New comparative genomics approach reveals a conserved health span signature across species

Comparing ranked lists of genes

We have developed a novel methodology to compare gene expression datasets across experiments. Differences in protocols, platforms, and number of replicate measurements can lead to significant differences in the stringency of the criteria used for detecting differential expression across experiments. The procedure most often used is to analyze the data and set an arbitrary p value cutoff for statistical significance and fold-change, then select sets of genes separately in the different datasets and compare the different sets to find the genes in their intersection. However, it is possible that the dynamic range in the different experiments is different, or that the experimental protocol has introduced more or less variance among replicates. We reasoned that if two gene lists are really related to each other because they share a large number of common genes, we could use the information in one list to inform the detection of significant genes in the other. This will then serve two purposes: to improve on the number of genes that are deemed significantly affected by the treatment, and at the same time to assess the degree of overlap between two experiments.

We used a non-parametric approach to compare lists of genes based on the rank statistics, where genes were ranked according to the magnitude of the change in expression between the treatment and control conditions. A similar approach has been successfully applied to the study of enrichment of functional categories of genes [23]. Starting from the genes with the highest ranks in the two conditions we want to compare, we looked at the probability that the observed overlap between such genes could be explained by chance alone. We then marched down the ranks in steps by adding genes until any of the following two criteria was not met: 1) the new overlap was not statistically significant; 2) the increase in the overlap from the previous step could be explained by chance alone. A detailed statistical derivation of the comparison method is given in the Supplemental Information.

First, we evaluated our method on simulated data generated as follows. We took a ranked list of genes (list A) and generated a new dataset by completely scrambling the list order (list B) (Figure 1A), such that the two lists would have no association between them. The inlet panel in Figure 1A shows the scrambling process. Then, we computed the percent overlap between the top 100 ranking genes in the two lists and repeated the process by adding 100 genes at a time from the ranked lists. Every time we added genes, we tested the statistical significance of the overlap and marked significant overlaps with red dots; we represented overlaps that could have been obtained by chance alone with blue dots. The two lists do not show any significant overlap for any of the ranks, because, as expected, the proportion of genes in the overlaps (y-axis) is roughly equal to the proportion of genes selected (x-axis).

Then, we simulated and tested the case in which two lists are related to each other. The rank of genes in list A was shuffled to generate list B, but this time we limited the number of ranks a given gene could jump to (see inlet in Figure 1B). In this case, the algorithm detected a significant overlap between the two lists (Figure 1B). When we limited the number of ranks a gene could change to between the two lists (inlet in Figure 1C) such that the two lists were very similar to each other, the algorithm selected a very large overlap (Figure 1C).

We then tested the algorithm on real data, using three different combinations of datasets from Bauer et al. [6]. In the first example, we compared the ranked gene list from decreased Dmp53 activity to a ranked list created by examining the differences in gene expression between two of the control strains used in the DR experiments in that study, specifically the yw, w¹¹¹⁸ and the Canton-S strains on high calorie (1.5N) food. Changes in expression induced by a gene mutation that extends lifespan are unlikely to be related to differences associated with specific genetic backgrounds.

Consistently, our algorithm did not detect any significant overlap between them (Figure 1D). However, when we compared the DR and the Sir2 datasets, we detected a significant overlap between the two lists (Figure 1E), as reported before [6]. As expected, the degree of overlap was even better when we compared the effect of the same intervention, DR, in two different fly strains (Figure 1F).

Life span extension in Drosophila melanogaster

Life span extension can be achieved by environmental interventions, genetic manipulations and drugs, and some longevity pathways are thought to be shared by these three modalities. One example is the small mole- cule resveratrol, which is believed to mimic the beneficial effects of dietary restriction by enhancing Sirt1 activity, especially in D. melanogaster and C. elegans [12]. We combined several available expression array datasets and used the list comparison algorithm to compare gene lists corresponding to DR, Sir2 overexpression and resveratrol treatment. Hence, the collection of experimental data we worked with contained data from different interventions, genetic backgrounds and tissues. Table I shows a breakdown of the datasets we used in D. melanogaster. Each dataset corresponds to a treatment/control pair of expression microarrays sets, with each set containing three biological replicates for both treatment and control. In what follows, overlaps between lists are reported if and only if such overlaps reached statistical significance. The complete list of genes identified using the list comparison algorithm can be found in the Supplemental Data.

