Rowan's docking workflow is designed to be automated, conservative, and interpretable. It couples multi-conformer ligand preparation with AutoDock Vina–style docking and post hoc refinement to ensure that the poses reported are not just low in the scoring function but also chemically sensible and structurally consistent with a ligand's intrinsic conformational preferences.
Rowan's docking workflow begins from a single input ligand and a protein structure. The ligand is first checked to ensure it corresponds to a single connected molecule. Ligand preparation can proceed along two closely related routes. If requested, Rowan performs a conformer search in water to generate a small ensemble of low-energy starting geometries. These conformers are optimized using the AIMNet2 neural network potential and are evaluated with a solvent-corrected single-point calculation (ALPB water) to obtain consistent relative energies. The resulting conformer ensemble, sorted by energy, defines a compact set of physically reasonable three-dimensional geometries that span the low-energy region of the ligand's conformational space. If a full conformer search is not requested, Rowan can still optimize the input geometry with AIMNet2 to ensure that even "single-conformer" docking starts from a relaxed and well-parameterized structure.
The receptor is taken from a user-provided PDB structure and the binding site is specified as a box center and size, either from prior structural knowledge or upstream workflow components; these values define the search region for docking. Docking itself is performed with AutoDock Vina or QVina2, depending on user settings, using the specified scoring function, exhaustiveness, number of output modes, and a fixed random seed for reproducibility. Rather than docking a single arbitrary ligand geometry, Rowan run an independent docking job for each member of the prepared conformer ensemble. Each conformer is docked against the protein and the resulting multi-model PDBQT output is parsed into individual poses. For each pose, Rowan also recovers the Vina-reported binding affinity.
Vina excludes hydrogens, making it difficult to use Vina-computed poses in downstream applications like molecular dynamics. Rowan reconstructs chemically valid three-dimensional structures by using RDKit to assign bond orders from the original ligand template and to add explicit hydrogens.
Because Vina frequently returns families of closely related poses, Rowan explicitly removes duplicates before interpretation or refinement. All poses are sorted by docking score from best to worst and then compared using both energy and geometry. Poses that share the same docking score and have a low heavy-atom RMSD (typically less than 0.5 Ă… by a direct coordinate comparison) are treated as redundant, and only the best representative is retained. This filters out trivial variants of the same binding mode and yields a non-redundant set of distinct poses that more faithfully reflects the diversity of plausible binding geometries.
For workflows with conformer search enabled, Rowan estimates a per-pose "strain energy" by comparing each docked pose's energy to the minimum energy found in the pre-generated conformer ensemble. To do this in a geometry-aware way, each pose is locally reoptimized with a harmonic restraint that keeps it close to the docked coordinates, using the same AIMNet2 neural network potential. Single-point energies on these constrained geometries, together with comparisons to the ensemble minimum, provide an estimate of how much higher in energy a docked conformation sits relative to the ligand's preferred solution-like states. For workflows without an explicit ensemble, Rowan instead applies the same restrained optimization and single-point protocol directly to the docked poses, yielding consistent energies but omitting the strain comparison. In both cases, the refinement is deliberately gentle: the harmonic restraint penalizes large deviations from the docked pose, so the optimization can relieve obviously unphysical features (bad clashes, distorted bonds) without erasing the binding mode that the search algorithm identified.
Each refined pose is subjected to a PoseBusters structural validity check against the protein to identify gross geometric pathologies like severe steric clashes or chemically impossible arrangements. The final output for each pose includes the docking score, the strain estimate when available, a boolean flag indicating whether it passes the PoseBusters checks.