A total of 3,865 paralogous protein families containing 21,998 proteins were identified [see Additional file 1] from the 42,653 total non-transposable element (TE)-related proteins predicted in the rice genome, leaving 20,655 putative singleton proteins encoded by single copy genes. On average, a rice family contained six family members, ranging in size from two to 214 family members (Fig. 1). A total of 11 paralogous protein families with more than one hundred member proteins were identified in rice which encoded proteins such as zinc finger proteins, protein kinases, Myb-like proteins, and transducins [see Additional file 2], similar to the largest protein families reported in Arabidopsis 30. Paralogous protein family genes of rice were distributed throughout the genome and within chromosomes in a pattern similar to the singleton genes [see Additional file 3A]. Although paralogous protein family genes were more frequently located in the euchromatic regions, this was consistent with previous reports that non-TE-related genes are found more prevalently in euchromatic regions. A comparison of segmentally duplicated genes with the paralogous protein family genes suggested that our classification pipeline was robust. Of the 2,403 segmentally duplicated gene pairs within 163 segmentally duplicated blocks, 1,570 duplicated gene pairs (65%) were classified in the same paralogous protein family. For the remainder of the segmentally duplicated genes, 175 pairs (7%) were classified in different paralogous protein families and 268 (11%) had one gene classified in a paralogous protein family and the other gene classified as a singleton. We observed that 390 segmentally duplicated gene pairs (16%) were not included in any paralogous protein family. Note that in our computational pipeline, four or more members were required to define a BLASTP-based domain. Consequently, a single pair of segmentally duplicated genes alone is insufficient to define a BLASTP-based domain. The lack of 100% correspondence between segmental duplication and paralogous family classification may be due to the acquisition of new domain(s) or loss of existing domain(s) within one of the duplicated genes as in our computational pipeline, only proteins with the identical domain composition were classified into the same paralogous protein family. Alternatively, the difference could be due to the different classification methods employed in each method. For example, LOC_Os08g37350 and LOC_Os09g28940 are segmentally duplicated genes from chromosomes 8 and 9, respectively. These two protein sequences had a 56% identity over 70% of the length of the longer sequence and were within a segmentally duplicated block of 43 collinear gene pairs. LOC_Os08g37350 has two Pfam domains (PF00443: Ubiquitin carboxyl-terminal hydrolase; PF01753: MYND finger) while LOC_Os09g28940 has only one Pfam domain (PF00443: Ubiquitin carboxyl-terminal hydrolase). As a consequence, these loci were classified in two different paralogous families (LOC_Os08g37350 is classified in Family 1545; LOC_Os09g28940 is in Family 3650). In a second example, LOC_Os11g03210 and LOC_Os12g02960 are from a segmental duplication event involving chromosomes 11 and 12 which includes 160 collinear gene pairs. LOC_Os11g03210 has a single Pfam domain (PF02798: Glutathione S-transferase, N-terminal domain) and thus is classified in Family 3362 while LOC_Os12g02960 is classified as a singleton as although it has two Pfam domains (PF02798: Glutathione S-transferase, N-terminal domain; PF00043: Glutathione S-transferase, C-terminal domain) no other protein has exactly the same domain profile. Note that in our computational pipeline, a paralogous family must have at least two members with identical domain profiles. In a third example, segmentally duplicated genes LOC_Os01g41900 and LOC_Os05g51160 are from chromosomes 1 and 5. These two genes were derived from full length cDNAs (FLcDNAs) and had a 59% identity over approximately three-quarters of the longer protein sequence. LOC_Os01g41900 has two Pfam domains (PF00249: Myb-like DNA-binding domain and PF00098: Zinc knuckle) while LOC_Os05g51160 has only one single Pfam domain (PF00249: Myb-like DNA-binding domain). As a consequence, they were classified in different families, Family 1452 and Family 3863, respectively. Manual inspection of these three sets of loci revealed that they were correctly annotated and that the lack of clustering into a single paralogous family could not be attributed to incorrect structural annotation which is another potential cause for lack of 100% correspondence between segmentally duplicated genes and paralogous families.

