Organic Crystal Structure Prediction (CSP) faces the challenge of having to find all crystal structures in a small energy window at the very bottom of a large and crowded search space. In the last three CSP blind tests held in 2007, 2010 and 2015, the highest success rates were scored by the GRACE software. Each crystal structure prediction starts with the generation of a tailor-made force field fitted from scratch to DFT-D reference data. Larger compounds are cut into chemically reasonable fragments. In the reference data generation, all favorable and unfavorable contacts, conformations and barriers are explicitly sampled to yield a balanced force field without any blind spots. The force field accuracy is benchmarked and fine-tuned against a diverse set of crystal structures obtained in a preliminary crystal structure prediction. The actual CSP is a three-step procedure in which the energy accuracy progressively increases and the number of considered structures decreases. In the first step, crystal structures are generated with the tailor-made force field using a Monte Carlo parallel tempering algorithm. In the second step, coarse DFT-D lattice energy optimizations are carried out, and statistical correlations between single-point energies, structural similarity and energy decrease upon minimization are exploited to reduce the number of structures to be explicitly minimized. In the final step, lattice energies are further optimized with more stringent convergence criteria. Convergence control mechanisms at all stages guarantee that all crystal structures in a user-defined target energy window are found with as user-defined degree of confidence. The approach has been shown to work for non-ionic compounds, co-crystals, hydrates and salts with one or more chemical unit per asymmetric unit. For pharmaceutical molecules in the 40 to 80 atoms size range, typical CPU time requirements range from 1 to 3 months of CPU time on 400 Intel Xeon cores. Looking back on over 30 confidential crystal structure prediction studies of pharmaceutical compounds, it appears that in 10-40% of the cases the thermodynamically stable form has not yet been experimentally observed, thus presenting a substantial risk of encountering a late appearing form or even a disappearing-polymorph event. Computer-assisted crystallization experiment design emerges as the new frontier in CSP. From the past CSP learning curve, future performance increases can be estimated

Organic crystal structure prediction for industrial applications by a convergence-controlled blend of DFT-D with tailor-made force fields

255th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.