Title: Site4Drug: Leveraging an AI Agent to Predict Drug-Binding Target Sites

Abstract: Identifying the precise location to intervene on a protein—essentially selecting a targetable site—often presents a more ambiguous and error-prone challenge than determining which molecule should bind. This difficulty is particularly acute for membrane proteins, where factors such as accessibility, topology, and post-translational modifications (PTMs) significantly restrict viable regions. To address this, we introduce Site4Drug, a modality-aware agent designed to generate a ranked list of targetable regions. This output includes explicit constraints, summaries of evidence, risk indicators, and a fully traceable decision log.

Unlike traditional approaches, Site4Drug does not require users to define the drug modality at the outset. Instead, it can recommend an appropriate binding modality—such as small-molecule versus antibody or peptide-like agents—by analyzing the same evidence base utilized for site discovery. This evidence encompasses topology, hydropathy, PTM propensity, disulfide bonds, domain context, and sequence data. Crucially, this evidence is applied uniformly across all modalities, including the discovery of small-molecule pockets, thereby preventing the selection of sites that may appear chemically feasible but are biologically inaccessible.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC