DeepConf: Leveraging ANI-ML Potentials for Exploring Local Minima with Application to Bioactive Conformations

Abstract

Here, we introduce a low energy conformer generation algorithm using ANI-ML potentials at the DFT accuracy and benchmark in reproducing bioactive conformations. We show that the method is efficient when the initial starting structure is far from equilibrium, when the ML potentials are stuck in nonsmooth regions, and when the quality of the conformers in a less conformer size is demanded. We specifically focus on conformations due to rotations around the single bonds. For the first time, we assess the performance of ANI-ML potentials using our conformer generation algorithm, DeepConf, in addition to previously reported Auto3D (J. Chem. Inf. Model. 2022, 62, 5373-5382) using the same potentials to reproduce bioactive conformations as well as providing a guideline for bioactive conformation evaluation processes. Our results show that the ANI-ML potentials can reproduce the bioactive conformations with mean value of the root-mean-square-deviation (RMSD) less than 0.5 & Aring;, outperforming the limit of conventional methods. The code offers several features including but not limited to geometry optimization, fast conformer generations via single point energies (SPE), different minimization algorithms, different ML-potentials, or high-quality conformers in the smallest amount of ensemble sizes. It is available free of charge (documentation and test files) at https://github.com/otayfuroglu/DeepConf.