Deliverome’s AI map flags a cancer-linked protein missed by five databases

A cancer-linked protein missing from five cell-surface reference databases is already the target of an experimental therapy in Phase 3 trials. An October 29 Hertz Foundation account highlights KLK2 as an example of what the Deliverome Project’s AI-assisted map captures: proteins that may become accessible from outside a cell only under particular conditions.

Published research describes a membrane-anchored form of KLK2 enriched in prostate cancer. Deliverome classifies it as contextually surface-accessible, rather than simply absent. The example illustrates the therapeutic relevance of the project’s paper, The accessible human surfaceome, published earlier in October.

Five databases, different definitions

The authors report 1,108 candidate surface proteins absent from UniProt, Gene Ontology Cellular Component, the Human Protein Atlas, the Cell Surface Protein Atlas and SURFY. These are classifications drawn from existing knowledge and literature—not newly discovered proteins.

Together, those databases list 5,546 surface proteins, but only 188 appear in all five. That gap is not simply a tally of mistakes. The resources use different definitions and evidence: expert curation, experiments or predictions from protein sequences. Deliverome instead asks whether a protein is reachable from outside the cell in at least one experimentally validated setting.

A benchmark advantage, with boundaries

The authors tested Claude Sonnet 4.6 on SurfaceBench, 147 manually labeled proteins deliberately selected to emphasize database disagreements. Using gene symbols without search tools, it scored 95.7% accuracy. UniProt reached about 90% after the researchers optimized each database’s inclusion rules against the benchmark.

From screening to inspectable evidence

1. A lightweight AI pass classified 19,324 protein-coding genes as surface, contextually surface or non-surface, giving a structured reason for each decision.
2. A deeper pipeline reviewed retrieved literature for proteins flagged by the AI or a reference database, anchoring records to exact quotations.

The deeper stage returned 1,783 canonical surface proteins and 1,252 with weaker or context-dependent evidence. Records separate confidence, cell-type restriction and state dependence, letting researchers filter classifications to suit their question.

The release includes free Open View access to classifications and supporting studies, a public API for software access, and open-source code. The authors identify access to full-text papers as a remaining limit: evidence-linked records depend on which literature the agents could read.