11.2 Docking and scoring

Principles of Docking

Molecular docking is a computational method that predicts how a small molecule (ligand) binds to a target protein. It answers two key questions: where does the ligand sit in the binding site, and how tightly does it bind? This makes docking a cornerstone of structure-based drug design, since you can virtually screen millions of compounds against a target without synthesizing any of them first.

Docking algorithms work by searching for the most energetically favorable binding pose of a ligand within a protein's binding site, then estimating the strength of that interaction using scoring functions.

Docking Algorithms

Docking algorithms explore the conformational space of the ligand-protein system to generate possible binding poses, then rank them by predicted binding affinity. The main search strategies include:

• Systematic search: Exhaustively samples all possible orientations and conformations. Thorough but computationally expensive.
• Stochastic methods (Monte Carlo): Randomly samples conformations and accepts or rejects them based on energy criteria. Good at escaping local energy minima.
• Genetic algorithms: Treat ligand poses as "individuals" in a population that evolve through mutation and crossover toward lower-energy solutions. Used in GOLD and AutoDock.
• Incremental construction: Breaks the ligand into fragments, docks a rigid core fragment first, then rebuilds the rest piece by piece. Used in FlexX.

Widely used docking programs include AutoDock, GOLD, Glide (Schrödinger), and FlexX, each implementing different combinations of these strategies.

Rigid vs. Flexible Docking

Rigid docking treats both ligand and protein as fixed shapes, allowing only translational and rotational movement. It's fast but can miss conformational changes that happen upon binding.

Flexible docking allows the ligand and/or protein to change shape during the docking process. This captures induced fit effects, where the protein adjusts its conformation to accommodate the ligand. In practice, ligand flexibility is routinely included (sampling rotatable bonds and ring conformations), while protein flexibility is harder to incorporate because of the enormous conformational space involved. Protein flexibility adds significant computational cost but can be critical for targets with flexible active sites.

Binding Site Identification

Before you can dock anything, you need to know where on the protein the ligand binds. Accurate binding site identification focuses the search space and improves both speed and reliability. Three main approaches exist:

• Knowledge-based: Uses known ligand-binding data from homologous proteins. If a closely related protein has a co-crystal structure with a ligand, you can infer where your target's binding site is.
• Geometric methods: Detect cavities and clefts on the protein surface that could accommodate a ligand. Tools like PASS and SURFNET map these pockets computationally.
• Energy-based methods: Place small chemical probes across the protein surface and calculate interaction energies to find favorable binding regions. Q-SiteFinder and FTMap are common tools for this.

Ligand Preparation for Docking

Ligands must be properly prepared before docking, or you'll get unreliable results. The typical preparation workflow involves:

1. Generate 3D coordinates from 2D chemical structures (most compound databases store molecules as 2D).
2. Assign protonation states and tautomers appropriate for physiological pH (~7.4). A carboxylic acid, for example, will typically be deprotonated.
3. Minimize geometry to remove steric clashes and achieve a reasonable starting conformation.
4. Generate multiple conformers to sample the ligand's flexibility, since the bioactive conformation may differ from the lowest-energy one.

Tools like LigPrep (Schrödinger) and OMEGA (OpenEye) automate these steps.

Docking Protocols

A docking protocol defines the complete set of steps and parameters for a docking experiment. Developing a robust, validated protocol is essential because docking results are sensitive to how the system is set up. The key components are protein preparation, grid generation, parameter settings, and handling of protein flexibility.

Protein Preparation for Docking

Raw crystal structures from the PDB are rarely ready for docking out of the box. Protein preparation typically involves:

1. Add missing hydrogens and optimize their positions (X-ray structures usually don't resolve hydrogen atoms).
2. Assign protonation states for ionizable residues (His, Glu, Asp, Lys) at the relevant pH.
3. Fix missing or incorrect residues/atoms, including incomplete side chains or loop regions.
4. Minimize the structure to relieve steric clashes introduced during model building.

Tools like the Protein Preparation Wizard (Schrödinger) and CHARMM-GUI streamline this process.

Grid Generation and Optimization

Most docking algorithms use a grid-based approach to speed up energy calculations. Instead of recalculating ligand-protein interactions from scratch for every pose, interaction energies are pre-computed at regularly spaced grid points covering the binding site.

