Skip to main content

A consensus database of the accessible human surfaceome.

Roughly 10–20% of human proteins reach the cell surface, yet over 65% of approved drugs target these molecules — accessibility on the extracellular face of the plasma membrane lets a drug act without crossing into the cell. Five public databases (UniProt, GO Cellular Component, the Human Protein Atlas, the Cell Surface Protein Atlas, and SURFY) each catalog the surfaceome, but with different definitions and methods, so their calls disagree more often than they agree.

This atlas reconciles them. An LLM triage agent scores every protein-coding human gene against the union of the five sources; a deep-dive agent then assembles the evidence behind each strong candidate — surface-localization methods, antibody validation, isoform topology, accessibility caveats — into per-gene records. Open accession, evidence-cited, agent-readable.

Curated lists

Curated shortlists over the deep-dive cohort. Canonical is the strictest tier, Likely is broader, Cell-state induced and Cell-type restricted are sub-buckets of Likely.

Applies only to genes with a deep-dive record. Non-deep-dive rows auto-exclude because the predicate reads fields the catalog row doesn't carry. The count badge on each chip is the population that survives the predicate alone — refine further with the existing search and facet chips below.

See the full filter definitions on the API page →

Strictest tier — direct evidence (single or multi-method), high/moderate confidence, surface-dominant or mixed, low / moderate / unclear state-dependence, high/moderate surface accessibility, high/moderate evidence density. The high-confidence surface shortlist.

Applies only to genes with a deep-dive record. Non-deep-dive rows auto-exclude because the predicate reads fields the catalog row doesn't carry.

Broader shortlist — adds supportive-but-indirect evidence, mostly-intracellular surface fractions (e.g. SRC via lysosomal exocytosis, HMGB1 via DAMP release), and high / unclear / null state-dependence.

Applies only to genes with a deep-dive record. Non-deep-dive rows auto-exclude because the predicate reads fields the catalog row doesn't carry.

Subset of Likely where surface presentation is induced by cell state (oncogenic transformation, stress, infection, immune activation). HSPA5, SRC, CD63, HMGB1, C3 land here.

Applies only to genes with a deep-dive record. Non-deep-dive rows auto-exclude because the predicate reads fields the catalog row doesn't carry.

Subset of Likely with constitutive surface in specific cell types only (KLK2 in prostate, etc.). Different cell types — not same cell across states.

Applies only to genes with a deep-dive record. Non-deep-dive rows auto-exclude because the predicate reads fields the catalog row doesn't carry.

19,324 genesTSV is verdicts + reason codes only. Free-text reasoning per run is on GET /v1/triage/{SYMBOL}; the full deep-dive SurfaceomeRecord is on GET /v1/genes/{SYMBOL}.

Reason
The deep-dive agent's surface verdict — does this protein reach the cell surface in at least one cell state (high, moderate, low, or no)? The levels are evidence strength for the surfaces-at-all call, not a steady-state magnitude. Present only for genes with a deep dive; click to open it.
How confident the deep-dive agent is in its surface verdict — high, moderate, or low.
Strength of the experimental surface evidence behind the call — from direct multi-method down to weak or conflicting.
How much the surface verdict depends on cell state or context (e.g. activation, stress) — low, moderate, or high.

Triage agent · a single Claude call with an NCBI context block, run on every protein-coding gene. Verdicts read as yes (constitutively on the surface), contextual (conditional or rare surface), no (no plasma-membrane presence). Deep-dive records linked from the yes column in the “Deep dive” slot override the triage call when present. The alternative models (Haiku 4.5, Opus 4.8) used in the manuscript comparison live on SurfaceBench.