Strikethrough metabolites are not included currently filtered dataset.Metabolites can be filtered by platform, fluid and time point (challenges) in the header dropdowns (top right). If the here selected metabolites isn't within the filtered set they are highlighted by strikethrough.
Modular framework structure to visualize aspects of dynamic changes
Metabolite selection
Search for your metabolite of interest using the available tables. Filter the results by platform, biofluid, pathway membership, and implemented distance measures to find metabolites with similar kinetics.
Temporal plots
Visualize your metabolic trajectories using mean or individual metabolite abundances in interactive plots. This visualization provides quick insight into metabolic kinetics over time.
Networks
Obtain a comprehensive overview of all metabolites measured using targeted and non-targeted MS-based platforms in plasma and urine. Networks can be animated over time to highlight temporal metabolic changes.
Statistical analysis
Perform statistical tests to identify metabolites with significant changes during a challenge of interest or analyze the observed metabolomics variance using PCA. All plots and generated data can be downloaded.
Futher reading
Published papers using the HuMet data
The dynamic range of the human metabolome revealed by challenges, Krug S., Kastenmüller G., Stückler F., Rist MJ., Skurk T., Sailer M., Raffler J., Römisch-Margl W., Adamski J., Prehn C., Frank T., Engel KH., Hofmann T., Luy B., Zimmermann R., Moritz F., Schmitt-Kopplin P., Krumsiek J., Kremer W., Huber F., Oeh U., Theis FJ., Szymczak W., Hauner H., Suhre K., Daniel H.,FASEB J. 2012 Jun;26(6):2607-19.
The showcases provide an easy access to display groups of metabolites with interesting curve trajectories and results from statistical analysis. All showcases were selected by the authors.
Distance info: the Fréchet distance was calculated on the averaged metabolite curves (averaged across subjects per time points) with the distFrechet function implemented within the
longitudinalData
R package
Distance info: The Fréchet distance was calculated with the distFrechet function implemented within the
longitudinalData
R package
Distance info: the
stats
R package was used to calculate the paired Pearson correlation on the trajectories averaged across all selected participants.
Distance info: the
stats
R package was used to calculate the paired Pearson correlation on the trajectories of each individual, with distances being averaged across all chosen participants.
Distance info: the z-scored dataset was used to calculate the Euclidean distance between two metabolite curves within one subject and subsequently aggregating the similarity measure across all chosen participants. We used the dist function implemented within the
proxy
R package to calculate the Euclidean distance
Distance info: the z-scored dataset was used to calculate the Manhattan distance between two metabolite curves within one subject and subsequently aggregating the similarity measure across all chosen participants. We used the dist function implemented within the
proxy
R package to calculate the Manhattan distance
Select a reference metabolite and optionally choose a distance measure
Loading...
KEGG pathways
Add all
Pathway information: the KEGG pathways are a collection of manually curated pathway maps which represents our current knowledge of molecular interactions. Selectable KEGG pathways were exctracted from
KEGG
.
Loading...
Add all
Info: to provide an overview of measured metabolites, we implemented a super-pathway and sub-pathway structure.
Metabolites identified from our in-house platforms were manually assigned to this structure and matched to knowledge-based platforms such as KEGG, PubChem, and HMDB.
For identified metabolites from vendor-based kits, we used annotation provided by the Metabolon HD4 and Biocrates p150 platforms to classify them into our super- and sub-pathway structure.
Single fluid GGMs are based on partial correlations. Network inference closely follows procedures established by
Do et al., 2015
but adding a dynamic partial correlation
Strimmer et al. 2005
, which takes the factor time into account.
The shrinkage estimator approach 'GeneNet' was used as the dataset includes fewer samples (840) than variables (2652).
Multi-fluid networks were constructed as described in
Do et al., 2015
termed as overlaid networks. Here we linked single fluid networks by metabolites that were measured on the Metabolon HD4 [nt-ms] platform.
