Serial Analysis of Gene Expression (SAGE) is a rapid method to study mRNA transcripts in cell populations 1. Two major principles underline SAGE: (1) short expressed sequenced tags (ESTs) are sufficient to identify individual gene products, and (2) multiple tags can be concatenated and identified by sequence analysis 12. With the ever-expanding sequence information available in public databases, identification of gene transcripts with SAGE tags has greatly facilitated transcriptome comparisons and gene identification 3.

SAGE data are usually analyzed with software packages like "SAGE300" or "SAGE2000". The majority of SAGE libraries use NlaIII or Sau3A (SalAI) as anchoring enzymes (AE) to create SAGE tags. Both of these enzymes have 4-bp palindromic recognition sequences (CATG for NlaIII and GATC for Sau3A) that flank individual ditags within concatemers. A major component of the software analysis is the identification of anchoring enzyme recognition sequences (AERS) that flank target sequences (SAGE ditags). After finding the first AE recognition sequence, the software continues reading the sequence until it finds the next one. The software then compares the distance between these recognition sequences with predicted ditag lengths (20–24 bp in the case of NlaIII or Sau3A), and ditags that are too short (<20 bp) or too long (>24 bp) are excluded. However, if the length of the AERS-flanked sequence satisfies the size criteria, it is identified as a ditag. This algorithm relies on the assumption that all sequences have correctly organized ditag concatemers; however, the cloning efficiency of SAGE rarely reaches 100%. In this report, we show that up to 5% of ditags from some SAGE libraries should be omitted from the final analysis. These false ditags (termed "quasi-ditags") result from genomic contaminants and apparently random combinations of nucleotides generated by cloning or sequencing errors. Using a mathematical model to simulate the frequency of quasi-ditags in DNA, we propose a method to exclude quasi-ditags from SAGE catalogs.

From twelve independent SAGE libraries, we analyzed numerous clones lacking organized ditag concatemers that would be excluded by SAGE software packages, including clones lacking inserts, clones with inserts containing bacterial or rodent genomic sequences, and clones with unidentifiable sequences (Figure 1). Depending on the quality of the SAGE library, examples of clones in Figure 1A,1B,1C can represent up to 50% of the total volume of clones sequenced 4, but generally range from 2–20%. A more typical example is taken from our R1 ES cell and AMH-II SAGE libraries 56, which contained 5,988 and 4,478 clones, respectively. The cloning efficiency was ~79% and ~76% (4,714 and 3,413 clones with inserts, respectively). Among these, 411 and 167 clones in the R1 ES SAGE library and 305 and 194 clones in the AMH-II library contained sequences with only 1 or 2 ditags, respectively.

During our sequence analysis of the clones that had produced a least number of ditags (1–2 per clone), we identified a subset of sequences (up to 40%) that contain ditags that may be false. Importantly, some of these "ditags" matched bacterial genomic sequences (Figure 2A), while others seemed to represent random combinations of nucleotides. Figure 2B show an example of a clone that contains a single ditag sequence embedded within a sequence of unidentifiable origin. Because most of this sequence is not composed of concatenated ditags, this embedded ditag may therefore represent a quasi-ditag, which should be excluded from further analysis. These two examples, among others, suggest that some inserts in pZErO-1 contain sequences that just by chance mimic SAGE ditags.

