Source-linked AI summary
Uncovering the hidden geometry behind metabolic networks
M. Angeles Serrano, Marian Boguna, Francesc Sagues
TL;DR
The paper addresses how metabolic network structure relates to function and how pathway interactions can be characterized. It embeds metabolic networks in a geometric model with degree-rescaled connection probabilities, finding close fit to E. coli and human networks and pathway backbones that reveal cross-talk.
Problem
The relationship between metabolic network structure and function, including how pathways interact beyond self-contained units, remains an open challenge.
Source: §II · p. 7Method
The authors embed bipartite metabolite–reaction networks on a circle and construct weighted pathway networks by summing connection probabilities across shared metabolites.
Source: A. Embedding algorithm and validation · p. 2Results
AUC = 0.96 for E. coli and AUC = 0.97 for human metabolism, and the inferred backbones show hub-like pathways connecting more localized pathways.
Source: A. Embedding algorithm and validation · p. 3, C. Cross-talk between pathways · p. 6Takeaways & Limitations
Pathways have disparate internal structures, and quantitative cross-talk reveals metabolic backbones that challenge viewing pathways as self-contained units.
Source: §II · p. 7
Abstract
from arXiv · showhide
Metabolism is a fascinating cell machinery underlying life and disease and genome-scale reconstructions provide us with a captivating view of its complexity. However, deciphering the relationship between metabolic structure and function remains a major challenge. In particular, turning observed structural regularities into organizing principles underlying systemic functions is a crucial task that can be significantly addressed after endowing complex network representations of metabolism with the notion of geometric distance. Here, we design a cartographic map of metabolic networks by embedding them into a simple geometry that provides a natural explanation for their observed network topology and that codifies node proximity as a measure of hidden structural similarities. We assume a simple and general connectivity law that gives more probability of interaction to metabolite/reaction pairs which are closer in the hidden space. Remarkably, we find an astonishing congruency between the architecture of E. coli and human cell metabolisms and the underlying geometry. In addition, the formalism unveils a backbone-like structure of connected biochemical pathways on the basis of a quantitative cross-talk. Pathways thus acquire a new perspective which challenges their classical view as self-contained functional units.
A. Embedding algorithm and validation
The model embeds metabolites and reactions on a circle and assigns connection probabilities that decrease with degree-rescaled geometric distance. It fits both E. coli and human metabolic networks closely and discriminates observed links from absent ones.
A. Embedding algorithm and validation
- The embedding represents metabolism as a bipartite network on a circle, with links only between metabolite and reaction nodes.The connection probability decreases with distance rescaled by the product of node degrees; higher-degree nodes can sustain farther-reaching links.Source: A. Embedding algorithm and validation · p. 2
- Metabolite coordinates are inferred first from a one-mode projection, then reaction coordinates are fitted by maximum likelihood using that embedding as a template.The projection links metabolites that participate in the same reaction.Source: A. Embedding algorithm and validation · p. 2
- The observed and predicted connection probabilities agree closely for both networks, with the same fitted β = 1.3.The calibration compares observed metabolite–reaction links with probabilities from the model.Source: A. Embedding algorithm and validation · p. 2, Figure 1 · p. 3
- AUC = 0.96 for E. coli and AUC = 0.97 for human metabolism, indicating strong discrimination between observed and absent links.The model’s ROC curves lie far above the random-guess diagonal.Source: A. Embedding algorithm and validation · p. 3, Figure 1 · p. 3
- The E. coli map shows unevenly occupied regions and empty gaps, suggesting hierarchical organization in the embedded network.Pathway names are placed at the average angular position of their constituent reactions.Source: A. Embedding algorithm and validation · p. 3
B. Pathway localization
Pathway reactions occupy diverse angular patterns in E. coli, from concentrated distributions to multi-peaked or broadly spread ones. Sector aggregation reveals functional groupings, while human pathways are more localized on average but more entangled across families.