Knowledge-based networks were constructed based on the super- and sub-pathway structure that we implemented. This structure provides a quick overview of available metabolites from different platforms.
Highlight bag:
Color by:
Show single nodes:
Fold change
p Value
Selected metabolite
Measured metaboliteNot measured
Statistical methods
Interactive visualization of selected statistical methods for longitudinal exploration of the HuMet dataset.
Dataframe containing all challenge time points used for the comparison of dietary intake challenges. Data was log2 transformed and missing values were imputed using RF-based imputation (missForest version 1.4)
Protein precipitation using Phenomenex ImpactTM protein precipitation plates. Details can be found here:
Fiamoncini J, Rist MJ, Frommherz L, Giesbertz P, Pfrang B, Kremer W, Huber F, Kastenmüller G, Skurk T, Hauner H, Suhre K, Daniel H and Kulling SE (2022) Dynamics and determinants of human plasma bile acid profiles during dietary challenges. Front. Nutr. 9:932937. doi: 10.3389/fnut.2022.932937
For the quantitative determination of BA a Nexera LC system (Shimadzu Europa GmbH, Duisburg, Germany) coupled to a 5500 Q-Trap mass spectrometer (Sciex, Darmstadt, Germany) was used with electrospray ionization in the negative mode . Source parameters were: 40 psi (curtain gas), 600°C (Source Temperature), -4500 V (Ion Spray Voltage), 50 psi/60 psi (Ion Gas 1 and 2, respectively). Data were recorded in the multiple reaction monitoring mode (MRM) with nitrogen as a collision gas. System operation and data acquisition were done using Analyst 1.5.2. software (AB Sciex).
We would also like to thank all research groups that generated the data integrated into the HuMet repository:
Krug S, Kastenmüller G, Stückler F, Rist MJ, Skurk T, Sailer M, Raffler J, Römisch-Margl W, Adamski J,
Prehn C, Frank T, Engel KH, Hofmann T, Luy B, Zimmermann R, Moritz F, Schmitt-Kopplin P, Krumsiek J, Kremer W,
Huber F, Oeh U, Theis FJ, Szymczak W, Hauner H, Suhre K, Daniel H.
The dynamic range of the human metabolome revealed by challenges.
FASEB J. 2012 Jun;26(6):2607-19.
Metabolite proflingMeasurement and identification of metabolites
To provide an overview of measured metabolites, we implemented a super-pathway and sub-pathway structure.
Metabolites identified from our in-house platforms were manually assigned to this structure and matched to knowledge-based platforms such as KEGG, PubChem, and HMDB.
For identified metabolites from vendor-based kits, we used annotation provided by the Metabolon HD4 and Biocrates p150 platforms to classify them into our super- / sub-pathway structure.
Metabolon HD4
In addition to the published metabolomics data, we acquired new data by profiling plasma and urine samples on the non-targeted LC-MS based platform Metabolon HD4 at Metabolon, Inc. (Durham, NC, USA).
First recovery standards were added to the samples for quality control purposes.
Thereafter the samples were prepared using the automated MicroLab STAR® system from Hamilton Company (Reno, NV, USA), to remove proteins.
The resulting sample was split into five portions, stored overnight under nitrogen and reconstituted in solvents compatible for the methods before each analysis.
Metabolite detection and quantification were performed on four different chromatography platforms:
Two separate reverse Phase (RP)/ ultra-high-performance liquid-phase chromatography (UPLC)-MS/MS methods with positive ion mode electrospray ionization (ESI)
HILIC/UPLC-MS/MS with negative mode ESI
(RP)/UPLC-MS/MS with positive ion mode ESI
RP/UPLC-MS/MS with negative ion mode ESI
Thereafter, organic solvent was removed by placing the samples on a TurboVap® (Zymark).
