How to Use MIF-ST

Overview. MIF-ST, short for Masked Inverse Folding with Sequence Transfer, is a structure-conditioned protein language model for inverse folding. The model reconstructs amino-acid sequence from backbone geometry, but extends structure-only inverse folding by transferring information from a pretrained sequence language model so it can benefit from much larger unlabeled protein sequence collections.

Use Cases. This makes MIF-ST useful for fold-preserving protein redesign, sequence exploration around a known scaffold, and structure-aware ranking of variants or homolog-style redesigns. It is especially relevant when researchers care about whether a sequence remains compatible with a specific three-dimensional backbone rather than only with a protein family in sequence space.

Workflow on Neurosnap. Researchers upload an Input Structure, choose how many sequences to sample, and then use masking and sampling controls to decide whether the run should stay close to the starting sequence or explore more broadly. The results page returns ranked sequences together with sequence-recovery and likelihood-based statistics for triage.

How It Works. The paper formulates the task as masked inverse folding on a protein graph. A structured graph neural network sees residue geometry from the backbone and predicts masked amino acids, which lets the model learn local and long-range structure-sequence relationships rather than only sequence regularities.

Methodological Innovation. The key advance is sequence transfer: outputs from a pretrained sequence-only masked language model are injected into the inverse-folding model. In the original study, that transfer improved pretraining perplexity and strengthened downstream protein representation learning, showing that structural pretraining and large-scale sequence pretraining are complementary rather than competing signals.

Workflow. On Neurosnap, Mask Probability Percentage, Temperature, Top k, and Top p define how aggressively sequence space is explored. Higher mean log-likelihood and higher sequence recovery usually indicate more conservative, scaffold-faithful proposals, whereas lower recovery can be useful when the goal is to search for more diverse but still structure-compatible variants.

• Structure-native protein design: MIF-ST starts from a solved or predicted backbone, which matches real fixed-backbone redesign workflows.
• Sequence-plus-structure learning: Sequence transfer lets the model use both geometric context and large sequence corpora rather than only the smaller structure-labeled set.
• Useful exploration controls: Masking and sampling parameters let researchers choose between conservative recovery and broader sequence diversification.
• Direct ranking signals: Mean log-likelihood and sequence recovery make the output immediately useful for candidate triage before folding, assay, or engineering follow-up.