Differential gene expression in an elite hybrid rice cultivar (Oryza sativa, L) and its parental lines based on SAGE data

Background It was proposed that differentially-expressed genes, aside from genetic variations affecting protein processing and functioning, between hybrid and its parents provide essential candidates for studying heterosis or hybrid vigor. Based our serial analysis of gene expression (SAGE) data from an elite Chinese super-hybrid rice (LYP9) and its parental cultivars (93-11 and PA64s) in three major tissue types (leaves, roots and panicles) at different developmental stages, we analyzed the transcriptome and looked for candidate genes related to rice heterosis. Results By using an improved strategy of tag-to-gene mapping and two recently annotated genome assemblies (93-11 and PA64s), we identified 10,268 additional high-quality tags, reaching a grand total of 20,595 together with our previous result. We further detected 8.5% and 5.9% physically-mapped genes that are differentially-expressed among the triad (in at least one of the three stages) with P-values less than 0.05 and 0.01, respectively. These genes distributed in 12 major gene expression patterns; among them, 406 up-regulated and 469 down-regulated genes (P < 0.05) were observed. Functional annotations on the identified genes highlighted the conclusion that up-regulated genes (some of them are known enzymes) in hybrid are mostly related to enhancing carbon assimilation in leaves and roots. In addition, we detected a group of up-regulated genes related to male sterility and 442 down-regulated genes related to signal transduction and protein processing, which may be responsible for rice heterosis. Conclusion We improved tag-to-gene mapping strategy by combining information from transcript sequences and rice genome annotation, and obtained a more comprehensive view on genes that related to rice heterosis. The candidates for heterosis-related genes among different genotypes provided new avenue for exploring the molecular mechanism underlying heterosis.


Background
Heterosis is defined as advantageous quantitative and qualitative traits of offspring over their parents, and the utilization of heterosis principles has been a major practice for increasing productivity of plants and animals [1].
A considerable amount of efforts have been invested in unraveling genetic basis of heterosis in rice (Oryza sativa, L) and it was explained mainly by mechanisms such as dominance [2] and epistasis [3]. Although many investigators favored one hypothesis over another, biological mechanisms of rice heterosis may not be fully characterized based on genetic approaches alone, especially based on classical genetic concepts.
Recently, it has been reported that differentially-expressed genes between hybrids and their parental inbreeds are correlated with heterosis [4,5]. In wheat, a variety of differentially-expressed genes including transcription factors and genes involved in metabolism, signal transduction, disease resistance, and retrotransposons were detected responsible for heterosis by using a differential display technique [6,7]. Even ribosomal proteins have been scrutinized since they are indicators of translation activities and plastid biogenesis [8]. Various techniques have been applied to pin down genes involved in heterosis, such as a variety of sequence-based and hybridization-based methods; some have yielded interesting candidates and others proposed expression patterns of these candidates [5,9]. For instance, a hybrid-specific expressed gene AG5 (a RNA-binding protein) in wheat was identified [10]. Another study on gene generated expression profiles of an elite rice hybrid and its parents at three stages of young panicle development by using a cDNA microarray consisting of 9,198 ESTs and the result pointed to a significant mid-parent heterosis [11]. Nevertheless, it is necessary to generate more data in large-scale, taking the advantage of the fast advancing genomic technology. SAGE technology is a sequence-based approach for investigating gene expression in large-scale and allows much deeper sampling than EST (expressed sequence tag)-based approaches. It has proven to be a very powerful method for large-scale discovery of new transcripts, acquisition of quantitative information of expressed transcripts, and the quantitative comparison between libraries [12][13][14]. The technique has been used extensively in animal systems including human and mouse, and more particular in cancer research where several hundred libraries and nearly 7 million SAGE tags have been obtained [13,15]. In plant, several studies have employed this methodology for transcript profiling in Arabidopsis [16,17] and rice [18,19]. However, a bottleneck of SAGE is tag-to-gene mapping, which refers to the unambiguous determination of the gene represented by a SAGE tag. Other limitations include lack of accurate genomic sequences and adequate amount SAGE data. Therefore, encouragements should be given to studies that generated publicly available data since heterosis is not simply a manifestation of a few seemingly important genes but many.
We have been studying the rice genome with a particular interest in the molecular mechanism of heterosis as part of the Super-hybrid Rice Genome Project (SRGP), focusing on an elite super-hybrid (Liang-You-Pei-Jiu, LYP9 [20]) and its parental lines, using gene expression technology, including EST and SAGE techniques. The objective of our current work was to recover more sequence tags (gene expression information) from our previous SAGE study [21]. In our new analysis, SAGE tags were mapped to two newly annotated genome assemblies, paternal cultivar (93-11) and maternal cultivar (Pei-Ai 64s, PA64s) (BGI unpublished data) [22,23]; the latter was not available when we carried out the first analysis. Prefect matches of SAGE tags to their own genome sequences allowed us to map more tags in a very significant way: twice as much tags were mapped as compared to the previous result. We also used three types of transcripts, including full-length cDNA (FL-cDNA) [24], expressed sequence tags (ESTs) [25,26], and UniGene data as well as a new strategy in the current analysis.