A parallel construction of paralogous protein families in Arabidopsis identified 3,092 paralogous protein families (18,183 proteins) and 8,636 single copy genes from a total of 26,819 protein coding genes from TAIR7 release 31. A similar size distribution of Arabidopsis protein families was observed, ranging from two to 182 (Fig. 1). In Arabidopsis, the largest families encode Myb-like proteins, zinc finger proteins, and protein kinases, consistent with what has been reported previously 30. Arabidopsis paralogous protein family genes distributed similarly to singleton genes and were more frequently located in the euchromatic regions [see Additional file 3B].

We examined the functional annotation of paralogous family and singleton proteins. A total of 21,403 and 23,081 genes were annotated as encoding known or putative proteins in rice and Arabidopsis, respectively, due to strong similarity with proteins with a known function or the presence of Pfam domains above the trusted cutoff. Genes with no known or putative function can be supported by experimental transcript evidence (i.e., encode an "expressed protein") or are predicted solely by an ab initio gene finder and lack expression support as well as sequence similarity to known proteins with the exception of other hypothetical proteins (i.e., encode a "hypothetical protein"). In rice, a total of 6,913 genes encode expressed proteins as shown by experimental transcript evidence from Expressed Sequence Tags (ESTs), FLcDNAs, Massively Parallel Signature Sequencing 32, Serial Analysis of Gene Expression, and/or proteomic data 33. In Arabidopsis, 2,270 genes encode expressed proteins as shown by experimental transcript in the form of ESTs and/or cDNA evidence (see Methods). The remaining 14,337 rice genes 33 and 1,468 Arabidopsis genes (see Methods) encode hypothetical proteins. A majority of rice paralogous family genes (73%) encode either a known or putative protein (Fig. 2). The remaining rice paralogous family genes encode expressed proteins (9%) and hypothetical proteins (18%). In contrast, rice singletons had a larger portion of hypothetical genes (50%) and a smaller portion of genes with a known or putative function (26%). Even though Arabidopsis overall has a smaller number of genes with unknown function than rice, a similar bias of genes with a known or putative function in paralogous family genes was observed in a parallel analysis in Arabidopsis (Fig. 2).

Using Plant GOSlim annotations 34, we compared the function of the proteins within rice paralogous families to that in the singletons. Within the 26 molecular function GOSlim categories identified in our analyses, rice paralogous protein families showed different patterns from singletons in a number of GOSlim categories (Fig. 3A). Although, the relative abundance of each GOSlim category varied with the size of the rice paralogous family, no obvious correlation was observed (Fig. 3A). For each category, a two-tailed two-sample binomial test was performed by comparing the abundance of that category in rice paralogous families with that in the singletons. Multiple testing was corrected using the Benjamini and Hochberg false discovery rate control at a level of 0.05 35. The statistical test revealed a substantial enrichment of 12 categories in rice paralogous family proteins including transcription factor activity, hydrolase activity, DNA binding, and transporter activity while a substantial reduction was seen in five categories including receptor activity, nucleotide binding and carbohydrate binding (Table 1). A similar skew in GOSlim categories was observed in a parallel analysis in Arabidopsis (Table 2 & Fig. 3B), consistent with a previous report in Arabidopsis 36 that non-random loss and retention of paralogous genes with different functions occurred after gene duplication.

Alternative splicing has been regarded as a mechanism to increase genetic novelty. In the rice genome, 6,253 non-TE-related genes have evidence of alternative splicing (see Methods) and we used this set of genes to examine alternative splicing in singleton versus paralogous protein family genes. The percentage of alternative splicing in single copy genes is 2,094/20,655 = 10.1%, while that in paralogous family genes is 4,159/21,998 = 18.9%; a statistically significant difference (χ² test, P < 1e-5). To remove any bias due to genes that lack transcript evidence, we restricted our analysis to genes with EST and/or FLcDNA evidence. The percentage of alternative splicing in singletons is 2,094/8,619 = 24.3%, while that in paralogous protein family genes is 4,159/14,072 = 29.6%; a statistically significant difference (χ² test, P < 1e-5). We further restricted our analysis to high confidence genes whose structures were completely supported by ESTs and/or FLcDNAs. The percentage of alternative splicing in singletons increases to 1,826/5,964 = 30.6%, while that in paralogous protein family genes increases to 3,765/11,235 = 33.5%; a statistically significant difference (χ² test, P < 1e-3).

