EHMN 2026: A Thermodynamically Refined Human Metabolic Network
A rigorously harmonised, SBML-standardised genome-scale reconstruction of human metabolism — engineered for QSP integration, reproducible constraint-based analysis, and multi-layer systems pharmacology.
Goryanin et al. · 2026
Systems Biology
QSP
Abstract: What Is EHMN 2026?
EHMN 2026 is an updated Edinburgh Human Metabolic Network featuring systematic identifier reconciliation (MetaNetX, ChEBI), duplicate reaction consolidation, thermodynamic directionality assessment, and structured Reactome pathway annotation. Encoded in SBML Level 3 Version 2 with FBC2, it ensures explicit GPR representation and compatibility with modern constraint-based modelling toolchains.
Compared with Recon3D and Human1, EHMN 2026 uniquely combines native Reactome reaction-level annotation, systematic MetaNetX harmonisation, documented thermodynamic cycle elimination (37 cycles, 0 remaining), and an 11-compartment architecture for QSP and multi-layer integration.
22,642
Reactions
14,321
Metabolites
3,996
Gene Products
2,194
Reactome IDs
Introduction: Why a New Reconstruction?
Legacy Limitations
Many existing GEMs accumulate heterogeneous metabolite identifiers, incomplete thermodynamic refinement, and ambiguous reaction directionality — constraining reproducibility and interoperability.
The EHMN Lineage
The original Edinburgh Human Metabolic Network (2007) emphasised literature-supported reaction structure. EHMN 2026 modernises this framework with MetaNetX/ChEBI harmonisation, Reactome mapping, and SBML Level 3 standardisation.
Refinement Over Expansion
Rather than competing on reaction count alone, EHMN 2026 prioritises annotation consistency, explicit Ensembl gene linkage, and hierarchical Reactome integration — a reproducible metabolic backbone for broader modelling infrastructures.
Methods: Multi-Stage Reconstruction Workflow
The workflow began with the original EHMN framework and proceeded through systematic identifier harmonisation, reaction consolidation, thermodynamic directionality refinement, and final SBML Level 3 FBC2 encoding — emphasising refinement and standardisation rather than reaction inflation.
Identifier Harmonisation & Annotation Coverage
MetaNetX & ChEBI Reconciliation
Metabolite reconciliation used four stages: exact identifier matching, structure-supported matching (InChIKey/formula), stoichiometric consistency checks, and manual conflict resolution. A total of 612 ambiguous mappings (4.3%) required rule-based disambiguation. Harmonisation removed 430 redundant entries and consolidated 276 stoichiometric duplicates.
  • 11,542 metabolites (80.6%) mapped to MetaNetX MNXref
  • 7,682 metabolites (53.6%) carry validated ChEBI identifiers
  • 3,873 metabolites (27.0%) lack ChEBI annotation (lumped species, transport pseudo-metabolites)
80.6%
MetaNetX Coverage
53.6%
ChEBI Coverage
42.6%
GPR Coverage (all rxns)
62%
GPR (enzymatic core)
Reaction Deduplication & Structural Consolidation
Before Deduplication
23,284 total reactions — including 438 reverse-direction duplicate pairs and 206 exact stoichiometric duplicates.
After Consolidation
22,642 reactions — a 2.8% reduction eliminating artificial flux loops and improving thermodynamic consistency. Zero GPR rules were lost.
Thermodynamic Cycles
37 Type-III infeasible energy-generating cycles detected and fully resolved via iterative LP (HiGHS solver). 0 remaining unconstrained ATP-generating loops in the deposited SBML.
GPR Standardisation & Gene Architecture
Gene–Protein–Reaction Encoding
Gene associations were standardised to Ensembl gene identifiers (ENSG) and encoded using SBML Level 3 FBC2 geneProduct structures with Boolean GPR logic (AND = multi-subunit complexes; OR = isoenzymes).
  • 3,996 gene products encoded in FBC2
  • 9,638 reactions with explicit GPR associations
  • 2,887 unique Ensembl gene identifiers (ENSG)
Why 42.6% Overall GPR?
