Skip to content

Asplund-Samuelsson/thermosampler

Repository files navigation

alt text

Thermosampler

Linear programming and Markov Chain Monte Carlo sampling tools for thermodynamic analysis of metabolism.

Contents

  1. Background
  2. System requirements
  3. MDF and NEM analysis
  4. Hit-and-run analysis
  5. Case study: Syntrophic communities
  6. Author

Background

Motivation

Tools such as NET analysis (Kümmel et al. 2006, Zamboni et al. 2008) allow calculation of the minimum and maximum feasible metabolite concentrations in a metabolic network, while ensuring that all reactions have a negative Gibbs free energy change, or, equivalently, a positive thermodynamic driving force. Similarly, Max-min Driving Force analysis (MDF; Noor et al., 2014) is another linear programming approach to finding an optimal set of metabolite concentrations in a metabolic network. However, the aforementioned approaches only identify extreme concentration values, which is why this thermosampler framework was developed to explore the full "solution space" of feasible metabolite concentrations (Fig 1). A random walk, or hit-and-run algorithm, allows assessment of metabolite concentration distributions through exploration of all combinations of concentrations that are thermodynamically feasible.

alt text
Fig 1. Thermosampler framework overview and example. One feasible combination of concentrations within the solution space serves as starting point ("Start") for a random walk. First, a random direction is picked (1). Stepping outside and back into feasible space over and over (2) determines the longest possible step size in the positive and negative directions (3). Then a random step length is performed (4). A new point in the solution space is thus reached and the process is repeated many times over. The right panel depicts 200 steps performed by the sampling.py script. Color indicates step number (fMCS; feasible metabolite concentration set). Note that the random walk occurs in logarithmic space, which makes it appear as if higher concentrations are more sparsely sampled.

Toy model

As an example, a clearly defined solution space was devised using a "toy" model. The toy model has two metabolites, A and B, and one reaction where A is transformed into B. The standard Gibbs free energy change was selected to be 5.705 kJ/mol, which corresponds to the constraint that [A] must be at least 10 times higher than [B]. Furthermore, [A] and [B] must be within the concentration range 0.01 to 1, and the sum of [A] and [B] must be 1 or lower. These three constraints define a solution space on the AB-plane that has the shape of a triangle (Fig 1). A real metabolic network has dozens of reactions and metabolites, and is therefore many magnitudes more complex and difficult to visualize.

Example sampling

A random walk applying the thermosampler algorithm was performed within the triangular AB solution space using the sampling.py script:

./sampling.py --reactions examples/toy.model.tab \
--std_drG examples/toy.model_drgs.tab \
--constraints examples/toy.concentrations.tab \
--max_conc 1 --steps 200 \
--outfile results/toy.sampling.tab

The results from the random walk were then plotted to yield the right panel in Fig 1:

examples/plot_toy_example.R
results/toy.sampling.pdf # Output PDF

Feasibility checks

In the thermosampler algorithm, the feasibility concept and how to check it is very important, as it ensures that the random walk occurs within the solution space, and thereby defines the solution space itself. Feasibility consists of adherence of the metabolite concentration vector to several constraints:

  1. Thermodynamic driving forces of all reactions must be positive, based on the user-supplied model reactions and directions, as well as the user-supplied standard reaction Gibbs free energy changes.
  2. Metabolite concentration ratios must be within the user-supplied ranges.
  3. Total metabolite concentration sum must be below the user-defined value.
  4. Metabolite concentrations must be within the user-supplied ranges.
  5. Sums of groups of metabolite concentrations must be below the user-supplied values.

Several functions exist within sampling.py to perform the feasibility checks, and are called by the "master" feasibility function named is_feasible. It should be relatively simple to add additional feasibility checks to constrain the solution space further.

Algorithm

The thermosampler algorithm starts at a feasible combination of concentrations within the solution space, i.e. one "feasible metabolite concentration set" (fMCS). This set is obtained through inefficient rejection sampling or from MDF analysis. More fMCSs are obtained along the random walk.

