Project · Feb 2026 · live data
Where a mutation lands in 3D
A pathogenicity score says how harmful one amino-acid change is likely to be. It does not say that two changes hundreds of positions apart in the sequence may sit side by side once the protein folds. This explorer, which won Best Project at Code for Life (VIT, Feb 2026), puts the scores on the structure so those clusters become visible.
- Award
- Best Project
- Sources
- UniProt · AlphaFold · AlphaMissense
- Runs
- In your browser, live
- Proteins covered
- Any reviewed human gene
Try it
- UniProt: gene to protein
- AlphaFold: predicted structure
- AlphaMissense: every substitution scored
Loading the structure…
Drag to rotate, scroll to zoom, click a sphere to select a residue.
How to read it
Each sphere is one amino acid, placed where AlphaFold predicts it sits in the folded protein. Coloured by predicted harm, red regions are where almost any substitution is likely pathogenic: usually the buried core and the active site, where the protein cannot tolerate change.
Switch to model confidence before trusting a distance. Orange regions are disordered or uncertain; AlphaFold is guessing their shape, so hotspots only use residues with pLDDT of 70 or more.
Hotspots are pairs of strongly harmful positions (average score 0.8 or more) at least 25 residues apart in the sequence but within 8 Å in 3D, close enough to touch. Variants at such pairs can damage the same structural region, which a one-dimensional score track hides.
Score a specific variant by typing it as original amino acid, position and new amino acid. Positions follow UniProt, which counts the first methionine; older papers often do not, so SOD1's well-known ALS mutation A4V is A5V here.
How it works
The gene symbol goes to the UniProt REST API for the reviewed human protein. Its accession fetches the AlphaFold DB model, read for the alpha-carbon of each residue and its pLDDT, and the AlphaMissense table of all 19 possible substitutions at every position. The page computes the per-residue averages, the distances and the hotspot search itself. Nothing is precomputed, so every reviewed human protein in AlphaFold DB works.
The hackathon version was a Streamlit app with the Mol* viewer; this is a browser rebuild of the same idea. AlphaFold and AlphaMissense predictions are for research and teaching, not clinical use.
Team project · original code on GitHub · Data: UniProt, AlphaFold DB (EMBL-EBI and Google DeepMind), AlphaMissense (Cheng et al., Science 2023).