B. Pathway localization
- E. coli pathways range from narrow angular concentrations to bimodal, multi-peaked, or nearly homogeneous distributions.Folate and oxidative phosphorylation are localized examples; histidine and glycolysis are multi-peaked examples.Source: B. Pathway localization · pp. 3–4
- Eight angular sectors aggregate reactions into functional groupings, including central metabolism, cell-membrane metabolism, nucleotide metabolism, and glycans.Murein-related pathways are separated in Sector 7, while Sector 6 combines mono- and polysaccharide pathways.Source: B. Pathway localization · p. 5, Figure 4 · p. 5
- Human pathways have average angular concentration 0.82 versus 0.79 in E. coli, while their average maximum peak size is 0.50 versus 0.36.The higher human localization coexists with greater entanglement among metabolic families.Source: B. Pathway localization · p. 5
- Some E. coli pathways that are transversal, such as Cofactor and Prosthetic Group or Transport, are split into more specialized pathways in human metabolism.The transversal-pathway category is minimally represented in the human metabolism analysis.Source: B. Pathway localization · pp. 5–6
C. Cross-talk between pathways
The paper builds weighted pathway networks by summing interaction probabilities for shared metabolites, then filters them to retain statistically relevant cross-talk. The resulting backbones show hub-like pathways connecting more localized pathways in both E. coli and human metabolism.
C. Cross-talk between pathways
- The backbones retain 91% of E. coli pathways with 9% of its links, and 69% of human pathways with 5% of its links.The corresponding retained total weights are 30% for E. coli and 20% for human metabolism.Source: Figure 5 · p. 6
- Pathway-pair cross-talk is quantified by summing connection probabilities for observed links between pathway reactions and their shared metabolites.The resulting weighted network is filtered with a disparity-based threshold to retain statistically relevant interactions.Source: C. Cross-talk between pathways · p. 6
- In E. coli and human metabolism, transversal pathways act as hubs that connect more localized pathways, often across metabolic families.Examples include Cofactor and Prosthetic Group Biosynthesis and Alternate Carbon in E. coli, and Folate, Fatty Acid Oxidation, and Keratan Sulfate Biosynthesis in human metabolism.Source: C. Cross-talk between pathways · p. 6, Figure 5 · p. 6
II. DISCUSSION
The discussion presents geometric maps as a way to relate metabolic network structure to system-level organization. Across E. coli and human metabolism, distance-based interaction probabilities recover pathway organization while revealing quantitative cross-talk and limits of treating pathways as self-contained units.
II. DISCUSSION
- Geometric cartography can help identify dominant nodes and relate network subparts, and may support prediction of missing or spurious interactions and optimal transport routes.These are proposed uses of geometrical cartographic maps, not reported evaluations in this discussion.Source: §II · pp. 6–7
- The proposed interaction law links metabolite–reaction connection probabilities to geometric distance and node properties, supporting a systems-level interpretation of chemical affinity.The authors describe the summed probabilities for a reaction as a network-based version of reaction affinity.Source: §II · p. 7
- The maps recover biochemical pathways as connected metabolic families while showing that pathways differ in internal structure and may not behave as self-contained units.Some pathways are more modular; others contain subunits with differentiated functionalities.Source: §II · p. 7
- Metabolic backbones represent higher-level interactions between pathways using quantitative cross-talk.The authors caution that pathway-specific protocols may underestimate connections that support the network's functioning.Source: §II · p. 7
- Human pathways are generally more modular and share fewer metabolites with other pathways than E. coli pathways, while human metabolic families are more entangled.The authors note that this entanglement may reflect greater functional complexity or reconstruction of reactions from differently differentiated cells in one network.Source: §II · p. 7
1. Hidden metric spaces and the S1 × S1 model.
The methods define a circle-based S1 × S1 model for metabolic networks and infer node positions from observed topology. They also describe how pathway backbones are filtered and how pathway angular position and concentration are measured.