The dataset
We obtained a total of 465,164 SAGE tags from nine SAGE libraries constructed in parallel from the three major rice tissues at distinct growth stages for the super-hybrid rice (LYP9) and its parental (93-11 and PA64s) cultivars. These libraries were made with mRNA isolated from (1) leaves at the milky stage of rice grain maturation, (2) panicles at the pollen-maturing stage, and (3) roots at the first tillering stage [21]. By using more stringent sequence-analysis criteria in a quality-improving protocol, we removed contaminated tags matched to cloning linkers, vectors, and simple repeats, and obtained 68,462 unique empirical tags; this number is 21 tags less than the previous dataset due to more stringent filters. Of these unique tags, 30,595 (44.7%) tags were observed more than once. The distribution of the mapped tags among different libraries is summarized in Table 1. We deposited all the original SAGE data in NCBI's Gene Expression Omnibus [27] and these data are accessible through GEO Series accession number GSE8048.

Evaluation dataset, virtual tags, and mapped tags
To obtain an evaluation dataset, we constructed a PCUE (Predicted genes, FL-cDNA, UniGene, and EST) database based on available genomic resources (see Materials and Methods). We classified 41,072 predicted genes of 93-11 into three sets: (1) 21,676 (53%) supported by one or more transcripts, i.e. by any of three pieces of supporting evidence (or types of transcripts) -FL-cDNA, UniGene, and EST, (2) 19,396 without supporting evidence, and (3) 10,702 supported by all three types of transcripts. This evaluation dataset contains 2,480 test tags from (3) and satisfies all five quality criteria (see Materials and Methods; Table 2).
In order to define virtual tags, we need to handle two classes of virtual transcripts based on predicted genes: (1) supported by transcripts that have actual 3'-UTR sequences ( Figure 1A) and (2) without supporting evidence but defined by adding an artificial 3'-UTRs ( Figure  1B). From the first class, we categorized 13 different groups of virtual tags based on variable 3' UTR sequence features (in Table 2). We also found that the virtual tags constructed from the longest UniGene (Unimax, 97.22%) and the longest EST (ESTmax, 74.92%) had better yield in matching the virtual tags to the test tags, largely due to their longer 3'-UTRs. As a comparison, the virtual tags constructed from the Uni-S and EST-S groups that possessing poly (A) signals had slightly poorer but significant yields -95.80% and 71.60%, respectively. For the second class, we need to choose a length range for artificial UTRs that are to be added to the predicted genes. For 19,079 non-redundant FL-cDNAs (see Additional file 1: UTR Size distribution), whose 3-UTRs have a distinct length distribution with a mean of 422 bp and a median of 295 bp, we decided to use a 100-bp window and an optimal length range of 300 bp. The four sets of virtual tags, including cDNA, Unimax, ESTmax, and predicted genes with 300 bp 3'-UTR, were used for further analyses (Table 2).
We assigned 20,595 unique tags to 19,961 predicted genes (Table 3) in three types: (1) 16,757 (81.36%) unambiguous tags, (2) 3,316 (16.10%) tags physically-mapped to 1,668 genes (two or more different tags assigned to the same predicted genes), and (3) 698 (3.39%) tags physically-mapped to 1,536 genes (each tag assigned to multiple genes). Among these mapped tags, 16,430 (80%) were supported by transcripts and 4,341 (20%) were not supported by known evidence; the latter are largely hypothetical transcripts that are either expressed at lower level or specific to certain tissues or developmental stages (based on microarray and EST analyses of our own data; data not shown). This process led to a more rigorous tag-to-gene assignment, allowing us to gain 10,268 additional tags, compared to our previous results. In addition, we found that 1,610 previously mapped tags were absent in the current data, and the missing tags were filtered out by the  more stringent criteria used in this study that resulted in a removal of 1,649 FL-cDNAs as compared to the previous data set. There were 45,025 unmapped tags that did not satisfy our stringent criteria (see Materials and Methods for details).

