Decisions with Confidence: Application to the Conformation Sampling of Molecules in the Solid State

Abstract

Accurate conformations of a molecule are critical for reliable prediction of its properties, so good predictive models require good conformations. Here, we present a method for conformer sampling based on distance geometry, implemented in our conformation generator OMEGA, which we apply to both macrocycles and druglike molecules. We validate it in the usual fashion, reproducing conformations from the solid state, and compare its performance in detail to other methods. We find that OMEGA performs well on three key criteria: accuracy, speed, and ensemble size. To support our conclusions quantitatively, particularly on accuracy, we developed a workflow for method comparison that uses parameter estimation, inference from confidence intervals, classical null hypothesis significance testing, Bayesian estimation, and effect size. The workflow is designed to be robust to the highly skewed performance data often found when validating tools in computational chemistry and to provide reliable, easy to interpret results. In this workflow, we emphasize the importance of confidently distinguishing between methods, with particular reference to a priori estimation of sample size and statistical power (false negative or Type II error rate), a topic almost completely ignored hitherto in computational chemistry.