Fig 1.
Combining metabolomics profiling with simulations of metabolism.
A: Experimental-based biomarker discovery produces metabolic profiles containing detected and annotated metabolites along with a concentration or fold-change value. B: Metabolic disruptions can be modelled to simulate metabolic profiles similar to those generated using metabolomics. C: By combining information from both types of metabolic profiles and improving both experimental annotation and in silico models, various approaches can be used to improve our knowledge of given biomarker sets, affected pathways, and patient disease classes.
Fig 2.
Methodology for the comparison of flux values and prediction of metabolite ranks.
Using a simple network (A) in two conditions (B), with single flux values (C). The methodology from Shlomi et al. [13] is shown in (D): each exported metabolite will have an associated change score for a given pair of conditions. (E) shows our methodology of scoring and ranking by absolute value among all of the metabolites in the network.
Fig 3.
Flux Variability Analysis (FVA) and sampling.
For simulating fluxes in different conditions (A). The resulting flux values to be compared differ depending on the method used. FVA generates minimum and maximum possible flux values, shown as intervals (B), whereas sampling generates many values within those bounds, shown as distributions (C).
Fig 4.
Z-score based ranks (left side of each plot) vs using the difference between means (right side of left plot) or medians (right side of right plot) to rank metabolites.
The metabolite labels are all ranked in the top 10 by each method. The metabolites highlighted in red are differentially abundant in the example dataset, whereas those in black are the rest of the top 10 metabolites which are not experimentally associated with the SNP.
Fig 5.
Flux bounds (FVA) and distributions (sampling) for urate and hypoxanthine.
The WT state is shown in light blue, and the MUT state in red. MUT here corresponds to the knock-out of the XDH gene. Highlighted in grey, red, and dark blue are the absences of shifts (=), decreases (-), and increases (+) respectively between WT and MUT.
Table 1.
Genes and reactions knocked-out to simulate Xanthinuria Type I in Recon2.
Fig 6.
Accuracy plot for every rank cut-off threshold for the IEM dataset.
Table 2.
Examples of mGWAS data.
Fig 7.
Observed and predicted changes for the five metabolites significantly associated with the rs603424 SNP.
The first column shows the observed change directions from the mGWAS study. The second column shows the predicted change direction using SAMBA (SAMBAdir). The third column shows the predicted change direction using FVA (FVAdir). The fourth column shows the SAMBA predicted rank out of the 1497 metabolites in the network (SAMBARank). The fifth column shows the SAMBA predicted z-score, with the colour scale as the absolute value of the z-score. The NAs represent metabolites for which SAMBA was unable to predict fluxes for one of the following reasons: (i) the metabolite is not in the network, (ii) the metabolite is in the network but has no exchange reaction, or (iii) the metabolite’s exchange reaction can carry no flux (=blocked). Sampling distributions and FVA predicted bounds for each metabolite’s exchange reaction in WT and MUT are shown on the right.
Fig 8.
Predicted ranks for the metabolites present in a ratio significantly associated with the rs603424 SNP.
The first column shows the predicted rank out of the 1497 metabolites in the network. The second column shows the SAMBA predicted z-score, with the colour scale as the absolute value of the z-score. The NAs represent metabolites for which SAMBA was unable to predict fluxes for one of the following reasons: (i) the metabolite is not in the network, (ii) the metabolite is in the network but has no exchange reaction, or (iii) the metabolite’s exchange reaction can carry no flux (=blocked). Sampling distributions for each metabolite’s exchange reaction in WT and MUT are shown on the right.
Fig 9.
SAMBA ranks for the metabolites involved in significant ratios for the ACADS SNP from Suhre et al. 2011.
[24] The SAMBARank column shows the predicted rank out of the 1498 metabolites in the network. The NAs represent metabolites for which SAMBA was unable to predict fluxes for one of the following reasons: (i) the metabolite is not in the network, (ii) the metabolite is in the network but has no exchange reaction, or (iii) the metabolite’s exchange reaction can carry no flux (=blocked). Sampling distributions for each metabolite’s exchange reaction in WT and MUT are shown on the right.
Fig 10.
Hypergeometric test p-values for different rank cut-off values for SCD.
Hypergeometric test p-values for different rank cut-off values for SCD. The left y axis (blue) shows the hypergeometric test p-values when using a given rank cut-off and the number of experimental metabolites predicted in that top ranking. The right y axis (green) shows the number of experimental metabolites predicted for each rank cut-off.
Fig 11.
Hierarchical BiNChE CHEBI graph.
Hierarchical CHEBI graph of the top 10 metabolites predicted to be differentially abundant (outlined in blue) predicted by SAMBA for SCD, extracted using BiNChE. The node colour corresponds to the BiNChE enrichment level. The metabolites in bold & italic were significant in the mGWAS dataset for SCD.
Fig 12.
ChemRich enrichment of the top 50 most changed metabolites for SCD.
The y-axis shows the most significantly altered clusters on the top. Each node reflects a significantly altered cluster of metabolites. Enrichment p-values are given by the Kolmogorov–Smirnov test. Node sizes represent the total number of metabolites in each cluster set. Cluster colours show the proportion of increased or decreased metabolites (red and blue respectively). The x axis represents a separation based on cluster order on the chemical similarity tree, and non-significant clusters are hidden.
Fig 13.
ChemRich enrichment of the top N most changed metabolites for SCD.
ChemRich using only experimentally significant metabolites (left) and using increasing numbers of highly ranked SAMBA metabolites (right) for SCD.