Evaluating transcriptome specialization in normal and cancerous tissues

The analysis of datasets A and B for estimating the transcriptome diversity (H_j,) and specialization (δ_j) indexes were done at three levels of cDNA library grouping with the following designation: individual cDNA libraries were defined as “ungrouped”; assembled cDNA libraries originating from the same kind of organ and tissue state (normal or cancer) were defined as “grouped”; and a further clustering of the cDNA libraries that only considered the tissue state and not the organ of origin were defined as “complete grouping” (see Materials and Methods). H_j measured the variability of the distributions of transcripts and δ_j assessed the average specificity of the genes expressed in the transcriptome. Visualization of the positions of the transcriptomes in the (H_j, δ_j) coordinates allowed us to effectively evaluate the relative differences in these significant parameters. Results for the ungrouped analysis, dataset D analysis, tables of confidence intervals for the relevant parameters, and dissection of the differences in transcriptome specialization are discussed in Supporting Text S1.

Figure 1 presents scatter plots for the levels of transcriptome diversity, H_j, and specialization, δ_j, at the grouped level for datasets A and B. When comparing 12 pairs of human analogous tissues, 11 cancerous tissues had significantly less specialization than their normal counterparts, with eye cancer being the only tissue that had an estimated specialization that was greater that its normal counterpart (Figure 1A, Table S4, and Supporting Text S1). However, after further analysis, we concluded that the eye tissue sample is most likely distorted due to the smaller sample size of the normal eye library (10,679 tags) compared to the cancerous eye library (42,029 tags) (see Supporting Text S1). This likely prevented the correct estimation of eye-specific genes in the normal library. All changes in specialization of the transcriptomes are statistically significant (Table S4 and Table S5; P<0.01). Figure 1A also shows that transcriptome diversity, measured by H_j, increased in the cancerous states of all tissues, with the exception of testis and placenta. The increase in H_j indicates a more even distribution of the transcription levels of expressed genes, which is most likely due to a decrease in the expression of genes prevalent in the normal tissues. As shown in Figure 1A, we observed that the specialization of the grouped ESC libraries is at the same level as the majority of the cancer tissues. This is in agreement with the low morphological specialization of ESCs.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 1. Scatter plot of H_j (Diversity) and δ_j (Specialization) in transcriptomes of normal tissues (blue), cancerous tissues (red), and stem cells (black).

Comparable data sets are linked by a discontinuous line. A - Human data from 53 libraries of 13 distinct tissues with a total of 671,197 tags for 28,087 genes; grouped analyses. B - Mouse data from 29 libraries of 5 distinct tissues and with a total of 541,453 expressed tags for 25,044 distinct genes; grouped analyses. Data for A and B are from the “Cancer Genome Anatomy Project” (http://cgap.nci.nih.gov/). Approximate 95% confidence intervals for diversity and specialization are plotted as continuous colored lines. See Supporting Text S1 as well as Figure S1, Figure S2, Figure S3, Figure S4, Figure S5 and Figure S6 that illustrate individual comparisons and details.

https://doi.org/10.1371/journal.pone.0010398.g001

In order to assess the drop in specialization of the cancerous tissues, we compared the average specificity of genes that were over-expressed in normal tissues to that in cancerous tissues. In general, we found that there was a significantly greater average specificity of genes over-expressed in normal tissues, suggesting that the decrease in specialization was due to the reduction or elimination of the expression of highly specialized genes in normal tissues during carcinogenesis (see Table S11 and Figure S11). We also analyzed the ten most influential genes that caused the reduction in specialization in all eleven tissues of dataset A. For each tissue, we found examples of organ-specific genes that were switched off in the corresponding cancer tissue, including Chondroadherin (CHAD) in bone, Uromodulin (UMOD) in kidney, the acid phosphatase prostate specific (ACPP) gene in the prostate, and a gene for the spermatogenesis-associated protein in the testis (Table S12).

