🌻 BOUQUET 💐

Generate beautiful ensembles of molecular geometries Bayesian Optimization Using QUantum Energy Tool

This repo contains code for optimizing conformers using Bayesian optimization for active learning and quantum chemistry computations.

Background

Conformers define the different geometries with the same molecular bonding graph but different coordinates. Finding the lowest-energy conformation is a common task in molecular modeling, and one that often requires significant time to solve. We implement optimal experimental design techniques to solve this problem following recent work that uses Bayesian optimization find optimize dihedral angles.

Installation

We recommend use of pixi:

pixi install

Many parts of the code are available through pip:

pip install '.[all]'

Use

bouquet provides a simple interface to the code. To optimize cysteine with default arguments.

bouquet --smiles "C([C@@H](C(=O)O)N)S"

This will produce a folder in the solutions directory containing the optimized geometry (final.xyz) and many other files for debugging.

Common options

Call bouquet --help for the full list. The most useful flags:

Input

Flag	Default	Description
--smiles	—	SMILES string of the molecule to optimize
--file	—	Read the starting structure from a file instead of a SMILES string
--conformer-file	—	File of multiple conformers to use as initial guesses (overrides --init-steps)
--name	SMILES / file name	Output name for the run directory
--seed	random	Random seed for reproducible runs

Search budget

Flag	Default	Description
--auto	off	Scale the number of steps to the number of detected dihedrals
--num-steps	32	Number of Bayesian-optimization steps
--init-steps	5	Number of initial guesses before optimization begins

Energy / geometry

Flag	Default	Description
--energy	gfn2	Energy method (aimnet2, ani, b3lyp, b97, gfn0, gfn2, gfnff, mmff, uff)
--optimizer	gfnff	Method used for geometry relaxation
--relax	on	Relax non-dihedral degrees of freedom before scoring each energy