Setting up the grid involves:

• Defining the grid box (size and center) so it fully covers the binding site with some buffer space.
• Setting grid spacing to balance accuracy against computational cost. Finer spacing gives more precise energies but takes longer.
• Pre-calculating interaction energies for different atom types at each grid point.

Grid optimization techniques like focusing (using a coarse grid first, then refining around promising regions) and softening (reducing repulsive penalties slightly to allow minor overlaps) help improve both accuracy and efficiency.

Docking Parameter Settings

Docking parameters control how the algorithm runs and directly influence your results. Key parameters include:

• Search algorithm settings: Number of independent runs, population size (for genetic algorithms), number of Monte Carlo steps.
• Scoring function: Which function to use and how to weight its individual terms.
• Ligand flexibility: Number of rotatable bonds to sample, ring conformations to explore.
• Protein flexibility: Whether to allow side-chain rotamers, induced fit, or ensemble docking.

Optimal settings are typically determined through benchmarking against known crystal structures. You dock ligands with experimentally known binding poses and check whether the algorithm can reproduce them accurately.

Handling Protein Flexibility in Docking

Accounting for protein flexibility remains one of the biggest challenges in docking. Several strategies exist, each trading off accuracy against computational cost:

• Soft docking: Reduces van der Waals repulsion penalties, allowing small overlaps between ligand and protein atoms. Simple but crude.
• Side-chain flexibility: Samples different rotamer states for selected residues in the binding site. Captures local rearrangements without full backbone flexibility.
• Ensemble docking: Docks ligands against multiple protein conformations (from crystal structures or MD simulations). Captures a range of receptor states but multiplies computational cost.
• Induced fit docking (IFD): Allows both ligand and protein to change conformation during the docking process. Most realistic but most expensive.

Scoring Functions

Scoring functions are mathematical models that estimate the binding affinity between a ligand and protein. They serve two purposes: selecting the best pose from a set of docked orientations, and ranking different compounds by predicted affinity during virtual screening. The central challenge is balancing accuracy with speed, since virtual screens may evaluate millions of poses.

Types of Scoring Functions

Four main categories exist:

• Force field-based: Use classical molecular mechanics force fields (AMBER, CHARMM) to calculate van der Waals, electrostatic, and bonded interaction energies. Physically grounded but can be slow and may miss solvation effects.
• Empirical: Express binding affinity as a weighted sum of individual interaction terms (hydrogen bonds, hydrophobic contacts, rotatable bond penalties, etc.). The weights are fitted to experimental binding data from a training set of protein-ligand complexes.
• Knowledge-based: Derive statistical potentials from databases of known protein-ligand structures. If a particular atom pair is found at a given distance more often than expected by chance, that distance is assigned a favorable energy. These functions capture interaction preferences implicitly.
• Machine learning-based: Train algorithms (random forests, neural networks, graph neural networks) on protein-ligand interaction data to predict binding affinity. Growing in popularity as more structural and affinity data become available.

Empirical vs. Knowledge-Based Scoring

Empirical scoring functions decompose binding free energy into a linear combination of energy terms, with coefficients fitted to experimental $\Delta G$ values. They assume additivity of contributions, which is a simplification. Examples: GlideScore, ChemScore, X-Score.

Knowledge-based scoring functions extract atom-pair distance preferences from structural databases and convert observed frequencies into pseudo-energies using Boltzmann statistics. They don't require experimental affinity data for training, only structural data. Examples: DrugScore, PMF (Potential of Mean Force), DSX (DrugScore eXtended).

The key trade-off: empirical functions directly target affinity prediction but depend heavily on training set quality, while knowledge-based functions are less biased by training data but may not correlate as well with actual binding affinities.

Consensus Scoring Approaches

Consensus scoring combines results from multiple scoring functions to improve prediction robustness. The rationale is that different functions have different strengths and weaknesses, so their agreement is more reliable than any single function alone.

Common strategies:

• Rank-by-number: Count how many scoring functions place a compound among the top hits. Compounds flagged by multiple functions rank higher.
• Rank-by-rank: Average (or weight) each compound's rank across all scoring functions.
• Rank-by-vote: Each scoring function "votes" for its top-ranked compounds; compounds with the most votes win.

Consensus scoring reduces both false positives and false negatives compared to using a single scoring function.