The headline figure covers all 22,642 SBML reactions — including 6,476 exchange/sink/demand boundary reactions (28.6%) for which no enzymatic gene exists by design, and 1,423 transport reactions. Restricting to the 12,969 MAR-curated enzymatic reactions yields 62.0% GPR coverage — applied to a core ~18% larger than Human1 or Recon3D.

EHMN 2026's 3,996 gene products exceed both Recon3D (3,288) and Human1 (3,628), confirming that annotation breadth and conservative philosophy are not in tension.
Reactome Pathway Mapping
Hierarchical Annotation
All Reactome pathway levels retained — enabling broad biological categorisation from "Metabolism" down to "Cholesterol biosynthesis" and cross-scale pathway aggregation.
Leaf-Only Layer
A secondary set using only terminal Reactome events (no child nodes) eliminates hierarchical duplication. 1,278 reactions linked to ≥1 leaf pathway; 642 unique leaf Reactome pathways represented.
Coverage Statistics
2,887 unique ENSG identifiers; 76.0% successfully mapped to Reactome. 2,194 unique Reactome pathway IDs spanning central carbon, lipid, amino acid, nucleotide, and mitochondrial bioenergetics pathways.
Biological Scope
Transport and exchange reactions excluded from pathway classification. Reactome coverage reflects database curation scope — not missing metabolic content in EHMN 2026.
Thermodynamic Directionality Assessment
Three-Phase Refinement
Phase 1: Biochemical irreversibility rules applied to 1,923 reactions across five classes — ATP-dependent ligases (ΔG′° ~−30.5 kJ/mol), decarboxylations, large-ΔE° NADH/NADPH reductions, OXPHOS/ETC (ΔΨ ~−180 mV), and fatty acid β-oxidation (~−69 kJ/mol per cycle). Cross-validated against KEGG, BRENDA, MetaCyc, and Human-GEM.
Phase 2: 37 infeasible cycles detected and resolved iteratively. Phase 3: Flux consistency analysis under closed-boundary conditions confirmed 0 remaining loops.
80%
OXPHOS/ETC Irreversible
59.1%
β-Oxidation Irreversible
54%
MAR Core Irreversible
43.2%
All Reactions Irreversible

A model with 43.2% irreversible reactions eliminates ~9,792 backward flux solutions, yielding more constrained, biologically realistic flux distributions — particularly in OXPHOS and fatty acid pathways.
SBML Encoding, Validation & Standards Compliance
01
SBML Level 3 Version 2 + FBC2
Explicit compartment definitions (11), 14,321 species, 22,642 reactions, 3,996 gene products, and reaction bounds consistent with constraint-based simulation. Zero embedded kinetic laws.
02
Structural Validation
Validated using validation.py (Supplementary S6): structural validity, correct FBC syntax, unique identifiers, no orphan references. Fully MEMOTE-consistent testing standards.
03
Platform Compatibility
Directly importable into COBRA Toolbox, COBRApy, Tellurium, CellDesigner, and all SBML-compliant modelling environments. MIRIAM-compliant annotation framework throughout.
Results: Structural Overview & Scale
Key Structural Statistics
EHMN 2026 contains +9,226 reactions (+68%) and +5,943 metabolites (+71%) compared to Human1, reflecting multi-compartment expansion to 11 compartments and chain-length-variant lipid representation. The gene product count (3,996) exceeds both comparators despite conservative GPR philosophy.
The additional reactions comprise: 6,476 exchange/sink/demand boundary reactions, 1,423 transport isoforms, and 855 fatty acid/sphingolipid chain-length variant reactions (C6–C26). None represent enzymatic knowledge absent from comparators — they reflect deliberate architectural choices.
Comparison with Human1 & Recon3D
GPR coverage difference reflects reconstruction philosophy: EHMN 2026 retains a larger boundary reaction set (28.6% of total) necessary for stoichiometric solvability. Restricting to the enzymatic MAR core yields 62% — like-for-like with comparators, applied to a larger enzymatic set.
