Rowan Docs

Conformer Search

Python API · stjames models · API example · Ensemble example

How it works

Find low-energy conformations for subsequent calculations and compare their relative energies. Rowan generates candidates, optionally clusters them, then optimizes and removes duplicates. A final single-point calculation, when selected, determines the energy ranking. Results include optimized structures, relative energies, and conformer and ensemble properties.

Settings

  • Conformer generator: openconf is the default; it combines ETKDG seeds with torsional Monte Carlo exploration. ETKDG uses distance geometry and force-field minimization; MCMM explores torsions with the selected energy method; iMTD uses CREST metadynamics. iMTD requires a 3D structure; the others accept SMILES or 3D inputs. Choose "Skip & input pre-generated conformers" to optimize an existing ensemble.
  • openconf controls: "Number of steps" defaults to 200 and "Energy window (kcal/mol)" to 10. "Parent strategy" defaults to "Softmax," favoring lower energies; "Uniform" samples more evenly, while "Best" always starts from the current lowest-energy conformer. "Final select" defaults to "Diverse"; "Energy" keeps the lowest-energy candidates. "Final refinement?" is enabled.
  • Other generators: ETKDG exposes "Num initial conformers" (300) and "Max MMFF iterations" (500). iMTD provides "MTD method," "Solvent," "Speed," "Energy window," and "NCI?" for noncovalent complexes. MCMM provides "Energy settings," "Num iterations" (250), "RMSD threshold (Å)" (0.5), and "Energy window (kcal/mol)" (20).
  • Max conformers: defaults to 200 in the form and limits generation with openconf, iMTD, and MCMM; increasing sampling effort is a separate choice.
  • Conformer clustering: disabled by default. When enabled, k-means groups candidates by shape and surface descriptors and keeps the lowest-energy representatives. "Num clusters" defaults to 5 and "Conformers per cluster" to 3. This loosely follows ReSCoSS. Clustering reduces optimization cost but can exclude relevant conformers; the form disallows it when generation is skipped.
  • Final method: choose optimization stages, final single-point settings, and solvent separately from generation. The form defaults to GFN2-xTB optimization with ALPB water, followed by g-xTB with CPCM-X water.

Notes

Searches do not guarantee the global minimum. Flexible molecules may need more sampling; energy rankings depend on the final method and solvent. openconf automatically widens its generation energy window to 100 kcal/mol for rings containing at least 12 atoms.

Ensemble properties use energy-based Boltzmann weights at 298.15 K, rather than frequency-derived free energies. Reported surface areas use GFN2-xTB COSMO in water, regardless of the optimization solvent.

For transition-state conformers, start from a fully optimized TS. In the form, select iMTD and enable "Transition state?" Supply generation constraints to preserve the reactive geometry; final TS optimizations remove those constraints.

In the API, use conf_gen_settings, conformer_clustering_settings, and multistage_opt_settings; supplying initial_conformers skips generation.

Submission video

Benchmarks and validation

The OpenConf Iridium benchmark compared recovery of experimentally observed, protein-bound conformations for 120 molecules. With a 200-conformer budget, OpenConf recovered more reference structures and ran faster than the ETKDG baselines:

Conformer generatorRecovery within 1 ÅMedian runtime
ETKDGv384.2%1.85 s
ETKDGv3 + MMFF94s82.5%3.99 s
OpenConf88.3%0.78 s

Recovery uses the best match across five runs, measured by heavy-atom RMSD to the experimental structure. Runtimes are for generation on a single Apple M2 Pro CPU core. OpenConf used a docking preset with an 18 kcal/mol energy window and uniform parent selection; these results do not include Rowan's subsequent optimization stages.

Further reading