Drybench Protein.

What the app shows, how it works and how it was built, in the words of its own write-up, followed by the parameter defaults and the data sources with their licences.

What it shows

Search UniProt by gene, protein name or accession. For the protein you pick, the page shows what UniProt records about its function and location, an experimental structure from the PDB beside the AlphaFold DB prediction, sequence tracks that line up domains, sites, modifications, structural coverage, prediction confidence and variants, and the drugs and clinical candidates that Open Targets lists against it.

A picked UniProt residue, domain or site follows SIFTS mapping into the experimental chain and unique exact sequence partners into the predicted model. Missing correspondence or unresolved residues are omitted and explained.

How it works

Everything is fetched live from public APIs in your browser; nothing is stored on a server. UniProt REST supplies search and the entry; a single-word search also asks for an exact human gene-symbol match, because UniProt's own ranking can put TP53TG5 above TP53. PDBe's best_structures API lists every PDB chain mapped to the protein, but its coverage figures describe the construct, not what was modelled, so PDBe's UniPDB endpoint is used to find the residues each entry actually resolves. The default structure is the entry that resolves the most residues, weighted by resolution (full weight to 2 Å, half at 4 Å, less for coarser maps and NMR); the leading candidates are checked against SIFTS sequence identity, and one at 95% identity or less counts for half, so an entry of another isoform is rarely the default; its title is shown so mutants and complexes are visible. The chosen chain is also compared with the UniProt sequence residue by residue, and stretches with a different sequence, or residues the chain has no counterpart for, are not counted as resolved.

AlphaFold DB's prediction API returns the model, isoform models and sometimes models from other providers; the page takes the entry for the exact accession and uses the file URLs the API gives, so a new database version needs no code change. Models are built for the UniProt sequence of their release, so the model's sequence is lined up with today's UniProt sequence before any per-residue value is drawn, and the page says when they differ. pLDDT colours use AlphaFold DB's cut-offs (90, 70 and 50; a score exactly on a cut-off counts in the higher band), taken from each residue's score rather than from the file's category letters, which differ between model providers. pLDDT is a per-residue estimate of local confidence, and low-confidence stretches are often disordered; it says nothing about how domains sit relative to each other, which is what the predicted aligned error (PAE) measures. The page does not show PAE; read it on AlphaFold DB before judging domain placement.

The variants track starts with UniProt's curated natural variants. The EBI Proteins API's fuller list, with ClinVar and other clinical classifications where they exist, can be added on request, because that response is tens of megabytes for well-studied proteins; records known only from COSMIC, and Ensembl copies of them, are left out, because COSMIC needs a paid licence for use on commercial websites. Drugs come from the Open Targets Platform GraphQL API through the Ensembl gene id in the UniProt entry. Approval is read from each drug's own highest stage and withdrawals from Open Targets' drug warnings: a drug counts as withdrawn from the market only when the warning is worldwide or covers both the United States and the European Union, and narrower withdrawals are listed with their countries. A drug with a narrower withdrawal that names the United States, the European Union or the United Kingdom, and no approval record from a regulator or drug label in Open Targets, is shown as withdrawn in those countries rather than approved, and so is a drug withdrawn anywhere that has no approval record at all. An approval stage that rests only on a classification or naming list (a WHO ATC code, a TTD entry, or a USAN or INN name) is marked as such; salt forms are grouped under one ingredient.

How it was built

A Next.js page with one client component tree. Mol* 5.11 (the prebuilt viewer, MIT licence) loads from the site's own vendor folder only once a protein is chosen and a viewer is near the screen, and each viewer is driven by a MolViewSpec scene built from the fetched data: pLDDT bands become four colour selections, and the chosen PDB chain, other copies of the protein, other molecules and ligands each get a colour. The two viewers are created once and reused for every entry and protein, because each new Mol* viewer costs a WebGL context and memory. The page downloads each structure file itself, with a time limit and a second host, and gives Mol* a local copy, so a stalled file host ends in an error with a retry rather than a spinner. Entries over 150,000 atoms load just the chosen chain from the PDBe ModelServer until you ask for the whole entry.

The tracks are hand-written SVG with their own zoom, scrolling inside the panel and a keyboard cursor. Parsing, ranking, residue mapping, drug grouping and scene building are plain TypeScript modules with unit tests on Node's built-in test runner. Field names for every API were checked against live responses in September 2026, including the Open Targets 26.09 schema, where drug data moved from target.knownDrugs to target.drugAndClinicalCandidates.

Parameter defaults

Default PDB entry

The entry that resolves the most residues of the protein, weighted by resolution: full weight to 2 Å, half at 4 Å, less for coarser maps and NMR.

Sequence identity

A leading candidate at 95% SIFTS sequence identity or less counts for half.

pLDDT bands

AlphaFold DB's cut-offs of 90, 70 and 50 (Varadi et al. 2022, read 24 October 2026); a score exactly on a cut-off counts in the higher band, read from each residue's score. Some viewers, including RCSB PDB's model pages, place a score of exactly 90 or 70 in the lower band.