All four chromatography methods were coupled with a Thermo Scientific Q-Exactive high resolution/accuracy mass spectrometer (MS/MS) applying standard protocols developed by Metabolon(
Evans et al. 2014
)
Two types of controls were measured to assess the mean relative standard deviation (RSD) per metabolites:
Pooled matrix samples (CMTRX) generated from all HuMet samples to assess biological variability versus process variability;
Vendor maintained human plasma samples (MTRX) to assess across Metabolon HD4 measured studies.
The area-under-the-curve was used to quantify peaks, giving rise to relative levels of a total of 595 known metabolites in the plasma and 620 in urine.
Metabolites of the Metabolon HD4 platform were assigned to 8 super-pathway classes (amino acid, carbohydrate, cofactors and vitamins, energy, lipid, nucleotide, peptide, xenobiotics), each being divided into two or more sub-pathways), resulting in a total of 78 and 68 sub-pathways for the plasma and urine metabolites, respectively.
Lipidyzer
Lipid concentrations in HuMet plasma samples of four participants were analyzed on the LipidyzerTM platform of AB Sciex Pte. Ltd. (Framingham, MA, USA) by Metabolon Inc., Durham, NC, USA. A detailed protocol for lipid quantification of the HuMet study samples is published in (
Quell et al., 2019
).
In brief, dichloromethane and methanol were used to extract lipids within a Bligh-Dyer extraction. For analysis, the lower, organic phase which includes internal standards was used and concentrated under nitrogen. 0.25 mL of dichloromethane:methanol (50:50) containing 10 mM ammonium acetate was used for reconstitution. The result was placed in vials for infusion-MS analysis which was performed on a Sciex 5500 QTRAP equipted with SelexIONTM differential ion mobility spectrometry. Phosphatidylcholines were detected in the negative MRM mode and quantified using 10 stable isotope labeled compounds.
This method allowed for absolute quantification of 965 metabolites of 3 annotated sub-pathways, including “neutral lipids”, “phospholipids” and “sphingolipids”. All metabolites were part of the “Lipids” super-pathway.
Metabolite proflingMeasurement and identification of metabolites
The metabolic raw area counts from the non-targeted LCMS platform were corrected for instrument inter-day tuning differences by setting the run-day medians equal to one. Additionally, data derived from urine samples were normalized by osmolality before run-day correction. The resulting transformation of these normalized datasets is termed “Concentrations and relative abundances” within the repository and can be downloaded in bulk.
Manual curation
Due to large fluctuations in metabolic concentrations, we applied manual data curation by systematically filtering single data points according to the following two criteria:
The value of the single data point was outside 4 times the standard deviation (SD) of this time point.
The data point was not measured within the first 30 minutes of a study challenge
A total of 163 data points fit both criteria, of which 92 were excluded after manual inspection.
The resulting data has been deposited as 'Concentrations and relative abundances' within the HuMet repository.
Data deposited in the repository does not contain manually excluded data points.
Imputation
Some of our statistical models require a complete dataset without missing values. This includes network inference with gaussian graphical models and principle component analysis. To address missing values, we used the manually curated dataset for imputation.
We used the machine learning algorithm
missForest
, which is implemented in the
missForest
R package, to impute missing metabolites of the urine and plasma Metabolon HD4 platforms that exhibited less than 30% missingness. A summary of metabolites that did not meet this criterion is provided in the table below.
Metabolites of the numares (Lipofit), Chenomx, In house biochemistry, PTR-MS, and FTICR-MS platforms were handled accordingly.
Overview of measured tissues, platforms and number of metabolites that were not imputed.
Fluid
Platform
Total number of metabolites
Metabolites >= 30% missingness
Plasma
Metabolon HD4
595
102
Lipidyzer
965
146
Biocrates p15
132
3
numares (LipoFIT)
28
0
In house Biochemistry
4
1
Urine
Metabolon HD4
619
25
Chenomx
6
0
Breath air
In-house PTR-MS
106
106
Breath condensate
In-house FTICR-MS
201
0
Data transformation
Users can select between two different types of data transformation based on the dataset with or without imputation. The following options are available:
Use z-scores to better compare across platforms
Choose fold changes calculated between user-defined time points.
