Site4Drug: Predicting Drug-Binding Target Sites with an AI Agent

Site4Drug is an AI agent that predicts ranked, modality-aware drug-binding target sites on proteins by integrating diverse biological evidence to overcome the ambiguity and failure rates associated with selecting actionable intervention regions, particularly for membrane proteins.

The Gist

Imagine you are a locksmith trying to pick a lock on a massive, complex safe (a protein). In the past, scientists often assumed they already knew exactly which part of the safe to turn (the "binding site") and just focused on making the key (the drug). But often, the real problem isn't making the key; it's figuring out where to even put it.

This is especially tricky for "membrane proteins," which are like safes built into a wall. Some parts are hidden inside the wall, some are covered in sticky tape (sugar coatings), and some are just too crowded to reach. If you try to pick a lock that's covered in tape or buried in the wall, your key won't work, no matter how good it is.

Site4Drug is a new AI "detective" designed to solve this "where" problem before you even start making keys.

The Detective's Toolkit

Instead of just guessing, Site4Drug acts like a super-smart agent that gathers clues from the protein's "ID card" (its amino acid sequence). It doesn't need a 3D blueprint of the safe; it can figure things out just by reading the text. Here is how it works:

1. Checking the Map (Topology): It looks at the protein to see which parts are sticking out (accessible) and which are buried inside the wall (transmembrane). It knows you can't pick a lock that's inside the wall.
2. Scanning for Sticky Tape (PTMs): It checks for "Post-Translational Modifications" (PTMs), which are like sticky notes or heavy tape (like sugars or phosphates) that might cover the lock. If a spot is covered in tape, the detective marks it as "risky" or "blocked."
3. Looking for Special Patterns (Motifs): It scans for specific patterns that are usually important for the protein's job. It knows that tampering with these might break the protein, so it flags them with caution.
4. Checking for "Sticky Pairs" (Disulfides): It counts "cysteines" (a type of amino acid) to see if the protein is tied together with internal knots. If a spot is part of a tight knot, it might be too rigid to bind a drug.

The "Agent" Approach

What makes Site4Drug special is that it doesn't just spit out a list of numbers. It acts like a team of specialists having a meeting:

• The Biologist checks if the spot makes sense for a biological drug (like an antibody).
• The Chemist checks if the spot makes sense for a chemical drug (like a pill).
• The Risk Manager points out all the red flags (e.g., "This spot is covered in sugar tape!").

The final "Decision Agent" listens to everyone and produces a ranked report. It doesn't just say "Here is the best spot." It says:

• "Here is the best spot."
• "Here is why we think it's good."
• "Here are the risks (like 'it's covered in sugar')."
• "Here is our confidence level."

This makes the process auditable. If the drug fails later, scientists can look at the report and say, "Ah, we picked a spot that was covered in sugar tape. That's why it failed," rather than just guessing.

How Well Does It Work?

The authors tested this detective on two types of locks:

1. Small Molecule Pockets: These are tiny holes inside proteins where chemical drugs fit. Site4Drug found these spots almost as well as traditional tools that require a 3D map of the protein, even though Site4Drug didn't use a 3D map at all!
2. Antibody Epitopes: These are the "handles" on the outside of proteins where antibodies grab on. Site4Drug successfully identified these handles by looking at the sequence clues.

A Real-World Check: Beyond these computer tests, there is an encouraging sign mentioned in the paper's conclusion: the authors note that one of Site4Drug's predictions was verified through actual laboratory experiments. While the specific details of this experiment were not disclosed in the paper, this single real-world confirmation is significant. It shows that the AI's "detective work" isn't just a theoretical success on a computer screen, but has at least one piece of proof that it can correctly identify a target in the real world.

The "Handoff"

Once Site4Drug finds a good spot, it doesn't stop there. It can hand off that information to other tools to actually design the drug.

• If it found a "pocket," it can send the coordinates to a tool that designs small molecules.
• If it found an "epitope," it can send the coordinates to a tool that designs antibodies or peptides.

The Bottom Line

Site4Drug is like a smart GPS for drug discovery. Instead of driving blindly and hoping you find a parking spot (a binding site), it analyzes the street signs, traffic, and road conditions to tell you exactly where to park, why that spot is good, and what potential hazards (like construction or no-parking zones) you should watch out for. It makes the first, most confusing step of drug discovery clearer, safer, and easier to understand.