GPR Coverage: Reconstruction Philosophy
The 13,004 reactions lacking GPR fall into five structurally distinct categories. Categories A and D will not gain GPR because none exists biologically. Category C is addressable in the next release. The 62% enzymatic-core figure is the appropriate metric for comparing annotation quality across models.
Reactome-Centred Functional Organisation
61% MAR Enzymatic Core Coverage
EHMN 2026 integrates Reactome pathway identifiers for 7,910 reactions (34.9% of all reactions), spanning 2,194 unique Reactome pathway/event IDs. Restricting to the 12,969 MAR-curated metabolic reactions yields 61% Reactome coverage — the highest among current human GEMs (Human1: ~31%; Recon3D: ~15%).
A further 159 gene-associated metabolic reactions lack Reactome entries because they describe confirmed human biochemical activities outside Reactome's current curation scope — flagged as candidates for future Reactome submission.
Pathway-Aware Encoding Enables
Subsystem-level flux interrogation and pathway-centric perturbation reporting
Modular pruning for disease-context modelling
Integration with signalling and regulatory knowledgebases (TRANSFAC, TRANSPATH)
Direct pathway-level flux summation without external mapping tables
Downstream Analysis: What EHMN 2026 Supports
Standard FBA
All 22,642 reactions. GPR associations not required. Theoretical aerobic ATP yield: 32.0 mmol/mmol glucose — identical to Recon3D, within accepted range of 31.45–32.
Gene Knockout & Essentiality
9,638 GPR-annotated reactions; 3,996 gene products screened. Conservative philosophy avoids false-positive essentiality calls from speculative homology transfers.
Transcriptomics & Proteomics
HGNC/ENSG namespace maps directly to RNA-seq and proteomics datasets. Supports FASTCORE, tINIT, iMAT extraction. Uniform MetaNetX namespace reduces ID-translation overhead.
Pathway-Level Flux Aggregation
Only model with native Reactome event IDs per reaction — enabling direct hierarchical flux summation (e.g., all "Cholesterol biosynthesis" reactions) without external mapping.
Organelle-Specific Modelling
11 compartments including inner mitochondrial space (20 species), lysosome (640 species), peroxisome (844 species) — enabling Complex I inhibition, lysosomal storage, and peroxisomal defect modelling.
QSP & Multi-Omics
Uniform MetaNetX namespace eliminates ID conflicts across model layers. Reactome IDs bridge to TRANSFAC and TRANSPATH. Thermodynamic consistency prevents infeasible flux propagation into pharmacological modules.
Stoichiometric & Chemical Consistency Validation
Four-Step Validation Workflow
1
Metabolite Annotation Harmonisation
Molecular formulas and charges assigned from ChEBI/KEGG at physiological protonation state (pH 7.3–7.4).
2
Core Metabolite Pool Completion
Standardised phosphate species, redox cofactors (NAD/NADH, NADP/NADPH), nucleotides, CoA derivatives, folate intermediates, CO₂/bicarbonate, and ammonia/ammonium.
3
Automated Reaction-Level Validation
Elemental composition and net charge compared for all substrates and products; minimal corrections applied preserving known biochemical mechanisms.
4
Manual Review of Residuals
Remaining imbalances annotated; retained only with curated biochemical evidence. All chemically verifiable reactions confirmed mass- and charge-balanced.
Functional Validation Results
Following SBML repair and flux-bound standardisation, the final model exhibits:
  • 100% of reactions with valid flux bounds
  • No malformed GPR associations
  • Complete stoichiometric connectivity
  • Thermodynamic directionality consistency
  • Zero SBML compliance errors
  • Theoretical ATP yield: 32.0 mmol/mmol glucose (within accepted 31.45–32 range)

37 thermodynamic cycles resolved; 9,792 reactions set irreversible; 227 blocked. Zero infeasible loops remain in the deposited model.
Architectural Feature Comparison
EHMN 2026 occupies a distinct niche: it prioritises identifier harmonisation, thermodynamic coherence, pathway-layer annotation, and SBML-native interoperability — properties specifically consequential for QSP and multi-layer modelling that are not simultaneously present in Recon3D or Human1.