In the
Time Course module
the default baseline time points are set to the first time point per block. In the
Network module
the default baseline is set to the baseline of each challenge. In the
Time Course module
the default is set to the first time point of the user-chosen time range.
Statistical analysisMethods used for data exploration
We offer several distance measures for ranking metabolites based on their similarity over time. Statistical calculation of these distance measures depends on the selected time range, included subjects, and imputation, which the user can choose individually.
Users can compare paired metabolite trajectories. Here, we use the z-scored dataset to calculate the distance (Euclidean, Manhattan) or correlation (Pearson) between two metabolite curves within one subject, and then sum up the distance matrices across all selected subjects. We use the dist function implemented within the
proxy
to calculate the Euclidean and Manhattan distance.
the
stats
R package is used to calculate the paired Pearson correlation.
Users can use distance measures capable of measuring the similarity between curves by taking the metabolite concentration and order of time points along the metabolite curves into account.
The Fréchet distance is set as the default within the similarity tool, as it is more suitable for our data type. Here, we use a window approach that allows the Fréchet algorithm to search for the smallest distance between curves in a defined timeframe. This timeframe is defined as follows: a maximum of +/- 30 minutes within all challenges except extended fasting. Within extended fasting, we allow for comparison of time points within a range of +/- 120 minutes. Metabolite curves are compared by subject within the user-defined time range, and subsequently, the mean is calculated across all subjects. To calculate the Fréchet distance, we use the
distFrechet
function implemented within the
kmlShape
R package.
Paired t-test
Metabolic changes were assessed by paired t-tests of non-imputed samples using the
t.test
function implemented in R statistical software (version 3.6.1).
The user can select either a group of metabolites to be analyzed or analyze all measured metabolites.
To adjust for multiple testing, the user can choose between FDR (q <= 0.05) or Bonferroni correction (nMetabolites / nTime points).
The levels of adjustment depend on the chosen time range and number of metabolites submitted for analysis.
Results are visualized in a volcano plot using the
plotly
R package.
Each data point in the volcano plot can be colored by corresponding super-pathways or measured platform.
Network generation
We constructed knowledge-based networks based on the super- and sub-pathway structure that we implemented.
This structure provides a quick overview of available metabolites from different platforms. The network inference of single fluid networks is based on estimating GGMs from the metabolite concentrations.
These models are based on partial correlations (
Opgenrhein & Strimmer 2006
) and have previously been shown to reconstruct biological pathways from metabolomics data (
Krumsiek et al. 2011
).
Our single fluid networks are based on the imputed and log2 transformed data.
Network inference closely followed the procedures reported in
Do et al. 2017
, termed 'fused networks' with a few changes due to the longitudinal nature of the dataset.
According to the user specification, one or more platform datasets are merged by vendor provided metabolite id.
We added the temporal dimension to the dataset and calculate dynamic partial correlation (
Opgenrhein & Strimmer 2006
), which takes the factor time into account by using the function dyn.pcor implemented within the
longitudinal
R package.
Furthermore, the shrinkage estimator approach “GeneNet” was used, as the dataset includes fewer samples (n samples = 840) than variables (n metabolites = 2652), which is available within the
GeneNet
R package.
If both dynamic partial correlation and Pearson correlation were statistically significant at alpha = 0.05/n samples, pairwise metabolite connections were integrated into the network. The user can choose to apply Bonferroni or FDR correction and cut metabolite links at specific dynamic partial correlations.
Multi fluid networks were constructed as described in
Do et al. 2017
termed as overlaid networks. Here, we linked single fluid networks by vendor provided metabolite id that.
The HuMet repository was built using the
Shiny R package
, which provides a basic structure for the interactive web application. All additionally used R packages are listed below. We wrote the majority of the code in R with custom tailoring using JS and CSS for data-specific individualization.
