Stability ΔΔG Predictor

Rank point variants of a protein by predicted change in folding stability. ThermoMPNN, a ProteinMPNN network trained on Megascale stability measurements, predicts ΔΔG in kcal/mol from one chain of the structure — negative is stabilizing, positive destabilizing — without building a mutant structure. Use it to order the variants of one protein, not as a measured value: on proteins the model was not trained on, its ranking of Megascale-type measurements has a rank correlation near 0.74, but only near 0.4 for literature mutations, where the typical error is about 1.2 kcal/mol. It flags destabilizing substitutions far more reliably than stabilizing ones. A variant with several substitutions sums its single-site predictions (not benchmarked), and an alanine scan ranks the residues of a chain by what replacing each would cost. Every row also carries how far its prediction moves under 0.2 Å of backbone noise; it grows with the size of the prediction and is not a confidence interval or an error estimate. Monomer chains and the 20 standard residues only.
0 variants ready
Estimated minimum cost: $—

Examples

Structure and parent

Upload the protein as a PDB file. Only one chain is scored, as a monomer: other chains in the file are ignored, including as context for the chain you score. Predicted structures work, and parent-relative ranking cancels much of their error.

or drop a file here

Chain to score (optional)

One letter or digit. Left blank, the first chain in the file whose sequence matches the parent below is scored, or, with no parent, the first chain in the file whose ID is a single letter or digit.

Parent chain sequence

One-letter codes or one FASTA record, up to 5 000 residues, exactly as the scored chain appears in the structure. Required for mutation strings and variant sequences; optional for a scan.

or drop a file here

Variants

Choose the substitutions to rank. Every variant is scored on the same backbone, so only the mutated residues move. Lists take at most 50 variants per run.

One variant per line, written as the residue you expect, its position, and the replacement — T315I — joined with + to combine several in one variant. Positions are the chain's residue numbers; each expected residue is checked against that chain when the job runs, so a numbering offset fails with the residue actually found. A variant with several substitutions reports the sum of its single-site predictions.

No variants yet.

Numbering and noise passes

First residue number of the parent sequence (optional). Left blank, it is read from the matched chain of the structure. When set, it must equal that chain's first residue number, so a parent numbered differently from the structure fails instead of mutating a different site.

Noise passes (2–32). Each pass adds 0.2 Å of Gaussian noise to the backbone, and the table reports the standard deviation of every prediction across them. It shows how far a prediction moves when the backbone moves, and it grows with the size of the prediction. In the benchmark it did not predict how wrong a prediction is, so it is not a confidence interval or an error bar. The prediction itself comes from the noise-free pass.