The thermosampler algorithm works as follows, with concentrations in logarithmic space:

  1. Create a step direction vector of the same length as the concentrations vector, but with random values.
  2. Hone in on the edge of the solution space by determining the maximum and minimum allowable factor theta to apply to the step direction vector before adding it to the current concentrations vector. First the maximum step size without violating any of the initial concentration range bounds is added and feasibility is checked. If feasible, the edge has been found. If not, it is necessary to hone in on the edge of solution space by iteratively increasing the inner step size (starts at zero) and decreasing the outer step size (starts at the previously mentioned maximum), until the difference is smaller than a user-defined, small value.
  3. Once the difference between inner and outer step size is small enough, the inner step size is taken as "touching" the edge of the solution space, and thereby constituting the maximum feasible step size. The minimum feasible step size (can be negative) is gained from performing step 2 in the opposite direction. Thereby the minimum and maximum step size is determined.
  4. Perform one step in the random walk by sampling one step size factor theta from the previously calculated range, multiplying the step direction vector by that factor, and adding the product to the current metabolite concentrations vector. The feasibility of the new vector, i.e. the new metabolite concentration set, is checked. The algorithm stops and prints an error if an infeasible point is reached, which could happen for example if the solution space would be noncontinuous or concave. The error is however more likely to occur from infeasible model parameters (the input files and variables).
  5. Repeat from step 1 until the desired number of steps through the solution space have been performed.

Practical examples of using the thermosampler algorithm with sampling.py are shown below.

System requirements

  • Operating system with Unix-like terminal (Tested on Ubuntu 18.04.5 LTS, 20.04.2 LTS, and MacOS 12.1)
  • bash (Tested with 3.2.57, 4.4.20 and 5.0.17)
  • Python ≥ 3.7 (Tested with 3.7.6 and 3.10.0)
  • R ≥ 4.1.1 (Tested with 4.1.1)
  • GNU parallel (Tested with 20161222 and 20220122)
  • Python libraries: numpy, pandas, scipy, tqdm, joblib
  • R libraries: doMC, foreach, ggridges, optparse, scales, tidyverse

Setting up the virtual environment

It is recommended to run the Python scripts within a virtual environment. The virtual environment can be created using the following command:

python3 -m venv venv
source venv/bin/activate
pip install -r requirements.txt

On Windows systems, replace "python3" with "py", and activate the virtual environment using the command venv\Scripts\activate.bat. Please note that this software has not been tested on Windows systems.

For improved performance om Mac systems with M1 processors, numpy can be built with support for the Accelerate framework. This can lead to roughly a doubling of performance in hit-and-run sampling. For more information, see this Stack Overflow question.

MDF and NEM analysis

mdf.py performs Max-min Driving Force (MDF; Noor et al., 2014) and Network-Embedded MDF (NEM; Asplund-Samuelsson et al., 2018) analysis. Run ./mdf.py -h to list all options of the script.

Example: NEM analysis on lysine biosynthesis in E. coli and Synechocystis

./mdf.py --min_conc 0.0000001 --max_conc 0.1 \
--constraints examples/E_coli.concentrations.tab \
--ratios examples/E_coli.Lys_opt_ratios.tab \
--pathway examples/E_coli.Lys_pathway.txt \
examples/E_coli.model.tab \
examples/E_coli.model_drgs.tab \
results/E_coli_Lys_nem.csv
./mdf.py --min_conc 0.0000001 --max_conc 0.1 \
--constraints examples/Synechocystis.concentrations.tab \
--ratios examples/Synechocystis.Lys_opt_ratios.tab \
--pathway examples/Synechocystis.Lys_pathway.txt \
examples/Synechocystis.model.tab \
examples/Synechocystis.model_drgs.tab \
results/Synechocystis_Lys_nem.csv

Example: MDF analysis on TCA cycle

./mdf.py --min_conc 0.0000001 --max_conc 0.1 \
--constraints examples/tca.concentrations.tab \
--ratios examples/tca.ratios.tab examples/tca.model.tab \
examples/tca.model_drgs.tab results/tca_mdf.csv

Hit-and-run analysis

sampling.py performs hit-and-run sampling of feasible metabolite concentration sets by a random walk through the solution space. Run ./sampling.py -h to list all options of the script.

