Deep analysis of cellular transcriptomes – LongSAGE versus classic MPSS

Hene L, Sreenu VB, Vuong MT, Abidi SH, Sutton JK, Rowland-Jones SL, Davis SJ, Evans EJ. (2007), BMC Genomics. 8, 333

Abstract

BACKGROUND: Deep transcriptome analysis will underpin a large fraction of post-genomic biology. ‘Closed’ technologies, such as microarray analysis, only detect the set of transcripts chosen for analysis, whereas ‘open’ e.g. tag-based technologies are capable of identifying all possible transcripts, including those that were previously uncharacterized. Although new technologies are now emerging, at present the major resources for open-type analysis are the many publicly available SAGE (serial analysis of gene expression) and MPSS (massively parallel signature sequencing) libraries. These technologies have never been compared for their utility in the context of deep transcriptome mining.

RESULTS: We used a single LongSAGE library of 503,431 tags and a “classic” MPSS library of 1,744,173 tags, both prepared from the same T cell-derived RNA sample, to compare the ability of each method to probe, at considerable depth, a human cellular transcriptome. We show that even though LongSAGE is more error-prone than MPSS, our LongSAGE library nevertheless generated 6.3-fold more genome-matching (and therefore likely error-free) tags than the MPSS library. An analysis of a set of 8,132 known genes detectable by both methods, and for which there is no ambiguity about tag matching, shows that MPSS detects only half (54%) the number of transcripts identified by SAGE (3,617 versus 1,955). Analysis of two additional MPSS libraries shows that each library samples a different subset of transcripts, and that in combination the three MPSS libraries (4,274,992 tags in total) still only detect 73% of the genes identified in our test set using SAGE. The fraction of transcripts detected by MPSS is likely to be even lower for uncharacterized transcripts, which tend to be more weakly expressed. The source of the loss of complexity in MPSS libraries compared to SAGE is unclear, but its effects become more severe with each sequencing cycle (i.e. as MPSS tag length increases).

CONCLUSION: We show that MPSS libraries are significantly less complex than much smaller SAGE libraries, revealing a serious bias in the generation of MPSS data unlikely to have been circumvented by later technological improvements. Our results emphasize the need for the rigorous testing of new expression profiling technologies.

Key figure: Comparison of tag sequences in three MPSS libraries produced from the same RNA sample

A. The three libraries were sampled to various sizes in a step-wise fashion to examine the effect of library size on the number of distinct tag sequences identified (as done for single SAGE and MPSS libraries in Fig. 1). Closed diamonds represent random sampling of tags from all three libraries combined. Open diamonds represent sampling of each library in turn. Clearly, although the number of distinct species identified by each library (with the possible exception of the third) appears to approach saturation, each library is sampling a different subset of sequences from the initial RNA pool. B. Venn diagrams showing the distribution of tag sequences between the three MPSS libraries. The library represented by the blue circle is the one used in most of the analyses presented in this study. Diagram (i) represents all the different tag sequences in the libraries. Diagram (ii) represents only those tags that match the genome; this reduces the influence of sequencing errors. In both comparisons, the majority of distinct sequences are found in only one library. Diagram (iii) represents known transcripts in the UTBS dataset found expressed in the sense direction. Here the pattern is less marked, but still only half the transcripts were observed in all three libraries (1,312/2,646). The improvement in the correlation of the libraries for known transcripts (i.e. those in the UTBS) was expected because more highly expressed transcripts are more likely to have been previously identified, and therefore known transcripts tend to be more abundant and have a greater chance of being sampled.