Upon session start, the repository loads preprocessed data, metabolite information, and sample information. The repository is then responsive to the user's options, allowing them to exclude data points based on time points, subjects, and platforms. The chosen data is then visualized in interactive plots, with customizable color options for the user.
Study designOverview over the HuMet (Human metabolome) study
The fasting challenge consisted of a 36-hour period of fasting, during which ten sampling time points were taken.
Standard liquid diet
The two standard liquid diets (SLD) on day two and three consisted of a defined fiber-free formula drink (Fresubin® Energy Drink Chocolate, Fresenius Kabi, Bad Homburg, Germany), providing one-third of the daily energy requirement. Within the SLD challenges in block one, participants were sampled at 13 time points. In block two, participants were sampled four times.
Oral glucose tolerance test
The oral glucose tolerance test (OGTT) on day three consisted of a 300 ml solution with mono- and oligosaccharides, equivalent to 75g of glucose after enzymatic cleavage (Dextro O.G.T., Roche Diagnostics, Mannheim, Germany). Samples were taken at nine different time points within the OGTT.
Physical activity test
Physical activity tested by a 30 min bicycle ergometer training (PAT).
Oral lipid tolerance test
The oral lipid tolerance test (OLTT) on day 4 combined two parts of the SLD and one part of a fat emulsion containing predefined long-chain triglycerides (Calogen®, Nutricia, Zoetemeer, Netherlands). Here, samples were taken at 11 time points within the 8-hour challenge. All challenge drinks were served at room temperature for ingestion within 5 minutes.
Cold stress
Participants were triggered by immersing one hand, up to wrist level for a maximum of 3 min in ice water.
Standardized chicken meal
Participants were provided with a standardized balanced chicken meal (FRoSTA Tiefkühlkost GmbH, Hamburg, Germany) at 7 pm one day prior to each to both blocks (before fasting and before OGTT).
Participants
Parameter
Mean
Standard deviation
Coefficient of variation
Age (y)
27.8
2.9
10.7
Height (cm)
180
10
3.5
Weight (kg)
77.5
7.1
9.1
Body mass index (kg/m2)
23.1
1.8
7.6
Fat mass (kg)
14.4
3.3
23.1
Fat free mass (kg)
59.5
5.9
9.9
Waist circumference (cm)
80.5
4.6
5.7
Hip circumference (cm)
90.1
4.7
5.2
Heart rate (min-1)
62
11.4
18.4
Blood pressure, systolic (mmHg)
116.6
5.9
4.9
Blood pressure, diastolic (mmHg)
81.9
5.9
7.3
24h resting metabolic rate (kcal)
1721.3
223.6
13
Respiratory quotient
0.8
0.1
6.5
A total of 15 healthy male participants were recruited. Participants did not take any medication, did not show any metabolic abnormalities and were non-smokers
Sample types
A total of four different types of samples were collected during the HuMet study. Detailed sampling information is described by
Krug et al., 2012. Time points are numbered from 1 to 56, with five extra time points (10.5, 11.5, 27.5, 39.5, 57) added in between the original time points. In total, samples were taken at 61 unique time points.
Plasma Time points: 56 Time range: 15 to 120 minutes
Spot urine Time points: 23 Time range: 2 to 4 hours
Breath air Time points: 56 Time range: 15 to 120 minutes
Breath condensate Time points: 32 Time range: every hour
The first step involved selecting a metabolite that is suitable and exhibits the expected trajectory (i.e., a decline over time). For this showcase, we chose
3-methylhisidine
as a reference metabolite, which is a known biomarker of meat intake [
Yin et al. 2017, J. Nutr.
].
2. Metabolite ranking:
The
Selection Module
was used to choose
3-methylhisidine
as the reference metabolite. The Fréchet distance was then calculated between the average trajectory of the reference metabolite and all other average metabolite trajectories.
We used the
distFrechetfunction implemented within the longitudinalData
R Package to calculate the Fréchet distance between our reference metabolite and all other metabolite trajectories.
3. Visualization of interesting patterns:
Metabolites with a Frechet distance < 1.6 were selected to be visualized in our
Time course module
.