Quick example: Hit-and-run sampling on TCA cycle starting from and maintaining MDF

./sampling.py --reactions examples/tca.model.tab \
--std_drG examples/tca.model_drgs.tab \
--outfile results/tca.mdf_sampling.tab \
--constraints examples/tca.concentrations.tab \
--concs results/tca_mdf.csv --steps 1000 --mdf 1.3

Advanced example: Hit-and-run sampling on TCA cycle with replicates

Here we sample TCA cycle metabolite concentrations five times, computed in parallel and always starting from the MDF, saving every 10th step. In the first sampling all reactions must have a driving force > 0 kJ/mol, while in the second sampling all reactions must have a driving force > 1.18 kJ/mol, which is 90% of the MDF.

# Sample 5 replicates of 10000 steps in parallel at feasible delta G
for i in {1..5}; do
  echo "./sampling.py --reactions examples/tca.model.tab \
  --std_drG examples/tca.model_drgs.tab \
  --ratios examples/tca.ratio_range.tab \
  --outfile results/tca_sampling/tca.Feasible.${i}.tab \
  --constraints examples/tca.concentrations.tab \
  --concs results/tca_mdf.csv --steps 10000 --max_conc 1.2 -n 10"
done | parallel --no-notice --jobs 5

# Sample 5 replicates of 10000 steps in parallel at optimum delta G (90% of MDF)
for i in {1..5}; do
  echo "./sampling.py --reactions examples/tca.model.tab \
  --std_drG examples/tca.model_drgs.tab \
  --ratios examples/tca.ratio_range.tab \
  --outfile results/tca_sampling/tca.Optimal.${i}.tab \
  --constraints examples/tca.concentrations.tab \
  --concs results/tca_mdf.csv --steps 10000 --max_conc 1.2 \
  --mdf 1.18 -n 10"
done | parallel --no-notice --jobs 5
Convert reaction table to stoichiometric matrix

stoich.py converts a model in reaction table format to a model in stoichiometric matrix format. Run ./stoich.py -h to list all options of the script. This representation of the model is needed for plotting in the next step.

./stoich.py examples/tca.model.tab results/tca.stoich.tab
Plot the output of multiple hit-and-run sampling replicates

plot_samples.R creates principal component analysis (PCA), concentration, and driving force plots from hit-and-run sampling experiments. Run ./plot_samples.R -h to list all options of the script. All sampling results files used as input must be stored without any other files in one directory (--indir option). Sampling results files must have the <path>/<indir>/<experiment>.<group>.<replicate>.tab name format (even if there is only one group and/or only one replicate).

./plot_samples.R -i results/tca_sampling -S results/tca.stoich.tab \
-G examples/tca.model_drgs.tab -c examples/tca.concentrations.tab \
-x C00001 -o results/tca_sampling_plots

Plotting yields the following output files:

results/tca_sampling_plots.sampling_pca.png
results/tca_sampling_plots.sampling_pca_combo.png
results/tca_sampling_plots.sampling_concs.pdf
results/tca_sampling_plots.sampling_concs_combo.pdf
results/tca_sampling_plots.sampling_dfs.pdf
results/tca_sampling_plots.sampling_dfs_combo.pdf
results/tca_sampling_plots.concs.pca.pdf
results/tca_sampling_plots.dfs.pca.pdf

The combo tag indicates that replicates have been combined in the plot and are not displayed individually, as in the other variant of each plot type. Combining multiple replicates improves coverage of the solution space.

