Originally published as Genetics Published Articles Ahead of Print on January 16, 2006.

Genetics, Vol. 172, 2113-2122, April 2006, Copyright © 2006
doi:10.1534/genetics.105.052688

Gene Expression From Random Libraries of Yeast Promoters

Martin Ligr*, Rahul Siddharthan{dagger}, Fredrick R. Cross* and Eric D. Siggia*,1

* The Rockefeller University, New York, New York 10021 and {dagger} Institute of Mathematical Sciences, Taramani, Chennai 600113, India

1 Corresponding author: Center for Studies in Physics and Biology, Rockefeller University, Box 25, 1230 York Ave., New York, NY 10021.
E-mail: siggiae{at}mail.rockefeller.edu

Genomewide techniques to assay gene expression and transcription factor binding are in widespread use, but are far from providing predictive rules for the function of regulatory DNA. To investigate more intensively the grammar rules for active regulatory sequence, we made libraries from random ligations of a very restricted set of sequences. Working with the yeast Saccharomyces cerevisiae, we developed a novel screen based on the sensitivity of ascospores lacking dityrosine to treatment with lytic enzymes. We tested two separate libraries built by random ligation of a single type of activator site either for a well-characterized sporulation factor, Ndt80, or for a new sporulation-specific regulatory site that we identified and several neutral spacer elements. This selective system achieved up to 1:104 enrichment of the artificial sequences that were active during sporulation, allowing a high-throughput analysis of large libraries of synthetic promoters. This is not practical with methods involving direct screening for expression, such as those based on fluorescent reporters. There were very few false positives, since active promoters always passed the screen when retested. The survival rate of our libraries containing roughly equal numbers of spacers and activators was a few percent that of libraries made from activators alone. The sequences of ~100 examples of active and inactive promoters could not be distinguished by simple binary rules; instead, the best model for the data was a linear regression fit of a quantitative measure of gene activity to multiple features of the regulatory sequence.