← research

StabilityDesigner

Find the stabilizing mutations worth testing — in seconds, from sequence alone.
Predict how a point mutation changes protein folding stability (ΔΔG), or scan a position for stabilizing substitutions. Sequence-only model — no crystal structure required. Trained on FireProtDB, benchmarked on S669 with state-of-the-art antisymmetry.
checking backend…
For: protein & enzyme engineers antibody / biologics developers biosensor & synthetic-biology labs
Examples: T4 lysozyme (164 aa) barnase (110 aa)
mutation
Format: wild-type, position, new residue — e.g. L15A.
position
Scores every substitution at one residue and ranks them. ~20 s.
⚠ The live prediction server isn't active right now.
You can still explore a validated saturation-scan report, read the code, and run the model locally on CPU.
→ Example report → Source & benchmarks → Run the CLI

What you can do with it

  • Triage thermostability: shortlist stabilizing mutations for a biocatalyst before wet-lab testing.
  • Harden a biologic: find stabilizing substitutions to improve antibody developability & shelf life.
  • Free first pass: prune obviously-destabilizing options before spending GPU/structure compute.

Where it fits

  • A transparent CPU baseline — interpretable features, no black box.
  • Sits in the published S669 accuracy band (DDGun-class) vs FoldX / DDGun / ThermoMPNN.
  • Best-in-class antisymmetry (forward/reverse bias ≈ 0) where FoldX & MAESTRO fail.
  • Open source (MIT). Structure features via --pdb push accuracy higher.
ΔΔG < 0 = stabilizing, ΔΔG > 0 = destabilizing (|ΔΔG| < 0.5 kcal/mol ≈ neutral). Predictions are computational estimates — confirm by thermal/chemical denaturation before drawing conclusions. Data: FireProtDB (Stourac 2021), S669 (Pancotti 2022), Ssym (Pucci 2018), splits via ThermoMPNN (Dieckhaus 2024).  ·  github.com/gianyrox/stabilitydesigner  ·  example report