Differentially-expressed genes among twelve distribution patterns
We defined differentially-expressed genes by calculating P values between any two libraries using a previously reported statistic method [28]; the process yielded 1,751 (8.5%) and 1,216 (5.9%) significant differentiallyexpressed genes with P values of < 0.05 and < 0.01, respectively (Table 4). In the process of summarizing overall expression profiles, regardless the origin of tissues, we found 781, 360, and 324 differentially-expressed genes from pair-wise comparisons of LYP9 versus PA64s (L vs. P), LYP9 versus 93-11 (L vs. N), and LYP9 versus both parental cultivars (both) at a less stringent threshold (P < 0.05), respectively. There is an obvious bias -the genes with paternal-like expression (PLE; L vs. P) are twice as much as those with maternal-like expression (MLE; L vs. N). This bias suggests that LYP9 possesses more differentially-expressed genes from PA64s than from 93-11, regardless whether they are up-regulated or down-regulated; in other word, LYP9 is more similar to 93-11 than to PA64s in its overall gene expression.
We further examined the profiles of differentiallyexpressed genes by classifying them into 12 different distribution patterns, displayed separately according to different tissues, and plotted the intensity of gene expression Description of the strategy used to construct the conceptual transcript Methods) were used for the construction of virtual transcripts. When the transcript sequences extend beyond the predicated coding sequence were available, the UTR sequences were aligned and determined (A). When the information was not available, the theoretical 3' UTR sequences were determined based on a stepwise (100-, 200-, 300-, 400-, and 500 bp) assessment of the genome sequences and added after the stop codons (B). Nearly 58.7% of the assigned tags have a 3'-UTR length of 300 bp.
as fold changes (less than 16-fold) at P < 0.05 and P < 0.01 ( Figure 2). There were 686, 568, and 413 genes differentially-expressed in panicles (see Additional file 2), leaves (see Additional file 3), and roots (see Additional file 4), among the triad at P < 0.05, respectively. The corresponding numbers were 599, 393, and 240 at P < 0.01. Genes that show changes of >16-fold and genes that only assigned to PA64s are also listed (see Additional file 5). In order to describe the gene distribution clearly according to their relationship between the hybrid and its parents, we partitioned the twelve distribution patterns into three basic categories: over-dominance (the top four slices), under-dominance (the bottom four slices), and mid-parent (the four slices divided by the horizontal line).
From the overall distribution of differentially-expressed genes with higher P values (P < 0.01), we made several observations among the samples. First, gene distribution pattern in panicles is rather distinct and more biased than that in the other two tissues, in such a way that most of the down-regulated genes are very paternal-like (or almost identical to 93-11, N = L < P) and the up-regulated genes are rather dispersive (not focused along the solid line of N = L > P). The dispersiveness suggested that most of these genes are roughly paternal-like but their expression levels are approximating toward either the hybrid (LYP9) or the mid-parent in a quantitative manner. We speculate that this obviously restricted distribution in panicles may be either due to one or both the following possible biases. One bias may come from thermo-sensitive male sterility unique to the maternal cultivar, PA64s, where germlinerelated genes may be crippled in their overall gene expression though epigenetic mechanisms. The other possible bias may be resulted from incompatibility between alleles from the parental lines, which may cause a rather major regulatory effect for the majority of genes, such as DNA methylation in germline tissues. Second, the distribution of genes in leaves and roots are somewhat similar, especially among the down-regulated genes, and fold changes of these down-regulated genes are not as apparent as those in panicles. However, the distributions of up-regulated genes in the two tissues are rather distinct, where the upregulated genes in leaves are biased toward over-dominant expression albeit a minority of the genes is found spreading toward mid-parent. In roots, the up-regulated a 1-1, one tag that was mapped to a single gene; n-1, multiple tags that were mapped to a single gene; 1-n, one tag that was mapped to multiple genes. b T-supported tags are those mapped to genes with known transcripts and P-supported tag are those mapped to predicted gene models. genes, though they are rather smaller in number as compared to panicles and leaves (101 genes, Table 4), are mostly over-dominant. Finally, in the process of summarizing gene distributions in the twelve patterns, we found that a minority of the differentially-expressed genes (25 to 45%) exhibited additive expression (P > L > N and N > L > P; genes that were plotted on the horizontal lines), whereas the majority of the genes, 380 (55%), 408 (72%), and 309 (75%), are non-additive in panicles, leaves, and roots, respectively. Among the sum of these non-additive genes in all three tissues, 552 genes showed over-dominant expression, and a smaller amount, 394 genes, were found under-dominantly expressed. In addition, 115 and 32 genes are expressed at the same level as their paternal line (93-11) and maternal line (PA64s), respectively; these genes are classified as dominant expression.