We first assessed the effect of different genetic backgrounds and tissues on the same treatment. While only 78 genes (Figure 2A) were selected in the intersection between DR datasets from different tissues - whole body versus head/thorax - a much larger number, 3,205 (Figure 2B) were shared by DR fruit flies from different genetic backgrounds (Canton-S and yw, w¹¹¹⁸). Hence, the tissue effect appears to have a predominant role in determining the set of genes responding to a longevity intervention. As expected, direct comparison of DR between datasets where both genetic background and tissue were different, revealed an even smaller overlap of only 25 genes (Figure 2C).

When we compared the Sir2 dataset with the resveratrol dataset, there were a significant number of genes (20) shared by the two longevity interventions (Figure 2D). Although this number was relatively small, the two datasets differed in both genetic background and tissue. If we take that into account and compare this intersection with the number of genes in common between the two DR datasets in Figure 2C, the small number likely reflects an underestimate of the true overlap between the significant genes in the two interventions.

Of the 20 genes in the intersection, 9 were annotated as belonging to the proteolysis gene ontology (GO:0006508). One of the genes, ninaD is involved in phototransduction and recently its mammalian homolog, the scavenger receptor SR-BI [24], has been shown to have a role in Alzheimer's disease pathogenesis [25].

Dietary restriction and resveratrol in mice

We next looked at a similar dataset from mice that reproduced, at least partially, our Drosophila dataset. Pearson et al. [17] have studied the expression changes induced by DR and two levels of resveratrol (high and low) in four different mouse tissues: heart, white adipose tissue, liver and skeletal muscle. The dataset however did not contain Sirt1 overpression and, to our knowledge, such a dataset has not been published yet. We applied our novel list comparison algorithm to all pair-wise comparisons between the treatment/control pairs in this dataset and clustered interventions using a hierarchical clustering algorithm with a metric based on the degree of overlap between lists (see Material and methods). The dendrogram resulting from this clustering procedure is shown in Figure 2E. The degree of similarity between interventions within the same tissue was higher than the degree of similarity between the same interventions across tissues. In fact, datasets were grouped by tissue first (same color in figure), and by treatment second, with the exception of liver, in which DR and high resveratrol were more similar to each other than to the low resveratrol. Although the mouse data did not contain expression profiles for different genetic backgrounds, the strong tissue effect displayed by the clustering algorithm confirms our previous findings in D. melanogaster.

Our findings are consistent with a good deal of experimental data indicating resveratrol is a DR mimetic in mammals and flies [12, 13, 17, 26, 27]. For example, Berger et al. [27] reported that in three different tissues - heart, skeletal muscle and brain - the expression profile induced by a low dose of dietary resveratrol partially mimics the DR expression profile. Consistently, when we applied the list comparison algorithm to this dataset we detected in all three tissues a large and significant overlap between genes responding to both DR and resveratrol with changes in the same direction (heart: 977 genes; skeletal muscle: 1,733 genes; brain: 912 genes). These sets of genes overlapped, at least in part, with the sets we identified in Pearson et al. [17], although the overlaps did not reach statistical significance.

Conserved health span signature between invertebrates and mammals

We wanted to determine if our new algorithm could detect a DR/resveratrol signature shared across species. First, we compared the gene list from DR to the gene list from resveratrol in Drosophila to define a resveratrol/DR signature. This identified 975 genes changing in the same direction in both DR and resveratrol treatment in fruit flies (761 up-regulated and 214 down-regulated) (Figures 3A and 3B). In mice, a resveratrol/DR signature identified 1,365 genes changing in the same direction, with 760 up-regulated and 605 down-regulated (Figures 3C and 3D). Then, we matched genes across species using the HomoloGene database (see Materials and methods). Of the 975 genes in the Drosophila signature, 229 genes had a mouse homolog according to the HomoloGene database, while of the 1,365 genes in the mouse signature, 342 had a fruit fly homolog. We then compared the direction of change of these genes in the two species and identified 18 genes that were up-regulated in both and 5 genes that were down-regulated in both (Figure 3E).

