Scoring Function

Classical scoring functions decompose the binding free energy into distinct physical contributions:

• Van der Waals (ΔG_vdw): Modeled by a Lennard-Jones potential to capture favorable steric contacts and unfavorable atomic clashes
• Electrostatics (ΔG_elec): Computed via a Coulombic potential with a distance-dependent dielectric function to approximate solvent screening
• Hydrogen Bonding (ΔG_hbond): A directional, angular-dependent term that rewards geometries matching ideal donor-acceptor distances
• Desolvation (ΔG_desolv): Estimates the energetic cost of stripping water molecules from the ligand and binding pocket, often using a solvent-accessible surface area model
• Entropic Loss (ΔS): Penalizes the restriction of rotatable bonds and translational/rotational degrees of freedom upon binding

These functions apply principles from molecular mechanics to compute the non-bonded interaction energy between protein and ligand atoms.

• DOCK Energy Score: Uses AMBER force field parameters with a Lennard-Jones 6-12 potential and Coulombic electrostatics
• GoldScore (GOLD): Employs a Tripos force field with a 4-8 van der Waals potential optimized for docking, plus an internal ligand strain term
• AutoDock Force Field: Incorporates a thermodynamic model with empirical weights calibrated against experimental binding constants

Key limitation: Force-field methods often neglect solvation effects and entropic contributions, requiring separate desolvation and entropy terms to be added ad hoc.

Empirical functions fit weighted structural descriptors to experimental binding affinity data using multivariate regression.

• ChemScore: Registers hydrogen bonds, metal interactions, lipophilic contact area, and a clash penalty, with coefficients derived from 82 protein-ligand complexes
• GlideScore (Schrödinger): Combines terms for lipophilic-lipophilic contacts, hydrogen bonding, metal ligation, and a penalty for buried polar groups, with a hydrophobic enclosure reward
• X-Score: Averages three scoring methods (HPB, HM, HS) to improve consensus accuracy, trained on 200 complexes

Advantage: Computationally fast and implicitly captures some solvation effects through the regression weights. Disadvantage: Accuracy depends heavily on the training set composition and may fail for novel chemotypes.

Also called statistical potentials, these functions derive energy terms from the frequency of atom-pair interactions observed in high-resolution protein-ligand crystal structures.

• PMF (Potential of Mean Force): Converts the radial distribution function g(r) of atom pairs into an energy potential using the inverse Boltzmann relation: ΔW(r) = -kBT ln[g(r)]
• DrugScore: Uses distance-dependent pair potentials for 17 atom types derived from the PDB, distinguishing solvent-accessible and buried environments
• ITScore: Iteratively refines the pair potentials by correcting the reference state until the predicted distribution matches the experimental one

Key insight: These potentials implicitly capture solvation, entropy, and many-body effects present in the training structures without explicit parameterization.

ML-based scoring functions learn complex, non-linear mappings from structural features to binding affinity, overcoming the functional form limitations of classical methods.

• RF-Score: A random forest trained on protein-ligand interaction fingerprints (atom-type pair counts within distance bins), achieving state-of-the-art performance without explicit energy terms
• NNScore: A neural network using AutoDock Vina energy terms plus atom-type pair counts as input features
• ΔVinaRF: An ensemble method that corrects the AutoDock Vina score using a random forest trained on the difference between Vina scores and experimental affinities
• OnionNet: A deep convolutional neural network that captures short-range and long-range contacts by organizing interatomic distances into hierarchical shells

Critical caveat: ML scoring functions are prone to overfitting and may memorize training complexes rather than learning transferable physics. Blind-test performance on truly novel targets remains a challenge.

Consensus scoring combines predictions from multiple orthogonal scoring functions to improve hit rate and reduce false positives.

• Rank-by-rank: Averages the rank position of each ligand across N scoring functions
• Rank-by-score: Averages the raw scores after normalization (Z-score or min-max scaling)
• Intersection-based: Retains only compounds that rank in the top X% by all scoring functions simultaneously

Empirical finding: Consensus strategies consistently outperform any single scoring function in retrospective virtual screening benchmarks. The improvement arises because different functions capture complementary aspects of binding, and their errors are uncorrelated.

Example: Combining GlideScore, ChemScore, and GoldScore in a rank-by-rank consensus can increase enrichment factors by 20-40% compared to the best individual function.