Complex prediction
Complex prediction answers the question: how exactly does a drug molecule sit inside a protein’s binding pocket? This 3D arrangement is called a binding pose, and knowing it is critical for designing better drugs.
What is molecular docking?
Imagine you have a lock (the protein pocket) and a key (the drug molecule). Molecular docking is the computational process of figuring out:
- Where the key fits into the lock (the binding site)
- How the key is oriented (the pose)
- How well the key fits (the binding energy)
Traditional docking programs (AutoDock, Glide) use physics-based scoring functions. AlphaFold 3 takes a fundamentally different approach: it uses deep learning trained on millions of known protein-ligand complexes.
How AlphaFold 3 works (simplified)
AlphaFold 3 uses a diffusion-based architecture. Here’s the simplified process:
- Input encoding — The protein sequence, ligand structure (SMILES), and any known templates are encoded into a joint representation.
- Pairwise attention — The model reasons about all pairwise interactions between protein residues and ligand atoms.
- Diffusion generation — Starting from random atom positions, the model iteratively refines the 3D coordinates through a denoising process (similar to how image diffusion models like DALL-E work, but in 3D).
- Confidence scoring — The model outputs confidence metrics for the prediction.
The semi-automated workflow
OpenDDE uses a semi-automated workflow for complex prediction due to AlphaFold Server’s terms of service:
- Prepare — Select a target and ligand in OpenDDE. Click “Predict complex”. OpenDDE generates an AlphaFold 3 job JSON file.
- Submit — Copy the JSON and submit it to the AlphaFold Server. This typically takes 5–30 minutes.
- Upload — Download the results (CIF file) from AlphaFold Server and upload them back to OpenDDE.
- Analyze — OpenDDE renders the complex in 3D and provides confidence metrics, contact analysis, and AI interpretation.
Why semi-automated? AlphaFold Server currently requires manual submission through their web interface. As API access becomes available, OpenDDE will automate this step entirely.
Understanding confidence scores
AlphaFold 3 provides several confidence metrics:
| Metric | Range | What it means |
|---|---|---|
| iPTM | 0–1 | Interface predicted TM-score. Measures confidence in the protein-ligand interface. Above 0.8 is high confidence; below 0.5 is unreliable. |
| pLDDT | 0–100 | Per-residue confidence. Above 90 is very high; 70–90 is confident; below 50 is low confidence (often disordered regions). |
| PAE | 0–31 Å | Predicted aligned error. Lower is better. Measures expected position error between residue pairs. |
Limitations of predicted structures
- Not experimental — Predicted complexes are computational models, not X-ray crystal structures. Always validate with experiments when possible.
- Single conformation — The prediction shows one possible binding mode. In reality, the drug may bind in multiple orientations.
- No water molecules — Water-mediated interactions are not explicitly modeled but can be important for binding.
- Novel scaffolds — Predictions may be less reliable for molecules very different from the training data.
- Flexible loops — Regions of the protein that are intrinsically disordered may be incorrectly positioned in the complex.