ghutchis / bouquet Goto Github PK

CREST allows gfnff minimization with gfn2 energies. This would be a useful option to balance speed and energy accuracy (e.g., also gfn2 geometries vs. dft energies, etc.)

Implement Cremer-Pople ring sampling techniques:
https://doi.org/10.1021/acs.jcim.0c01144

• Initial implementation for simple (non-fused rings)
  • add ring parameters to BO acquisition (i.e., different kernels)
  • convert to bond torsions (i.e., cut a bond, change ring torsions, re-bond)
• Gather better data from conformer archive
• Work on ring decomposer for fused rings

While GFN-FF, GFN2, and ANI are probably the best options for accuracy, it might help for benchmarking to have MMFF94, GAFF, and UFF methods:

• Implement energy calculator for Open Babel
• Integrate dihedral constraints

The BOKEI paper (https://doi.org/10.1039/C9CP06688H) shows the success of using correlated dihedral information.

• Consider how to implement correlated dihedrals as a weak prior (e.g., πBO)
• Consider implementation using constrained EI (e.g., adapting KEI from paper)
• Gather more distributions from conformer set

botorch implements batch acquisition functions. Many tools would benefit from batch evaluation (e.g., 4, 8, 10, etc. energy and relaxation in parallel).

• generate batch of candidates
• separate TorchANI from ASE (may need constrained dihedral optimization)
• evaluate if ASE can run multiple threads
• return batch energies

Ideally, there would be an "automatic" mode that adjusts the initial sampling and total sampling to ensure a full ensemble.

• how many iterations needed based on structure / number of dihedrals
• when is the ensemble complete?

One way to make progress is to have a set, continue to sample (i.e., to see if there's a better global minimum and if the ensemble is complete) and extract some heuristics.

The script should be able to save all geometries (in XYZ or SDF or PDB) format within X kcal/mol of the minimum (e.g. default threshold ~6-10 kcal/mol).

We have recently established GFN2-derived gas phase preferences.
https://doi.org/10.26434/chemrxiv-2022-h9gjq

• Assign dihedrals based on SMARTS in the Torlib patterns / data
• Assign initial random points from the normalized probability distributions
• Use πBO acquisition functions to use the probability distributions as soft priors

For now, the code uses UCB with a formula favoring increasing exploration, which might make sense for finding all minima.

Benchmark # of steps required and lowest minima across a subset of the BOKEI test set:

• UCB
• Expected Improvement
• log Expected Improvement
• Probability of Improvement
• log Probability of Improvement

Other interesting acquisition functions in BoTorch:

• noisy EI (should be similar to EI, but worth checking)
• q-variants to analytic acquisition functions
• SAASBO
• GIBBON
• Knowledge gradient (botorch.acquisition.knowledge_gradient.qKnowledgeGradient)
• Entropy Search (MES) (botorch.acquisition.max_value_entropy_search.qMaxValueEntropy)

Right now, the initial sampling seems to be a normal distribution with ±30° range.

Evaluate:

• uniform random distribution
• larger ranges, including 45°, 60°, and 90°