To predict the potential frequency of randomly occurring quasi-ditags, we employed a stochastic model system to generate random sequences. We then used both computer-generated random sequences and true genomic DNA sequences to test this possibility. Random sequences were generated and analyzed with a Visual Basic program designed to mimic SAGE software analysis of ditags. The simulated sequences varied in length from 600 to 1200 nucleotides, which corresponds to the average sequence lengths generated by automated sequence analyzers. One million random sequence strings with L = 600, 700, 800, 900, 1000, 1100, and 1200 nucleotides were generated. Table 1 shows expected frequencies of quasi-ditags according to the model (equation (5)) and the observed frequencies based on computer simulations. The line plots of the expected (model) and observed (computer simulation) quasi-ditag frequencies are almost identical (Figures 3 and 4). Fragmented Saccharomyces cerevisiae genomic DNA that lack SAGE ditag concatemers was also employed for in vivo / in silico model validation, and a number of quasi-ditags was detected in these (Figures 4 and 5). When compared to Saccharomyces cerevisiae genomic DNA 7, quasi-ditag frequencies were somewhat less abundant than those generated by the computer, potentially due to the presence of nucleotide repeats and unequal frequencies of individual nucleotides in the Yeast genomic sequences. These data, however, support our hypothesis that quasi-ditags can be generated randomly from potential sequencing errors or from genomic contaminants. This analysis furthermore underscores the limited extent that quasi-ditags occur: the distribution of expected number of quasi-ditags per clone is clearly bimodal, with peaks at 1 and 2 ditags (Q₁ and Q₂, respectively). At the same time, the frequency of occurrence for three quasi-ditags (Q₃) is extremely low (0.01% for L = 600 to 0.02% for L = 1200), such that the value of P³ _(20–24) effectively converges to zero for the majority of the SAGE catalogs (i.e. <3000–5000 clones) (Figure 4). Accordingly, the clones that include ditag concatemers of higher length should lack quasi-ditags.

Clones containing only one or two ditags/quasi-ditags, however, could be excluded from SAGE analyses, without adversely affecting the data set (Figure 6). As an example, we extracted sequences from clones that produce 1–2 total ditags from AMH-II and R1 ES cell libraries. This reduced the total number of tags by 1.06% for ES R1 and 1.94% for AMH-II, but it effectively removed all contaminating bacterial sequences and improved the data reliability. However, the total AMH-I library (2,365 clones, ~78% cloning efficiency; 6) had a larger proportion of ditags extracted as being too long (>24 bp), as indicated by lower tag per clone ratio (average insert size of 12.2 tags/clone vs. 22.6 in AMH-II library) amid the same average sequence length, suggesting higher proportion of quasi-ditags. Analysis of the AMH-I SAGE library has revealed 353 and 52 clones that contained just 1 or 2 ditags, respectively. Exclusion of these sequences decreased the total number of tags by 5.21% (calculated after duplicate dimer exclusion), and proved critical to our subsequent quantitative SAGE comparisons 6. Failure to remove these quasi-ditag sequences decreased the quantitative reproducibility (R values) between AMH-I and AMH-II SAGE libraries, showing that quasi-ditags can adversely affect the reliability of SAGE libraries.

SAGE is an important tool of modern molecular biology widely used in a number of applications. We hypothesized that actual SAGE catalogs could be contaminated by false ditags ("quasi-ditags") of various origins. Although SAGE software packages are designed to ignore sequences that lack 20–24 bp sequences flanked by two anchoring enzyme recognition sites, it does not exclude quasi-ditags originating from genomic contaminants or unknown sequences that may arise as cloning or sequencing artifacts (Figure 2). Negative controls (self-ligated vector) do not produce any colonies after Zeocin selection and cannot account for the appearance of background clones and quasi-ditags in Zeocin-resistant bacteria. Since some quasi-ditags, however, originate directly from E. Coli, we suggest that one probable source for these contaminanting tags is from recombination events that occur in E. Coli. Indeed, such a mechanism has already been documented 8 and has led to the development of Stbl2 bacteria that are mcrA^-/mcrBC^-hsdRMS^-mrr^-. Since pZErO-1 was not translated into recombination deficient bacteria (DH10B), large-scale amplifications of this plasmid within bacteria would be expected to lead to some random recombinations, and the generation of quasi-ditags (e.g. Figure 2A).

Some of the ditags derived from the clones that had produced a least number of ditags (1–2 per clone) do not match genomic sequences and thus might be originated from sequencing errors. We therefore suggested a model that provides a mathematical basis for the hypothesis that such a possibility exists. The mathematical model presented in the manuscript is an attempt to predict the frequency distribution of quasi-ditags in random sequences. The phenomenon itself is rather complex and there is no simple model that would capture it in full complexity. We, however, believe that we have selected a reasonable level of model complexity that captures the major pattern of frequency distribution.