1. Hidden metric spaces and the S1 × S1 model.
- The S1 × S1 model places metabolites and reactions in a circle, assigns expected degrees, and links metabolite–reaction pairs probabilistically using their geometric distance.The selected connection function generates maximally random networks given the model constraints.Source: §1 · p. 7
2. Inverse problem
- The inverse problem estimates node positions that maximize the model's likelihood of reproducing observed topology, using a two-phase embedding for the bipartite network.The method first embeds a one-mode projection, then uses inferred metabolite angles to position individual reactions.Source: §2 · p. 8
3. The disparity filter
- The disparity filter extracts a connected backbone by retaining edges with statistically significant intensity for at least one endpoint, accounting for local weight heterogeneity.It preserves more nodes and weights with fewer links than a global threshold filter.Source: §3 · p. 8
4. Average angular position and concentration of pathways.
- A pathway's mean angular position is the angle of its average reaction-position vector, and the vector's modulus measures angular concentration.A modulus of 1 indicates identical reaction angles; 0 indicates a perfectly homogeneous distribution around the circle.Source: §4 · p. 8
Appendix B: The S1 model and its extension to bipartite networks
The S1 framework places nodes in a metric space and assigns links probabilistically using distance and hidden degree. Its bipartite extension models metabolite–reaction connections, while metabolic-specific parameterization and finite-size corrections adapt it to reconstructed networks.
Appendix B: The S1 model and its extension to bipartite networks
- S1 places nodes on a circle with hidden expected degrees and connects pairs probabilistically according to effective distance, using a Fermi–Dirac connection probability whose β parameter controls clustering.The model’s hidden variable κ equals a node’s expected degree, so its degree distribution follows the chosen power law at large k.Source: Appendix B · pp. 8–9
1. The S1 × S1 model
- The S1 × S1 extension places metabolites and reactions on a circle and connects only cross-type pairs according to their angular distance and hidden variables.The framework permits any integrable connection function and allows mean hidden variables to match observed mean degrees.Source: §1 · p. 9
2. Specific model for metabolic networks
- Metabolic-network parameters are chosen to represent power-law metabolite degrees and Poisson-like reaction degrees, with ⟨κr⟩, ⟨κm⟩, β, and γ as free parameters.The metabolite degree exponent is approximately 2.6; β controls clustering through the chosen connection probability.Source: §2 · p. 10
3. Parameters estimation and finite size effects
- Finite-size networks require corrections because the thermodynamic-limit formulas omit zero-degree nodes and observed power-law samples are truncated at their maximum degree.The estimation procedure accounts for zero-degree nodes and finite-sample truncation when estimating model parameters.Source: §3 · p. 10
- The corrected parameter-estimation equations are solved numerically to recover hidden-degree means and zero-degree probabilities.The section also compares empirical and model complementary cumulative degree distributions for metabolites and reactions.Source: §3 · p. 11, Figure 6 · p. 11
4. Parameters of the real metabolisms
The authors fit the S1 × S1 model to reconstructed E. coli and human metabolic networks, then use likelihood-based methods to estimate embedding coordinates. The model matches metabolite degree distributions well but overestimates reactions involving five or more metabolites.
4. Parameters of the real metabolisms
- The analyzed reconstructions contain 1512 reactions in E. coli and 2201 in human metabolism, with estimated hidden-variable and network-size parameters for each.For E. coli, the estimates are ⟨κm⟩=4.06, ⟨κr⟩=2.65, Nm=1123, and Nr=1720; for human, they are ⟨κm⟩=4.22, ⟨κr⟩=2.73, Nm=1646, and Nr=2326.Source: §4 · p. 11
- The model’s metabolite degree distributions agree very well with both metabolic networks, but it overestimates reactions involving five or more metabolites.The comparison uses β = 1.3 for both networks.Source: §4 · p. 11
Appendix C: Embedding algorithm and validation
- The embedding algorithm estimates hidden variables and node angles by maximizing the likelihood of the observed adjacency matrix under the S1 × S1 model.The likelihood combines the model prior over hidden variables with the probability of observing the network given those variables.Source: Appendix C · p. 12
1. MLE for expected metabolites’ degrees κm
- The maximum-likelihood estimate for a metabolite’s expected degree equates its model-implied expected degree with its observed degree, subject to the stated lower-bound handling.The method first derives the optimum κm,l and then applies the minimum-degree constraint.Source: §1 · p. 12
MLE for angular coordinates θ
Direct angular optimization is unreliable because low-degree reactions have uncertain positions that propagate uncertainty to neighboring metabolites. The method therefore embeds a metabolite projection first, then locates reactions against the fixed metabolite coordinates; synthetic-network tests show close agreement with known coordinates and connection probabilities.
