---
name: alphafold3-drug-target-protocol
description: Predict small molecule drug-target binding modes by combining AlphaFold 3 protein structure prediction with ligand pose analysis and binding site assessment.
allowed-tools: WebFetch, Bash(python *), Bash(mkdir *), Bash(cp *), Bash(ls *), Bash(jq *), Bash(cd *)
---

# AlphaFold 3 Drug-Target Structure Predictor Protocol

## Purpose

Predict how a small molecule drug or compound binds to its target protein by combining AlphaFold 3 structure prediction with binding site identification and ligand orientation analysis. This workflow is for structural hypothesis generation, not rigorous docking.

## Inputs

Create an `inputs/` directory containing:

- `inputs/target.json`: AlphaFold 3 JSON for the target protein.
- `inputs/ligands.sdf` or `inputs/ligands.smiles`: Ligand structures to analyze (SDF preferred, SMILES acceptable).
- `inputs/known_binders.md` (optional): Known active compounds, co-crystal structures, or literature binding data.
- `inputs/metadata.md`: Target name, UniProt ID, known active site residues, known allosteric sites, therapeutic indication.

## Pre-Run Checks

1. Confirm research use is permitted.
2. Validate protein sequence uses standard amino acid codes.
3. Verify ligand molecules have valid SMILES or SDF format.
4. Check for covalent warheads in ligands (cysteine-reactive compounds) - note limitation.
5. Confirm compound names/IDs are consistent across files.

## Step 1: Target Structure Prediction

### Route A: AlphaFold Server

1. Create job with target protein.
2. Submit and wait for completion.
3. Download results to `outputs/target/`.
4. **Important**: For ligand-containing predictions, ensure ligand is included in the AF3 input if supported.

### Route B: Local AlphaFold 3

```bash
mkdir -p outputs/target
python run_alphafold.py \
  --json_path=inputs/target.json \
  --output_dir=outputs/target
```

## Step 2: Analyze Protein Structure

From the predicted structure:

1. **Extract confidence metrics**:
   - Overall pLDDT score
   - Per-domain pLDDT (identify high/low confidence regions)
   - PAE matrix for multi-domain proteins

2. **Identify binding sites**:

   Based on `inputs/metadata.md`:
   - Active site residues (e.g., catalytic triad positions)
   - Known allosteric sites
   - Literature-known binding pockets

   Or predict pockets using:
   ```python
   # Example usingFpocket (if available)
   from fpocket import Fpocket
   pockets = Fpocket().run(predicted_structure)
   ```

3. **Characterize binding site**:

```json
{
  "target": "KINASE_A",
  "predicted_binding_sites": [
    {
      "site_id": 1,
      "type": "active_site",
      "residues": [40, 41, 42, 80, 81, 82, 120, 121],
      "pLDDT_mean": 85.3,
      "pocket_volume_A3": 1200,
      "hydrophobic_fraction": 0.45,
      "notes": "Contains catalytic lysine"
    }
  ]
}
```

## Step 3: Ligand Preparation

1. Parse ligand structures from SDF or SMILES.
2. Generate 3D conformations if not present.
3. Optimize geometries.
4. Calculate basic properties:
   - Molecular weight
   - LogP (hydrophobicity)
   - H-bond donors/acceptors
   - Rotatable bonds

```json
{
  "compound_id": "COMPOUND_A",
  "smiles": "CC(C)Cc1ccc(cc1)CC(NCCc2ccc(Cl)cc2)=O",
  "mw": 432.3,
  "logp": 4.2,
  "hbd": 2,
  "hba": 3,
  "rotatable_bonds": 8
}
```

## Step 4: Binding Mode Prediction

**Note**: AlphaFold 3 can predict some ligand-containing complexes, but for detailed pose prediction, use specialized tools. This protocol focuses on structural hypothesis generation.

