(623d) Pharmaceutical Crystal-Structure Prediction Via Near-Exhaustive Landscape Searching and Advanced Machine-Learning Intermolecular Potentials

In the application of molecular simulation to the age-old problem of Crystal-Structure Prediction (CSP), the two Grand Challenges are, firstly, the accurate representation of inter- (and intra-) molecular interactions, and secondly, how to search - exhaustively - the entire crystal (free-) energy landscape. Nowhere is this formidable dual-challenge more acute than in the pharmaceutical industry, where API flexibility and chirality are, of course, further complicating factors.

In this seminar, the intellectual machinery of novel CSP approaches are described - which hinge, in essence, on the use of proprietary permutation-invariant neural networks (PINNs) to for fitting to high-quality electronic-structure calculations under periodic boundary conditions - thus fashioning a generalisable and high-fidelity fit to a diverse range of polymorph-packing arrangements for APIs - tailor-made for the fit. In turn, and in direct contrast to the entire prevailing intellectual infrastructure of Artificial Intelligence in the Pharma Industry, we carry out our further proprietary exhaustive search of the polymorph landscape with our high-quality, 'good-fit' PINNs using a biaised-Hamiltonian global optimisation basin-hopping search approach based on efficient supercell treatment of pairwise interactions under periodic boundary conditions - implementing all space-group symmetry and chirality constraints. Having constructed a shortlist of plausible polymorph candidates, we then correct their thermalised energy for temperatures of interest, e.g., manufacture or storage, and we rank by either lattice energy or matching simulated spectral features to Powder X-Ray Diffraction data - or both, if desired.