The correlation between mRNA and protein expression is typically not strong, as previously reported [2]. On the other hand, transcription is one of the most important steps in gene regulation, and information about transcript levels is important to estimate gene activity and to characterize a molecular signature for a cellular phenotype. In this context, new methods have been developed for transcriptome analysis and gene expression profiling [3].

Serial Analysis of Gene Expression (SAGE) [4] is one of the widest used techniques for this purpose. SAGE method allows a quantitative and parallelized analysis of a large number of gene transcripts in any particular cDNA library derived from cells or tissues [5], without prior knowledge of the genes. The SAGE technique is based on the isolation of short sequence tags that are extracted from defined positions of the transcript (3'-most anchoring enzyme restriction site; NlaIII is the most commonly used enzyme for this purpose). These tags, in theory, are strongly associated with individual genes [6]. The tags are concatenated into long DNA molecules, which subsequently are identified by conventional DNA sequencing. Sequence tags are extracted from raw sequence concatamers using a parsing program that tabulates the occurrence of each tag and creates an abundance level report file.

Although being advantageous in many aspects [3], SAGE protocol can be also subject to sequence errors, mainly introduced by the polymerase chain reaction (PCR) and sequencing steps. The occurrence of these errors produces sequence-based artifacts, introducing noise into the sampled transcriptome profile. Massively Parallel Signature Sequencing (MPSS) [7] is another sequencing-based method that, similarly to SAGE, provides quantitative gene expression data. This technique can generate millions of short signature sequences, providing a better coverage of the transcriptome in a single cell population. It has a routine sensitivity on a low-molecular level of mRNA per cell.

Another recent technology with high sensitivity is DeepSAGE [8], which is similar to the initial steps of LongSAGE [9] in conjunction with an ultra-high throughput DNA sequencing [10].

In recent years, significant efforts have been directed towards the development of new and reliable DNA sequencing technologies [11]. All three commercially available technologies today are characterized by the production of millions of reads in a single experiment [12,13], reducing the cost of sequencing.

These technologies are available for quantitative measurement of gene expression and, as in SAGE, individual mRNAs can be identified through short tag sequences, immediately adjacent to the 3'-most anchoring enzyme restriction site. The throughput reached by these new sequencing technologies allows also a shotgun strategy for the transcriptome.

Although these represent technological advances, there are still many factors that potentially contribute to the production of artifact sequence tags, such as errors introduced through reverse transcription, PCR, and others, inherent to the cloning (when necessary) and sequencing steps, leading to spurious molecular signatures. In order to mitigate these negative traits, several studies were made to define and explore the inherent implications, mainly to the SAGE protocol. In addition, these studies disclosed the impacts of the spurious sequences and proposed efficient algorithms for detection and removal of these tags [14-17]. In this work, we propose a new methodology to index the sequence tags according to their reliability, using a score system. We also present the implementation of this methodology in a web-based tool, initially available for human (Homo sapiens) and mouse (Mus musculus) data. The main applications of this tool are the evaluation of the sequenced library, and the identification and selection of the reliable tags, which can be considered more effective for gene expression profiling, according to the scoring system. To rank the tags, we established some empirical rules merging relevant parameters, one based on absolute tag frequency and the other, on matches in a set of virtual tag databases. There is still another component associated with the previous, based on a concept of neighborhood [15], i.e. abundant tags can contaminate those whose sequence is very close. The virtual tags are predictions of a real tag that might be produced by a SAGE (10 bps for short SAGE, 17 bps for long SAGE), MPSS (13 bps for short MPSS, 16 bps for long MPSS) or SBS (16 bps or 17 bps) experiments using different anchoring enzymes (NlaIII or DpnII). They are extracted from gene transcript sequences from various sources with some relevant attributes [18], considered in each score rule elaboration. The other rule component, based on the absolute frequency, assumes that the abundance and representation in the set of libraries presupposes more reliability. The automated library analysis process involves a series of searches on S3T database to retrieve relevant source information from corresponding virtual tag for the application of a specific classification rule set (Table 1). Another approach to rank tags according to their nature, using a relational database and a set of different transcript sequence sources, was previously described [19]. However it does not consider the possibilities of tags derived from SAGE adapters or cloning vectors, or tags matching mitochondrial or nuclear genomes, or tags that can be technology-related artifacts, such as generated by sequencing errors. The approach of S3T comprises these analyses, using additional virtual tag datasets and still considers the tag frequencies to rank the tags.

