Paper of the week: Visualizing associations between paired data sets

This week's paper of the week is by González, et al., entitled "Visualising associations between paired `omics' data sets," and published in BioData Mining (vol 5:19). The pdf of this report can be found here.

The authors demonstrate that graphical outputs such as Correlation Circle plots, Relevance Networks and Clustered Image Maps are useful in the visualization and interpretation of output from integrative analysis tools. The goal is to facilitate an understanding of systems as a whole when complex data often force donning of blinders to not observe the whole forest.

The graphical tools described in the report are implemented in the freely available R package mixOmics and in its associated web application.

As an example of what the authors have built, consider their presentation of Nutrimouse data showing correlations (or not) between between large data sets, in case gene expression and metabolite levels in liver, as taken from their figure 5.The Nutrimouse data are from a nutrigenomic study in which 40 mice from two genotypes (wild-type and Ppara -/-) were fed five diets with different fatty acid compositions. Details are in the Methods section. Expression of 120 genes in liver cells was obtained with microarrays and concentrations of 21 hepatic fatty acids were measured by gas chromatography. Hence, the data matrices are of size (40 × 120) for the gene expression and (40 × 120) for the fatty acids measurements.

The Authors write: The Correlation Circle plot (above) displays all fatty acids and the genes selected on each component (100 in total in this plot). Highlighted are subsets of variables important in defining each component. For example, C18:2ω6, C20:2ω6 and C16:0 are fatty acids for which variation allows the definition of the sPLS component 2 (top and bottom of the y-axis). Similarly, genes such as Car1, Acoth, Siat4c, Scarb1 (SR.BI) and Slc10a1 (Ntcp, or Ntop [sic]) are positively correlated to each other, and to the fatty acid C16:1ω9 and their variation participate in defining the sPLS component 1 (left-hand side of the x-axis).

I find such analysis and depiction of results useful and look forward to trying this with our GWAS data.

Proteomics of fasting

It was with keen interest that I read the article by Bouwman, et al. entitled “2D-electrophoresis and multiplex immunoassay proteomic analysis of different body fluids and cellular components reveal known and novel markers for extended fasting,” which appeared in BMC Medical Genomics last week. As we are interested in the genetic basis for the differential response to diet, we view perturbations to the system, either by a high-fat intervention or fasting/calorie restriction, as instrumental in deciphering this response. Overall, I found this to be nice work and deserving of a wide audience.

The authors report that “[p]rincipal component analysis applied to the multiplex immunoassay (RBM) data set revealed that each of the subjects could be identified based on levels of 89 plasma proteins (see figure 3). It appears that such data can be used to provide a metabolic fingerprint of the individual volunteers participating in this intervention study. However, this demonstrates that the between-subject effects are larger than that of the fasting effect.”

This is not surprising given that of the 44 different proteins identified as responding to extended fasting (see tables 2 & 3, figure 4), nine are encoded by genes harboring variants responding differentially to environmental factors such as dietary intake and physical activity. The dietary component most often modulating the association between those genes (and their variants) and a phenotype pertinent to metabolic syndrome is fat. Physical activity is also a wide-reaching modulator of the association between genetic variation and various phenotypes pertinent to metabolic syndrome. In other words, a combination of genetic variation between study participants in combination with each individual’s lifestyle choices (say, more or less exercise) could indeed influence the levels of certain proteins found to respond to the fasting intervention.

At the same time, I cannot dispute, as the authors write, that the between-subject variation may have arisen from heterogeneity of the study cohort “with regard to various parameters, including gender and BMI.” This is logical, but again other factors such as habitual diet and exercise, even sleep patterns could be at work here. Another source of between-subject variation is certainly genetic.

The authors observe that “[m]ost interesting biomarkers are involved in metabolic pathways, as well as those related to inflammation and oxidative stress.” This is where my quite minor complaint with the work arises – I would have liked to see more interpretation of the results from a biological or even medical perspective. Thus, I note that IL10, IL1B, TNF, SERPINE1, INS and CCL2 respond to extended fasting and are members of the Insulin resistance inflammation network (Olefsky, Glass (2010) in a review of Macrophages, Inflammation and Insulin Resistance (Annu Rev Physiol 72:219-46)). Furthermore, VCAM1, APCS, CRP, IL1B, TNF, IL18 and CCL2 are assigned an inflammation role within the set of PPARA target genes (Rakhshandehroo, Kersten 2010 PPAR Research pii: 612089).

A second comparison I undertook was to look at the number of genes responding to the fasting intervention and to an intervention termed AIDM: Anti-inflammatory dietary mix (Bakker, et al 2010 Am J Clin Nutr 91:1044). Large-scale assays of genes, proteins, and metabolites in plasma, urine, and adipose tissue showed that an intervention with selected dietary components influenced inflammatory processes, oxidative stress and metabolism in humans. Eight genes are in common and we’d expect about one by chance. These eight genes are FABP3, VCAM1, IL12A, AFP, FTH1, IL18, APOA1 and F7. Most of these eight were described in the AIDM article as down-regulated (lower levels) in plasma by the dietary intervention, similar to the response to fasting. This raises the intriguing hypothesis that the AIDM diet at least partially mimics fasting.

Adipokines are signaling proteins that are secreted from adipocytes. It is an interesting observation, then, that four of the altered proteins seen during fasting are described by Rosenow, et al as adipokines. These are SERPINE1, SERPINF1, C3 and TIMP1. Perhaps fasting-induced changes to the signaling potential of adipose tissue should focus on these four proteins.