STING-ExositeDB: An AI-assisted curated database of protein exosites for drug discovery

Protein exosites are secondary binding sites that recruit macromolecular partners and control molecular recognition across protein families. Despite their therapeutic potential demonstrated by approved drugs such as bivalirudin, systematic characterization remains limited by fragmented literature and inconsistent reporting. Existing databases lack exosite-specific curation, residue-level contact mapping, and structure-function integration.

We present ExositeDB, a curated database of protein exosites built through AI-assisted curation combining large language model extraction with expert validation. Our three-pass pipeline acquires papers from multiple databases, extracts data using GPT-4o with confidence scoring, validates findings against PDB structures, and links all claims to source text. The current release contains 525 expert-validated exosite records spanning 280 unique proteins, each annotated with residue-level positions, partner types, and functional roles. Quality is maintained through multi-tier confidence scoring integrating experimental methodology (35% weight), structural validation (25%), literature consistency (20%), and terminology precision (20%). The database follows FAIR principles through structured metadata, versioned releases, persistent identifiers, and an open REST API. Structural coverage includes 395 unique PDB structures supporting structure-function validation. ExositeDB provides training data for AI-driven exosite prediction, supports rational design of selective modulators, and enables analysis of exosite-mediated regulatory networks.