results/tca_sampling_plots.sampling_pca.png
alt text
The PCA divided by individual samples shows how each random walk in the hit-and-run algorithm has traversed the solution space. The point color indicates the index of each feasible metabolite concentration set (fMCS) produced by the random walk. The figure below shows a cutout with a single replicate, where the left panel is the random walk at a driving force > 0 kJ/mol ("Feasible"), and the right panel has a driving force > 1.18 kJ/mol ("Optimal"). Note that the actual output file contains all replicates and is provided in the results directory.
results/tca_sampling_plots.sampling_pca_combo.png
alt text
The combined PCA plot layers all replicate random walks, giving an indication of the coverage of the solution space. Each replicate up to 12 in total is given a unique color. The transparency attempts to show the step index.
results/tca_sampling_plots.sampling_concs.pdf
alt text
Concentration distributions obtained from the random walk are first visualized divided by replicate. "Feasible" is a driving force > 0 kJ/mol and "Optimal" is a driving force > 1.18 kJ/mol. Note that "Feasible" and "Optimal" are the Group variable names we gave to the two experiments through their file names. Up to 9 such names/groups can be selected freely to accomodate different types of experiments. In the concentration plots below, the vertical line marks on the x axes indicate the boundaries of the input concentration ranges.
results/tca_sampling_plots.sampling_concs_combo.pdf
alt text
As the random walks are quite noisy, it is best to combine multiple replicates to obtain smoother distributions. It can also be helpful to perform more steps, for example in the millions. The combined plot visualizes all values from all replicates of each group. Distributions are compared between Group variables with the Kolmogorov-Smirnov algorithm (function ks.test), providing the D statistic that measures how different two distributions are. The D statistic is indicated by the color ("Difference") of vertical lines connecting two distributions, situated in each facet to the right of the distributions. Any two distributions with D ≤ 0.15 are considered equal and therefore do not get a connecting vertical line.
results/tca_sampling_plots.sampling_dfs.pdf
alt text
Finally, the plotting script calculates the driving forces for each reaction based on the sampled metabolite concentrations. The driving forces are first plotted per replicate. Note that the x axis uses a square root scale.
results/tca_sampling_plots.sampling_dfs_combo.pdf
alt text
Then, driving forces are plotted with replicates combined to yield a smoother representation of the thermodynamic solution space. It appears that R01082, i.e. fumarate hydratase, is somewhat of a thermodynamic bottleneck in this experiment. As for the concentrations, the Kolmogorov-Smirnov D statistic ("Difference") is used to highlight differences in the distribution shapes.
results/tca_sampling_plots.concs.pca.pdf results/tca_sampling_plots.dfs.pca.pdf
alt text alt text
PCA of all metabolite concentrations reveals co-variation between metabolites. PCA of all driving forces reveals co-variation between reactions.
Use a plot config file to filter, order, and rename reactions, metabolites and groups

The plot_samples.R script can be supplied with a config file to filter, order, and rename reactions and metabolites. The example config file examples/tca.plot_config.tab is provided. Let's look at the contents:

cat examples/tca.plot_config.tab | column -tn -s $'\t'
Id        Name                      Type
R00351    Citrate synthase          Reaction
R01325    Citrate hydro-lyase       Reaction
R01900    Isocitrate hydro-lyase    Reaction
R00709    Isocitrate dehydrogenase  Reaction
R08549    2OG dehydrogenase         Reaction
R00405    Succinyl-CoA synthetase   Reaction
M00148    Succinate dehydrogenase   Reaction
R01082    Fumarate hydratase        Reaction
R00342    Malate dehydrogenase      Reaction
C00122    Fumarate                  Metabolite
C00149    Malate                    Metabolite
C00311    Isocitrate                Metabolite
C00036    Oxaloacetate              Metabolite
C00042    Succinate                 Metabolite
Feasible  Free                      Group
Optimal   MDF                       Group

The config file has three columns:

  • Id is the identifier used for reactions, metabolites, or groups in the model and results files.
  • Name is the alternative name we want to display in the plots, replacing the identifiers that are otherwise displayed.
  • Type is one of Reaction, Metabolite, or Group, and defines what variable the identifier and name refers to. At least one line for each of the three types must be present in the config file to create meaningful plots and avoid crashes.