Functional analyses of differentially-expressed genes
We annotated 217 (22.8%) and 850 (89.3%) differentially-expressed genes on the basis of two general databases, KEGG (Kyoto Encyclopedia of Genes and Genomes) [29] and InterPro/Network [30], respectively. The genes were further classified into 20 categories according to KEGG Gene Ontology (KOG) classification scheme ( Figure 3); among them, genes involved in carbohydrate metabolism are the most abundant (16%), followed by energy metabolism (10%), and amino acid metabolism (8%). For instance, differentially-expressed genes in the hybrid are mostly related to enhancing carbon assimilation, energy metabolism, and biosynthesis of secondary metabolites; this effect is not due to simple distribution bias in the overall gene distribution since other categories were found decreased in the hybrid, such as protein sorting/folding/degradation in leaves ( Figure 4). Dramatic down-regulation was also seen in metabolisms of co-factors and vitamins in panicles.
Although the overall comparison to the previous results that were based on less number of tags led to similar conclusions, we feel that our current data allowed us to further look into more pathways and molecular details, which were not thoroughly exploited in the previous analysis. We divided carbon metabolism into three cellular compartments: the chloroplast, the mitochondrion, and the cytoplasm ( Figure 5). The genes involved in photosynthesis in chloroplast were all up-regulated both in leaves and roots but down-regulated in panicles; this trend was readily observed in the overall distribution ( Figure 2). Among them, 12 genes encode chlorophyll a/b binding proteins, 17 are photosystem I/II component genes, and ribulose bisphosphate carboxylase that is a key enzyme mediating the initial reaction of CO 2 fixation. Details of genes involved in light reaction are listed (see Additional file 6). We also observed three key enzymes involved in Expression patterns and fold changes of differentially-expressed genes five other selected key pathways (glycolysis/gluconeogenesis, citrate cycle, anaerobic respiration, glycolic acid oxidate, and fatty acid β-oxdidation) in the mitochondrion and cytoplasm. The first enzyme, alcohol dehydrogenates involved in the anaerobic respiration, is the most up-regulated gene in all three tissues. The second enzyme, fructose-1,6-bisphosphatase involved in gluconeogenesis, is up-regulated only in leaves. The last, pyruvate kinase that catalyzes phosphoenolpyruvate to form pyruvate and ATP (or decomposition of carbohydrate) is down-regulated both in leaves and panicles but not in roots. In addition, we observed that catalase, known to be involved in glycolic acid oxidate pathway (one of the three respiration pathways and unique to rice for better adapting its watery environment), is significantly up-regulated. Furthermore, along the pathway of synthesizing sucrose and its storage form (starch), we identified four genes, encoding betaphosphoglucomutase, 1,4-alpha-glucan branching enzyme, sucrose phosphate synthase, and sucrose synthase, which are also up-regulated in leaves and panicles. These enzymes are believed to contribute to high grain yield in the super-hybrid rice.
There were many other functionally annotated genes found to be significantly up-regulated, including rapid alkalinization factor, proteinase inhibitor, and MADS-box transcription factors; all appeared to be relative to the traits for photoperiod sensitive genic male sterility, male fertility restoration, and pollen fertility, according to the quantitative trait loci (QTL) database (Gramene [31]; see Additional file 7). Among them, the MADS-box (9311_Chr06_3092 and 9311_Chr01_4641) and rapid alkalinization factor (9311_Chr12_1510) genes were found highly expressed in the hybrid as compared to its parental lines despite the fact that the expression of these genes are already higher in its paternal line 93-11 than in its maternal line PA64s. This result indicated that these genes may play important roles directly or indirectly in flower morphogenesis and fertility of hybrid LYP9.
We also identified a large number of down-regulated genes that were not obvious in the previous analysis, largely due to more mapped tags and subtleties in data analysis protocols. These expression-suppressed genes belong to different functional categories among the three tissues; most of them are involved in energy metabolism, lipid metabolism, and glycan biosynthesis and metabolism in panicles, amino acid metabolism and protein processing in leaves, and biosynthesis of secondary metabolites in roots (Figure 4). The top-one down-regulated genes in panicles, leaves, and roots are metallothionein, peptidase M48, and glutathione S-transferase respectively. Metallothioneins are cysteine-rich proteins that can bind to heavy metals and scavenging reactive oxy-Functional categories of differentially-expressed genes (P < 0.05) among the three cultivars Figure 3 Functional categories of differentially-expressed genes (P < 0.05) among the three cultivars.
gen to protect plants from oxidative damage. Although it is the most down-regulated gene in panicle, it is up-regulated in root which plays an important role in assimilating, filtrating, and concentrating metal irons especially in screening heavy metal irons. Peptidase M48 is a family of proteins that function in protein degradation. We also found some other down-regulated genes related protein degradation, such as ubiquitin and ubiquitin-conjugating enzyme. Glutathione S-transferase is an enzyme to metabolize toxic exogenous compound that utilizes glutathione in the detoxification, for chemical defense in plants. We speculate that both of these up-and down-regulated genes represent a significant fraction of the genes regulating panicle development, rapid growth, stress tolerance, and grain yield in LYP9. Obviously, further verification and functional examination of these differentially-expressed genes are of essence in understanding their precise roles in heterosis.