Using the computer simulation we show that random combinations of nucleotides generated could be indeed recognized by SAGE software as valid SAGE ditags. We also demonstrate that quasi-ditags may constitute a non-negligible proportion of SAGE catalogs. Our model, which simulates the frequency of quasi-ditags in DNA (equations (1–6)), suggests that single or double ditags may represent quasi-ditags; however, the results of the in silico experiments show that the probability of finding more than two quasi-ditags in the same sequence converges effectively to zero (Table 1 and Figure 4). Based on these findings, we suggest that additional steps be performed with SAGE libraries. We recommend removing clones with sequences containing ≤ 2 ditags at a pre-processing step ("clean-up"). The removal of clones containing 1 or 2 ditags can effectively remove bacterial genomic sequences and potential sequencing artifacts from SAGE libraries. The overall number of SAGE tags excluded by this additional step (authentic and quasi-ditags) is usually low, and generally does not exceed 1.0–1.8% of the total number of sequenced SAGE tags 56910; however, the frequency of potential quasi-ditags could be high (>5%) in some SAGE libraries. In AMH-I library, for example, the fraction of clones lacking appropriate ditag concatemers was >20%. In these instances, quasi-ditags significantly contribute to the final SAGE tag count, and should be removed.

Chart in Figure 6 plots values for ditag distribution from both the model-based simulations (L = 800 bp) and actual clones from the SAGE libraries that had sequences of the same mean length (L ≈ 800 bp). The expected maximum frequency of 1–2 quasi-ditags in the plotted model data approximated the observed frequency of clones with 1–2 total ditags detected in the pool of the actual SAGE clones. Contrary to that, the frequency of occurrence of three or more quasi-ditags predicted by the model is extremely low, demonstrating a divergence in the distribution of expected quasi-ditags and valid SAGE ditags for higher number of ditags per clone. Note that owing to the gel-purification of concatemers the majority of clones in the representative samples belong to the clusters of higher ditag numbers (AMH-II and ES R1 libraries, 13–26 total ditags; AMH-I library, 4–11 total ditags).

Comparing values of observed frequencies of the actual SAGE clones that produce 1–2 total ditags with those of expected quasi-ditag frequencies for the sequences of given length might be indicative on the possible contribution of cloning and sequencing artifact-derived quasi-ditags (Figure 6). The possible contribution of quasi-ditags to the final tag yield in SAGE libraries cannot be accurately predicted in advance but a failure to report the cloning efficiency and the number of clones with 1 or 2 ditags precludes an evaluation of potential false tags present in SAGE catalogs. Current SAGE protocols do not ensure 100% accurate size fractioning of concatemers: some of the smallest concatemers could therefore be cloned and sequenced. We recognize that some authentic tags (representing valid, but extremely short inserts that were not extracted during gel-purification of concatemers) will be excluded by removing all clones containing only 1 or 2 ditags. Nevertheless, we suggest that any potential loss of authentic ditags in the clean-up procedure is negligible compared to the advantage of having more reliable SAGE results.

SAGE protocols are extremely complex technologically and every possible mean should be employed to ensure qualitative and quantitative accuracy of catalogs on both the experimental and analytical steps. Evaluation of the cloning efficiency and precision (e.g. with RAST-PCR 11) and sequencing accuracy are therefore essential on the stage preceding large-scale sequencing of the clones. Nonetheless, introduction of the simple pre-processing step eliminating false ditags would further improve the accuracy of the method resulting in its wider application.

We have hypothesized that actual SAGE catalogs could be contaminated by false ditags (termed "quasi-ditags") of various origins and employed a mathematical model to predict the frequency of quasi-ditags in random nucleotide sequences. Cloning and sequencing artifacts contaminating SAGE libraries could be eliminated using simple pre-screening procedure to increase the reliability of the data.

SVA developed the hypothesis, overall plan and performed SAGE, computer simulations, and analysis of Yeast genome fragments. AAS developed and implemented the mathematical model predicting the appearance of "quasi-ditags" in random sequences of given length. Both authors have contributed to the writing and approved the final manuscript.