Clustering is a useful starting exploratory tool to visualize highly dimensional data, and has been widely used to microarray data since the seminal paper of Eisen and cols. 17. Here we applied two clustering methdos, the classical one based on the UPGMA criterion 17, and the neighbor-joing (NJ) clustering. In both cases, we used the distance one minus the squared correlation (1-r²) between the samples, after normalizing the raw data with the RMA procedure 18, as detailed in Material and Methods. Results are drawn in Figure 1, where it can be seen that samples were clearly grouped by tissue, next by breed in both trees. This was neatly observed for ileum, liver, thyroid gland, adeno and neurohypohysis and olfactory bulb. Muscle samples were clustered by tissue (diaphragma vs. M. biceps femori) but less clearly within each muscle. As for fat, the similarity was larger between tissues than between breeds, and samples of both back and abdominal fat origins were clustered together. The same was observed between cortex and medulla from adrenal gland. In this case, contamination between both tissues cannot be ruled out because of the irregular limits of the medulla that make not easy to separate that region neatly from the cortex collecting rapidly enough amount of tissue for analysis. Other authors have described previously contamination of medulla in the cortex sample when mechanical separation is performed19. Thus, this resemblance was not completely unexpected. The only outlier sample seemed to be the pineal gland of the Large White male (PING_LWM), which clustered with the rest of hypothalamus microarrays. Here contamination can be discarded in all likelihood because the two regions, hypothalamus and pineal gland, are in distinct areas of the brain. However, the pineal gland works in harmony with the hypothalamus. The former produces melatonin, which directly influences the function of various brain centers, including the hypothalamus. In stomach, less clearly in blood, samples were grouped by sex rather than by breed.

As expected, both UPGMA and NJ methods provided identical clustering at the first coalescence level. However, there were some interesting differences at higher levels. NJ identified four groups of tissues. The first group comprised brain tissues, including hypophysis; the second group, thyroid and adrenal endocrine glands; the third, muscle and fat tissues; whereas stomach, ileum, blood and liver were in the last group. In contrast, UPGMA clustered blood, liver and muscles in separate groups, while brain tissues were grouped together with fat, thyroid and adrenal glands; blood and liver were in distinct groups. It is well known that NJ have better properties than UPGMA when reconstructing evolutionary trees because it does not assume a constant evolutionary rate 20. But is not evident how to translate this advantage in relation to the performance for classifying transcriptomes. The UPGMA method is equivalent to clustering algorithms implemented in popular microarray packages 17, whereas NJ have not been used so far, to our knowledge, in microarray studies. Here we found that, overall, NJ provided more biologically sound results than UPGMA: the first two groups in the NJ-tree comprised brain tissues and endocrine glands; muscle and fat tissues have a predominantly structural function in the body, while the rest of tissues (liver, stomach, ileum and blood) have an important metabolic role. In contrast, UPGMA results were not always so easily interpretable, e.g., it grouped hypophysis closer to fat tissues than to other brain organs like hypothalamus or the olfactory bulb.

The trees depicted in Figure 1 were obtained with the original data (after RMA transformation). A better alternative to study relationships between tissues is to consider the Probe × Tissue solutions obtained from model (2), as these solutions are corrected for 'noisy' factors such as sex, breed and global tissue effects. Thus we constructed a NJ-tree of tissue transcriptomes using the 1-rT2 MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemOCai3aa0baaSqaaiabdsfaubqaaiabikdaYaaaaaa@2F92@ distance (see methods). This tree is shown in Figure 2. Although the pattern was similar to the NJ-tree in Figure 1, this tree provided a clearer picture of relationships between tissues. For instance, it emphasized that blood is the most distant tissue, corroborated also by the fact that blood harbored the largest number of extreme probes (Table 1). We defined a tissue's extreme probe as a probe for which all four mRNA levels of the tissue were either smaller or larger than the remaining mRNA levels (Material and Methods). In contrast, the two adrenal tissues and both fats were the closest pair of tissues. The brain is not a uniform organ, and this well known fact 212223 was corroborated by a relevant variability within the different brain tissues sampled. Olfactory bulb was neighbor to hypothalamus. The two pituitaries, adenohypohysis and neurohypohysis, exhibited neighbor but distinct transcriptomes. This result agrees with the fact that both organs carry out very different physiological functions and have separate embryonic origins from the ectoderm layer (from an ectodermal outpocketing of the stomodeum and neural ectoderm, respectively). In fact, it was more interesting to note that neurohypophysis and hypothalamus were relatively distant. A close relationship might have been expected because neurohypophysis consists primarily of axonic projections of the hypothalamus. The pineal gland exhibited an intermediate transcriptome between hypophyses and hypothalamus – olfactory bulb.

