Scoring Function

The ultimate biological truth a scoring function seeks to approximate. Binding affinity is the quantitative strength of a non-covalent interaction, typically expressed as the dissociation constant (Kd), inhibition constant (Ki), or Gibbs free energy change (ΔG).

• Experimental gold standards: Isothermal Titration Calorimetry (ITC) and Surface Plasmon Resonance (SPR) provide measured ΔG values for training and benchmarking.
• Prediction challenge: A perfect scoring function would perfectly correlate its computed score with experimentally measured binding affinity across diverse chemotypes.
• Key metric: Pearson correlation coefficient and RMSE between predicted and experimental pKd/pKi values.

The primary metric for evaluating pose prediction accuracy—the scoring function's ability to identify the correct binding geometry. RMSD measures the average atomic distance between a predicted docked pose and the experimentally determined binding mode.

• Success threshold: A pose is typically considered 'correct' if RMSD ≤ 2.0 Å from the crystal structure.
• Scoring power vs. ranking power: A function may achieve low RMSD (good pose identification) but poor correlation with affinity (bad ranking), or vice versa.
• Limitation: RMSD is a symmetric metric and can be disproportionately penalized by a single highly flexible moiety.

Scoring functions fall into distinct classes based on their underlying physical assumptions and mathematical formulation.

• Force field-based: Compute interaction energy using molecular mechanics terms (van der Waals, electrostatics). Physically rigorous but computationally expensive and often neglect solvation entropy.
• Empirical: Sum weighted terms (hydrogen bonds, hydrophobic contacts, entropy penalty) fitted to experimental binding data. Fast and interpretable but heavily training-set dependent.
• Knowledge-based: Derive statistical potentials from the frequency of atom-pair contacts in structural databases like the Protein Data Bank (PDB). Assumes observed geometries represent favorable interactions.

Modern approaches that learn complex, non-linear mappings from protein-ligand complex features to binding affinity, overcoming the additive linear assumptions of classical functions.

• Input features: Intermolecular interaction fingerprints, voxelized 3D grids, or graph representations from Message Passing Neural Networks (MPNNs).
• Δ-VinaRF20: A random forest scoring function trained to correct AutoDock Vina scores, demonstrating that classical functions can be boosted by ML.
• Equivariant networks: Architectures like SE(3)-Transformers guarantee that predicted scores are invariant to rotation and translation of the input complex, a critical physical constraint.

The rigorous physics-based alternative to scoring functions. FEP computationally 'mutates' one ligand into another through a series of non-physical intermediate states, calculating the relative binding free energy (ΔΔG) with high accuracy.

• Accuracy: Typically achieves RMSE of 0.8–1.5 kcal/mol, significantly outperforming docking scoring functions.
• Cost: Requires extensive Molecular Dynamics Simulation sampling, making it orders of magnitude slower than a single scoring function evaluation.
• Synergy: Scoring functions are used for high-throughput virtual screening; FEP is reserved for rigorous lead optimization of the top-ranked hits.

A retrospective metric that evaluates a scoring function's ability to prioritize true active compounds over decoys in a Virtual Screening campaign.

• Calculation: EFx% = (Actives found in top x% / Total actives) / (x% / 100). An EF1% of 10 means the scoring function is 10 times better than random at finding actives in the top 1% of the ranked library.
• BEDROC: A more robust alternative to EF that weights early recognition more heavily and is bounded between 0 and 1.
• Decoy bias: Enrichment metrics are highly sensitive to the physicochemical similarity of the decoy set to the actives; poorly chosen decoys inflate apparent performance.