Figure 4
Functional Categories of up-regulated and down-regulated genes in panicles, leaves, and roots.

Cross-referencing SAGE data to Microarray-based results
We have compared our SAGE data with those from microarray-based experiments in a limited way where only data from one tissue, the leaf, were eligible for legitimate comparison, since the mRNA sample was harvested from leaves at the milking stage, identical to what we used for the SAGE experiment. The microarray data were acquired by using a custom-designed oligoarray that contains 60,727 oligonucleotide probes representing all predicted genes from the genome assembly of 93-11 [22]. From this grand total, we identified 3,355 informative data points that were found in both microarray and SAGE data, and 2,312 (69%) of them showed a consistent trend between the two types of experiments (the spearman coefficient is Differentially-expressed genes that are involved in selected key metabolic pathways among three major cellular compartments Figure 5 Differentially-expressed genes that are involved in selected key metabolic pathways among three major cellular compartments. Genes involved in photosynthesis, glycolysis/gluconeogenesis, citrate cycle (TCA cycle), anaerobic respiration, glycolic acid oxidation, and fatty acid β-oxdidation pathways are shown. The enzymes ( # denotes key or rate-limiting enzymes) are: E1 # , fructose-1,6-bisphosphatase; E2, fructose-bisphosphate aldolase; E3, glyceraldehyde 3-phosphate dehydrogenase; E4, phosphoglycerate kinase; E5 # , pyruvate kinase; E6 # , alcohol dehydrogenase; E7, catalase; E8, acyl-CoA dehydrogenase; E9, succinyl-CoA ligase; E10, malate dehydrogenase; E11 # , ribulose bisphosphate carboxylase; E12, transketolase; E13, ribulose-phosphate 3-epimerase; E14, phosphoribulokinase; E15, beta-phosphoglucomutase, 1,4-alpha-glucan branching enzyme; E16 # , sucrose phosphate synthase; E17 # , sucrose synthase. Proteins and enzymes in the light reaction complex are plastocyanin, ferredoxin [2Fe-2S], chlorophyll A-B binding protein, photosystem II protein PsbX, photosystem II protein PsbW, photosystem II protein PsbY, photosystem II oxygen evolving complex protein PsbP, photosystem II protein PsbR, photosystem II manganese-stabilizing protein PsbO, photosystem II oxygen evolving complex protein PsbQ, photosystem I reaction centre (subunit XI PsaL), photosystem I psaG/psaK protein, photosystem I reaction centre subunit N, photosystem I reaction center protein PsaF (subunit III), NADH:flavin oxidoreductase/NADH oxidase, and cytochrome b ubiquinol oxidase. The ratios of up-(+) or down (-) -regulated tags are indicated. Detailed information for light reaction complexes is listed in Additional file 6. Note that the key enzymes are either up-or down-regulated in three tissues; this behavior suggests active yet unique regulations in the hybrid. 0.497, P < 0.0005). We found that the consistent trend among genes with a moderate-to-high expression between the two datasets correlated fairly well (the spearman coefficient is 0.743, P < 0.0005; data not shown). Of these genes, 222 (39%) were differentially-expressed according to the SAGE data with significance (P < 0.05). We listed 23 genes with a fold change of greater or equal to 2 in Table 5. These confirmation rates are not much different from reported comparative analyses between these two types of experiments since the reasons for systematic errors are multifold, including sampling time, experimental procedures, and data normalization [13].

Discussion