Background
Fig. 1. Study design with standardized chicken meal.
Before each of the two blocks in the HuMet study, a standardized chicken meal (Frosta) was given. Thus
3-methylhisidine
was expected to demonstrate a clear excretion curve. Further intake of
3-methylhisidine
could be ruled out, as none of the highly controlled challenge meals contained chicken.
What are putative dietary biomarker?
Dietary biomarkers are metabolites that should exhibit steady decreasing excretion curves after food intake, with minimal interferance by environmental stimuli.
How many markers have been validated?
Validation of biomarkers is one of the onging efforts in nutritional metabolomics. Several databases integrate results from various biomarker studies including
FOODB
and the
Exposome Explorer
.
Which metabolite did we use as reference biomarker?
We used the known chicken meat intake biomarker
3-methylhisidine
to identify further dietary biomarkers by distance ranking.
Results (plot)
Fig. 1. Ten selected metabolites with similar kinetics to3-methylhisidineranked by Fréchet distance.
Metabolites depicting a "washout-like" trajectory
A total of 21 metabolites showed a distance lower than 1.5 to the reference metabolite.
10 xenobiotics, 8 amino acids, 3 cofactors and vitamins, and 1 lipid in blood and urine.
Many have been declared as or may be putative dietary biomarkers of food intake.
Results
Table.1. Metabolites that depict similar trajectories as the reference metabolite 3-methylhisitidine.(Mean Frechet <=1.6)
Network reconstruction Gaussian graphical models (GGMs) establish connections between metabolites based on significant partial correlation, which corrects for indirect correlations that are mediated by a third metabolite. These pairwise correlations are assembled as edges into a network, generating the GGM in a data-driven manner from metabolomics data. The resulting network has been shown to reconstruct biochemical reaction structures, resembling known metabolic pathways.
Our network
Here, we use the single fluid GGM (Metabolon HD4 Plasma, Biocrates p150 plasma and in-house biochemistry, Pcor >= 0.12). We mapped metabolic changes of the OGTT (60 min vs. baseline), SLD (60 min vs. baseline) and OLTT (60min vs baseline) and focused on two selected pathways.
Patterns of special interest: Bile acids We can visualize time-dependent responses to metabolic challenges utilizing networks. The holistic network approach depicts changes across distinct biological pathways across the individual challenges. The most well-studied included challenge is the oral glucose tolerance test (OGTT), where we commonly evaluate glucose disposal, ß-cell function and insulin sensitivity for diagnosis. In the OGTT, we focus on the temporal response of glucose, insulin and insulin regulated pathways. However, recent studies have suggested an important role of bile acids in glucose regulation and energy homeostasis. Using our holistic temporal network approach, we can investigate the regulation of bile acids across all dietary challenges.
Results
Networks provide an overview of changes upon nutritional changes
Observed differential responses of N-acetyltyrosine
Similar responses of bile acid metabolism in all three challenges
Method
1. Network generation
Network inference of single fluid networks is based on estimating GGMs from the metabolite concentrations.
These models are based on partial correlations and have previously demonstrated to reconstruct biological pathways from metabolomics data.
Our single fluid networks are based on the imputed and log2 transformed data. Network inference closely followed the procedures reported in Do et al.,
termed “fused networks,” with a few changes due to the longitudinal nature of the dataset.
2. Statistical results
The bubbles within the network represent metabolites. Here, we added bubble color and size. Color displays the log2 foldchange between challenge baseline and chosen time point with red indicating an increase in metabolite concentration and blue indicating a decrease in metabolite concentration. Size depicts the -log10 p-value of metabolic changes between challenge baseline and chosen time point. Thereby, node size increases with lower p-value.
3. Cluster selection
Interesing clusters of metabolites were selected manually.
References
Opgenrehein & Strimmer 2006
Krumsiek et al., 2011
Krumsiek et al., 2012
Do et al., 2015
Repository implementation
Individualized results of the pairwise t tests can be viewed in the
Selection Module
. T test results are used within the time-resolved networks for animation.
