What is AutoDock-GPU?

AutoDock-GPU is a GPU-accelerated implementation of AutoDock4, one of the most widely-cited docking programs in computational drug discovery. It predicts how small molecules bind to protein targets using the classic AutoDock4 physics-based force field, with reported speedups of up to 56x over single-threaded AutoDock4.

The key distinction from AutoDock Vina is the scoring function: AutoDock-GPU uses the original AutoDock4 force field with explicit terms for van der Waals interactions, hydrogen bonding, electrostatics, and desolvation. This physics-inspired approach behaves differently than Vina's empirically-optimized scoring, particularly for metal-containing binding sites and systems with complex electrostatics.

AutoDock-GPU is developed by the Forli Lab at Scripps Research and is open-source under the GPL license.

How does AutoDock-GPU work?

AutoDock-GPU combines pre-calculated grid maps with the Lamarckian Genetic Algorithm (LGA) to efficiently search conformational space on GPUs.

The AutoDock4 scoring function

The scoring function estimates binding free energy ($\Delta G$) through five physics-based terms:

$\Delta G = \Delta G_{vdW} + \Delta G_{hbond} + \Delta G_{elec} + \Delta G_{desolv} + \Delta G_{tor}$

• Van der Waals ($\Delta G_{vdW}$): Lennard-Jones 12-6 potential modeling steric interactions
• Hydrogen bonding ($\Delta G_{hbond}$): 12-10 potential with a directional term favoring optimal H-bond geometry
• Electrostatics ($\Delta G_{elec}$): Coulombic interactions with a distance-dependent dielectric
• Desolvation ($\Delta G_{desolv}$): Energy penalty for displacing water from hydrophobic surfaces
• Torsional entropy ($\Delta G_{tor}$): Penalty proportional to rotatable bonds lost upon binding

These interaction energies are pre-calculated by AutoGrid at 0.375 Å resolution and stored as 3D grids. During docking, energies are rapidly looked up via trilinear interpolation rather than computed pairwise—this is what enables ~250 million scoring function evaluations per docking job.

The Lamarckian Genetic Algorithm

The LGA is a hybrid global-local search strategy:

1. Population initialization: Random poses (position, orientation, torsion angles) are generated
2. Genetic evolution: New poses are created through crossover and mutation of parent poses
3. Local refinement: ~6% of the population undergoes local search optimization each generation
4. Lamarckian inheritance: Improved poses from local search re-enter the population (unlike Darwinian evolution where acquired traits aren't inherited)
5. Termination: Search stops when maximum evaluations or generations are reached

Multiple independent LGA runs (controlled by Number of LGA runs) explore different regions of conformational space. The best pose across all runs is returned by default, or multiple ranked poses can be requested.

Local search methods

AutoDock-GPU supports five local search optimizers:

ADADELTA (gradient-based): Uses scoring function gradients to descend toward energy minima. Converges faster for flexible ligands (8+ rotatable bonds) and typically requires 6-23x fewer evaluations than Solis-Wets. Recommended for most applications.

Solis-Wets (random): Applies random perturbations to pose variables, accepting improvements probabilistically. Better for small, rigid ligands (≤7 rotatable bonds) where gradient information is less beneficial.

Steepest-Descent, FIRE, and ADAM (gradient-based): Additional native minimizers for controlled comparisons and specialized searches. AutoDock-GPU's ligand-based evaluation heuristics support only ADADELTA and Solis-Wets, so these three methods use the configured maximum evaluation count directly.

GPU parallelization

AutoDock-GPU exploits three levels of parallelism:

• High-level: Multiple LGA runs execute simultaneously across GPU threads
• Medium-level: Individuals within a genetic generation are processed in parallel
• Low-level: Fine-grained tasks (rotation, scoring) are distributed across work-items

This hierarchical approach produced speedups of up to 56x in the published benchmark, depending on the GPU and system size. A single GPU can screen 2,000-40,000 ligands per day.

Grid maps

Grid maps pre-calculate interaction energies between each ligand atom type and the receptor at every point in space. This upfront cost (handled automatically by ProteinIQ via AutoGrid4) enables extremely fast scoring during the docking search—~250 million scoring evaluations per job by looking up cached values rather than computing pairwise interactions. Grid maps are cached per receptor, so docking additional ligands against the same protein reuses existing maps.

How to use AutoDock-GPU online

ProteinIQ handles Meeko preparation, grid map generation, and GPU scheduling automatically—no local installation of AutoDock-GPU, AutoGrid, Meeko, or CUDA toolkits required.

Inputs

Settings

Docking parameters

Advanced settings

Results

Results include a ranking.csv ranking compounds by best predicted affinity (most negative first), complete torsion-aware PDBQT pose files, the Meeko-prepared receptor and ligands, the AutoGrid parameter file, raw DLG/XML records, and pinned runtime provenance. When multiple poses are requested, each ligand's top N poses are returned ranked by binding energy. Optional --npdb files represent poses from the final LGA populations; they are not cluster representatives.

Understanding the results

Binding affinity interpretation

AutoDock-GPU reports binding affinity in kcal/mol. More negative values indicate stronger predicted binding:

AutoDock4 and Vina scoring functions are calibrated differently—affinities should not be compared directly between the two programs. Scores below -12 kcal/mol may be scoring artifacts rather than true predictions, particularly for ligands with many rotatable bonds or poses with unphysical geometry. Visual inspection of such poses in the 3D viewer is recommended.

Limitations

• AutoDock-GPU handles molecules up to about 32 rotatable bonds effectively. For larger peptides or macrocyclic compounds, the search space becomes too large for the LGA to explore adequately. DiffDock handles flexibility differently and may be more suitable.
• AutoDock-GPU v1.6 has fixed limits of 256 combined ligand and flexible-residue atoms, 32 ligand-system atom types, and 57 combined ligand and flexible-residue rotatable bonds. ProteinIQ applies the stricter 32-rotatable-bond product limit to each ligand before docking.
• The protein backbone remains rigid during docking. Sidechain flexibility can be modeled by providing a flexible residues PDBQT file, but large-scale conformational changes are not captured.
• AutoDock-GPU's atom type parameterization does not cover metal coordination in ligands. For metalloligands, use GNINA.

When to use AutoDock-GPU vs alternatives

AutoDock-GPU is the best choice when AutoDock4 scoring is specifically needed (e.g., for consistency with published results) or when screening large compound libraries where GPU acceleration provides the biggest advantage. For individual dockings, runtime is comparable to Vina (1–5 minutes). The speed advantage emerges in batch screening: once grid maps are generated for a receptor, subsequent ligands dock very quickly, enabling 2,000–40,000 compounds per day per GPU.

Comparison to other docking tools