Tag-to-gene mapping procedures SAGE and related sequencing-based techniques are very effective for studying gene expression in organisms where well-characterized genome sequences are available, and they have been applied to a number of eukaryotic species [17,19,32] and the merits and success have been discussed very recently by Marco Marra and his colleagues with ample experimental data [12], albeit pitfalls do exist [13]. In our previous SAGE study, we utilized the available FL-cDNA sequences [24] for tag-to-gene mapping [21], as these FL-cDNA sequences best represent the rice transcriptome albeit in a rather limited amount. However, a large proportion (83%) of the SAGE tags was not found in this cDNA data collection that is known not covering all the genes of the rice genome. To overcome this limit, we utilized a new strategy for tag-to-gene mapping based on newly annotated genes of the two rice genome assemblies and other transcript sequences (FL-cDNA, UniGene, and ESTs). This process led to a significant improvement in gene identification, resulting in 10,268 additional tags and 68.85% extra differentially-expressed genes at a higher P value (P < 0.01), as compared to the previous collection.
Aside from the success of mapping SAGE tags to annotated genes in the genome, there are a couple of important points that are worthy of further discussion. First, we always have tags that are mapped to ambiguous positions, and they may belong to multiple loci (such as gene families and splicing variants) in the genome sequence, especially when the length of SAGE tags is as short as 14 bp. There were 4,014 (20%) such tags in our case, we assigned these tags to the genomes and used them for functional analysis. For example, despite the fact that a tag with a sequence of "AACAAGCTCA" was assigned to two different loci (9311_Chr04_1718 and 9311_Chr05_1829), the two were evidenced by two different FL-cDNA sequences (AK0ah71547 and AK061050), allowing us to identify them as members of the fructose-bisphosphate aldolase gene family. These two genes were found down-regulated in roots of the hybrid line, and they are involved in glycolysis/Gluconeogenesis pathways. Therefore, it is critical to map these seemingly ambiguous genes, especially when they are differentially regulated in the hybrid. It is possible to design experiments to distinguish these genes with locus-specific primers since most of these duplicated (or closely related) genes may be not identical in their UTR and genomic sequence, especially when genome sequences are readily available. As we have reported previously, the rice genome has enormous number of duplicated genes [23] that some of them may actually hold pivotal information in hybrid vigor.
The second point has to do with the fact that a fraction (often more than 40%) of the experimental tags remains unassigned to genes so we need to figure out the possible reasons. When comparing unassigned tags to virtual tags based on predicted NlaIII sites in the nuclear and organellar (mitochondrial and chloroplast) genome sequences, we found that 2,500 tags out of 47,867 (5%) were absent in the genome sequence assembly of 93-11, and 342 tags (0.6%) were derived from either the mitochondrial (491 kb) or chloroplast genomes (134 kb). These unassigned tags are most likely due to sequencing errors, sequences interrupted by introns, un-assembled sequences (including those in the sequence gaps), and organelle-specific sequences. In addition, we have technically implemented an artificial 300-bp UTRs for predicted genes without transcript-based evidence and only extracted the 3' most (canonical position) tags from virtual transcripts. This procedure is certainly incapable of including all UTR length variants, largely due to the absence of canonical polyadenylation signal for the accurate determination of the 3' UTR length in plant genomes [33]. To estimate the result of such a procedure, we compared the remaining total unassigned tags to a cumulative virtual tag dataset constructed by varying the artificial UTR lengths in a 100bp interval, from 100 to 500 bp, resulting in a further assignment of 3,119 (6.5%) additional tags. However, these tags were considered unreliable and were not included in this analysis. Nevertheless, the UTR-derived anomaly seems contributing to the impaired tag assignment in a similar way as the sequence anomaly. Other obvious factors resulting in unassigned tags, such as experimental artifacts (incomplete enzyme digestions and ligations, as well as inefficient cloning procedures), are not discussed here in details.