Generated networks can be viewed within our
Networks Module
.
Time-resolved networks"Which areas of metabolism change after extended fasting (36 h) compared to standardized overnight fasting in the reconstructed metabolic network?"
Network reconstruction Gaussian graphical models (GGMs) establish connections between metabolites based on significant partial correlation, which corrects for indirect correlations that are mediated by a third metabolite. These pairwise correlations are assembled as edges into a network, generating the GGM in a data-driven manner from metabolomics data. The resulting network has been shown to reconstruct biochemical reaction structures, resembling known metabolic pathways.
Our network Here, we use the multi-fluid network (Metabolon HD4 plasma pcor >= 0.12, Metabolon HD4 urine pcor >= 0.09). On this backbone, we mapped statistical results comparing extended fasting (36h postprandial, TP 10) versus standardized overnight fasting (12h postprandial, TP 1).
Fasting challenge Extended fasting is of special interest as it depicts part of the catabolic state which the body undertakes several times per day. Here, we can find known and novel patterns within the dynamic adaption.
Results
Fig. 1. Temporal multi fluid networks based on the plasma and urine Metabolon HD4 platform. This network combines two single fluid networks: (i) plasma single fluid GGM, Metabolon HD4 (dynamic partial correlation >=0.12), and (ii) urine single fluid GGM (dynamic partial correlation >= 0.08). Here, metabolic changes in biological pathways are compared between fasting baseline vs 28 hours of fastingColor displays the log2 foldchange between challenge baseline and chosen time point with red indicating an increase in metabolite concentration and blue indicating a decrease in metabolite concentration. Node size depicts the -log10 p-value of metabolic changes between challenge baseline and chosen time point. Thereby, node size increases with lower p-value.
Network view provides a holistic overview of typical metabolic responses upon perturbation
Multiple biological processes are involved in adaption to 36h extreme fasting
Multiple biological pathways affected by extended fasting
(A) Upregulated bile acid metabolism
(B) downregulated benzoate metabolism
(C) downregulated xanthine metabolism
(D) upregulated ketone bodies
(F) upregulated levels of free fatty acids
Method
1. Network reconstruction
Network inference of single fluid networks is based on estimating GGMs from the metabolite concentrations.
These models are based on partial correlations and have previously demonstrated to reconstruct biological pathways from metabolomics data.
Our single fluid networks are based on the imputed and log2 transformed data. Network inference closely followed the procedures reported in (Do et al. 2015),
termed “fused networks,” with a few changes due to the longitudinal nature of the dataset.
2. Statistical results
The bubbles within the network represent metabolites. Here, we added bubble color and size. Color displays the log2 foldchange between challenge baseline and chosen time point with red indicating an increase in metabolite concentration and blue indicating a decrease in metabolite concentration. Size depicts the -log10 p-value of metabolic changes between challenge baseline and chosen time point. Thereby, node size increases with lower p-value.
3. Cluster selection
Interesing clusters of metabolites were selected manually. Each of cluster represents a unique biological pathway that is part of the metabolic responses to extended fasting.
References
Gaussian Graphical Model inferrence
Opgenrehein & Strimmer 2006
Krumsiek et al., 2011
Krumsiek et al., 2012
Do et al., 2015
Repository implementation
Individualized results of the pairwise t tests can be viewed in the
Selection Module
. T test results are used within the time-resolved networks for animation.
Generated networks can be viewed within our
Networks Module
.
Showcase: Platform comparisonObjective: Comparison pf metabolites across platforms
Here we use 38 metabolite pairs measured on the non-targeted LCMS and targeted LCMS matched by
Yet et al., 2016
comprising of 43 metabolites that were measured both using the Biocrates p150 kit and using an older version of the Metabolon HD4 platform in a british cohort. Out of the 43 listed metabolite pairs, we could match 38 within our dataset.
