3D protein structure
The 3D structure view places a cohort's mutations on the protein's three-dimensional structure. It answers questions a lollipop plot cannot: which mutations sit together in space even when they are far apart in sequence, whether they are buried in the protein core or exposed on the surface, whether they line a binding pocket, and which ones touch a bound ligand in an experimental structure.
Structures come from AlphaFold DB (predicted models, CC-BY 4.0) and the Protein Data Bank (experimental structures, CC0). Everything is fetched and cached by the platform's analysis service — your browser never contacts these services, so the genes you look at are not disclosed to third parties.
Prerequisites
- A completed Mutation landscape run
- A parseable protein change column (
HGVSp_ShortorProtein_Change, e.g.p.G12Dorp.Gly12Asp) - A human gene with a reviewed UniProt entry (the 20 curated driver genes, or any other gene resolved through the bundled UniProtKB table)
Opening the 3D view
On the Mutation landscape results, scroll to Gene detail, choose a gene, and click View in 3D. The structure opens below the lollipop plot.
- Spheres mark mutated residues, sized by number of events and coloured by the dominant variant class (the same colours as the lollipop).
- Linked selection: click a lollipop head, a sphere, or a residue chip to select that residue in both views. The camera zooms to it and the lollipop rings it.
- Navigation: drag to rotate; Ctrl/⌘ + scroll or pinch to zoom (plain scrolling scrolls the page).
How positions are placed
Every mutation is validated before it is drawn: the reference amino acid in the protein change must match the structure's residue at that position. Changes that fail are listed under not placed with a reason, never silently plotted:
| Reason | Meaning |
|---|---|
| Reference amino acid does not match | The upload likely uses a different protein isoform than UniProt canonical |
| Position is outside the structure | The residue number is beyond the modelled sequence |
| Not residue-anchored | Missing protein change, splice-site placeholder, or range notation |
| No reference amino acid | Placeholder reference (e.g. p.X5?) |
When more than 10% of residue-anchored changes fail, an isoform warning appears and 3D hotspot statistics are withheld. Mutation landscape keeps the MAF Transcript_ID and ENSP columns when present, for future isoform resolution.
Structure sources
Use the Structure selector to switch between:
- AlphaFold model (default) — the full-length predicted structure, coloured by model confidence (pLDDT: very high ≥ 90, confident 70–90, low 50–70, very low < 50). Low-confidence regions are often disordered; the residue card warns when a selected residue has pLDDT < 70.
- Experimental PDB entries — the best entries covering the protein (ranked by coverage and resolution via SIFTS), labelled with method, resolution, chain, and coverage.
For PDB entries, positions are translated through SIFTS residue-level UniProt numbering. Residues where the crystallized construct differs from UniProt (engineered or variant residues — for example KRAS G12D mutant structures) are never used for placement, and the panel explains which residues were not placed and why. When AlphaFold DB has no model (proteins longer than its length limit, such as BRCA2 and NF1), the view opens on the best PDB entry automatically and positions are validated against the UniProt canonical sequence.
3D hotspot analysis
Click Find 3D hotspots to run the structure_mutation_map job (inline progress below the button). It runs on the AlphaFold model and adds, for each mutated residue:
| Metric | Method |
|---|---|
| Secondary structure | DSSP (helix / strand / coil) |
| Solvent accessibility | Shrake–Rupley SASA, relative to the residue type's maximum; buried when RSA < 0.25 |
| Domain context | Containing curated domain, or Cα distance to the nearest one |
| Hydrogen bonds and clashes | Geometric heavy-atom criteria (no explicit hydrogens); clashes at van der Waals overlap ≥ 0.6 Å |
| Surface potential | Coulombic estimate (formal charges at pH ~7, distance-dependent dielectric) |
| Pocket membership | Grid-based pocket detection (buriedness), confident residues only |
| Hotspot q-value and cluster | See below |
Hotspot test. Using residues with pLDDT ≥ 70, each mutated residue is scored by the mutation events within 10 Å of it. The observed counts are shuffled across eligible residues of the same protein 10,000 times (recorded seed) to build the null; p-values are corrected with Benjamini–Hochberg (significant at q < 0.05), and significant residues within 10 Å are merged into clusters. At least 5 confidently modelled mutated residues are required.
Results include:
- Clusters — residues, events, q-value, and sequence span; click to focus the whole cluster. KRAS G12, G13, and Q61 typically form one cluster despite spanning 50 residues in sequence.
- Pockets — volume, lining residues, mean pLDDT, and which mutated residues lie inside; click to highlight in violet.
- Colour modes — mutation density (orange), burial (blue), electrostatics (red/blue surface).
- Residue details — the full per-residue table; click a row to select.
Results persist: reopening the gene loads the latest completed run for that gene and mutation run. The job is deterministic for its seed, so re-runs and retried deliveries give identical results. If AlphaFold DB has no model for the protein, the hotspot analysis is unavailable and the button explains why.
Ligands in experimental structures
In PDB mode, ligands are drawn as ball-and-stick and ions as spheres. The Ligands in this entry panel lists each component with the protein residues within 4 Å (UniProt numbering), flags mutated residues in contact, and marks likely crystallization additives (glycerol, sulfate, PEG, …) so they are not read as biology.
Ligand–receptor complexes
From Cell-cell communication results, the complex explorer finds the best experimental structure of a ligand bound to its receptor and highlights the interface residues on each side. See that page for details.
AI interpretation
Structure runs are included in AI interpretation. Additional guardrails apply: no structural-effect claims for residues with pLDDT < 70, no interface or complex claims from single-chain AlphaFold models, and any hypothesis using structure evidence must cite a computed structure metric (for example a cluster q-value or the buried fraction).
Reports
Each gene's latest structure run appears in the Report Builder's oncology library as 3D structure — {gene}, rendering the 3D view, summary metrics, hotspot clusters, parameters, and attribution.
Limitations
- AlphaFold models are predictions; treat low-confidence regions with caution.
- Hydrogen bonds, electrostatics, and pockets are geometric estimates, not simulations or Poisson–Boltzmann solutions.
- Hotspot analysis and the structure metrics use the AlphaFold model; experimental entries are for viewing, ligands, and complexes.
- Positions assume UniProt canonical numbering; isoform mismatches are reported rather than corrected.
- Research use only — not a clinical variant interpretation.