AI Pushes the First Question in Drug Design Upstream: Where to Target a Protein

In drug discovery, the earliest and often most underestimated question is not how to design the molecule, but which part of the protein is actually worth touching. If the chosen site is on the inner side of the cell membrane, obscured by glycosylation or other modifications, or has an attractive pocket that is nevertheless inaccessible in a biological environment, then even the most sophisticated later design of antibodies, peptides, or small molecules may be built from the wrong starting point.

A preprint posted on arXiv on June 1 proposes Site4Drug, an AI agent for predicting drug-bindable regions of proteins. The research team positions it as a "modality-aware" site-selection system: it not only lists candidate targets, but also ranks actionable regions on a protein and provides constraints, evidence summaries, risk flags, and traceable decision records.

This setup targets a fairly upstream step in the AI drug discovery workflow. Many generative models can propose antibody, peptide, or small-molecule candidates, but if users must specify the drug modality from the outset, the system may lock into the wrong strategy too early. Site4Drug's design instead attempts to start from the same set of biophysical and sequence evidence and judge whether a given region is more suitable for antibody or peptide binding, or more likely to support a small-molecule pocket.

According to the abstract, the clues incorporated by Site4Drug include protein topology, hydrophobicity, post-translational modification tendencies, disulfide bonds, domain context, and amino acid sequence. This information is especially important for membrane proteins, because membrane proteins often face constraints involving intracellular versus extracellular orientation, transmembrane regions, glycosylation, and degree of surface exposure at the same time; a site that appears reasonable in a structural model may not necessarily be truly reachable by a drug.

Another focus is "auditability." In experimental and drug development settings, researchers need not only a ranked result, but also to know why the model excluded certain regions, why it favored a particular drug modality, and which judgments rely on inference rather than direct evidence. If such records can be generated consistently, they would help human experts scrutinize AI recommendations instead of merely accepting a score that is difficult to interrogate.

However, the paper is currently an arXiv preprint and notes that it has been accepted by the ICML 2026 "Generative and Agentic AI for Biology" workshop; the public abstract does not provide clinical cases, wet-lab validation, or large-scale prospective testing results. Therefore, at this stage, Site4Drug is better viewed as a methodological framework: it proposes how to place multiple forms of evidence for target-site selection into a single inspectable workflow, rather than proving that it can already reliably improve the success rate of drug candidates.

The real test will come later. To enter drug R&D practice, systems of this kind must be repeatedly compared across known targets, failure cases, and different protein families to confirm that they can avoid sites that are biologically inaccessible or carry excessively high development risk; if used to influence major investments or preclinical decisions, they will also need clearer version control, records of data sources, and boundaries of responsibility. AI can make early judgments more orderly, but in drug discovery, order must still withstand experimental testing.