Binding Affinity Prediction

These functions approximate the binding free energy by summing individual energetic contributions. They are the mathematical core of docking engines.

• Force-field terms: Calculate van der Waals and electrostatic complementarity using potentials like AMBER or CHARMM.
• Empirical terms: Weight hydrogen bonds, hydrophobic contacts, and metal coordination based on regression against known affinities.
• Desolvation penalties: Account for the energetic cost of stripping water from polar groups, often using implicit solvent models like Generalized Born (GB) or Poisson-Boltzmann (PB).
• Example: The Vina scoring function combines a steric term, a hydrophobic term, and a directional hydrogen bond term into a single master equation.

Instead of hand-crafted features, equivariant neural networks operate directly on the 3D coordinates of protein-ligand complexes, respecting rotational and translational symmetries.

• SE(3)-equivariance: Guarantees that rotating the input complex rotates the predicted forces identically, a critical physical constraint.
• Tensor field networks: Build representations using spherical harmonics to capture complex angular geometries within binding pockets.
• Interaction graphs: Atoms are nodes; edges are constructed via distance cutoffs or k-nearest neighbors to capture non-covalent contacts.
• Example: EquiBind and DiffDock use SE(3)-equivariant architectures to predict binding poses and affinities without relying on pre-calculated docking grids.

The gold standard for relative binding affinity prediction. Free Energy Perturbation (FEP) computationally mutates one ligand into another through a series of non-physical intermediate states.

• Thermodynamic cycle: Calculates the relative binding free energy (ΔΔG) by alchemically transforming Ligand A to Ligand B in both solvent and protein environments.
• Lambda windows: The transformation is split into discrete steps (λ = 0 to 1), with independent MD simulations run at each window.
• Bennett Acceptance Ratio (BAR): A statistical estimator that combines forward and reverse perturbation data to minimize variance.
• Accuracy: Routinely achieves chemical accuracy (< 1 kcal/mol error) when protocols are carefully executed, making it suitable for lead optimization.

A fixed-length vector representation capturing the specific non-covalent contacts between a ligand and its target. This bridges structural data and machine learning.

• Bit-string representation: Each bit corresponds to a specific residue and interaction type (e.g., 'PHE92_arene-H', 'ASP189_salt-bridge').
• Structural Interaction Fingerprints (SIFt): Encode interactions from a docked pose, enabling QSAR-like modeling on 3D information.
• Resilience to pose errors: Interaction fingerprints are often more tolerant of small docking inaccuracies than raw Cartesian coordinates.
• Application: Used to train random forests or gradient boosting machines to distinguish true binders from decoys in virtual screening post-processing.

A machine learning paradigm that jointly models the ligand space and the target space, enabling predictions for previously unseen protein-ligand pairs.

• Target descriptors: Encode the binding pocket using sequence-based features (e.g., Z-scales) or structure-based features (e.g., interaction profiles).
• Ligand descriptors: Use molecular fingerprints, physicochemical properties, or graph embeddings.
• Cross-term modeling: The model learns the interaction tensor, capturing that a specific chemical group on the ligand interacts favorably with a specific amino acid type in the pocket.
• Kinase panel example: A single PCM model can predict the affinity of a library of inhibitors against the entire kinome, identifying selectivity cliffs.

Accounts for protein flexibility by docking a ligand against an ensemble of receptor structures rather than a single rigid snapshot.

• Multiple receptor conformations (MRC): Structures are harvested from Molecular Dynamics simulations or different crystal forms.
• Ensemble scoring: The final affinity is computed as a Boltzmann-weighted average across the ensemble, capturing induced-fit effects.
• Cryptic pocket identification: MD simulations can reveal transient binding pockets absent in static crystal structures, expanding the druggable proteome.
• Markov State Models (MSMs): Partition the conformational landscape into discrete states, allowing efficient sampling of rare events relevant to binding kinetics.