To confirm that our observation was not restricted to rice, we performed a parallel analysis with Arabidopsis. Using data on alternative splicing as provided with the TAIR7 release (see Methods), the percentage of alternative splicing in Arabidopsis single copy genes is 943/8,636 = 9.8%, while that in paralogous protein family genes is 2,856/18,183 = 15.7%. This difference is also statistically significant (χ² test, P < 1e-5), similar to that observed in rice. Restricting the analysis to only those Arabidopsis genes with EST and/or cDNA support as provided in the TAIR7 release revealed that the percentage of alternative splicing in singletons is 942/6,663 = 14.1%, while that in paralogous family genes is 2,852/15,369 = 18.6%; a statistically significant difference (χ² test, P < 1e-5). Our findings are contradictory to previous reports in model animal species in which duplicated genes tend to have fewer alternative spliced isoforms thereby supporting the 'function-sharing model' that alternative splicing and gene duplication are two mechanisms that are complementary with respect to proteomic function diversity 3738. Our results suggested that plants may employ multiple mechanisms for proteomic complexity, gene duplication and alternative splicing.

While there are previous reports on gene duplication in rice 1516171819, they utilized alternative assemblies and annotation datasets of the rice genome. To provide information on the age of paralogous families identified in this study, we estimated the age of a paralogous family from the maximum value of the distribution of pairwise d_S calculated among all members of that protein family (see Methods). We found that the origin of most paralogous families dates back to over 115 Million Years (MY), the point at which synonymous sites are saturated and dating becomes unreliable (d_S ~1.5) [see Additional file 4A]. Among protein families for which the maximum pairwise d_S value is less than 1.5, the distribution of maximum d_S is fairly flat, with the exception of a recent peak at d_S between 0 and 0.1 [see Additional file 4B]. This suggests that paralogous families have been arising at a relatively constant pace within the past 115 MY, but that a burst of duplication took place within the last 7.5 MY. Alternatively, paralogous families arise at a rate similar to that observed for the first few million years, but about 2/3 of them revert to single-gene status soon thereafter, accounting for the quick decline after the first 7.5 MY. The fairly constant number of older paralogous families can be due to selective constraints maintaining the elevated copy number or if the loss of paralogs is dependent on sequence similarity, such that after ~10% sequence divergence, paralog loss is negligible. Finally, for each family we identified the largest peak below 1.5 (if there was one) in the distribution of all pairwise d_S values. The distribution of this peak value across all families is bimodal [see Additional file 5], and it confirms the presence of a large number of recently duplicated genes (0 ≤ d_S < 0.1). In addition, the peak at 0.7 ≤ d_S ≤ 1 most likely results from the large-scale segmental duplication event that occurred ~70 MYA.

We further examined the expression patterns of the paralogous families using MPSS data from 18 libraries 32. MPSS tags were searched against our release 4 pseudomolecules and cDNA sequences of all annotated gene models to ensure that all MPSS tags would be identified even if they spanned the intron(s). We found 11,619 genes within the paralogous protein families that were associated with unique, reliable, and significant MPSS tags, which were referred as MPSS-qualifying genes.

Suitable summary statistics of correlation for expression divergence of a gene family can be found in Gu 39 and Gu et al. 40, though microarray data were the primary focus in these studies. To be concise, we restricted our analysis of expression correlation in the libraries and tissues to paralogous families with exactly two MPSS-qualifying genes (674 protein families). To measure the expression correlation, the Pearson's Correlation Coefficient (r) of their expression was computed for each pair of MPSS-qualifying genes from each of the 674 protein families across all 18 MPSS libraries. It is important to note that we excluded MPSS tags which mapped to multiple locations, as most of these are likely to match to closely-related paralogs and could have confounded our analyses. We employed the method used by Blanc and Wolfe 36 to determine a minimum cutoff value for Pearson's Correlation Coefficient (r) to classify two duplicated genes as having divergent expression. Basically, a total of 10,000 gene pairs were generated by random shuffling of the singleton genes and the Pearson's Correlation Coefficient (r) was calculated similarly for each pair. Ninety five percent of the random shuffled gene pairs had a correlation value r < 0.59. As random shuffled gene pairs should have divergent function and expression patterns, we utilized r < 0.59 as an indicator of divergent expression. Our results show that the expression correlation value (r) of the paralogous protein family genes ranged from -0.6 to 1.0 although the majority of the gene pairs had little correlation with r peaking at -0.2 ~0, similar to that observed with the singletons (Fig. 4). Using the correlation cutoff (r = 0.59), a total of 598 (89%) paralogous protein families with two-qualifying MPSS genes exhibited divergent expression patterns, consistent with what has been reported in Arabidopsis 36 and in yeast in which more than 80% of the older duplicated gene pairs (ds > 1.5) showed divergence in expression 41.