Discussion: From Network Expansion to Structural Integrability
Design Philosophy
As models increase in size, structural coherence, annotation density, and interoperability become more decisive than reaction count alone. EHMN 2026 was designed to address this transition — prioritising harmonisation, pathway-level integration, and thermodynamic consistency within a standards-compliant SBML Level 3 + FBC framework.
Thermodynamic Coherence
Infeasible energy-generating cycles can distort predictions when models are dynamically coupled to regulatory or pharmacological modules. Structured directionality refinement reduces such artefacts — especially important for QSP contexts where metabolic states influence immune cell proliferation, cytokine secretion, or drug-target engagement. Full ΔG′° parameterisation is planned once formula coverage extends to ~3,313 remaining metabolites and CYTOCON DB concentration data are incorporated.
Comparative Context
Recon3D and Human1, with near-complete GPR coverage and tightly curated enzymatic sets, are better suited to high-throughput gene essentiality screening. EHMN 2026 is designed for complementary tasks where metabolite granularity, identifier coherence, and pathway-level interpretability matter more than GPR density.
Three Intentional Distinctions
  • Metabolite granularity: 11 compartments, 855 chain-length-specific FA/sphingolipid reactions (C6–C26)
  • Annotation architecture: Uniform MetaNetX namespace + Reactome event IDs + HGNC/ENSG — all three simultaneously in SBML
  • Thermodynamic completeness: 1,923 reactions re-constrained, 37 cycles eliminated, 0 remaining
EHMN 2026 as a QSP Integration Platform
CYTOCON / CYTOCON DB
Curated in vivo baseline concentrations of human immune cells and mediators — enabling calibration of metabolic flux states influencing immune activation dynamics and cytokine signalling.
SABIO-RK & BRENDA
Selective kinetic augmentation at key metabolic control points, drug targets, or disease-relevant pathways — preserving stoichiometric scalability while introducing dynamic behaviour where biologically necessary.
TRANSFAC & TRANSPATH
Regulatory databases providing transcription factor control and signal transduction cascades — enabling signalling → transcription → metabolism coupling for drug-induced or inflammation-driven metabolic rewiring.
Toward AI-Assisted Continuous Model Evolution
The Vision
Because EHMN 2026 is fully standards-compliant and annotation-dense, it is structurally amenable to AI-assisted updating workflows. Semi-automated pipelines leveraging literature extraction, identifier harmonisation, and regression validation can maintain model relevance as biomedical knowledge expands.
This evolutionary architecture shifts genome-scale reconstruction from a static artefact toward a continuously curated modelling infrastructure.
Limitations & Future Work
  • Primarily stoichiometric — kinetic embedding not globally implemented
  • Semi-quantitative thermodynamic heuristics do not yet capture concentration-dependent directionality shifts of near-equilibrium reactions
  • Full ΔG′° parameterisation planned once formula coverage extended to ~3,313 remaining metabolites
  • Systematic benchmarking across multi-cohort metabolomics and clinical QSP case studies remains a future step
  • Category C legacy KEGG-lineage reactions (4,730) targeted for Reactome mapping in next release cycle
Conclusions
EHMN 2026 advances human metabolic reconstruction beyond reaction expansion toward structurally engineered integrability — combining thermodynamic coherence, identifier harmonisation, pathway hierarchy integration, and SBML Level 3 FBC compliance.
Rigorous Harmonisation
80.6% MetaNetX coverage, 53.6% ChEBI validation, uniform namespace across 14,321 species enabling clean cross-layer integration.
Thermodynamic Integrity
37 infeasible cycles eliminated, 43.2% irreversible reactions, 0 remaining ATP-generating loops — stricter than any published comparator.
Pathway-Resolved Architecture
2,194 Reactome IDs natively embedded in SBML; 61% MAR enzymatic core coverage — highest among current human GEMs.
QSP-Ready Foundation
Designed as a scalable metabolic backbone for future multi-layer systems pharmacology, regulatory integration, and AI-assisted continuous model evolution.
The model will be publicly deposited in BioModels (EMBL-EBI) and at www.iqanova.org. Available on request: [email protected]