To confirm our hypothesis that specialization decreases in cancerous tissues, we examined a completely independent model of mouse tissues (dataset B). In this analysis, all of the normal tissues showed significantly greater specialization than the corresponding cancerous tissues (Figure 1B, Table S6 and Table S7; P<0.01). Dataset B also included a library of HSCs obtained from bone marrow. These cells showed a level of specialization comparable with normal lung and skin even when undifferentiated (Figure 1B). In four of the five organs studied in dataset B, the average specificity of the genes that were over-expressed in normal tissues was significantly greater than the corresponding value for cancerous tissues, with the exception of the mammary gland (Table S11). However, genes related to milk production, which were within the most influential genes of the mammary gland and have a high specificity of expression in this tissue (Table S13 and Table S14), showed an extreme drop in expression in the cancerous tissue. These results explain the general drop in specialization seen in the mammary gland. In addition, the scatter plots of gene frequency change between normal and cancerous tissues vs. specialization showed a prevalence of highly specific, over-expressed genes in normal tissues from the five organs studied (Figure S12). We conclude that highly specific genes that have diminished expression in the cancerous tissues drive the drop in specialization, similar to that seen in dataset A.

The data from the “Human Transcriptome Map” (dataset C) consist of a collection of SAGE gene tags that belong to heterogeneous normal and tumor tissues that are grouped by chromosome. We disregarded the obvious differences in transcription profiles between distinct organs and only tested the hypothesis that specialization diminishes in the tumor transcriptomes. It is worth noting that in contrast to the analyses of datasets A and B, where gene specificity was estimated for the combination of tissue and condition, in dataset C the specificity is estimated only with regard to the state of the tissue (normal vs. tumor) and disregards the tissue of origin. Therefore, gene specificity in this dataset only refers to normal or tumor tissues and implies that a much lower estimated specialization would be observed. A large and significant change in transcriptome specialization between the normal and tumor tissues was seen for all chromosomes (Figure 2, Figure S7, Figure S8, Figure S9 Figure S10, Table S9, and Supporting Text S1), with the exception of chromosome Y, for which the difference is not significant. Most of the differences that are significant (23 out of 24) are in the expected direction and have less specialization in the tumor transcriptomes. One exception to this was chromosome 18, for which the change in specialization is in the opposite direction (see Supporting Text S1). However, the analysis of all loci together (Figure 2 and Table S9) strongly supports the hypothesis that cancer reduces specialization of the tissues.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Figure 2. Scatter plot of H_j (Diversity) and δ_j (Specialization) in transcriptomes of normal (blue) and tumor (red) tissues in dataset C.

Human expression data are from the “Human Transcriptome Map” project (http://bioinfo.amc.uva.nl/HTMseq/controller), datasets “All tissues normal” and “All tissues tumor”. Data consist of 18,609,073 tags for a total of 62,916 loci by chromosome. See Figure S7, Figure S8, Figure S9 and Figure S10 that amplify the boxes of this figure presenting the 95% confidence intervals for the estimates.

https://doi.org/10.1371/journal.pone.0010398.g002

In our previous study, we showed that the estimated rank of variation of diversity and specialization in the human transcriptome is much smaller when using microarrays than when counting gene tags [14]. This is due to the relatively narrower dynamic range of microarrays compared with tag counting strategies [18], which distorts both high and low expressed genes. Despite these shortcomings, the analysis of normal (TDLUs) and precancerous (HELUs) tissues paired by patient (dataset D) showed a large drop in specialization in precancerous tissues in seven of the eight cases studied (Figure S13).

Genes detected only in cancer

The information theory approach for studying the transcriptome has the advantage of allowing an estimation of the degree of global gene specificity, S_i, of each gene studied, as well as its target specificity, TS_ij, a parameter that measures the specificity of a given gene for a selected transcriptome (see Mathematical Addendum in Supporting Text S1). These tools permit the easy selection of genes that are preferentially expressed in cancer tissues and therefore have the potential to serve as molecular markers of malignancy. In addition, these indexes may aid in identifying genes that are specific to a particular kind of cancer or genes that are not significantly altered during the development of cancer and therefore can serve as marker controls when measuring genes of varying expression. It is important to note that when a gene in a particular dataset is detected in only cancer tissues, it cannot be inferred to be exclusively in cancer, since it could also be present in normal tissues at undetectable levels. However, genes with high levels of expression that are only found in cancer tissues are good candidates for being significantly up-regulated in cancer.