Interestingly, there was a relation between embryonic origin and clustering. Most tissues with the same embryonic origin clustered together in the NJ-tree (Figure 2). All brain tissues are of ectodermic origin, as is hypophysis, and these tissues were clustered together in the first group of tissues. The second group comprised the adrenal and thyroid glands, tissues with double embryonic origin. The adrenal medulla derives from ectoderm (neural crest cells), while the cortex develops from mesoderm. The thyroid gland develops from two distinct embryonic lineages: follicular cells (which produce thyroxine) and parafollicular C-cells (which produce calcitonin) and are of endodermal and ectodermal neural crest origins, respectively 24. Mesoderm gives origin to fat and muscle. Liver, and stomach and ileum epithelia are derived from endoderm. Finally, blood, the most separate and outlier tissue, is of mesoderm origin. Thus, it seems that embryonic development leaves a standing footprint on the transcriptome in each of the tissues.

We sought to investigate more in detail the genetic basis of the tissue arrangement by embryonic layers, and to pick up genes that can be differentially expressed in concerted action within each embryonic layer. To that end, we obtained the extreme gene probes differentiating the ectoderm tissues (olfactory bulb, hypothalamus, pineal gland, adenohypophysis and neurohypophysis), mesoderm (muscle and fat), and endoderm (liver, stomach and ileum). An extreme probe for each embryo layer was defined as for individual tissues, i.e., a probe whose all expression levels for that group of tissues were either smaller or larger than the levels for the rest of tissues (Material and Methods). We excluded thyroid and adrenal glands for being mixed tissues and blood, for outlier. The complete list in additional file 1. A pie-plot with differential gene ontology (GO) frequencies obtained with onto-express 25 is in Figure 3. A wide variety of GO were over represented in each layer; this fact, together with a large percentage of unknown or unannotated genes makes it somewhat difficult to interpret the results in detail.

Neverheless, some interesting results appeared. For instance, genes with ontologies like nervous development and ion transport – clearly related to central nervous system – were more frequent than expected within ectoderm extreme genes, as were genes involved in development (Figure 3A). In particular, some of the most overexpressed genes like INA (internexin neuronal intermediate filament protein), ASTN (astrotactin) or NAPIL2 (nucleosome assembly protein 1-like 2) are involved in neuron development, whereas others like GABRG2 (GABA A receptor) or SYT4 (synaptotagmin IV) are involved in normal neuron functioning. The case of NAP1L2 (nucleosome assembly protein 1-like 2) is particularly interesting. Although, according to Pubmed, 'the function of this family member is unknown, mouse studies suggest that it represents a class of tissue-specific factors interacting with chromatin to regulate neuronal cell proliferation'. Our results strongly suggest that NAP1L2 is involved in neuronal system and that the function is maintained across species.

