CSD&E: Expanding Efficient Conformer Sampling to Diverse Charged and Neutral Molecules

With support from the Chemical Theory, Models, and Computational Methods program in the Division of Chemistry, Drs. Geoffrey Hutchison and David Koes at the University of Pittsburgh are developing data-driven techniques to study the fundamental geometries and flexibility of neutral and charged molecules. This research seeks to advance the predictive ability of molecular computational chemistry in the areas of polymers, molecular materials, and drug discovery via an inverse property-driven design approach.

Most molecules, ranging from proteins to pharmaceuticals and plastics, are flexible and at room temperature can exist in multiple geometries called conformers. As a molecule increases in size, complexity, and flexibility, the number of possible stable conformers increases exponentially. Moreover, while rules for common geometric motifs exist for neutral molecules, charged states, whether positively or negatively charged, may have very different structures.

The project draws on expertise in data science, machine learning, and optimization theory to build both experimental and computational data resources to predict likely conformers. Techniques will be validated across multiple experimental and computational benchmarks and applied to key areas of chemistry, including finding new targets for drug design and plastic electronic materials.

The project will provide opportunities for the education and active training of high school, undergraduate, and graduate students in data science, machine learning, and chemistry.

The project draws on a connection between statistical thermodynamics and Bayesian statistics. Using experimental and computational data, one can estimate distributions of dihedral angles for most molecules. From such probabilities, Bayesian optimization can accurately explore and sample Boltzmann-weighted ensembles of the potential energy surface.

By building data repositories of neutral and charged species, one can efficiently train new recurrent machine learning methods, as well as design new “few-shot” geometry optimization methods. In turn, these methods will draw on improvements in accurate quantum chemical methods to predict thermochemistry and free energies for a large selection of neutral and charged organic species.

Through generating these large data repositories, Hutchinson and Koes and their co-workers will create a new curriculum for data science lessons, machine learning workshops, and classes in chemistry. The project will also provide substantial educational training to high school, undergraduate, and graduate students in the interdisciplinary combination of data science, statistics, machine learning, and computational chemistry.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.