Technical Summary

Technical Summary: Site4Drug – Predicting Drug-Binding Target Sites with an AI Agent

1. Problem Statement

The selection of a specific protein region for therapeutic intervention (target site selection) is often a more ambiguous and failure-prone bottleneck than the subsequent selection of a binding molecule. This challenge is particularly acute for membrane proteins, where accessibility, topology, and post-translational modifications (PTMs) constrain actionable regions.

Current discovery pipelines frequently assume the binding site is known, focusing instead on docking, screening, or binder generation (e.g., BoltzGen, DrugCLIP). However, teams often stall at the upstream question: where and with what modality (e.g., small molecule vs. antibody/peptide) should the protein be targeted?

• Limitations of Existing Resources: Curated databases like IEDB aggregate immune epitope evidence from specialized experimental settings that are context-dependent and incomplete for therapeutic site selection. Similarly, "known pockets" are often operationally defined by residues contacting co-crystallized ligands; this definition is brittle for proteins with no prior ligand-bound structures (e.g., cancer neoantigens).
• Limitations of Conventional Tools: Geometric tools (e.g., fpocket) and learned structural representations (e.g., RAPID-Net) primarily identify canonical binding pockets but struggle to incorporate diverse metadata or define sites for alternative drug modalities.
• The Gap: There is a lack of auditable, evidence-integrated systems that can propose candidate regions, rank them based on feasibility constraints, and explicitly log the reasoning behind site selection to distinguish between binding model failures and incorrect site choices.

2. Methodology: The Site4Drug Agent

Site4Drug is a modality-aware AI agent designed to reframe site selection as a constraint-first, evidence-integrated decision problem. It operates via a two-module pipeline:

Module 1: Evidence Aggregation and Region Discovery

This module constructs a residue- and region-level evidence summary from a protein sequence alone, without requiring 3D structural input. It processes three classes of signals:

1. Topology & Accessibility Priors: A sliding-window Kyte–Doolittle hydropathy profile and a heuristic transmembrane (TM) detector derive a coarse accessibility prior. Regions overlapping TM segments are labeled tmd/restricted, while others are outside/exposed.
2. PTM Prediction & Neighborhood Masking: PTM sites are extracted using a MusiteDeep-backed extractor with rule-based augmentation. Each site is expanded into a typed local mask (e.g., residues 208–214 for a phosphoserine at 211). The system records overlaps between candidate regions and these masks, including PTM type and local density.
3. Motif Hits & Cysteine/Disulfide Proxy: The system queries ScanProsite for motif spans, flagging overlaps as "motif-overlap" due to potential functional constraints. Cysteine counts are tracked as a lightweight proxy for disulfide-constrained segments.

Candidate Generation & Ranking:
Conditioned on the sequence and the compact evidence summary, a Large Language Model (LLM) proposes a ranked JSON list of candidate regions.

• Scoring Logic: Candidates are scored using a mode-specific heuristic: $S_0(r) = s_{mode}(r) - p_{TM}(r) - p_{PTM}(r) - p_{motif}(r)$.
  • Epitope mode: Favors polar, non-TM, PTM-light windows.
  • Pocket mode: Places more weight on hydrophobicity and applies weaker PTM penalties.
• Risk Flags: A separate risk vector generates typed annotations (e.g., TM-overlap, glyco-mask-overlap, disulfide-constrained) alongside the score.
• Specialist Adjudication: A panel of agents (BioAgent, ChemAgent, RiskAgent) provides secondary critique on claims and evidence. A DecisionAgent synthesizes these critiques into a final modality decision and adjusted ranking, strictly adhering to the evidence present in the context.

Module 2: Modality-Specific Design Handoff

Ranked outputs from Module 1 serve as an interface to downstream design tools:

• Epitope Mode: Top-ranked spans are passed to peptide or antibody-binder design tools (e.g., BoltzGen).
• Pocket Mode: Top-ranked pocket regions are fed into small-molecule design or scoring tools (e.g., DrugCLIP, BindCLIP).

3. Key Contributions

• Constraint-First Site Discovery: Site4Drug introduces a framework that prioritizes feasibility constraints (topology, PTMs, motifs) before proposing candidates, ensuring that chemically plausible but biologically occluded sites are avoided.
• Modality Agnosticism: Unlike tools that assume a specific drug type, Site4Drug can recommend a binding modality (antibody/peptide vs. small molecule) based on the same evidence used for site discovery.
• Auditability: The system outputs a structured report including recommended modality, ranked candidates, explicit evidence summaries, risk flags, and a traceable decision log, making the site-selection step debuggable.
• Sequence-Only Operation: The system achieves structural plausibility without direct protein structure input, relying on sequence-derived evidence aggregation.

