AutoDock Vina

Vina employs a stochastic global optimization algorithm that combines local BFGS optimization with a Metropolis acceptance criterion. The algorithm performs a series of random Markov chain mutations to escape local minima, followed by local optimization steps. This hybrid approach achieves a success rate of predicting near-native poses within 2 Å RMSD that rivals more computationally expensive methods, while requiring only a fraction of the CPU time.

The scoring function approximates the binding free energy as a sum of intermolecular and intramolecular contributions:

• Steric interactions: Modeled by a repulsive term and an attractive Gaussian term
• Hydrophobic effect: Proportional to the buried solvent-accessible surface area
• Hydrogen bonding: A directional term dependent on donor-acceptor geometry
• Torsional entropy: Penalty proportional to the number of rotatable bonds frozen upon binding This function was trained on a diverse set of protein-ligand complexes from the PDBbind database.

Vina natively supports multithreaded parallelism to exploit multi-core CPU architectures. The global optimization algorithm runs multiple independent search trajectories concurrently, each starting from a different random seed. This embarrassingly parallel design yields near-linear speedup with the number of available cores, enabling high-throughput virtual screening of large compound libraries on standard workstation hardware without requiring specialized cluster management software.

While the protein backbone remains rigid, Vina allows selected receptor sidechains to be treated as flexible during docking. Users specify which residues to make flexible in the PDBQT input file. The optimizer simultaneously samples ligand poses and sidechain rotameric states, capturing induced-fit effects that are critical for targets with plastic binding pockets. This partial flexibility model balances computational tractability with biological realism.

Vina uses the PDBQT format, which extends the standard PDB format with partial charges and atom type assignments. Key features include:

• AutoDock atom types: Assigns each atom a type (e.g., C, OA, HD) for scoring function lookups
• Gasteiger partial charges: Computed using an empirical method for electrostatic interaction calculation
• Rotatable bond flags: Marked with BRANCH/ENDBRANCH keywords to define torsional degrees of freedom This format ensures seamless interoperability with the broader AutoDock ecosystem.

The exhaustiveness parameter controls the trade-off between speed and accuracy by setting the number of independent global optimization runs. A higher value increases the probability of finding the global minimum but linearly increases computation time. Typical values range from 8 for quick screening to 32-64 for high-accuracy pose prediction. This tunable knob allows researchers to adapt Vina's behavior to the specific requirements of virtual screening versus lead optimization workflows.

The process of exploring the rotatable bond space of a flexible ligand to generate diverse 3D shapes. Vina employs an iterated local search algorithm that combines stochastic global exploration with local BFGS optimization.

• Each ligand typically has 5–15 rotatable bonds, creating an enormous search space
• Vina's algorithm efficiently escapes local minima to find the global energy minimum
• Sampling quality directly determines whether the correct binding pose is recovered

The gold-standard metric for evaluating docking accuracy. RMSD measures the average atomic distance between a predicted ligand pose and the experimentally determined pose from X-ray crystallography.

• A prediction is considered successful if RMSD ≤ 2.0 Å
• Vina typically achieves success rates of 70–80% on redocking benchmarks
• Lower RMSD values indicate near-native pose prediction quality

The high-throughput computational evaluation of large chemical libraries (10⁵–10⁹ compounds) against a target protein. Vina's speed—typically < 1 second per ligand on a single CPU core—makes it a workhorse for this application.

• Used to filter commercial compound catalogs before expensive physical screening
• Often combined with pharmacophore filters to pre-select drug-like molecules
• Success measured by enrichment factor: how many actives appear in the top-ranked fraction

The quantitative estimation of interaction strength between ligand and receptor. While Vina's scores correlate with experimental binding affinities, they are approximate and best used for ranking rather than absolute free energy calculation.

• For rigorous ΔG prediction, methods like Free Energy Perturbation (FEP) are preferred
• Vina's scoring function was trained on the PDBbind refined set of experimentally measured complexes
• Typical correlation with experimental pKi/pKd: R ≈ 0.5–0.6