Exact confidence transfer

Source-model scores stay in model coordinates. Current-UniProt confidence and predicted focus use only identical standard residues paired in every optimum of the fixed global/free-end BLOSUM62 alignment, gap cost 10 + 0.5(k − 1). Other positions stay missing; range means state contributing n over selected N. Unverified fragment domains are withheld. The mapping runs in a cancellable worker with a 10-second operational limit; each sequence is at most 10,000 residues and the length product at most 10 million.

Capture versions and historical records

Version 4.0.0 applies to new experimental structure captures and 3.0.0 to new predicted-model captures. Historical records retain their recorded procedure and original mapping policy; opening this overview does not upgrade them.

Large entries

Entries over 150,000 atoms load only the chosen chain until you ask for the whole entry.

Withdrawn from the market

Only when an Open Targets drug warning is worldwide, or covers both the United States and the European Union.

Checking a substitution

One amino-acid substitution in UniProt numbering (V600E, p.V600E, Val600Glu or p.(Val600Glu)) is checked against the current UniProt sequence; a different residue there stops the check and names the residue. The AlphaMissense score for the substitution is shown only when the AlphaFold model's sequence equals the UniProt sequence or is a unique stretch of it (or the other way round) with no differing residue, every row of the AlphaMissense file names the model's own residue, and the row's reference residue equals both the model's and UniProt's at that position; otherwise the panel says which of these failed. The score is shown as the file writes it, with AlphaMissense's own class and thresholds. The mean over the substitutions at the residue is calculated here and shown apart, never as a substitution score.

Comparing two entries

Two UniProt sequences (an entry or an isoform, or the picked residues of this entry) are aligned in full by Gotoh's affine-gap dynamic programming with BLOSUM62: a gap of k residues costs 10 + 0.5 (k − 1), and gaps at either end are free. Where gaps in the two sequences sit next to each other at an end, only the one that reaches its own sequence's end is free; the other costs as an internal gap, as in Biopython's PairwiseAligner. Scores are computed in doubled integers, so ties are exact. Identities, similar pairs (BLOSUM62 above 0) and gap columns are counted over all alignment columns, including end gaps; identities are also given over each sequence's length. Where several alignments share the best score, one is shown in a fixed order, and a residue pair counts as shared only when it appears in every one of them. The matrix is NCBI's BLOSUM62 file, checked entry by entry against Biopython's copy. More than 10 million comparison cells (the product of the two lengths) are refused before any table is made; a partial alignment is never shown. No E-value or significance figure is computed.

Retaining a structure view

Save the reference task, prepare one visible viewer's image, then confirm the displayed source and settings. Capture procedure protein.structure-evidence-capture uses 4.0.0 for new experimental views and 3.0.0 for new predicted views; historical records retain their recorded procedure. It retains the PNG exactly as the viewer captured it, actual loaded-byte SHA-256, source receipts, native camera and rendering settings, selection qualifications and your explanation. Images recorded with 1.0.0, from 18 to 20 October 2026, also carry title, time, source and provenance text embedded before recording, and say so when opened. Provider versions remain unavailable when not supplied. Coordinate files are not retained; a saved view does not establish structural validity, biological activity or physical experimental execution.

Data sources, licences and versions

Times on screen are local time with the UTC offset; times in exports are UTC. Dates with no time are shown at the precision the source gave them.

Recorded evidence offers Download retained PNG and Download source receipt. Capture embeds its source provenance before freezing and hashing the PNG; these downloads preserve those image bytes and input/result information. Reopening does not fetch today's source or reconstruct an interactive scene. Text-only Study views state when the image is omitted and its bytes are not verified there. Ordinary live-view PNG exports continue to embed source provenance.

The Export menu in the top bar offers the PDB entries, the drugs and clinical candidates, and the features and variants on the tracks as CSV with per-row provenance columns; the canonical sequence as FASTA with a provenance file beside it; the protein map and the sequence tracks as SVG or PNG; each 3D view as PNG; on the workbench, a checked substitution as CSV, and an alignment as aligned FASTA (each header with the sequence version and UniProt's checksums) and as a column-by-column correspondence CSV. Every file names its sources, their licences and versions, and when this browser retrieved them. A copied link reopens the same protein, PDB entry, view and pick.

Release gate: before the app leaves the preview, its exports and printed reports will carry the data versions, parameters, limits and a timestamp needed to reproduce a result, as the regulatory position says. This section will describe each format as it ships.

