BOKEI: Bayesian Optimization Using Knowledge of Correlated Torsions and Expected Improvement for Conformer Generation

In previous blog post, we introduced the idea of Bayesian optimization and its application in finding the lowest energy conformation of given molecule[1]. Here, we extend this approach to incorporate the knowledge of correlated torsion and accelerate the search.

The torsion rules are typically derived from the experimental crystal structure databases, such as Protein Data Bank (PDB)[2] and Cambridge Structural Database (CSD)[3]. The nearest neighbour effects are generally ignored in these rules — each dihedral is treated as a free rotor. In fact, information about the correlated torsion is crucial for conformer generation, as adjacent torsion angles are naturally constrained, in order to reduce steric clashes, retain \pi conjugation, align intramolecular hydrogen bonds or other similar non-covalent interactions.

In our work[4], we use a bivariate von Mises distribution to capture the correlations, and use it to constrain the search space. We validate the performance of our Bayesian optimization with knowledge of correlated torsion (BOKEI) on a dataset consisting of 533 diverse small organic molecules, using a force field (MMFF94) and a semi-empirical method (GFN2). We show that in 70(± 2.1)% of the cases examined, our new approach, BOKEI, finds lower energy conformations than global optimization with BOA-EI[1] or genetic algorithm.

Reference

  1. L. Chan, G.R. Hutchison, G.M. Morris. Bayesian Optimization for Conformer Generation.  Journal of Cheminformatics, 2019,11, 32  
  2. H. M. Berman et al. The Protein Data Bank. Nucleic Acids Research 2000 
  3. C. R. Groom et al. The Cambridge Structural Database. Acta Cryst. 2016
  4. L. Chan, G.R. Hutchison, G.M. Morris. BOKEI: Bayesian Optimization Using Knowledge of Correlated Torsions and Expected Improvement for Conformer Generation. ChemRxiv 2019

Author