The exact order of reactions, metabolites, and groups in the config file is used to order the plots in the output accordingly. Thus, we use the plot config file to get nicely ordered TCA cycle plots with meaningful names:

./plot_samples.R -i results/tca_sampling -S results/tca.stoich.tab \
-G examples/tca.model_drgs.tab -c examples/tca.concentrations.tab \
-C examples/tca.plot_config.tab -o results/tca_cfg/tca_cfg_sampling_plots

These are the plots made with the help of the config file:

results/tca_cfg/tca_cfg_sampling_plots.sampling_pca.png
results/tca_cfg/tca_cfg_sampling_plots.sampling_pca_combo.png
results/tca_cfg/tca_cfg_sampling_plots.sampling_concs.pdf
results/tca_cfg/tca_cfg_sampling_plots.sampling_concs_combo.pdf
results/tca_cfg/tca_cfg_sampling_plots.sampling_dfs.pdf
results/tca_cfg/tca_cfg_sampling_plots.sampling_dfs_combo.pdf
results/tca_cfg/tca_cfg_sampling_plots.concs.pca.pdf
results/tca_cfg/tca_cfg_sampling_plots.dfs.pca.pdf

The concentrations now display only the five selected metabolites, C00122, C00149, C00311, C00036, and C00042, using their names Fumarate, Malate, Isocitrate, Oxaloacetate, and Succinate. Note that the groups of samples have been renamed to Free and MDF.

alt text

The driving forces are now displayed under more informative reaction names ordered according to their position in the TCA cycle.

alt text

The metabolite concentrations and reaction driving forces PCA plots show the data of interest according to the definitions in the plot config file.

alt text alt text

Case study: Syntrophic communities

One application of thermosampler is the study of the metabolite concentrations and thermodynamics of metabolic reactions in microbial communities. Syntrophic propionate oxidation and methanogenesis was considered in this case study.

Syntrophic propionate oxidizer model

A model was prepared for a syntrophic propionate oxidizing bacterium (SPOB) based on Candidatus Syntrophopropionicum ammoniitolerans (Fig 2). The model was parameterized using data from literature and Equilibrator.

Model component File
Reactions describing metabolic network data/methylmalonyl.model.tab
Stoichiometric matrix from stoich.py data/methylmalonyl.stoich.tab
Standard reaction Gibbs free energy changes data/methylmalonyl.model_drGs.tab
Default metabolite concentration ranges data/methylmalonyl.concentrations.tab
Fixed concentration ratios for MDF data/methylmalonyl.ratios.tab
Concentration ratio ranges for sampling data/methylmalonyl.ratio_range.tab

alt text
Fig 2. Model of syntrophic propionate oxidizing bacterium. Protons (H+) are not shown. Fdred, reduced ferredoxin; Fdox, oxidized ferredoxin; MK, menaquinone.

Fixed extracellular high and low propionate (input metabolite) and acetate (output metabolite) concentrations were defined to allow thermosampler analysis regarding the thermodynamic effects of such conditions.

Concentrations File
Default data/mmcoa_fixed_prp_ac/methylmalonyl.concentrations.free.tab
High acetate data/mmcoa_fixed_prp_ac/methylmalonyl.concentrations.acHi.tab
Low propionate data/mmcoa_fixed_prp_ac/methylmalonyl.concentrations.prpLo.tab
High prp/Low ac data/mmcoa_fixed_prp_ac/methylmalonyl.concentrations.prpHi_acLo.tab
Low prp/High ac data/mmcoa_fixed_prp_ac/methylmalonyl.concentrations.prpLo_acHi.tab

The SPOB model was used to sample thermodynamically feasible metabolite concentrations for all the conditions above as described in this Bash script:

source/methylmalonyl_fixed_prp_ac.sh

The results were plotted using plot_samples.R and yielded the following output:

