AlphaFold Structures for Docking

The predicted local distance difference test (pLDDT) score is a per-residue confidence metric output by AlphaFold2, critical for docking triage. Residues with pLDDT < 50 should be treated as disordered or unreliable for rigid docking. Binding site pLDDT is a stronger predictor of docking success than overall model confidence. A structured approach involves:

• Masking low-confidence loops during grid generation
• Using the predicted aligned error (PAE) matrix to assess domain-domain orientation certainty
• Prioritizing targets where the orthosteric pocket has a mean pLDDT > 80

AlphaFold2 is trained on the Protein Data Bank (PDB), which is enriched for holo (ligand-bound) conformations. Consequently, predicted structures often adopt a binding-ready state, which can be advantageous for docking but may overestimate pocket openness. Key considerations include:

• AlphaFold structures frequently outperform experimental apo (unbound) structures in retrospective docking benchmarks
• The model may hallucinate a collapsed pocket for targets that require significant induced fit
• Side-chain rotamer accuracy in the binding site is the primary determinant of docking pose quality

AlphaFold-Multimer enables the prediction of protein-protein and protein-peptide complexes, extending docking to interfaces. For virtual screening, this allows targeting of:

• Protein-protein interaction (PPI) interfaces, historically considered undruggable
• Allosteric pockets revealed at domain-domain junctions
• Oligomeric assembly states that may expose cryptic binding sites absent in monomeric predictions
• The interface predicted TM-score (ipTM) metric quantifies the confidence of the predicted complex stoichiometry and orientation

A single static AlphaFold model cannot capture conformational heterogeneity. Ensemble docking mitigates this by using multiple structures:

• Subsampling the multiple sequence alignment (MSA) by reducing depth or diversity generates structurally diverse models
• Clustering output structures by backbone RMSD identifies distinct conformational states
• Docking against a representative ensemble of 3-5 diverse pocket conformations improves hit rate over any single structure
• This approach partially compensates for the lack of true protein flexibility in rigid docking protocols

In rigorous benchmarks, AlphaFold-derived structures demonstrate near-experimental performance for structure-based virtual screening when the binding site is well-resolved. Critical findings include:

• Enrichment factors for AlphaFold models can match those of high-resolution crystal structures for targets with pLDDT > 80 in the pocket
• Performance degrades sharply for metalloproteins and targets with complex cofactor coordination
• Retrospective docking of known actives is an essential validation step before committing to a prospective screen
• Cross-docking accuracy—the ability to dock ligands from other structures—is a stringent test of model utility

AlphaFold predictions are not a wholesale replacement for experimental structures but a complementary starting point. Hybrid workflows combine:

• AlphaFold models as initial phasing models for molecular replacement in X-ray crystallography
• Predicted structures as docking templates for fitting into cryo-EM density maps
• Iterative refinement where experimental density is used to correct side-chain rotamers in the predicted model before docking
• This synergy accelerates structure determination for novel targets while grounding computational predictions in empirical data

The computational workhorse that uses a 3D receptor structure to dock and rank candidate ligands. When the receptor is an AlphaFold prediction, the workflow shifts from experimental to in silico structural biology. Key steps include:

• Target preparation: Adding hydrogens, assigning protonation states, and defining the binding site on the predicted model.
• Library preparation: Generating tautomers, stereoisomers, and low-energy 3D conformers for each ligand.
• Docking: Sampling ligand poses within the binding pocket and scoring them to estimate binding affinity.

A mathematical model that approximates the free energy of binding for a protein-ligand complex. Scoring functions are the critical decision-makers in docking. They must be robust to the subtle inaccuracies in AlphaFold side-chain positions. Common classes include:

• Force-field-based: Calculate van der Waals and electrostatic interaction energies (e.g., DOCK, GoldScore).
• Empirical: Sum weighted terms for hydrogen bonds, hydrophobic contacts, and entropy penalties, trained on known affinity data (e.g., ChemScore, GlideScore).
• Knowledge-based: Derive statistical potentials from the frequency of atom-pair contacts in crystal structures (e.g., PMF, ITScore).

A single static AlphaFold structure represents a frozen snapshot of a dynamic protein. To account for conformational selection, use ensemble docking:

• Generate multiple receptor conformations via molecular dynamics (MD) simulations or by sampling different AlphaFold multimer seeds.
• Dock the ligand library against each receptor conformation independently.
• Combine results to identify ligands that bind consistently across the ensemble, enriching for true positives that tolerate receptor flexibility.

The primary metric for validating a docking campaign. EF quantifies how effectively a scoring function concentrates known active compounds in the top fraction of a ranked database compared to random selection.

• EF1%: The ratio of actives found in the top 1% of ranked results vs. a random 1% sample. An EF1% of 10 means the method is 10x better than random.
• BEDROC: A more robust metric that weights early recognition more heavily, penalizing late discovery of actives.
• Always validate an AlphaFold-based docking protocol by its ability to enrich known ligands before screening novel compounds.

A deep learning methodology that accelerates billion-scale virtual screening by training a neural network surrogate model on a small, randomly sampled subset of docking results. The workflow:

• Iterative training: Dock a small fraction of the library, train a model to predict docking scores from molecular fingerprints, then use the model to predict scores for the remaining compounds.
• Recall-focused: The model is tuned to prioritize high recall, ensuring that promising candidates are not erroneously discarded.
• This approach is essential when using computationally expensive, physics-based scoring functions on AlphaFold structures to explore ultra-large chemical spaces.

AlphaFold outputs per-residue and inter-residue confidence scores that are critical for docking triage:

• pLDDT (predicted Local Distance Difference Test): A per-residue confidence score (0-100). Low pLDDT regions (<50) are often disordered and should be treated with caution as binding sites.
• PAE (Predicted Aligned Error): Measures the expected position error between residue pairs. High PAE between domains indicates uncertain relative orientation.
• Best practice: Prioritize targets with high pLDDT (>80) in the binding pocket and low inter-domain PAE for reliable docking results.