The differentially-expressed genes in multiple expression patterns
Over the years, differential gene expression between the hybrid and its parental cultivars has been hypothesized to attribute to heterosis [5,34]. As having partitioned the differentially-expressed genes into twelve patterns as conventionally done, we found only 25% to 45% or minorities of the genes were additively expressed in the rice hybrid; Among all differentially-expressed genes, we identified up to 70% of them (P < 0.01) exhibiting paternal-like expression (PLE) profiles, especially in panicles, which are at least in part attributable to two plausible mechanismsmolecular imprinting and defective expressions of the maternal alleles -as often observed in panicles harvested at the pollen maturing stage, where thermo-sensitive male sterility of the maternal line (PA64s) may be relevant [38]. For instance, two MADS-box transcription factors related to pollen fertility have been consistently observed as upregulated in the hybrid, but they do not express in the male-sterility plant [39,40]. The rapid alkalinization factor, a polypeptide hormone that was suggested to be related to nuclear sterility and development [41], was observed to be up-regulated and located in photoperiodsensitive and genic male sterility trait based on our QTL analysis. Although we have not been able to plot plausible functional scenarios on the precise roles of these genes, the findings undoubtedly provide useful clues for future molecular studies.

Putative regulation mechanisms of differentially-expressed genes
Differential gene expression in plants is known to be mainly regulated by two forms of mechanisms -cis-and trans-regulations at transcription levels as well as epigenetic and post-transcription modulations [6]. For instance, differential methylation in CpG or CNG islands [9,42]  An increasing number of studies have reported that functional divergence in duplicated gene is accompanied by gene expression change although the evolution mechanism behind this process remains unclear. There was a report that 7% of duplicated gene pairs co-express in yeast [46], and we know that gene and chromosomal segment duplications widely exist in the rice genome, including an ancient whole genome duplication, recent segmental duplications, and massive ongoing individual gene duplications that cover 65.7% of the genome [23]. We found 7 of our 698 ambiguous assigned tags are mapped to the duplicated gene pairs, which we suspected the duplication with a high homology may affect gene expression including silencing and up-or down-regulation of one of the duplicated genes after hybridization [47]. When looking into the possible molecular assays in distinguishing the different alleles, we found that it is actually possible to design allele-specific primers to detect the expression level of duplication pairs.

Conclusion
We improved the tag-to-gene mapping strategy by combining information from transcript sequences and rice genome annotation and obtained over 10,000 new tags for a more comprehensive view of genes that related to rice heterosis. These heterotic expression genes among different genotypes provided new avenues for exploring the molecular mechanisms underlying heterosis, including variable gene expression patterns.