2. Calculation of curve similarity
The curated, imputated and z-scored Metabolon HD4 plasma and Biocrates p150 plasma dataset is used to calculate the paired Pearson correlation between two metabolite curves within one subject and the resulting distance matrices are summed up across all subjects.
Background
Why compare overlapping metabolite signals across profiling platforms?
There are multiple solutions to measuring and identifying metabolites. Therefore, metabolites reported across different studies are measured on different profiling platforms. It is thereby essential to compare a unique metabolite that is measured on different platforms, to estimate the precision of each platform.
Which overlapping signals were measured in the HuMet study?
We could match 38 metabolites that were measured on both the Metabolon HD4 and Biocrates p150 platform. We aimed to judge whether matchable pairs of metabolites measured on Metabolon HD4 and Biocrates p150 can be justified. Our study setup provides good conditions as the platforms measured the same samples and the factor time adds in dimension.
Results
Fig. 1. Comparison of 38 overlapping metabolites measured on both non-targeted (Metabolon HD4) and targeted (Biocrates p150) platform matched by expert knowledge.
Pearson correlations were calculated on the individuals metabolite trajectories and subsequently averaged across all subjects to assess the synchronicity. Metabolites were ranked according to the super-pathway and Pearson correlation.
Summary
Strong correlations between most overlapping metabolites sums (Biocrates p150 [t-ms]) and single metabolite signals (Metabolon HD4 [nt-ms])
Results
Fig. 2. Depiction of Laurylcarnitine/C12 (Laurylcarnitine).Laurylcarnitine/C12 (Laurylcarnitine) depict the strongest correlation (r = 0.95).
Fig. 3. Butyrylcarnitine/C4 (butyrylcarnitine).Butyrylcarnitine/C4 (butyrylcarnitine) show the weakest correlation (r = 0.18) and the isoform isobutyrylcarnitine/C4 (butyrylcarnitine) demonstrating a high correlation (r = 0.82). Therefore, repository exploration reveals high contribution of isobutyrylcarnitine [nt-ms] to C4 (butyrylcarnitine) sum [t-ms].
The
Module Selection
can be used to calculate the Pearson correlation between metabolite trajectories.
The
Module Time course
can be used visualize metabolites with similar trajectories.
Paper figures
currently under development
Fasting vs. Non-Fasting
Explore metabolites that are informative of the fasting status, which were identified via an iterative random forest model using backward variable selection.
this site is currently under developement
Kinetic Patterns
Early peaks in metabolic levels
Find pattern in challenge:
Table.1: Ranking of metabolites with early peaks by euclidean distance
Washout metabolites
Find pattern in challenge:
Table.1: Ranking of metabolites with early peaks by euclidean distance
Late peaks in metabolic levels
Find pattern in challenge:
Table.1: Ranking of metabolites with early peaks by euclidean distance
Not measured
_small.png){kind=logo}
_small.png){kind=logo}
Standard liquid diet
Oral glucose tolerance test
Physical activity test
Oral lipid tolerance test
Cold stress
Individualized results of the pairwise t tests can be viewed in the
Selection Module
. T test results are used within the time-resolved networks for animation.
Generated networks can be viewed within our
Networks Module
.
Fig. 1. Temporal multi fluid networks based on the plasma and urine Metabolon HD4 platform.
This network combines two single fluid networks: (i) plasma single fluid GGM, Metabolon HD4 (dynamic partial correlation >=0.12), and (ii) urine single fluid GGM (dynamic partial correlation >= 0.08). Here, metabolic changes in biological pathways are compared between fasting baseline vs 28 hours of fasting
Color displays the log2 foldchange between challenge baseline and chosen time point with red indicating an increase in metabolite concentration and blue indicating a decrease in metabolite concentration. Node size depicts the -log10 p-value of metabolic changes between challenge baseline and chosen time point. Thereby, node size increases with lower p-value.
The
Module Time course
can be used visualize metabolites with similar trajectories.