Protein Engineering vol. 16 no. 8 pp. 589-597, 2003
© 2003 Oxford University Press
Optimizing the search algorithm for protein engineering by directed evolution
1Maxygen, Inc., 200 Penobscot Drive, Redwood City, CA 94063, 3DNA 2.0, Inc., 1455 Adams Drive, Menlo Park, CA 94025 and 4Bioren, Inc., 100 Glenn Way, Suite 1, San Carlos, CA 94070, USA
2 To whom correspondence should be addressed. e-mail: richard.fox{at}maxygen.com
An in silico protein model based on the Kauffman NK-landscape, where N is the number of variable positions in a protein and K is the degree of coupling between variable positions, was used to compare alternative search strategies for directed evolution. A simple genetic algorithm (GA) was used to model the performance of a standard DNA shuffling protocol. The search effectiveness of the GA was compared to that of a statistical approach called the protein sequence activity relationship (ProSAR) algorithm, which consists of two steps: model building and library design. A number of parameters were investigated and found to be important for the comparison, including the value of K, the screening size, the system noise and the number of replicates. The statistical model was found to accurately predict the measured activities for small values of the coupling between amino acids, K 
1. The ProSAR strategy was found to perform well for small to moderate values of coupling, 0 K 3, and was found to be robust to noise. Some implications for protein engineering are discussed.
Received January 2, 2003; revised May 13, 2003; accepted June 19, 2003.