Each module below is a set of automated tests in the app's source code, with the expected results written into the tests. They check the app's logic against recorded inputs and stored reference records; none of them calls a live service.

src/lib/protein/confidence.test.ts

The frozen independent Biopython 1.88 reference supplies strict source parsing, exact-only transfer, precision at band edges, sparse and zero scores, contributing-count means, split focus targets and source identity/domain refusals. Recorded EGFR assets establish representative full-model coordinates, without claiming runtime structural sequence verification.

src/components/tools/protein-explorer/residue-map-client.test.ts

Worker cancellation during hashing and execution; exactly-once teardown after success, timeout, malformed/stale response, worker, constructor or posting failure; bounded closed messages, actual sequence/confidence bindings and separate legacy-only admission. No main-thread alignment fallback.

src/apps/protein-explorer/evidence-exact-map.test.ts

New predicted schema2/adapter2/capture3 receipts preserve exact paired runs, source confidence and actual loaded-file identity. Source-independent rereading rejects mixed versions, contradictory statuses, altered hashes, bridged holes and pending correspondence; historical v1 semantics and lazy/full projection parity remain strict.

src/apps/protein-explorer/evidence-record.test.ts

Retained image and receipt integrity, actual source identity, finite native camera values, capture limits, field-named refusals and text-projection parity. Independent Python fixtures supply byte hashes, PNG boundaries, scene identities and the 3IKA point substitution; browser checks cover actual rendering separately.

src/components/tools/protein-explorer/uniprot.test.ts

Accession formats and isoform suffixes, search queries (including gene symbols shaped like accessions, such as P2RY12), the ranking of suggestions, curated variants and the reading of a UniProt entry.

src/components/tools/protein-explorer/structures.test.ts

PDBe best structures and UniPDB ranges, the resolution weighting, the default-entry ranking, SIFTS residue mapping, and recorded cases: the K-Ras4B stretch in 5TAR, an engineered point mutation, and 6GOD at exactly 95% identity.

src/components/tools/protein-explorer/alphafold.test.ts

Choosing the AlphaFold DB model, placing a model built for an older UniProt sequence (RAB12-, SLC3A2- and POLR2A-like cases), the pLDDT cut-offs, the boundary rule, and AlphaMissense averaging.

src/components/tools/protein-explorer/variants.test.ts

COSMIC-only records are dropped, clinical classes are ranked, variants are counted per residue and named in UniProt numbering (HBB residues 7 and 8 among the cases).

src/components/tools/protein-explorer/drugs.test.ts

Open Targets drug parsing: salt forms merged, stages ranked, approvals that rest only on an ATC code not counted (efaproxiral), and withdrawals by market (voxelotor, ketorolac, metamizole, troglitazone).

src/components/tools/protein-explorer/substitution.test.ts

The substitution check against a Python oracle's write-once fixtures (BRAF V600E in five notations, V599E refused, EGFR L858R and T790M, mature-numbered L834R refused, models offset either way, withheld entries, every refusal); each score also found in AlphaMissense's source file; every score's text as the file writes it.

src/lib/protein/align/align.test.ts

The aligner against Biopython 1.88's PairwiseAligner on UniProt inputs (TP53 human and mouse, TP53 and its isoform 2, a tagged EGFR kinase construct, a one-residue insertion, identical sequences, B, Z and X codes, short peptides), runs and repeats with tied best alignments, and gaps in opposite sequences side by side at an end, in three modes: the optimal score exactly, the shown alignment re-scored to it, the identity, similarity and gap counts, the number of co-optimal alignments and the residue pairs shared by all of them; refusals over 10 million cells and for selenocysteine; BLOSUM62 entry by entry.

src/components/tools/protein-explorer/mvs.test.ts

The MolViewSpec scenes: one colour per pLDDT band, unresolved residues faded, the chosen chain highlighted, and local copies of structure files.

Release gate: before the app leaves the preview, each release will publish a verification dataset with expected outputs on this page, so you can run your own checks.

Scientific computation runs in your browser, and individual apps need no account. Optional Bench accounts and backups use the site server and configured providers. The privacy statement, the hosts it contacts and the headers its workspace is served with are generated from its registry entry, the same record that sets the browser's policy.

What leaves your browser

Hosts the app contacts

• rest.uniprot.org

  Operator

  UniProt Consortium (EMBL-EBI, SIB and PIR)

  What for

  Search, and the protein's entry: names, function, location, sequence and features.

  What it receives

  your search text and UniProt accessions

  When

  on lookup

• www.ebi.ac.uk

  Operator

  EMBL-EBI

  What for

  PDBe: the PDB entries mapped to the protein, the residues each one resolves, residue numbering and structure files. The EBI Proteins API: the full variant list, when you ask for it.

  What it receives

  UniProt accessions and PDB ids

  When

  on lookup

• data.rcsb.org

  Operator

  RCSB PDB (United States)

  What for

  Atom counts, so that very large entries load one chain first.

  What it receives

  PDB ids

  When

  on lookup

• models.rcsb.org

  Operator

  RCSB PDB (United States)

  What for

  A second source for structure files when PDBe does not answer.

  What it receives

  PDB ids

  When

  on lookup

• alphafold.ebi.ac.uk

  Operator

  EMBL-EBI (AlphaFold DB)

  What for

  The predicted model, its per-residue confidence, and AlphaMissense scores when you ask for them.

  What it receives

  UniProt accessions

  When

  on lookup

• api.platform.opentargets.org

  Operator

  Open Targets

  What for

  Drugs and clinical candidates for the target, with mechanisms, stages, indications and warnings.

  What it receives

  the target's Ensembl gene id, or its UniProt accession when the entry gives no Ensembl id

  When

  on lookup

Headers the workspace is served with