Serial Analysis of Gene Expression (SAGE) is a method for measuring the relative abundance of gene transcripts in different mRNA samples. It identifies a short mRNA tag from each individual transcript and concatenates them into long DNA molecules, which are then sequenced. By counting these tags one can estimate, for example, the expression of genes in a cell [1]. SAGE popularity is growing fast and many public data are accessible from the Internet [2].

Processing of SAGE sequences is described in [1], [2] and [3]. The usual length of SAGE tag is 10 bases. In practice, this length is not sufficient to uniquely identify each gene: several genes share the same tag. The SAGE method uses tag-pairs to avoid bias by PCR amplification. As pointed out in [2], the observed length of the di-tags is not constant, it varies between 20 and 26 (see Table 1) due to a certain flexibility in the enzyme used. The usual processing of SAGE sequences does not take advantage of these longer di-tags. Here we present a new method to predict an 11^th base for sufficiently abundant tags, hence increasing precision in gene identification (the number of possible genes to which a tag is mapped is divided by 4 on average).

We use di-tags of sufficient length to compute the frequencies of the four possible extra bases (A, C, G and T) for every tag. Then we use contingency tables and hypothesis testing [4] to determine relevant extra bases. The null hypothesis we apply is that every possible extra base has the same probability to be sequenced.

In the publicly available data set [5], we used as an example a SAGE library made for Homo sapiens normal white matter [6]. We used the di-tag list of this SAGE library to exemplify the usefulness of our method. [6] contains 51640 di-tags of length between 20 and 26 bases. We rejected 3856 suspect repeated di-tags (see [2] and [3]). From the remaining 47784 di-tags we extracted 32668 different tags. The number of tags for which we could predict an 11^th base by applying our method is given in Table 2.

For illustration purpose, we identified these tags by extracting SAGE tags of UniGene [7] clusters (build 108). We only considered tags at the end of the UniGene sequences, i.e. we consider UniGene sequences as 5' oriented. Other identification strategies are possible, see for instance [2]. An example of a tag is CAAGCATCCC, observed 1510 times with 5 extra As (the base A was observed at the 11^th position in the di-tag), 1426 extra Cs, 13 extra Gs and 13 extra Ts. We uniquely identified this tag in UniGene as Hs.250444 small inducible cytokine A7 (monocyte chemotactic protein 3). The 11^th base found in the UniGene cluster sequence matches with the dominant extra C we mention above. According to the method we propose (see Materials and Methods), the prediction of C as an extra base is relevant at the 99.9% level.

An example of a tag shared by two genes, one of which is apparently not expressed, is provided by GGGCTGGGGT, observed 86 times with 5 extra As and 80 extra Cs. GGGCTGGGGTA is identified in UniGene [7] as Hs.90436 sperm acrosomal protein (SPAG7) and GGGCTGGGGTC as Hs. 183 698 ribosomal protein L29 (RPL29). According to the extra bases observed, it seems that only RPL29 is expressed (99.9% relevant, SPAG7 is possibly weakly expressed).

The special situation of several expressed genes sharing the same 10-base tag is illustrated by GTGAAACCCC, observed 422 times with 161 extra As, 22 extra Cs, 202 extra Gs, and 12 extra Ts. According to the null hypothesis (equiprobability of every extra base), both A and G are relevant at a higher probability than 95% (99.9% in this case). We can estimate a count of 161/(161+202).422=187 for GTGAAACCCCA and 202/(161+202).422=235 for GTGAAACCCCG. We subsequently found that this tag is shared by many UniGene[7] clusters: 49 clusters with extra A, 7 cluster with extra C and 54 clusters with extra G.

[2] proposes the assignment of a score to each identification, in order to characterize its reliability. If a tag comes with a predicted extra base, the latter should be checked with the database sequence and the result included in the score computation.

The complex situation of tag GTGAAACCCC above suggests a possible extension of our method. We test the relevance of predicted extra bases by comparing (hypothesis testing) the observed frequencies with the hypothetical situation of equiprobability. Another possible null hypothesis would be (1) to chose a method for identifying tags, as we did with UniGene [7], and (2) to estimate the relevance of the possible extra bases according to this new null hypothesis. Returning to the example of tag GTGAAACCCC, none of the four possible extra bases significantly departs from the distribution obtained from UniGene. This implies that no extra base can be selected reliably and, consequently, every possible extra base should be considered. We cannot obtain any simplification of the data in that case, contrary to what we found with the equiprobability null hypothesis.

We do not apply the latter extension of the method in practice for two reasons: first, this extension is dependant on the method for identifying tags and, second, considering the difficulty in analyzing SAGE data, we prefer to concentrate on dominantly abundant extra bases for the sake of simplicity.

We presented a method that allows for the prediction of one extra base for sufficiently abundant tags (at least 7 occurrences). The method applies to every SAGE library, without any special preparation. The predictions may be controlled in terms of relevance by using appropriate hypothesis testing techniques. The longer tags permit a better identification of expressed genes.

We assume that, in the case of di-tags of length 20 or more, the first 10 bases belong to the first tag and the last 10 bases belong to the second tag. Since the tags are linked into di-tags randomly, the extra available bases, in the middle of a di-tag of length 22 or more (see Figure 1), belong to each tag with a probability that is symmetrical. Accordingly, we propose a new di-tag processing.