Of the 23 genes in common between fly and mouse, 9 (39%) are involved in stress response functions such as response to starvation (Asl and Sord), oxidative stress (Aco1, Prdx6, Dhdh, Gstt1, Ldha and Acy1) and cellular protection (Ruvbl1), while 8 (35%) are involved in metabolism and growth (Umps, Dpp4, Gnmt, Slc7a2, Slc19a2, Riok3, Gtpbp2 and Acy1). Response to stress, oxidative stress in particular, has been identified as one of the possible mediators to the longevity response [28]. The intersection also contained two aquaporins, both up-regulated in all health span-extending interventions we studied. Interestingly, Lee et al. have recently shown that in C. elegans glucose shortens lifespan by down-regulating the expression of an aquaporin gene [29].

Conclusions

Our findings demonstrate it is possible to identify preserved gene signatures by pair-wise comparison of gene lists from different health span extending interventions. The list comparison method we have developed can identify shared genes as well as the degree of similarity between two interventions. We have shown that both genetic background and predominantly tissue play an important role in defining a signature. However, our novel algorithm was sufficiently sensitive to detect common genes across different tissues and genetic backgrounds.

By comparing gene expression changes induced by Sir2 overexpression and resveratrol in D. melanogaster we have shown for the first time a significant overlap between the two datasets. This is consistent with a model in which Sir2 and resveratrol share a common longevity pathway. The overlap, while significant, was relatively small; however, this could be explained by the fact that the two datasets, Sir2 and resveratrol, were collected using different tissues and genetic backgrounds. Because our results show that these differences can drastically reduce the size of the overlap, it is likely that the intersection we detected underestimates the true size of the overlap. It would be very interesting to know if and how many of these genes are shared by Sirt1 overexpression and resveratrol treatment in mice. However, no Sirt1 overexpression dataset in mouse is currently available.

Our novel gene expression dataset of DR and resveratrol treatment in D. melanogaster has revealed a large degree of similarity in the gene expression response to the two longevity interventions. In addition, by applying the list comparison algorithm to a similar dataset from mouse, we confirmed and extended the findings already reported in [17] about the close association of the two interventions in mouse. Finally, by comparing homologous genes across species we identified a preserved signature comprised of 23 genes that change in the same direction in response to DR and resveratrol treatment in both D. melanogaster and mouse, the majority of which are involved in stress response, metabolism and growth.

Gene Expression datasets

The D. melanogaster dataset was obtained by combining gene expression datasets from [17] and [30]; mouse gene expression data were obtained from [6]. GCRMA [31] was used for quantile normalization and summarization of the data to generate expression scores in the log₂ scale. Probesets with an expression score below the 25^th percentile (compared to the rest of the mean expressions for that condition) in both treatment and control were removed from further analysis. Fold-changes used in the list comparison algorithm were computed as the ratio between the average expression scores of the three biological replicates in the treatment and control cohorts. The fold-change of genes with more than one probeset on the array were computed as the average fold-change between all corresponding pobesets.

Comparison of gene lists

The probability of observing at least k* genes in the intersection of two lists generated by randomly choosing two sets of m genes out of a total of N was computed as: $P(X\ge k^{*})={\displaystyle \sum _{k=k^{*}}^{m}P}(X=k)$ where $P(X=k)=\frac{\left(\begin{array}{c}m\\ k\end{array}\right)\left(\begin{array}{c}N-m\\ m-k\end{array}\right)}{\left(\begin{array}{c}N\\ m\end{array}\right)}$ is the hypergeometric distribution. A mathematical derivation of the list comparison algorithm can be found in the Supplementary Information. R code to run the list comparison algorithm and its documentation is available at http://www.physics.brown.edu/physics/researchpages/ibns/listcomp.htm.

Clustering

Hierarchical clustering was performed using a complete linkage algorithm and using m_ij = max_(l,m)=1…N (k_lm) − k_ij as metrics, where N is the total number of experimental conditions to be compared, and k_ij is the number of genes in the intersection when comparing the i-th experimental condition to the j-th experimental condition.

Probesets identification numbers in the fruit fly and mouse microarrays were converted to the corresponding gene symbol and matched across species using the HomoloGene database from NCBI.