MLE for angular coordinates θ
- Low-degree reactions have uncertain angular positions, so direct optimization can spread localization uncertainty across the network.The method addresses this by embedding the metabolite-only projection, where high-degree nodes can be located more accurately and used as a template.Source: MLE for angular coordinates θ · pp. 12–13
- After fitting metabolite angles in the one-mode projection, the algorithm estimates each reaction’s angle independently with metabolite coordinates fixed.This two-stage procedure uses the S1 model for the metabolite projection and then maximizes the reaction-coordinate likelihood.Source: MLE for angular coordinates θ · p. 13
- On synthetic networks, inferred angles closely match their true values up to fluctuations and rotational symmetry, and inferred connection probabilities agree excellently with the generating probabilities.Figure 7 compares inferred and real angles on the left and empirical and theoretical connection probabilities on the right.Source: MLE for angular coordinates θ · p. 13, Figure 7 · p. 13
Appendix E: Pathways crosstalk and the disparity filter
Pathway cross-talk is quantified from observed connection probabilities involving metabolites shared between pathways, and the resulting weighted network is filtered into a backbone. The disparity filter reduces edges while retaining much of the network’s weight and node structure.
Pathway cross-talk
- Cross-talk sums connection probabilities for observed links involving metabolites shared by two pathways.The resulting pathway network uses pathways as nodes and weighted links to represent their interactions.Source: Appendix E · p. 13
- 82.00% of 561 possible E. coli pathway pairs (460) and 38.64% of 4,278 human pathway pairs (1,689) have nonzero cross-talk.Human metabolism also has one isolated pathway, Limonene and Pinene Biosynthesys, with no cross-talk.Source: Appendix E · p. 13
Disparity filter
- The multiscale backbone balances maximum weight against the number of included nodes, whereas a global threshold loses more nodes for the same links and backbone weight.Figure 8 compares the disparity backbone with the global-threshold backbone.Source: Appendix E · pp. 13–14, Figure 8 · p. 14
Pathway classification
- E. coli pathways are classified by angular concentration as localized, bimodal, multi-peaked, or transversal, with adjacent-bin cases potentially changing classification at another resolution.The classes use 75% concentration thresholds, with transversal pathways having no bin above 25%.Source: Table I · p. 14
Disparity filter
- The disparity filter significantly reduces edges while retaining almost all nodes and a large fraction of total weight, as well as the degree cutoff, weight distribution, and clustering coefficient.It tests whether each edge’s normalized weight is unusually intense under a null model of uniformly distributed local weights.Source: Appendix E · p. 14
Appendix F: Results for human cells metabolism
The human metabolism results include a map of reactions and metabolites and angular distributions for its pathways. These views encode node connectivity and the distribution of pathway reactions around the circle.
Human metabolism map
- Figure 9 maps human metabolism with reactions as yellow circles and metabolites as blue squares, radially placed by degree and sized by logarithmic degree.Connections are black or grey according to whether the model assigns them probability above or below 0.5.Source: Appendix F · p. 14, Figure 9 · p. 15
Pathway angular distributions
- Figure 10 bins the full angular domain into 50 bins of 7.2° and shows the fraction of each pathway’s reactions in each bin.Each pathway has a separate graph, with the pathway families distinguished by color.Source: Appendix F · p. 14, Figure 10 · p. 18
- Figure 10’s three panels distinguish pathway families by color, including amino-acid, nucleotide, carbohydrate, cofactor, glycan, and lipid metabolism.Pathway names are abbreviated in standard forms where possible.Source: Figure 10 · p. 18