In parallel to results for ecctoderm tissues, the most significant enriched ontology was muscle development in the mesoderm (Figure 3B). Here, some of the most extreme genes were Hox genes involved in limb development and myogenesis (e.g., PRRX1, HOXD8 and MEOX2). In particular, it has been suggested that PRRX1 and HOXD8 are also involved in the urogenital tract development, also of mesoderm origin. The two top extreme genes for mesoderm tissues gamma-sarcoglycans (SGCD and SGCG) are also involved in muscle development. As for the endoderm (liver, ileum and stomach), an excess of genes involved in general metabolism was observed, although we also found a significant enrichment in nervous development genes (Figure 3C). Two of the most overexpressed extreme genes, HNF4A (hepatocyte nuclear factor 4, alpha) and FOXA3 (hepatocyte nuclear factor 3 or forkhead box A3) are known to be involved in liver development, and it is suspected that HNF4A can play a role also in intestine development, which is supported by our data. Intriguingly, some of the most extreme overexpressed genes were involved in the complement pathway (C5, BF, SDC1, C4BPA), may be as a consequence of the defense function of the ileum. Finally, when all extreme genes across the three embryo layers were considered jointly (Figure 3D), the most significant ontologies were nervous system development, cell differentiation, cell adhesion and multicellular organism development, which again suggests a sound biological interpretation.

We also obtained the extreme probes for each of the individual tissues. The main over represented biological process per tissue is in Table 1, and a complete list of the genes is in additional file 2. Results confirmed, overall, previous biological knowledge. brain tissues (olfactory bulb, hypophyses, pineal gland and hypothalamus) exhibited a large number of genes in signalling pathways (see also Figure 3A). The highly oxidative muscle diaphragma was enriched in genes involved in ATP synthesis through citrate cycle and oxidative phosphorylation, whereas biceps femori (glycolytic muscle) extreme genes were often related to the production of ATP by glycolysis and sugar pathway. The liver extreme genes were in the urea cycle pathway and bile acid and amino acid metabolisms; blood genes pertained to hematopoietic processes and signaling pathway systems.

Mixed model analyses were highly successful in modelling transcriptome variability (Table 2). For instance, for model (2) the fraction of unexplained variance was very small, 1 - hP2 MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemiAaG2aa0baaSqaaiabdcfaqbqaaiabikdaYaaaaaa@2F76@ - hPT2 MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemiAaG2aa0baaSqaaiabdcfaqjabdsfaubqaaiabikdaYaaaaaa@30A7@ = 0.03, and ~15% in the rest of models. This means that the set of expression levels can be very accurately modelled as a random normal variate in the log scale. Note that models (1 – 4) were highly parsimonious: a few parameters, i.e., the variance components and the fixed effects tissue, sex and breed were needed to model the data. By far, the largest variance component was that of the probe, which explained overall at least 85% of the total variability (Table 2). The interaction probe × tissue (σPT2 MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaeq4Wdm3aa0baaSqaaiabdcfaqjabdsfaubqaaiabikdaYaaaaaa@3111@) accounted for 11% of phenotypic variance (model 2), whereas the total variances explained by probe × breed or probe × sex were marginal, only about 1% (models 3 and 4). These were the global variance estimates, when all probes were considered jointly. It does not follow that sex or breed were not associated to changes in gene expression. The relative importance of sex or breed did increase when we restricted the analysis to a data subset. For instance, the percentage of total variance explained by breed or sex increased to 25% or 48% when we used the 100 most differentially expressed genes between breeds or between sexes, respectively (last two rows of Table 2). Symmetrically, the relevance of probe effect decreased for these data subsets.

At a false discovery rate FDR < 0.05, we identified 19 and 33 genes differentially expressed between sexes and breeds, respectively. The complete lists together with tentative annotations 26 are in Tables 3 and 4. Two of the genes, orthologous to human genes in the HLA-A and HLA-B complex, were differentially expressed simultaneously between breeds and between sexes. However, there was no correlation overall between tissue, breed or sex effects. This was evident from the pattern in Figure 4, which displays the relation between z^b and z^s scores (top figure) and between z^b scores and the standard deviation of the probe × tissue solutions (bottom figure). If there were a relation, we would see the dots around a diagonal rather than close to the axes. Thus, each of the factors studied here, sex and breed, influence the transcriptome through different genetic programmes.

