Using targeted measurements to improve the accuracy of predictions of molecular physical properties

COMP 193

Robert S. DeWitte, Advanced Chemistry Development, 90 Adelaide W, Toronto, ON M5H 3V9, Canada and Eduard Kolovanov, Advanced Chemistry Development, Inc, 90 Adelaide St.West, Suite 600, 90 Adelaide St.West, Suite 600, ON M5H 3V9, Canada.
Even with the state of the art prediction technology available today for LogP, pKa and LogD, there are compound classes (specifically proprietary chemical classes) that are under-represented in the algorithmic "training space". This talk focuses on a combined system for determining physical properties that uses prediction when the algorithm has a high confidence of accuracy, and uses measurement when this confidence is low, and then employs these measurements to expand the scope of chemical classes for which it can confidently predict physical properties. By employing this hybrid approach, thermodynamic quality can be achieved with high accuracy and very high throughput. Specific aspects to be addressed include: how to know when you need to measure; Overcoming the quality/throughput trade-off; continually improving the accuracy of predictions; and minimizing total cost of measurements.