There is an online service platform, publicly available at S3T website, for accessing and performing the referred analysis. The source code is also available for a local installation.

According to the initial analysis using 359 public SAGE libraries, it is possible to observe the large diversity among proportions of tags throughout the established scores (Figure (Figure1).1). The high percentage of unique tags with -2 score (see Table 1) is evident; they may have arisen from errors in abundant tags whose sequence is very close. This indicates that errors in the tag sequence are the most prevalent. The correlation between the total of sequenced tags in a library and the number of unique tags with -2 score is strong (0.91), i.e. the number of unique tags with -2 score increases according to the number of sequenced tags. The correlations between the total of tags in a library and the unique tags in the other scores were obtained and are available at the supplemental section of the S3T website.

These results can illustrate the overall quality of tag libraries. Those samples observed with a high percentage of tags classified with negative scores, cannot be considered reliable for further analysis. Box plots were used to visually summarize and compare the groups of data [22] (Figure (Figure1).1). This diagram contains the distributions of unique tags along the proposed scores. It is crucial to focus particularly on those with positive scores that represent the effective contribution for the biological information, the others, with negative scores, possibly represent contamination or technique errors inherent to the experimental process. However, the tags with negative scores that had matches in the genome sequences, or did not match on anything, could also come from unknown transcripts [23]. But it is also important to emphasize that these unknown transcripts will probably be characterized through the constant updates of the current transcript sequence databases, especially with the use of new high-throughput sequencing technologies, which has high sensitivity to detect rare transcripts. The hierarchical clustering analyses using the selected histological groups (14) of SAGE libraries were also performed. The clusters of samples with known histological information, in a general way, were more concordant with their predefined histological sample classes (see Implementation), after filtering, in some cases (32.65%). In the others, the filtering did not affect the overall quality of cluster set (58.16%) and, in a few cases (9.18%), the clusters after filtering were less concordant with their previously defined classes (Table 2). The brain group is one of them, once there are many heterogeneous samples into subgroups, leading to difficulties to distinguish them among other samples. The F-Measure [24] was used to evaluate the cluster quality. This measure cluster evaluation method combines the precision and recall ideas. The F-Measure values are in the interval [0–1] and larger F-Measure values indicate higher clustering quality. The clusters formed before and after S3T filtering were then compared using the Overall F-measure (Table 2), the expectation was to find clusters more congruous after filtering, properly grouping samples from the same cellular phenotypes together and closer than others (Figure (Figure22).

In this analysis, applying the proposed set of rules (Table 1), the results (Figure (Figure3)3) point that the tags with score -2 is the most abundant among those who received negative scores. This was observed for almost all libraries, considering the count of unique tags and the sum of their frequencies (95.26% and 99.16%, respectively). This obviously affects the difference between the raw data and the S3T-filtered data, and indicates which libraries do not have enough tags with good quality; the criterion for this consideration depends on the analysis stringency parameters.

1. Collecting experimental data, virtual tag datasets and related information, and store them into the relational database;

2. Setting the score classification rule parameters (Table 1);

3. Carrying out the searches on database and applying the rules, that is a match of the experimental data with one of the tag scoring rules.

4. Quantitating the unique tags and tag frequencies for each score to perform a library quality evaluation.

Project home page: http://gdm.fmrp.usp.br/s3t/

Operating system(s): UNIX-like Platforms

Programming language: Perl

Other requirements: Perl v5.8.0 or higher, CPAN modules, MySQL 4.1.13 or higher

License: GNU GPL