(323e) Optimizing Antifreeze Protein Activity Via an Elitist Genetic Algorithm: A Comparison of High and Low Activity Variants

Kozuch, D. J., Princeton University
Stillinger, F. H., Princeton University
Debenedetti, P. G., Princeton University
Given the existence of a suitably efficient fitness function, genetic algorithms offer a promising method for systematically optimizing the properties of biological materials. In our recent paper [1], we established a neural network based method for predicting the antifreeze activity of different proteins from molecular dynamics simulation. Using this method as our fitness function, we designed a simple, but effective, elitist genetic algorithm for selecting amino acid mutations that should increase the activity of native antifreeze proteins. We apply this algorithm to two antifreeze proteins: (1) a low activity Type III antifreeze protein from the winter flounder, and (2) a hyperactive antifreeze protein from the longhorn beetle. We show that the algorithm finds structures with a predicted freezing point depression 20 - 200% greater than that of the native protein. We hope that these results inspire further investigation of high activity antifreeze biologics that could someday be used in more effective tissue preservation.

[1] Kozuch, D. J., Stillinger, F. H., & Debenedetti, P. G. (2018). Combined molecular dynamics and neural network method for predicting protein antifreeze activity. Proceedings of the National Academy of Sciences, 115(52), 13252–13257