We performed a clustering of probes and samples using only the differentially expressed genes from Tables 3 and 4. Results were drawn in Figure 5, where the discrimination between sexes (top) and breeds (bottom) is neat and clearly visible. This clustering contrasts with that in Figure 1 where all probes were employed. The discrepancy occurs because the expression pattern of most genes is primarily affected by the tissue and not so much by sex or breed, as discussed. However, as Figure 5 clearly shows, this does not mean that a careful selection of probes does not allow us to discriminate samples between according to sex or breed.

The list of differentially expressed genes between sexes comprised eight X or Y linked genes, based on the localization of orthologous human genes (Table 3). Note that three X-linked genes were up-regulated in males (ZFX, UTX and TMSB4X), a phenomenon also observed by Yang et al. 14 although for different genes. Male up-regulated genes were predominantly involved in spermatogenesis (DDX3Y), cell proliferation, migration, and differentiation (GSTA1, TMSB4X), as well as genes genes from the histocompatiblity complex (HLA-A, HLA-B). Female up-regulated genes had functions related to positive regulation of the transforming growth factor beta receptor signaling pathway (HIPK2), proteins up-regulated in gonadotropinomas (FAM5C) or affecting the circadian rhytm (CLOCK). Up-regulation of CLOCK in females is consistent with evidence showing that this gene is related to the cyclic nature of the reproductive process in females, both in mammal and in non mammal species 272829.

We found a good concordance with the list of differentially expressed genes obtained with the same microarray in lung and mesenteric lymph node by our group in a separate experiment (Fernandes et al., unpublished results). Thirteen genes (mesenteric lymph node) and ten genes (lung) were also in Table 3, despite the fact that corresponded to different tissues, treatments and breeds. Yang et al. 14 compared microarrays of 169 female and 165 male mice in four tissues (whole brain, liver, skeletal muscle and gonadal adipose), the largest study to date of sex differentially expressed genes. They reported 27 genes that were differentially expressed in all four tissues, four of them (DDX3Y, EIF2S3Y, JARID1D and UTX) were also found here; DDX3Y and EIF2S3Y were the most differentially expressed genes found here and in 14. It is interesting to note that these authors also reported that immune genes, e.g., histocompatiblity genes, were overexpressed in males with respect to females, in agreement again with our results 1430. Differential expression between sexes for SSCY genes can be considered as validated.

We also validated probe (Ssc.4897.1.A1_at), annotated as CLOCK gene 26, by real time quantitative RT-PCR (QRT-PCR) in backfat samples of 27 pigs. QRT-PCR results clearly confirmed the microarray data (Table 5). The average relative expression was 1.88 ± 0.68 in females, whereas we were not able to reliably detect expression in males at 1/2000 cDNA dilution because we were below the limit of detection. To verify whether this gene product is actually expressed in males, we repeated the assay at 1/20 cDNA dilution. We could detect low levels of expression in four out of the 13 males studied. Iberian male 27, that with largest relative quantification (RQ), exhibited nonetheless a expression level ~100 fold lower than in females.

The list of probe sets corresponding to differentially expressed genes at FDR < 0.05 between LW and IB breeds is in Table 4. The thirty three probes corresponded to 27 genes and 5 hypothetical proteins; 12 were over expressed in Large White (LW-biased) and 21 over expressed in Iberian (IB-biased). The LW-biased list comprised genes involved in ATP biosynthetic process, nervous system development, lipid metabolism, RNA processing, and control of growing and cellular division. The IB-biased genes affected the regulation of cell growth, RNA splicing factor activity or muscle development. These results hint that transcriptome differences between breeds seem to be manifold. No particular single biological process was predominantly affected. It is to be noticed, nonetheless, that there were several genes involved in nervous development (SEMA3A, GPM6B, SYN3) or growth and cell cycle (DATF1, LYZ, RAB18, ZNF12, DLC1, BVES). As there is currently very little information on the transcriptome differences between pig breeds, it is not possible to know how general these results will be. Some recent results have compared muscle in Duroc vs. Taoyuan breeds 8, in Landrace vs. Tongcheng 10 and liver and muscle in Duroc vs. Pietrain 79. While these studies have focussed on a particular tissue and with different techniques and microarrays, all suggest that breeds differ in genes related to a variety of functional categories. In some cases, the microarray included predominantly genes involved in muscle structure and development or energy metabolism 7 so their results may not be representative of the whole genome.

