BIOINFORMATICS Vol. 20 Suppl. 1 2004, pages i31–i39 DOI: 10.1093/bioinformatics/bth924 Statistical modeling of sequencing errors in SAGE libraries Tim Beißbarth 1, * , Lavinia Hyde 1 , Gordon K. Smyth 1 , Chris Job 2 , Wee-Ming Boon 2 , Seong-Seng Tan 2 , Hamish S. Scott 1 and Terence P. Speed 1 1 Walter and Eliza Hall Institute of Medical Research, Genetics and Bioinformatics, 1G Royal Parade, Parkville, Vic 3050, Australia and 2 Howard Florey Institute, Brain Development Laboratory, University of Melbourne, Parkville, Vic 3010, Australia Received on January 15, 2004; accepted on March 1, 2004 ABSTRACT Motivation: Sequencing errors may bias the gene expression measurements made by Serial Analysis of Gene Expression (SAGE). They may introduce non-existent tags at low abund- ance and decrease the real abundance of other tags. These effects are increased in the longer tags generated in Long- SAGE libraries. Current sequencing technology generates quite accurate estimates of sequencing error rates. Here we make use of the sequence neighborhood of SAGE tags and error estimates from the base-calling software to correct for such errors. Results: We introduce a statistical model for the propagation of sequencing errors in SAGE and suggest an Expectation- Maximization (EM) algorithm to correct for them given observed sequences in a library and base-calling error estim- ates. We tested our method using simulated and experimental SAGE libraries. When comparing SAGE libraries, we found that sequencing errors can introduce considerable bias. High abundance tags may be falsely called as significantly differ- entially expressed, especially when comparing libraries with different levels of sequencing errors and/or of different size. Truly, differentially expressed tags have decreased signific- ance as ‘true’-tag counts are generally underestimated. This may alter if tags near the threshold of differential expres- sion are called significant. Moreover, the number of different transcripts present in a library is overestimated as false tags are introduced at low abundance. Our correction method adjusts the tag counts to be closer to the true counts and is able to partly correct for biases introduced by sequencing errors. Availability: An implementation using R is distributed as an R package. An online version is available at http://tagcalling.mbgproject.org Contact: beissbarth@wehi.edu.au ∗ To whom correspondence should be addressed. 1 INTRODUCTION Serial Analysis of Gene Expression (SAGE) is a gene expression profiling technique that estimates the abundance of thousands of gene transcripts (mRNAs) from a cell or tissue sample in parallel (Velculescu et al., 1995). SAGE is based on the sequencing of short sequence tags that are extracted at defined positions of the transcript. As opposed to DNA microarray technology (Schena et al., 1995; Lockhart et al., 1996), SAGE does not require prior knowledge of the tran- scripts, and results in an estimate of the absolute abundance of a transcript. However, due to sequencing errors a propor- tion of the low-abundance tags do not represent real genes altering the ability of SAGE to estimate the number of tran- scripts that have been observed. Moreover, loss of ‘true’-tags due to sequencing errors will result in altered numbers for the abundance of genuine transcripts. Stollberg et al. (2000) have studied the effects of various sources of errors on SAGE results by simulating libraries. Pre- viously sequencing errors have been minimized by removing low-abundance tags or tags with low sequence quality from the libraries (Margulies and Innis, 2000). Velculescu et al. (1999) first attempted to join low-abundance tags to their neighborhood. A more refined approach, that assumes con- stant error probabilities and uses matrix inversion to correct for sequencing errors, has been introduced by Colinge and Feger (2001). Another recently developed approach by Blades et al. (2004a,b) uses a linear relation between the copy number of observed tags and the number of neighbors with one-base sub- stitutions to estimate the average rate of sequencing errors and eliminate unreliable tags. Akmaev and Wang (2004) use such error estimates to correct for sequencing errors and PCR based artifacts. They estimate that in LongSAGE libraries 3.5% of the tag sequences have errors resulting from PCR artifacts and 17.3% of the tag sequences have errors resulting from sequencing errors. These approaches, however, do not take into account estimates for sequencing errors of the individual bases. Bioinformatics 20(Suppl. 1) © Oxford University Press 2004; all rights reserved. i31 by guest on March 9, 2016 http://bioinformatics.oxfordjournals.org/ Downloaded from