4. Experimental Results

The authors evaluated Site4Drug using a curated dataset of 89 targets (55 small molecules, 26 antibodies, 8 mixed-modality) rather than a single fixed benchmark, acknowledging the lack of definitive ground truth for novel site discovery.

Small-Molecule Targets (Pockets)

• Setup: Evaluated on 63 targets with known co-crystal structures (RCSB). Ground truth was defined as residues within 4 Å of the ligand, remapped to the input sequence.
• Performance: Site4Drug achieved pocket-site localization performance comparable to fpocket (a structure-based baseline) even without direct structure information.
• Significance: Statistically significant top-1 results were primarily kinases, which possess characteristic recurrent pockets.
• Ablation: A sequence-only baseline (without explicit TM/PTM/motif evidence) yielded significantly lower overlap (3/63 cases at top-1) compared to the full pipeline, demonstrating the value of the explicit evidence aggregation.

Antibody Targets (Epitopes)

• Setup: Evaluated on 26 targets from the ABCD database with annotated epitope positions.
• Performance: Site4Drug produced significant overlap ($p < 0.05$) with annotated epitopes in 8/26 cases (top-1) and 11/26 cases (top-5).
• Implication: Despite the sparsity and noise of the benchmark, the system successfully recovered meaningful epitope-like regions from sequence-derived evidence.

Structural Reliability

• Validation: Using AlphaFold3 to predict structures for the 63 pocket-mode targets, the authors analyzed the pLDDT (predicted Local Distance Difference Test) values of the predicted sites.
• Finding: The top-1 predicted sites generally exhibited higher mean pLDDT (indicating higher structural confidence) than the broader top-5 set. This suggests that the sequence-based evidence aggregation implicitly recovers signals correlated with structural reliability.

Experimental Validation

• Confirmation: In the paper's conclusion, the authors note that a single Site4Drug prediction was experimentally validated, confirming the system's potential to identify actionable targets; however, the specific experimental details and target identity were not disclosed in the report.

End-to-End Design (Proof of Concept)

• Pocket Mode: Using the top-ranked EGFR pocket, DrugCLIP retrieved molecules structurally similar to known EGFR inhibitors (e.g., lapatinib), with a hypergeometric test showing high structural overlap ($p < 10^{-11}$).
• Epitope Mode: Using top-ranked EGFR epitopes, BoltzGen generated peptide binders. The resulting binder rankings correlated with known structural features (e.g., Domain III targeting) and membrane topology constraints.

5. Significance and Limitations

Significance:
The paper claims that Site4Drug improves the reliability and transparency of target-site selection by producing constraint-aware region proposals and audit-ready decision logs. It serves as an interpretable upstream interface for downstream binder or small-molecule design workflows. By identifying targetable regions alongside the evidence supporting each decision, it addresses the upstream bottleneck in drug discovery, particularly for targets with limited experimental information. While the system includes a single undisclosed experimental confirmation, its outputs should generally be treated as hypotheses for experimental follow-up.

Limitations:

• Data Scale: The validation datasets are modest in size, particularly for antibody-relevant targets where reliable site-level annotations are difficult to curate.
• Baseline Comparison: Direct comparison with existing structure-trained ML models is challenging due to potential data leakage from overlapping training databases (e.g., scPDB/RCSB).
• Structural Gating: While the system works without structure, pocket mode inherently depends on structural context. Some models require structural input, which can be token-intensive and less scalable.
• Quaternary Structure: The current model accepts single protein sequences and does not account for quaternary structures or multi-domain assemblies common in channel proteins.
• Supervision: Preliminary experiments with Supervised Fine-Tuning (SFT) showed shortcut behaviors (e.g., repeated N-terminal selections), suggesting that future work requires biologically grounded reward signals rather than simple formatting supervision.
• Concentration Dependency: The model does not currently account for concentration-dependent target engagement.

The authors emphasize that Site4Drug outputs should be treated as hypotheses for experimental follow-up, not as validated therapeutic candidates, and that downstream molecules require independent validation for binding, specificity, and safety.