TY - JOUR
T1 - Thresholding rules for recovering a sparse signal from microarray experiments
AU - Sabatti, Chiara
AU - Karsten, Stanislav L.
AU - Geschwind, Daniel H.
N1 - Funding Information:
The first author thanks Emmanuel Candes for suggesting the relevance of FDR procedures in this context; additionally she thanks organizers and participants of the IPAM session on functional genomics during the fall of 2000 for their generosity in sharing ideas. We also thank Stan Nelson and Zhun Chen of the UCLA microarray core for array fabrication. Grant support from NIHM (D.H.G.) and NIA (S.L.K. and D.H.G.) is acknowledged for microarray work.
PY - 2002
Y1 - 2002
N2 - We consider array experiments that compare expression levels of a high number of genes in two cell lines with few repetitions and with no subject effect. We develop a statistical model that illustrates under which assumptions thresholding is optimal in the analysis of such microarray data. The results of our model explain the success of the empirical rule of two-fold change. We illustrate a thresholding procedure that is adaptive to the noise level of the experiment, the amount of genes analyzed, and the amount of genes that truly change expression level. This procedure, in a world of perfect knowledge on noise distribution, would allow reconstruction of a sparse signal, minimizing the false discovery rate. Given the amount of information actually available, the thresholding rule described provides a reasonable estimator for the change in expression of any gene in two compared cell lines.
AB - We consider array experiments that compare expression levels of a high number of genes in two cell lines with few repetitions and with no subject effect. We develop a statistical model that illustrates under which assumptions thresholding is optimal in the analysis of such microarray data. The results of our model explain the success of the empirical rule of two-fold change. We illustrate a thresholding procedure that is adaptive to the noise level of the experiment, the amount of genes analyzed, and the amount of genes that truly change expression level. This procedure, in a world of perfect knowledge on noise distribution, would allow reconstruction of a sparse signal, minimizing the false discovery rate. Given the amount of information actually available, the thresholding rule described provides a reasonable estimator for the change in expression of any gene in two compared cell lines.
KW - False discovery rate
KW - High dimension
KW - Minimax
KW - Sparcity
KW - Thresholding
UR - http://www.scopus.com/inward/record.url?scp=0036128656&partnerID=8YFLogxK
U2 - 10.1016/S0025-5564(01)00102-X
DO - 10.1016/S0025-5564(01)00102-X
M3 - Article
C2 - 11867081
AN - SCOPUS:0036128656
SN - 0025-5564
VL - 176
SP - 17
EP - 34
JO - Mathematical Biosciences
JF - Mathematical Biosciences
IS - 1
ER -