To gain a better understanding of the expression patterns of paralogous protein family members in different organs/tissues, we classified the 18 MPSS libraries 32 into four groups by organs/tissues: roots, leaves, reproductive organs/tissues, and "other tissues". Within the 674 paralogous families with exactly two MPSS-qualifying genes, 239, 168, 223, and 200 paralogous families had only a single member of the pair expressed in roots, leaves, reproductive organs/tissues, and "other tissues", respectively, which demonstrated their diverged expression patterns, and possible tissue-specific expression. To further examine the tissue-specific or stress-induced expression patterns of paralogous protein family members, we calculated the Preferential Expression Measure (PEM) for each of the 1,348 genes from the 674 paralogous families (see Methods) in the 18 MPSS libraries. The PEM shows the base-10 log of ratio of the observed expression level in a given tissue/treatment to the expected expression level assuming uniform expression across all tissues/treatments. A PEM value of 1 means the observed expression level in a given tissue/treatment is 10 times that of expected and indicates strong tissue specific expression. For each gene, tissue(s) with a stringent cutoff of PEM ≥ 1 were compared with the other member of the duplicated gene pair. A total of 375 (375/674 = 55.6%) of the paralogous families showed little tissue-specific expression as none of the associated PEMs had a value equal to or greater than 1. Two hundred ninety-nine families showed strong tissue specific expression patterns; 19 families were preferentially expressed in the same tissue or treatment, 49 families were preferentially expressed in different tissues or treatments, and 231 families had only one of the duplicated genes with preferential tissue-specific expression.

We further examined the correlation between expression divergence and sequence divergence. For each family, we calculated the Pearson's Correlation Coefficient (r) for all possible pairs of the MPSS-qualifying genes to measure expression divergence. We then used ds as a proxy of divergence time for each gene pair. We restricted our analysis to d_S ≤ 1.5 so that the synonymous sites are not saturated. The Pearson's Correlation Coefficient (r) values were plotted against the ds values for each interval of 0.1 to gain better resolution. That is, we plotted for gene pairs with 0 <d_S ≤ 0.1, 0.1 <d_S ≤ 0.2, 0.2 <d_S ≤ 0.3, and so on. We found no correlation between d_S and correlation of expression except for gene pairs with 0 <d_S ≤ 0.1 (R = 0.33, P < 1e-4) where duplicated genes were relatively young [see Additional file 6]. The number of non-synonymous substitutions per site (dN) was also calculated for each gene pair and plotted against correlation of expression. No correlation was observed between dN and correlation of expression (data not shown). This is consistent with reports in Arabidopsis in which expression divergence is not strictly coupled with sequence divergence as shown by no appreciable change for the majority of gene duplicates with highly diverged amino acid sequences in expression pattern in developing roots 42.

Positive correlation of expression patterns among paralogous protein family members would suggest that similar transcriptional regulation was retained in both members and possibly, similar functions. However, we observed a large number of gene pairs with little expression correlation which could be an indication of subfunctionalization or neofunctionalization after gene duplication. The duplication-degeneration-complementarity (DDC) model proposed by Force et al. 3 and Lynch and Force 4 suggests that subfunctionalization is a major mechanism for retention of duplicated genes as a result of differential expression caused by accumulation of mutations in regulatory regions rather than protein coding regions. The 49 families with preferential expression in two different tissues or treatments, along with the 231 families having only one member of the paralogous pair preferentially expressed, is a strong indicator of subfunctionalization. As our paralogous protein family classification required that each family member have the same domain profile, the differential expression may be attributable to mutations in regulatory regions rather than gene coding regions, consistent with the DDC model.