Full dataset plots.
# Random walk PCA
results/mmcoa_fixed_prp_ac_plots.sampling_pca_combo.png
results/mmcoa_fixed_prp_ac_plots.sampling_pca.png

# Concentrations and driving forces PCA
results/mmcoa_fixed_prp_ac_plots.concs.pca.pdf
results/mmcoa_fixed_prp_ac_plots.dfs.pca.pdf

# Concentration distributions
results/mmcoa_fixed_prp_ac_plots.sampling_concs_combo.pdf
results/mmcoa_fixed_prp_ac_plots.sampling_concs.pdf

# Driving force distributions
results/mmcoa_fixed_prp_ac_plots.sampling_dfs_combo.pdf
results/mmcoa_fixed_prp_ac_plots.sampling_dfs.pdf
Reduced dataset plots.
# Random walk PCA
results/mmcoa_fixed_prp_ac_plots_reduced.sampling_pca_combo.png
results/mmcoa_fixed_prp_ac_plots_reduced.sampling_pca.png

# Concentrations and driving forces PCA
results/mmcoa_fixed_prp_ac_plots_reduced.concs.pca.pdf
results/mmcoa_fixed_prp_ac_plots_reduced.dfs.pca.pdf

# Concentration distributions
results/mmcoa_fixed_prp_ac_plots_reduced.sampling_concs_combo.pdf
results/mmcoa_fixed_prp_ac_plots_reduced.sampling_concs.pdf

# Driving force distributions
results/mmcoa_fixed_prp_ac_plots_reduced.sampling_dfs_combo.pdf
results/mmcoa_fixed_prp_ac_plots_reduced.sampling_dfs.pdf

Syntrophic methanogenesis model

The SPOB was given company by an acetoclastic methanogen and a hydrogenotrophic methanogen to produce a multi-organism syntrophic model. To distinguish the different organisms' reactions, a prefix was introduced; po_ for the SPOB, and ac_ and hm_ for the acetoclastic and hydrogenotrophic methanogens. The three organisms affected eachother through the overlap in metabolites, i.e. hydrogen, formate, acetate, methane, and carbon dioxide. The model was parameterized using data from literature and Equilibrator.

Model component File
Reactions describing metabolic network data/syntrophic.model.tab
Stoichiometric matrix from stoich.py data/syntrophic.stoich.tab
Standard reaction Gibbs free energy changes data/syntrophic.model_drGs.tab
Default metabolite concentration ranges data/syntrophic.concentrations.tab
Individual concentration sums for each organism data/syntrophic.conc_sums.tab
Fixed concentration ratios for MDF data/syntrophic.ratios.tab
Concentration ratio ranges for sampling data/syntrophic.ratio_range.tab

Additionally, a file with a fixed high methane concentration was prepared to make contrast with free methane concentration thermodynamics:

data/syntrophic.concentrations.high_CH4.tab

The syntrophic model was used to sample thermodynamically feasible metabolite concentrations in free and high methane concentration conditions, as described in these Bash scripts:

source/syntrophic_2x10x10M.sh
source/syntrophic_2x10x10M.high_CH4.sh

The results were plotted using plot_samples.R and yielded the following output:

Methane comparison plots.
# Random walk PCA
results/syntrophic_CH4.sampling_pca_combo.png
results/syntrophic_CH4.sampling_pca.png

# Concentrations and driving forces PCA
results/syntrophic_CH4.concs.pca.pdf
results/syntrophic_CH4.dfs.pca.pdf

# Concentration distributions
results/syntrophic_CH4.sampling_concs_combo.pdf
results/syntrophic_CH4.sampling_concs.pdf

# Driving force distributions
results/syntrophic_CH4.sampling_dfs_combo.pdf
results/syntrophic_CH4.sampling_dfs.pdf

Author

Johannes Asplund-Samuelsson ([email protected])