### Option A: AlphaFold 3 with Ligands

If ligand is in the supported list (ions, simple molecules):

```json
{
  "name": "kinase_with_ligand",
  "sequences": [
    {
      "protein_chain": {
        "sequence": "MVVFG...",
        "id": {"value": "A"}
      }
    }
  ],
  "ligands": [
    {
      "chemical_components": ["ATP"],
      "modeling": {
        "mode": "polymer",
        "id": {"value": "LIG"}
      }
    }
  ]
}
```

### Option B: Ligand Placement

For complex ligands, note that AF3 predictions may not accurately place the ligand. Document this limitation.

## Step 5: Binding Site Assessment

Assess druggability of identified sites:

| Metric | Good Druggability | Moderate | Poor |
|--------|-------------------|----------|------|
| Pocket volume | > 1000 ų | 500-1000 | < 500 |
| Hydrophobic fraction | > 40% | 20-40% | < 20% |
| pLDDT at site | > 80 | 60-80 | < 60 |

## Step 6: Generate Report

Write `outputs/drug_target_analysis.md`:

```markdown
# Drug-Target Binding Mode Analysis Report

## Target Protein
- Name: [name]
- UniProt ID: [ID]
- Source: [organism]
- Length: [N] residues
- Therapeutic relevance: [indication]

## Structure Prediction Quality
- Overall pLDDT: [mean]
- Active site pLDDT: [value]
- Confidence assessment: [High/Medium/Low]

## Known Binding Sites

### Active Site
- Key residues: [list with positions]
- Catalytic mechanism: [if known]
- Conserved residues: [if known]

### Allosteric Sites (if identified)
- Location relative to active site
- Known modulators: [if available]

## Predicted Pockets (if analyzed)

| Pocket ID | Volume (ų) | Hydrophobic % | Druggability |
|-----------|-------------|---------------|-------------|
| 1         | [value]     | [value]       | [Good/Mod/Poor] |

## Ligands Analyzed

### [Compound ID]
- Structure: [SMILES or filename]
- Properties: MW=[value], LogP=[value]
- Known activity: [if available]

## Binding Hypothesis
[Describe predicted binding orientation and key interactions]

## Structural Insights
- Binding site well-structured (pLDDT > 80): [yes/no]
- Flexible regions near binding site: [yes/no, locations]
- Known resistance mutations: [list if available]

## Limitations
- AlphaFold 3 ligand placement is limited to supported molecules
- Binding pose prediction is hypothesis-generating, not definitive
- Affinity/IC50 cannot be predicted from structure alone
- Does not account for:
  - Water-mediated interactions
  - Conformational changes upon binding
  - Protein dynamics
  - Solvent effects

## Recommendations
1. Use dedicated docking software (AutoDock, Glide, GOLD) for pose refinement
2. Validate predicted binding mode with mutagenesis
3. Check known resistance mutations against binding site
4. Consider molecular dynamics for binding kinetics
5. If polypharmacology expected, analyze off-target binding sites

## References
- AlphaFold 3: Abramson et al., Nature, 2024
- Druggability: Krasowski et al., Bioinformatics, 2011
- Docking: Eberhardt et al., Aust J Chem, 2021 (AutoDock Vina)
```

## Success Criteria

- Target protein structure is predicted with interpretable confidence.
- Binding sites are identified and characterized.
- Ligand properties are documented.
- Report provides testable hypotheses for experimental validation.
- Limitations are explicitly stated regarding predictive accuracy.

## Failure Modes

- Target has very low pLDDT in binding region → binding site may be disordered
- Ligand not supported by AF3 → note limitation, use alternative approach
- No clear binding pocket identified → consider alternative conformations or dynamics
- Prediction contradicts known SAR → investigate other binding modes

## References

- AlphaFold 3: Abramson et al., Nature, 2024
- Druggability prediction: Krasowski et al., Bioinformatics, 2011
- AutoDock Vina: Eberhardt et al., Aust J Chem, 2021
- Binding site detection: Le Guilloux et al., BMC Bioinformatics, 2009 (Fpocket)