Dataset for "Bayesian Optimization Alleviates the Conformational Bottleneck in Mechanistic Modeling"

Data collection method:

The data sets of the force-field searched and CREST-searched molecule conformers were sourced from their previous publications (see below). The four new data sets of transition state conformers were computed according to the following procedures: The "BPA" transition state was conformationally searched using the OPLS3e force field in MacroModel v12.0, before being subjected to a redundant conformer elimination with a RMSD threshold of 0.6 Angstrom and an OPLS3e energy threshold of 21.0 kJ/mol to give 308 conformers. DFT energies were then computed at the B3LYP-D3(BJ)/Def2-TZVPP-IEFPCM(toluene)//B3LYP/6-31G(d) level of theory using Gaussian 16, Revision A.03. Single-point energies on the low-level geometries were computed at the B3LYP-D3(BJ)/Def2-TZVPP-IEFPCM(toluene) level of theory. The "NTOB" transition state was conformationally searched using OPLS3e in MacroModel v12.0 and redundant conformers were eliminated using a RMSD threshold of 0.5 Angstrom and an OPLS3e energy threshold of 21.0 kJ/mol, with 81 conformers in total remaining. DFT calculations were performed at the B3LYP-D3(BJ)/Def2-TZVPP-IEFPCM(dichloromethane)//B3LYP/6-31G(d) level of theory using Gaussian 16, Revision A.03, but unfortunately 6 conformers could not be converged to a correct transition state, and these are recorded as NaN values in the NTOB DFT NumPy array file. Single-point energies on the low-level geometries were computed at the B3LYP-D3(BJ)/Def2-TZVPP-IEFPCM(dichloromethane) level of theory. The "cinchona" transition state was conformationally searched using OPLS4 in MacroModel v13.8 and a redundant conformer elimination was performed with a RMSD threshold of 1.0 Angstrom and an OPLS4 energy threshold of 25 kJ/mol to give 110 conformers. Each of these transition state conformers were optimized using the AM1 semi-empirical method in Gaussian 16, Revision C.01. Single-point energies on the low-level geometries were computed at the AM1 level of theory. The "BIMP"/"adsbimp" transition state was conformationally searched using the GFN-FF force field in CREST v2.12 and a redundant conformer elimination was performed with a RMSD threshold of 1.5 Angstrom to yield 625 conformers in total. Each of the transition state conformers were optimized using the AM1 semi-empirical method in Gaussian 16, Revision C.01. Unfortunately, 233 conformers could not be converged to a correct transition state, and these are recorded as NaN values in the adsbimp DFT NumPy array file. Single-point energies on the low-level geometries were computed at the AM1 level of theory. The "nitroaldol" R (labelled 1 in the files) and S (labelled 2) transition states were conformationally searched using the GFN2-xTB method in CREST v2.12 and a redundant conformer elimination with a RMSD threshold of 1.0 Angstrom was performed on each set to yield 468 R-forming transition state conformers and 436 S-forming transition state conformers. These structures were first optimized using the HF/3-21G method with Gaussian 16, Revision C.01, and then using the B3LYP-D3(BJ)/def2-TZVP//B3LYP-D3(BJ)/6-31G(d) level of theory. Single point-corrected quasi-harmonic Gibbs free energies of all structures were computed using the GoodVibes Python package, version 3.2, at a temperature of 298.15 K and a concentration of 1.0 mol/L.

Data processing and preparation activities:

The computational details of the creation of the force field-searched and CREST-searched data sets of molecule conformers may be found from the original publications (DOI: 10.1021/acs.jcim.3c00649 and DOI: 10.1021/acs.joc.2c00066 respectively). The full data sets are available from their data archive repositories (DOI: 10.17863/CAM.88197 and DOI: 10.15125/BATH-01065). From each of the two data sets, we extracted all molecules that had 80 or more conformers available and performed a redundant conformer elimination with a RMSD threshold of 0.5 Angstrom using MacroModel v13.8 and discarded all corresponding DFT-optimized structures that did not survive this elimination.

Technical details and requirements:

The software used included MacroModel v12.0; MacroModel v13.8; Gaussian 16, Revision A.03; Gaussian 16, Revision C.01; Python 3.12.4; GoodVibes 3.2; CREST v2.12.

Additional information:

The objects of primary interest are the files "molecules.zip", "crest_molecules.zip", "transition_states.zip" and "single_points.zip", which contain the necessary data to directly reproduce our results. Each of these archive files should be extracted and used as input to the programs in the experiments folder of the associated GitHub repository (the-grayson-group/finding_confs) as detailed there. In the archive files "ts_structures.zip" and "ts_spe_outputs.zip" we provide the Gaussian output files for the optimized conformers of the four transition state data sets, the single-point energies on the low-level geometries of the four transition state data sets. The Gaussian output files for our mechanistic case study are stored in "nitroaldol.zip". This archive contains the output files for the HF- and DFT-optimized transition state conformers. Files named "nitroaldol1*" refer to the R-forming diastereomeric transition state conformers and files named "nitroaldol2*" refer to the S-forming diastereomeric transition state conformers.

Methodology link: