How to Use ThermoMPNN

ThermoMPNN is a structure-based model for predicting how point mutations affect protein fold stability. It is useful in protein engineering campaigns where the main decision is which substitutions should be tested to stabilize a scaffold, rescue a problematic region, or avoid destabilizing edits before experimental work begins.

The method supports both single and double mutations, which makes it more useful than a simple saturation-mutation screen when a project needs to reason about combinatorial designs or potential epistatic interactions between nearby sites.

On Neurosnap, researchers upload an Input Structure, optionally restrict analysis with Input Chain(s), and choose Mode as Single, Additive, or Epistatic. Additional controls such as Threshold, Distance, Include Cysteine, and Penalize determine which mutations are reported and how disulfide-related cases are handled.

ThermoMPNN uses transfer learning to improve prediction of mutation-driven stability change from structural context. The recent ThermoMPNN work extends that foundation to double mutations and introduces a Siamese neural-network design so double-mutant predictions are order invariant rather than dependent on how a mutation pair is written. That is scientifically important because true thermodynamic effects should not depend on mutation-list ordering.

On Neurosnap, Single mode screens individual substitutions, Additive mode approximates double mutants by summing single-site effects, and Epistatic mode attempts to capture interaction effects between paired mutations directly. Distance is especially relevant for double-mutant workflows because nearby residues are more likely to show nonadditive behavior, while cysteine and disulfide controls matter when redox-sensitive sites could otherwise produce misleadingly favorable scores.

Researchers should interpret ThermoMPNN comparatively. The workflow is strongest for ranking mutation ideas on the same scaffold, distinguishing clearly stabilizing from clearly destabilizing edits, and identifying whether a double-mutant campaign likely needs explicit epistasis-aware evaluation rather than simple additive reasoning.

• Structure-centered mutation ranking: ThermoMPNN evaluates substitutions in the context of the actual protein scaffold rather than from sequence alone.
• Explicit double-mutant support: Additive and epistatic modes make it useful for combinatorial engineering, not only single-site scans.
• Scientifically relevant controls: Threshold, distance, and disulfide-related options help keep the screening aligned with real stability questions.
• Good engineering triage: The model helps narrow variant sets before expression, thermal assays, or deeper biophysical characterization.