Scoring Function

The primary goal is to estimate the Gibbs free energy of binding (ΔG). This is a thermodynamic quantity governing the spontaneity of protein-ligand complex formation. Scoring functions decompose this into components:

• Enthalpic terms: Hydrogen bonds, van der Waals contacts, electrostatic interactions.
• Entropic terms: Desolvation effects, loss of ligand and protein conformational freedom.
• Mathematical form: Often a linear sum of weighted terms, e.g., ΔG ≈ Σ(ΔG_vdw) + Σ(ΔG_hbond) + Σ(ΔG_elec) + ΔG_desolv.

These functions use classical molecular mechanics to compute the non-bonded interaction energy between the protein and ligand. They are derived from physics-based potential energy functions like AMBER or CHARMM.

• Van der Waals energy: Modeled by a Lennard-Jones 12-6 potential.
• Electrostatic energy: Calculated using Coulomb's law with partial atomic charges.
• Limitation: Pure force-field scores often neglect solvation and entropic effects, requiring supplementary terms for accurate ranking.

These are the workhorses of high-throughput docking. They sum a set of weighted, uncorrelated structural terms, with weights calibrated by fitting to experimental binding affinities of a training set.

• Examples: ChemScore, GlideScore (SP/XP), GoldScore.
• Terms include: Hydrogen bond count, lipophilic contact area, metal coordination, and penalty terms for frozen rotatable bonds.
• Advantage: Extremely fast computation, enabling the evaluation of millions of poses per hour.

Also known as statistical potentials, these are derived from the inverse Boltzmann relationship. They analyze the frequency of atom-pair interactions in experimentally determined protein-ligand complex structures (e.g., from the Protein Data Bank).

• Principle: A distance-dependent interaction observed more often than random chance indicates a favorable contact.
• Examples: PMF (Potential of Mean Force), DrugScore.
• Benefit: Implicitly captures complex effects like solvation and entropy without explicit parameterization.

Modern approaches replace the linear additive assumption with non-linear models trained on structural interaction fingerprints and experimental data.

• Classical ML: Random Forest and Support Vector Machines trained on protein-ligand interaction features (e.g., RF-Score, NNScore).
• Deep Learning: Convolutional Neural Networks (CNNs) operating on 3D voxel grids of the complex, or Graph Neural Networks (GNNs) processing the molecular graph.
• Key challenge: Avoiding overfitting to training data and ensuring generalizability to novel chemotypes.

A strategy that combines the predictions of multiple distinct scoring functions to improve hit rates and reduce false positives. The assumption is that the intersection of independent errors is smaller than the error of any single function.

• Methods: Rank-by-rank, score-by-vote, or averaging normalized scores.
• Effectiveness: Often rescues true binders that are ranked poorly by one function but highly by others.
• Trade-off: Can reduce the number of identified hits if functions disagree, requiring careful calibration of the consensus threshold.