This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility: More than 65% of approved medicines work by binding to proteins on the surface of human cells. Yet scientists developing new treatments face a fundamental problem: Major biological databases disagree about which proteins are actually found on cell surfaces.
Hertz fellow Becca Carlson and her team at the Deliverome Project are using artificial intelligence to help resolve those discrepancies. By analyzing published research, they've developed an open, interactive map of the human surfaceome—the collection of proteins found on the surface of human cells—that could help researchers identify new targets for drug development. The AI-powered map identifies 1,108 candidate cell surface proteins missing from all five major reference databases, with supporting evidence from published studies.
The findings could expand the number of potential drug targets available to researchers developing treatments for cancer and other diseases. "My hope is that this will help people develop drugs that are more diverse and serve more patients and more indications," Carlson said in June, describing the Deliverome Project. To build the map, Carlson and her co-founder, biochemist Bobby Hollingsworth, compared five widely used biological databases: UniProt, GO CC, the Human Protein Atlas, the Cell Surface Protein Atlas and SURFY.
Together, the databases identify 5,546 surface proteins, but all five agree on only 188. The team then developed an AI agent using Claude Sonnet 4.6 to review evidence from published studies and evaluate whether individual proteins are found on the surface of human cells. The AI system correctly classified 95.7% of proteins in SurfaceBench, a hand-labeled test set of 147 proteins on which existing databases frequently disagree.
UniProt, the best-performing database in the comparison, achieved approximately 90% accuracy. Among the newly identified candidates are proteins that lack the structural features typically associated with cell surface proteins. Some may appear on cell surfaces only under certain conditions, including disease states such as cancer, making them potential targets for more precise drug delivery.
One example is KLK2, a protein that all five reference databases classify as absent from the cell surface. However, research has found that a membrane-anchored form of KLK2 is enriched in prostate cancer, and an experimental therapy targeting the protein is being studied in phase 3 clinical trials. The Deliverome Project classifies KLK2 as a protein that can appear on the cell surface under certain conditions, demonstrating how AI-assisted analysis of existing research could help identify drug targets overlooked by conventional databases.
Extract — continue reading at the source.