Limitations of Scoring Functions

Scoring functions have well-known limitations that you should keep in mind:

• They use simplified representations of complex physicochemical phenomena. Real binding involves polarization, charge transfer, and many-body effects that most scoring functions ignore.
• Entropic contributions are poorly captured. Conformational entropy loss upon binding and desolvation entropy are difficult to estimate accurately.
• Performance depends on the quality and diversity of training data. Functions trained on a narrow set of protein families may not generalize well.
• Novel chemotypes that fall outside the training set's chemical space are often scored inaccurately.

Ongoing efforts to address these limitations include incorporating more physics into the models, leveraging larger datasets with machine learning, and developing hybrid approaches.

Evaluating Docking Results

Evaluating docking performance is essential for knowing whether your protocol is trustworthy before using it to make real drug design decisions. Evaluation focuses on two questions: Can the method reproduce known binding modes? Can it distinguish active compounds from inactive decoys?

Binding Pose Analysis

Binding pose analysis compares predicted ligand orientations to experimentally determined structures (usually from X-ray crystallography).

• RMSD (Root-Mean-Square Deviation): The most common metric. It calculates the average distance between corresponding atoms in the predicted and experimental poses. An RMSD below 2.0 Å is generally considered a successful prediction.
• Tanimoto Combo (TC) score: Quantifies overlap based on both atomic positions and molecular shape. Useful when RMSD alone doesn't capture the quality of the prediction well (e.g., symmetric molecules).

Visual inspection remains important. Always check for reasonable hydrogen bonding patterns, absence of steric clashes, and proper orientation of key pharmacophoric groups.

Interaction Fingerprints

Interaction fingerprints encode the specific contacts between a ligand and protein as binary (present/absent) or count-based vectors. They capture interactions like hydrogen bonds, hydrophobic contacts, salt bridges, and $\pi$-stacking.

These fingerprints are useful for:

• Comparing binding modes across different ligands docked to the same target
• Assessing whether a docking protocol reproduces the correct interaction pattern, not just the correct geometry
• Clustering docked poses by interaction similarity

Common methods include SIFt (Structural Interaction Fingerprints) and PLIF (Protein-Ligand Interaction Fingerprints).

Enrichment Factors and ROC Curves

These metrics evaluate how well docking and scoring can prioritize active compounds over inactive decoys in a virtual screen.

Enrichment factor (EF) measures how much better the method performs compared to random selection. For example, if 1% of a database consists of actives, and docking places 10% actives in the top 1% of the ranked list, the $EF_{1\%} = 10$. Higher is better.

ROC curves plot the true positive rate (sensitivity) against the false positive rate (1 - specificity) across all possible rank thresholds. The area under the ROC curve (AUC) summarizes overall performance:

• AUC = 0.5 means random performance (no discrimination)
• AUC = 1.0 means perfect separation of actives from decoys

An AUC above 0.7 is generally considered useful, though the early enrichment (top 1-5% of the ranked list) often matters more in practice than overall AUC.

Experimental Validation of Docking Predictions

Computational predictions ultimately need experimental confirmation. Key validation techniques include:

• X-ray crystallography: Solves the 3D structure of the protein-ligand complex, providing direct evidence of the binding mode. The gold standard for pose validation.
• Isothermal titration calorimetry (ITC): Measures thermodynamic binding parameters ($\Delta G$, $\Delta H$, $\Delta S$, $K_d$) in a single experiment. Confirms whether predicted binders actually interact with the target.
• Surface plasmon resonance (SPR): Measures binding kinetics ($k_{on}$, $k_{off}$) and affinity ($K_D$) in real time. Useful for confirming hits from virtual screens.

Experimental validation feeds back into the computational workflow by helping refine docking protocols and identify where scoring functions fail.

Applications of Docking

Docking is used across multiple stages of drug discovery, from initial hit finding through lead optimization. It's most powerful when combined with other computational and experimental methods.

Virtual Screening for Lead Discovery

Docking-based virtual screening involves docking a large compound library (often millions of molecules) into a target's binding site and ranking compounds by predicted affinity. The top-scoring compounds are then purchased or synthesized and tested experimentally.

This approach enables rapid, cost-effective exploration of chemical space. Notable successes include the discovery of HIV-1 protease inhibitors and influenza neuraminidase inhibitors through docking-based virtual screens. Typical hit rates from well-designed virtual screens range from 1-20%, compared to ~0.01-0.1% for random screening.

Structure-Based Drug Design

Structure-based drug design (SBDD) is an iterative cycle: solve or model the target structure, dock and design compounds, synthesize and test them, solve the co-crystal structure of the best hits, then use that structural information to design the next round of compounds.

