AI-driven mapping reveals 1,108 potential drug targets on the human cell surface

AI-driven mapping reveals 1,108 potential drug targets on the human cell surface

Contents
  1. Databases that disagree
  2. How the tool works
  3. A case in point: the KLK2 protein
  4. An open resource
  5. References

Over 65% of approved medicines function by attaching to proteins on the exterior of human cells, although primary biological databases fail to agree on their actual presence. The Deliverome Project, which operates as an open research initiative, applied artificial intelligence to analyze existing literature and construct a comprehensive map of the human surfaceome, representing the complete array of cell surface proteins.

The map flags 1,108 candidate proteins that are missing from all five major reference databases but are supported by published studies. The findings were posted in October 2026 as a preprint on bioRxiv, so they have not yet been peer reviewed, and the proteins are candidates rather than confirmed targets.

Databases that disagree

The team compared five databases: UniProt, GO CC, the Human Protein Atlas, the Cell Surface Protein Atlas and SURFY. Together they list 5,546 surface proteins, but all five agree on only 188, according to the Hertz Foundation.

How the tool works

Co-founders Becca Carlson and Bobby Hollingsworth built an AI agent using the Claude Sonnet 4.6 model. The agent reviews the scientific literature to determine whether a protein sits on the cell surface.

On SurfaceBench, a hand-labeled collection consisting of 147 proteins, the agent achieved a 95.7% classification accuracy, whereas UniProt, which stood out as the top-performing database, reached roughly 90%.

A case in point: the KLK2 protein

All five databases classify the protein KLK2 as absent from the cell surface. However, research has found a membrane-anchored form enriched in prostate cancer, and an experimental therapy targeting it is in Phase 3 trials. Deliverome classifies KLK2 as able to appear on the surface under certain conditions.

An open resource

The map is free to explore in Open View, an interactive tool that shows each protein alongside its supporting research. The project also offers a public API and open-source code.

“My hope is that this will help people develop drugs that are more diverse and serve more patients and more indications,” said Carlson, a 2017 Hertz Fellow and the project’s co-CEO.

Deliverome launched on May 27, 2026, with $5 million in seed funding from the Astera Institute. It is set up as a Focused Research Organization, a model designed to deliver public research tools within about five years. Researchers from Coral Biosciences also co-authored the preprint.

References

By William Parker