In release 4 of the TIGR Rice Genome Annotation 33, a total of 55,890 genes were annotated, of which 13,237 were related to TE. The TE-related genes were excluded from all further analyses. As alternative splicing occurs in the rice genome and some genes have multiple splice forms, the largest peptide sequence was used whenever alternative isoforms existed. Short protein sequences (<50 amino acids) were excluded from this analysis. A total of 42,653 rice protein sequences were used to classify paralogous protein families using protein domain compositions as described in Haas et al. 30. The basic approach for generating the protein families involved identification of the domains followed by organization of the families based on domains. Two different types of domains were used for the generation of paralogous families: Pfam/HMM domains and BLASTP-based domains. For the Pfam/HMM domains, the predicted rice proteome was searched against the Pfam HMM domain database 63 using HMMER2 64 and proteins with scores above the trusted cutoff value were retained. For the BLASTP-based domain, peptide regions that were not covered by the Pfam HMM profiles were then clustered based on homology derived from an all versus all BLASTP search 65. Links were made if two peptides had an >45% identity over >75 amino acids with an E-value <0.001. To prevent multi-domain proteins that are not related from artificially clustering due to single linkages, the Jaccard coefficient of community 66, also known as link score, was used in the clustering process. As described in Haas et al. 30, a link score was calculated for the pairs of linked peptide sequences a and b as follows:

J a , b = # d i s t i n c t   s e q u e n c e s   m a t c h i n g   a   a n d   b   i n c l u d i n g ( a , b ) # d i s t i n c t   s e q u e n c e s   m a t c h i n g   e i t h e r   a   o r   b MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=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@A7D8@

Peptides with a link score above the cut-off value (0.66) were selected to generate single linkage clusters. Clustered peptides were then aligned using CLUSTALW 6768 and used to develop BLASTP-based domains, which were used to build the families if the domain alignments contained four or more members. Protein families were then organized based on the domain composition that refers to the type and number of the domains, which included both Pfam HMM domains and BLASTP-based domains. Proteins with identical domain composition were then classified into putative protein families. Paralogous protein families in Arabidopsis were constructed similarly with a total of 26,819 protein coding genes from the TAIR7 release of the predicted proteome 31.

Segmentally duplicated genes in the rice genome were defined in Release 4 as described previously 69. In brief, similar gene pairs were identified by all versus all BLASTP search (WU-BLASTP, parameters "V = 5 B = 5 E = 1e-10 -filter seg") 65, which were then used to define segmentally duplicated blocks by running DAGchainer 70 with parameters "-s -I -D 100000".

A total of 26,819 Arabidopsis protein coding genes were downloaded from the TAIR7 release of the predicted proteome 31 and searched against an in-house non-redundant amino acid database that contains all publicly available protein sequences (e.g. GenBank, Swissprot, etc.) using BLASTP 65 and the Pfam HMM domain database 63 using HMMER2 64. BLASTP matches to Arabidopsis sequences were excluded unless they were from Swissprot. BLASTP matches to conserved hypothetical or hypothetical proteins were excluded as well. Arabidopsis proteins with a BLASTP match (< 1e-10 and > 30% identity over 50% coverage) or Pfam domains with scores above the trusted cutoff value were classified as known or putative proteins. The remaining Arabidopsis genes were classified as expressed genes or hypothetical genes according to the gene set downloaded from TAIR7 release 31 which had at least one supporting cDNA and/or EST.

To assign Gene Ontologies (GO) 71, the predicted rice proteome was searched against the predicted Arabidopsis proteome (TAIR6 Genome Release) 31 using BLASTP. Using an E-value cutoff of 1e-10, plant GOSlim annotations 34 were transitively annotated using the GO terms from Arabidopsis. Hypothetical/expressed proteins, TE-related proteins, and proteins assigned with GO terms with "unknown" definitions were excluded from this analysis. The GOSlim assignment of Arabidopsis proteins was obtained form TAIR7 release 31.