Docking contributes by revealing which interactions drive binding and suggesting specific modifications to improve potency, selectivity, or drug-like properties. Drugs developed with significant SBDD input include the kinase inhibitor imatinib (Gleevec) for chronic myeloid leukemia and the HIV-1 protease inhibitor nelfinavir (Viracept).

Protein-Protein Interaction Inhibitor Design

Protein-protein interactions (PPIs) are notoriously difficult drug targets because their interfaces are typically large, flat, and lack deep pockets. Docking can help identify small molecules that bind to hotspot residues, which are the few key residues that contribute most of the binding energy at the PPI interface.

This application requires careful attention to binding site definition, flexibility handling, and scoring function choice, since PPI interfaces differ substantially from traditional enzyme active sites. Successful PPI inhibitors developed with docking support include venetoclax (Venclexta), a Bcl-2 inhibitor for chronic lymphocytic leukemia, and RG7112, an MDM2-p53 interaction inhibitor.

Docking in Fragment-Based Drug Discovery

Fragment-based drug discovery (FBDD) screens small molecular fragments (typically 150-300 Da) that bind weakly to the target, then optimizes them into potent drug candidates through fragment growing, linking, or merging.

Docking supports FBDD by:

• Identifying fragment binding sites on the target
• Predicting fragment binding modes to guide optimization strategies
• Evaluating whether two fragments can be linked or merged into a single, more potent molecule

Because fragments are small and bind weakly (often $K_d$ in the mM range), docking in FBDD demands high pose prediction accuracy. Drugs developed through FBDD with docking support include vemurafenib (Zelboraf), a BRAF kinase inhibitor for melanoma, and venetoclax.

Challenges and Advancements

Despite significant progress, several fundamental challenges limit docking accuracy. Active research in each of these areas is pushing the field forward.

Accounting for Water Molecules in Docking

Water molecules in the binding site can mediate critical hydrogen bonds between ligand and protein. Ignoring them can lead to incorrect pose predictions and inaccurate affinity estimates. The challenge is that water molecules are dynamic and their positions are hard to predict.

Three strategies exist:

• Explicit water docking: Include specific crystallographic water molecules as part of the receptor. Works well when water positions are known and conserved, but requires prior knowledge.
• Implicit water models: Use scoring functions that account for desolvation and water-mediated effects without placing individual water molecules. Less accurate but more general.
• Hybrid approaches: Combine explicit placement of high-confidence waters with implicit treatment of bulk solvent. Balances accuracy and generality.

Handling Protein Flexibility and Induced Fit

Many proteins undergo significant conformational changes upon ligand binding. A docking protocol that treats the protein as rigid may fail entirely for such targets. Recent advances include:

• Ensemble docking against multiple conformations from X-ray structures or MD trajectories
• Induced fit docking (IFD) protocols that alternate between ligand docking and protein minimization
• Selective flexibility approaches that allow movement only in key regions (active site loops, specific side chains) while keeping the rest of the protein rigid

The trade-off is always between conformational sampling quality and computational cost.

Improving Scoring Function Accuracy

Predicting absolute binding affinities ($\Delta G_{bind}$) remains the single hardest problem in docking. Current efforts focus on:

• Better physics: Incorporating polarization, charge transfer, and explicit solvation effects into energy models.
• Machine learning: Training on the growing volume of structural and affinity data. Deep learning models like those using 3D convolutional neural networks or graph neural networks show promise for both pose prediction and affinity estimation.
• Hybrid approaches: Combining physics-based energy terms with data-driven corrections to get the best of both worlds.

Integrating Docking with Molecular Dynamics Simulations

MD simulations model the time-dependent behavior of protein-ligand systems at atomic resolution, capturing conformational dynamics that static docking cannot. Combining the two methods improves both pose prediction and affinity estimation.

Integration approaches include:

1. Generating conformational ensembles: Run MD on the apo protein or a protein-ligand complex, cluster the resulting conformations, and use them for ensemble docking.
2. Refining docked poses: Take top-scoring docked poses and run short MD simulations to relax the complex. Poses that remain stable are more likely to be correct.
3. Free energy calculations: Use docked poses as starting points for rigorous free energy methods like free energy perturbation (FEP) or thermodynamic integration (TI) to obtain more accurate binding affinity predictions.

This integration is computationally demanding but represents one of the most promising directions for improving the predictive power of computational drug discovery.