Methods and data provenance

Dataset coverage, search rules and interpretation of the recorded evidence.

What is included

All 4,878 genes are retained. The default curated scope contains 2,525 targets selected by the source rules: plasma-membrane proteins, GPI-anchored proteins, nuclear receptors and explicitly retained families. Switch to All targets to inspect the remainder.

Pharmacological evidence comes from ChEMBL, DrugBank and TTD. The workbook also contains Open Targets disease evidence, development summaries, Russian-market indicators and company-interest assessments. The atlas contains fixed versions of these datasets, not live database results. On a target page, record counts refer only to that target in the included data. Search results refer to the selected target set; comparison panels explicitly name their reference cohort.

Search semantics

Simple search is case-insensitive. Space-separated terms must all match somewhere in the selected fields; quotation marks keep a phrase together. Advanced groups use AND or OR. Conditions within one related-evidence group must hold for the same record. Separate groups may match different records. “Is one of” accepts values separated by |; semicolon-separated statuses are matched individually.

Numeric comparisons preserve the supplied units. Do not compare activity values without specifying units and measurement type. Missing values do not satisfy numeric comparisons or “is not one of”; use the explicit Missing operator.

Evidence boundaries

TTD links targets to diseases and clinical statuses. The supplied rows do not identify drugs, so TTD stages are never assigned to individual DrugBank drugs.
ChEMBL records were collected with pChEMBL ≥6. Measurements for protein complexes are labelled separately from single-protein measurements. pChEMBL does not make all assay types interchangeable.
DrugBank status groups may coexist, including approved, investigational and withdrawn. Phase I–III numbers are not supplied by DrugBank.
The normalised Open Targets maxClinicalStage field is a clinical-precedent score, not a clinical phase.
FIC, BIC, NICFF and NICMT are preserved source strategy scores. The atlas does not recalculate them or create a composite ranking.
Historical source flags are kept separate from newly verified structure availability. Missing evidence is not evidence of absence.

Visual summaries and reference cohorts

Bars retain the original source scales. Distribution markers compare a target with all in-scope targets or an informative family. Percentiles exclude missing values, use midpoint ranks for ties and require at least 20 reference values. Numeric position is not therapeutic superiority. Disease-group evidence maxima may describe different diseases. Generic or unclassified families are not comparison cohorts.

Protein names are cached from human UniProt records, retaining recommended or explicitly labelled submitted names, accessions and retrieval dates. Family labels are presented separately from protein names.

Structures

Experimental mappings are retrieved from PDBe/SIFTS using UniProt accessions, retaining human structures. Representatives prioritise sequence coverage, then reported resolution and stable PDB ID order. AlphaFold models retain their accession, coverage, version and confidence. Explicit reviewed overrides may replace the automatic representative.

One experimental structure and one predicted model per target are saved when available. The complete matching catalogue is listed, but additional entries require following an external link. Fragments and predicted models are labelled; AlphaFold confidence is not experimental validation. Failed lookups and downloads are reported separately from confirmed empty results.

Membrane topology colours come from UniProtKB Transmembrane, Intramembrane and Topological domain features, retrieved through the EMBL-EBI Proteins API. PDB residues are mapped individually using SIFTS, retaining author chain IDs and insertion codes. AlphaFold residue numbers are checked against the exact UniProt accession and sequence. Uncertain feature boundaries, isoform-specific features, conflicting annotations and unmapped residues remain grey. Lumenal and other compartments are distinguished from extracellular and cytoplasmic regions. These colours do not define a geometric membrane plane. Source dates and evidence codes are retained; see the topology coverage report.

The optional Membrane layer uses existing OPM / PPM coordinates for the exact PDB entry. The two translucent boundaries mark the lipid carbonyl levels of a computationally placed planar bilayer. Coordinates are fitted to matching target-chain Cα atoms (at least 30 atoms and 50% chain coverage, RMSD at most 2 Å). Curved or multiple membranes, missing entries and failed fits are not displayed; orientation is not transferred to AlphaFold models. See the OPM coverage report and OPM / PPM documentation. Base view clears selections, hides membrane and surface, restores ligands, target chain, default colours and the initial camera view.

Provenance and dates

Recorded retrieval or processing dates, workbook-file dates and documentation dates have different meanings. A documentation date does not imply that a database was queried that day. Every field retains its date category. Supplied source files are checksummed and unchanged.

Local operation and downloads

All libraries, search data, 2D drawing and saved 3D models run locally. External database links open only when clicked. Start a static server in the website directory with python -m http.server 8000, then open http://localhost:8000. No application server or npm build is needed.

Download target dataset · CSV · Dataset manifest · Quality report · Structure coverage & unresolved mappings · Distribution checksums · Library versions & licences

Interpretation references

Terminology was checked against NCATS Pharos development-level definitions and Open Targets prioritisation documentation. ChEMBL activity definitions and AlphaFold confidence guidance explain assay and model scales. Historical values are preserved; current documentation is not used to retrospectively assign new phases or classifications.