To identify genes that are differentially expressed in cancer tissues, we determined gene specificity (S_i) and target specificity (TS_ij) in dataset A using the complete grouping analysis. Table 1 shows examples of genes represented in cancer tissues at the highest proportion of expression level (greater than 1 in 10,000) and absent from all normal tissues. These genes were only detected in the cancer tissues, with a number of tags (ranging from 54 to 535) in cancer tissues and no tags in normal tissues (maximum S_i in the analysis). To statistically validate the upregulated frequency of these genes, we applied the Fisher exact test [19] with the Bonferroni correction for multiple testing [20] (see Methods). Table 2 presents genes that were detected in only one type of cancer. The identification of these kinds of genes was possible (trough S_i and TS_ij) due to the inclusion of various types of cancer tissues in the analysis. Table S10 shows examples of genes exclusively expressed at relatively high rate in tumor tissues in the analysis of dataset C.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Genes exclusively detected by the fully grouped analysis in cancer tissues in dataset A (human data) at high expression frequency (>0.0001).

https://doi.org/10.1371/journal.pone.0010398.t001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Examples of genes highly expressed in only one type of cancer in dataset A.

https://doi.org/10.1371/journal.pone.0010398.t002

General conclusions

Our present data advance the current hypothesis that cancer diminishes the specialization of affected tissues. This suggests that although cancer tissues gain specialized functions in cell cycle control, angiogenesis, and metastasis, they concomitantly exhibit a loss of specialization and deregulation of sets of genes with tissue-specific functions.

The application of gene specificity and target specificity provides a powerful, practical tool for data mining of the numerous studies that have already compared gene expression in normal and cancerous tissues. We have shown how this method has the capability to recover not only genes well known to be associated with cancer, but also less understood genes that are highly expressed and up-regulated in neoplasia. This approach will help us understand the process of tumor development and hopefully provide new possibilities for intervention with drugs and genomic medicine. These same tools can be used to identify genes that are down-regulated in cancer or control genes that remain constant during the process. In addition, these information tools can be easily adapted to existing web pages that show transcription results from counting gene tag strategies as well as microarrays experiments.

Datasets

Information theory and statistical analyses

If we consider the relative frequencies of transcription p_ij for the i-th gene (i = 1, 2, …, g) in the j-th tissue or transcriptome (j = 1, 2, …, t), then the diversity of the transcriptome of each tissue can be quantified by an adaptation of Shannon's entropy formula,

H_j will vary from zero when only one gene is transcribed up to log₂(g) where all g genes are transcribed at the same frequency: 1/g. If we consider the average frequency of the i-th gene among tissues, say,

we can define gene specificity as

S_i will attain a value of zero if the gene is transcribed at the same frequency in all tissues and a maximum value of log₂(t) if the gene is exclusively expressed in a single tissue. Tissue specialization is then measured, for the j-th tissue, as the average gene specificity, say

These formulas were previously described [14]. For further details see Mathematical Addendum in Supporting Text S1.

To apply the analysis of information parameters of the transcriptome, estimating its diversity, H_j, and specialization, δ_j, as well as the gene specificities, S_i, and target specificities, TS_ij, it is necessary to have estimates, p_ij, of the relative frequency of expression of the i-th gene in the j-th transcriptome, where i = 1, 2, …, g, the total number of genes, and j = 1, 2, …, t, the total number of transcriptomes (tissues) studied [14] (see also Mathematical Addendum in Supporting Text S1). For datasets A, B, and C, which are the product of counting strategies, the values of p_ij were obtained by dividing the number of tags for gene i in the j transcriptome (library or set of libraries) by the total number of tags obtained in the transcriptome j. For microarray data (dataset D; see Supporting Text S1), the p_ij were obtained by dividing the normalized signal obtained for gene i in the j transcriptome by the sum of normalized signals in the transcriptome j (microarray slide j). Using the matrices, {p_ij}, the analysis was performed by using the statistical environment R [31], by use of a set of functions developed to such effect.