PCUE database
We constructed a PCUE database for rice (Oryza sativa) on the basis of available genomic resources that contain (1)  We aligned the collected transcript sequences to the two genome sequences by using BLAT [51] to obtain a dataset for tag annotations. The threshold parameters set for aligned transcripts are (1) at least 90% identical to their genomic sequences and (2) covering ≥ 90% transcript sequences. When a transcript has more than one hit to genomic sequences, the longest consensus was selected as the best-aligned (true) locus. We further selected sequences that span the 3' end of a predicted gene but do not extend to the next with ≥ 100-bp overlapping sequences. As a result, our predicted genes were partitioned into two sets: supported by one or more transcripts and without supporting data.

The evaluation dataset
In order to evaluate the accuracy of tag-to-gene mapping methodology, we built a test dataset that contains 2,480 FL-cDNA sequences that satisfied all five criteria: (1) ORF length > 300 bp, (2) with poly(A) signal (AATAAA/ ATTAAA) or poly(A) tails (with a minimal number of five A) [15], (3) alignable to a unique predicated gene with homolog (based on 50% protein sequence similarity or 100 residues) to Arabidopsis, (4) a unique CATG tag and experimental data, and (5) alignable to a unique predicted gene and corresponding UniGenes or ESTs. We further divided this dataset into three categories: UniGene, EST, and predicted gene. In the Unigene and EST categories, we have twelve subsets. Eight of those were sequences with poly(A) signal (Uni-S and EST-S), with poly(A) tails (Uni-A and EST-A), with both poly(A) signal and tail (Uni-B and EST-B), without poly(A) signal and tails (Uni-N and EST-N). The other four subsets contained the longest and the best transcripts that were best validated by either UniGenes or ESTs (Unibest or ESTbest). To know the length of 3'-UTR, we used 19,079 non-redundant FL-cDNA to determine the length distribution and found that 95% of these genes have UTR length shorter than 1280 bp, with an average size of 422 bp and a median of 295 bp. We therefore added five different lengths (100-, 200-, 300-, 400-, and 500-bp) to construct virtual UTRs for the predicted genes. We finally built virtual tags from each of the above-mentioned subsets by extracting a 10-bp tag from the immediate downstream sequence of the last (3'most) NlaIII (CATG) site. We evaluated the success rates of virtual tags that match the test set.

Virtual tags and tag-to-gene mapping
Since predicted genes do not have UTRs, we extracted consecutive exons together to form gene models from the two genome assemblies and added to them either UTR sequences based on information from known transcripts or artificial UTRs in a length of 300 bp. We obtained four groups of tag data, including those based on cDNA, Unimax, ESTmax, and predicted genes (P-300). We mapped 68,462 unique empirical tags from our data [21] to the four groups of virtual tags after filtering cloning linkers, vectors, and simple repeats. We excluded 47,867 tags from further processing and their outcomes from our analysis protocol were summarized (see Additional file 8). These tags were regarded as unmapped tags although 45,025 of them were actually mapped to the nuclear genome but in unexpected range of correct positions of exon and UTR sequences. Most of them were believed to fragmented mRNAs that were co-processed during library construction procedures.
We annotated all our SAGE tags based on InterPro/Network and KEGG for protein families, domains, and functions. We chose the best scoring primary (sequence similarity-based) annotations from family-type categories first, followed by domain-type and others. If the gene had no primary annotation then we used a network-based annotation [52]. P values between copy numbers among libraries were calculated based on Audic-Claverie (or AC) statistics [28] by using IDEG6 software [53,54]. The significance of the differentially-expressed genes was defined with P values less than 0.05 or 0.01. Ratios of up-regulated and down-regulated genes were calculated according to ratio = L/[(P+N)/2] (≥ 2) and [(P+N)/2]/L (<2), respectively.

Microarray and QTL data
We used microarray data from the leaf tissue at the milky stage, which were generated in our laboratory. The microarray contains 60,727 oligonucleotide probes representing all predicted genes from the genome sequence of 93-11 [22]. We physically mapped the oligonucleotides to the most up-to-date version of the genome assembly [48] with the threshold that each oligonucleotide must match to one unique gene with 90% or higher sequence identity. We also used rice QTL data with physical position on TIGR4 genome from Gramene [31]and mapped differentially-expressed genes to nine QTL categories.