Approximately 780,000 rice EST sequences were released subsequent to the generation of the Release 4 gene models 33. Thus, we utilized the PASA program 72 to re-annotate the gene models and comprehensively identify alternatively spliced genes with the latest set of rice transcript data. Alternative splicing information on Arabidopsis was obtained from TAIR7 release 31.

A multiple protein sequence alignment was obtained for each family using CLUSTALW with default parameter settings 6768. From each protein family of size n, all (n²-n)/2 pairwise alignments were extracted from the global family alignment, maintaining the position and length of all gaps. A maximum likelihood estimate of the number of synonymous substitutions per synonymous site (d_S) was obtained for all pairwise alignments. All calculations were performed using the codon-based substitution model of Goodman and Yang 73 implemented in codeml, of the PAML package, version 3.15 74, running in pairwise mode (runmode = -2), with codon equilibrium frequencies estimated from average nucleotide frequencies at each codon position (codonFreq = 2).

The age of a paralogous protein family is defined by the duplication that gave rise to its second member, and can be approximated by the divergence between the most distantly-related pair of genes in the family. Given the rate of synonymous substitutions in grasses, estimated to be ~6.5 × 10^-9 per site per year 75, the number of synonymous substitutions per site (d_S) between the most divergent gene pair in a family can be converted into a divergence time, provided synonymous sites are not saturated (d_S < ~1). In addition, peaks in the distribution of intra-family pairwise d_S values suggest periods of family diversification. For each family, the distribution of pariwise d_S values was determined, plotted within the range of 0 to 1.5, with bin size of 0.1. Both the modal bin of each distribution (usually resulting from the most ancient split in the family tree) and the largest modal value of d_S < 1.5 (reflecting a burst in diversification within the last 100 MY) were recorded.

A total of 106,521 significant (>3 TPM) and reliable (observed in more than one sequencing run) MPSS 32 tags were obtained from the Rice MPSS Project 3276. These MPSS tags are derived from nine treated or untreated organs/tissues including callus, leaf, seed, crown vegetative meristematic tissue, ovary, stigma, pollen, panicle and stem. To reduce background noise, the method of Haberer et al. 77 was used to remove tags if the total minimal abundance across all libraries was ≤ 10 TPM or if the tag was not detected at ≥ 5 TPM in at least a single library, resulting in a total of 74,748 tags for subsequent analyses. The final set of MPSS tags were searched against TIGR rice pseudomolecules 33 using the Vmatch program 78. As tags can span an intron(s), MPSS tags were also searched against all the cDNA sequences of the annotated genes. MPSS tags that mapped to the anti-sense sequence of the annotated genes or that mapped to multiple locations of the genome were excluded, which is important to minimize false correlations among closely related paralogs. If a gene was associated with multiple MPSS tags, only the most 3' tag was used for the expression analysis. Paralogous genes that were associated with unique, reliable, and significant MPSS tags were analyzed. Pearson's Correlation Coefficient (r) was calculated for each gene pair to determine the expression correlation using the following formula 79:

r = n ∑ i = 1 n x i y i − ∑ i = 1 n x i ∑ i = 1 n y i [ n ∑ i = 1 n x i 2 − ( ∑ i = 1 n x i ) 2 ] [ n ∑ i = 1 n y i 2 − ( ∑ i = 1 n y i ) 2 ] , MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=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@8B64@

Where n is the number of DNA libraries. X_i and Y_i represent the expression level of the gene pair in the i-th library.

To determine if a gene was preferentially expressed in a specific tissue, we employed the PEM devised by Huminiecki et al 80. PEM is defined as log₁₀(O/E). Basically, it compares the observed (O) expression level in a given tissue with that of expected (E) level, assuming uniform expression across all tissues. The PEM value of the i-th gene in the j-th tissue was calculated as followed:

P E M i , j = log ⁡ 10 ( x i , j / ( ∑ k = 1 m x k , j ∑ l = 1 n x i , l / ∑ k = 1 m ∑ l = 1 n x k , l ) ) MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xI8qiVKYPFjYdHaVhbbf9v8qqaqFr0xc9vqFj0dXdbba91qpepeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=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@70B1@

Where m and n represent the total number of MPSS-qualifying genes and tissues, respectively.x_{i, j} is the expression level of the i-th gene in the j-th tissue.