To obtain confidence intervals for the information parameters in the datasets A, B, and C, we applied the bootstrap procedure [32]. This assumed a multinomial distribution for the number of tags in the transcriptome, and obtained B = 2000 bootstrap replicates, which are samples from the multinomial distribution where the total number of tags in the transcriptome, n_j, and the observed relative frequencies, {p_ij}, are assumed as parametric values. The assumption of multinomial distribution for the number of tags representing each gene in a transcriptome appears reasonable even if it most likely underestimates the real biological variation and possible dependencies in sets of genes present in the transcriptome. Having obtained the set of B = 2000 bootstrap replicates of each parameter, we applied the Bootstrap Percentile Method [32] at 95% confidence level to obtain the confidence intervals plotted in the figures presented (Figure 1A, Figure 1B, and supporting Figure S1, Figure S2, Figure S3, Figure S4, Figure S5, Figure S6, Figure S7, Figure S8, Figure S9 and Figure S10).

Using the same bootstrap approach, we obtained approximated 99% confidence intervals for the true differences between the specialization of pairs of transcriptomes, for example Δ_mn = δ_m−δ_n, where m and n are the transcriptomes of interest (normal and cancer tissues), for which the hypothesis of equal specialization (H0: Δ_mn = 0) needed to be tested. The null hypothesis of identical specialization was rejected when the approximated 99% confidence interval did not include the value of 0. We also tested the assumption of normality for the bootstrap estimates by means of the Shapiro-Wilks test [33] (Table S4, Table S5, Table S6, Table S7 and Table S8).

The estimation of confidence intervals and hypothesis testing was not feasible for the microarray data (dataset D), since in that case, the original data belonged to an unknown continuous distribution, and the assumption of a given distribution is difficult without having a very good estimate of the (unknown) variances and covariances for the gene expression levels. The absence of true replicates in these data (dataset D) made it impossible to obtain confidence intervals for the parameters or to perform statistical tests on the specialization differences.

Dataset Grouping

The analysis for datasets A and B were done at three levels of grouping. First, the “ungrouped” analysis was performed by considering each of the individual libraries (Table S1 for dataset A and Table S2 for dataset B), with comparisons between the levels of specialization being done between the normal and cancerous states in libraries of the same tissue (Figure S1, Figure S2, Figure S4, and Figure S5). The second level, designated here as “grouped analysis”, corresponds to grouping libraries that were from the same tissue and state (normal or cancerous). This was done by adding the tags for each gene in the corresponding datasets. This is a valid procedure since the libraries are independently obtained from the same organ in approximately the same tissue state (normal or cancerous). In addition, the variations that are disregarded correspond to biological differences between individuals from which the samples were obtained, putative differences in methodology of sequencing, and random variation. The third level of grouping, “complete grouping”, added tags from all distinct tissues in the same state (normal or cancerous) to form only two groups: normal and cancerous states. This “complete grouping” disregards or confounds the transcriptome variation given by the different nature of the tissues, but allows for the analysis of differences between cancerous and normal tissues, which is the main source of this study (Figure S3 and Figure S6). This procedure is analogous to the collapse of contingency tables routinely performed in the statistical analysis of discrete data [19].

To evaluate the statistical significance of genes exclusively detected in cancer in dataset A (Table 1) we used the Fisher's exact test [19] on the 2×2 contingency table produced by grouping the tags belonging to the categories “normal” or “cancer” and “gene I” or “gene no-I” for each one of the 6,220 genes found to be exclusively detected in cancer. Given that we were performing 6,220 tests, the Bonferroni correction for multiple testing [20] was applied by dividing the desired experimental-wise error Type I (α = 0.05) between the number of tests to be performed, for example 0.05/6,220≈7.5e–6. Only genes with P-value less than 7.5e–6 in the Fisher's exact test are presented in Table 1. The same procedure was performed to obtain the significance (P-value) for the data of Table 2, except that in this case, no correction was needed since only one gene was tested for each (independent) set of tags in a given type of cancer.

For dataset C, which corresponds to human data for normal and tumor tissues grouped by chromosome, an individual analysis was performed for the loci in each chromosome followed by a complete analysis for all loci together. This analysis was performed on the non-normalized tags, since the information analysis normalizes the data by taking the relative frequencies of expression of each gene (p_ij). As in the cases of datasets A and B, a bootstrap analysis was performed to obtain confidence intervals for the parameters of interest and to test the differences in specialization between normal and tumor tissues.