional score or probability arrays for downstream analysis.

Inputs

The input structure determines what HyperMPNN can preserve. Fixed catalytic residues, metal-binding residues, disulfide cysteines, interface hot spots, and experimentally required mutations should be constrained before running broad redesigns.

Settings

Core settings

Design controls

Results

HyperMPNN returns both a human-readable table and native files. The FASTA output is the primary scientific output because it preserves ProteinMPNN headers and metadata.

For multichain designs, the Results view separates each sampled design by chain while the Data table keeps one row per complete sampled design. The primary FASTA file preserves ProteinMPNN's combined multichain records and headers; workflows use chain-specific records from the designed-only FASTA so the input sequence is not treated as another design.

Jobs are limited by total padded structure residues multiplied by requested sequences: up to 250,000 residue-samples normally or 100,000 when probability arrays are requested. Fixed context chains and PDB numbering gaps count because ProteinMPNN featurizes the complete structure. This keeps large sequence and probability outputs within the execution and result-transfer envelope.

Interpreting HyperMPNN designs

HyperMPNN scores rank sequences by model likelihood, not by measured melting temperature. A low Score or Global score indicates that the sequence is compatible with the input backbone under the HyperMPNN model, but experimental thermostability still depends on expression, folding kinetics, oligomerization, cofactors, and the validity of the backbone.

Global score is the better broad ranking field for complete designs. Score is more useful when the redesign is local and fixed residues dominate the structure. When both scores are similar across candidates, mutation patterns usually matter more than small score differences.

Sequence recovery has a different meaning in thermostability redesign than in native-sequence recovery benchmarks. A low value is not automatically bad. HyperMPNN is expected to move sequences toward hyperthermophile-like composition, so useful designs may have substantial substitutions on the surface, in the core, or around flexible regions.

Practical ranking workflow

• Initial sampling: 8 to 10 designs at Sampling temperature 0.1 or 0.15 gives enough diversity for a first pass.
• Functional filter: Designs that mutate protected functional residues should be discarded unless those residues were intentionally left redesignable.
• Model ranking: Strong candidates combine low Global score, reasonable mutation load, and no obvious disruption of active-site chemistry or interface contacts.
• Enzyme designs: Catalytic residues usually belong in Fixed positions, followed by stability or function-specific evaluation of the redesigned structures.
• Mutation-focused studies: ThermoMPNN is more direct for single-site or saturation mutagenesis stability questions because it predicts mutation-level thermostability changes rather than full sequences.

How HyperMPNN works

HyperMPNN fine-tunes the ProteinMPNN inverse-folding architecture on structures from hyperthermophilic organisms. The published workflow started from 96,738 sequences, clustered them at 50 percent identity, filtered AlphaFold2-predicted structures by pLDDT above 70, and trained on 29,042 protein structures from organisms adapted to extreme heat.

During inference, HyperMPNN receives the same information as ProteinMPNN: backbone geometry, chain selection, fixed-position masks, tied positions, amino acid omissions, and optional amino acid biases. The difference is the learned sequence distribution. HyperMPNN shifts sampling toward amino acid patterns observed in thermostable proteins rather than the broader Protein Data Bank distribution.

The HyperMPNN paper reports several composition-level trends in hyperthermophilic proteins and HyperMPNN designs:

These shifts are consistent with common thermostability mechanisms, including tighter hydrophobic packing and stronger electrostatic networks. HyperMPNN should not be read as a salt-bridge maximizer. The reported median salt bridge count in native hyperthermophilic proteins was close to the mesophilic comparison set, while HyperMPNN designs recovered hyperthermophile-like salt bridge counts better than standard ProteinMPNN in the authors' analysis.

The experimental validation case in the HyperMPNN preprint used the I53-50B pentamer from an icosahedral nanoparticle system. Its published command sampled 200 sequences at temperature 0.001; the Published validation sampling preset reproduces that sampling envelope. Designs selected from the workflow retained stability at 95 degrees C, while the parent sequence had a melting temperature of 65 degrees C. That result is a strong proof of concept, not a guarantee that every redesigned backbone will gain 30 degrees C of thermal tolerance.

When to use HyperMPNN vs alternatives

For structures that come from sequence rather than experiment, backbone prediction or refinement with AlphaFold2 should come before HyperMPNN. Low-confidence or flexible regions need cautious interpretation because HyperMPNN can only design against the geometry it receives.

Design caveats that affect experiments

• Expression can drop: Thermostable sequence patterns may reduce soluble expression in mesophilic hosts such as E. coli. The HyperMPNN validation work observed that expression host and purification strategy matter.
• Function is not guaranteed: Active sites, binding residues, post-translational motifs, and interface residues should be fixed when function must be preserved.
• Backbone errors propagate: A predicted or low-resolution structure can produce plausible sequences for the wrong geometry. Confidence in the backbone should be checked before interpreting model scores.
• No ligand awareness: Bound ligands, metals, and cofactors are not part of HyperMPNN inference. Ligand-sensitive redesign should use LigandMPNN or fix residues that coordinate the non-protein atoms.
• Composition bias is global: Amino acid biases affect sampling across designed positions. They are useful for broad constraints, but position-specific design intent is better expressed with fixed or redesigned residue lists.