It should be noted that the differentially expressed genes reported here (Tables 3 and 4) were those that exhibited largest differential expression across all tissues. This follows from the specification of models (1–4), where all probes were analyzed simultaneously and where the tissue effect was included in the models as just an additional effect. Thus, it is important to bear in mind that other genes could show larger differential expressions in a specific tissue than those listed in Tables 3 and 4. The experimental design of this study did not allow us to characterize with reasonable FDR the specific expression of a gene in a single tissue. Nevertheless, we can get a glimpse of the effect of tissue in differential expression by subdividing the tissues in groups. We reasoned that the probe × tissue interaction should be larger between tissues that clustered far apart, as their genetic programmes are more different that tissues that are very similar, e.g., back and abdominal fats. We defined the following tissue groups base on the NJ-tree shown in Figure 2: brain (olfactory bulb, hypophises, hypothalamus and pineal gland), endocrine (adrenal and thyroid glands), structural (fat and muscle) and metabolic tissues (stomach, ileum, liver and blood). To a large extent, these groups reflect also a common embryonic origin (Table 1).

We reanalyzed each of the groups with models 3 and 4, obtaining new z^b and z^s scores for each of the tissue groups. Table 6 presents the correlation between the z-scores obtained with all tissues analyzed simultaneously and each of the groups. In order to focus on the most relevant genes, the correlations shown were those obtained with the 100 genes with largest absolute z-scores. The pattern shown was highly illuminating. For sex, there seemed to be little interaction, as z^s-scores were highly correlated across groups of tissues. Correlation coefficients were always > 0.9. In contrast, correlations between z^b scores were much more variable and, importantly, much lower overall. Thus, whereas the correlation between brain and metabolic tissues was 0.91 for sex z-scores, it was only 0.34 for breed z-scores. This means that most differentially expressed genes between sexes were shared across tissues whereas this was much more unlikely when comparing two distant breeds like Iberian and Large White. It is interesting to note that the clusters in Figure 5 also hinted that heterogeneity was larger within breed differentially expressed genes than within sex differentially expressed genes. Note that the color patterns of sex differentially expressed genes were more uniform across samples than for breed differentially expressed genes. Besides, clustering was much 'flatter' – i.e., more uniform – within sexes than within breeds. This result may have important implications. It suggests that physiological changes responsible for breed differences have targeted differentially the transcriptome across tissues. Not all tissues have been equally affected. It remains to be studied on which of the tissues the effect of breed differentiation has been the largest. Our current data set does not allow us to respond to this question accurately: the probe × tissue heritabilities (hPB2 MathType@MTEF@5@5@+=feaafiart1ev1aaatCvAUfKttLearuWrP9MDH5MBPbIqV92AaeXatLxBI9gBaebbnrfifHhDYfgasaacPC6xNi=xH8viVGI8Gi=hEeeu0xXdbba9frFj0xb9qqpG0dXdb9aspeI8k8fiI+fsY=rqGqVepae9pg0db9vqaiVgFr0xfr=xfr=xc9adbaqaaeGaciGaaiaabeqaaeqabiWaaaGcbaGaemiAaG2aa0baaSqaaiabdcfaqjabdkeacbqaaiabikdaYaaaaaa@3083@) were very similar in all tissue groups and no clear pattern emerged from Figure 1.