MetaboGraph: A Framework for Metabolomics and Lipidomics Annotation and Pathway Network Analysis

Abstract

Untargeted metabolomics and lipidomics generate high-dimensional data sets whose biological interpretation remains challenging, particularly at the pathway and network levels. Here, we present MetaboGraph, a standalone Python-based workflow for end-to-end metabolomics and lipidomics analysis, enabling pathway-level interpretation from small-molecule data. MetaboGraph integrates automated data cleaning, comprehensive multidatabase metabolite and lipid annotation, pathway mapping, and direction-aware pathway inference. A central feature of the platform is its ability to predict pathway direction by integrating metabolite/lipid-level fold changes with pathway membership structure, supporting biologically interpretable pathway and network analyses beyond conventional enrichment approaches. MetaboGraph supports multiomics integration and comparative analysis, enabling consistent pathway-level interpretation across metabolomics, lipidomics, and multiple studies. We demonstrate the platform using untargeted LC-MS/MS metabolomics and lipidomics data comparing two breast cancer cell lines with distinct metastatic potential, MCF7 (HTB22; less metastatic) and MDA-MB-453 (HTB131; more metastatic). Relative to HTB22, the HTB131 cells exhibited coordinated metabolic remodeling, including altered amino acid and nitrogen metabolism, increased nucleotide biosynthetic demand, lipid remodeling, and changes in energy-associated pathways. These pathway-level alterations are consistent with established metabolic adaptations associated with increased cancer aggression. MetaboGraph expands the analytical toolbox for small-molecule biology and facilitates reproducible, biologically grounded insights from metabolomics and lipidomics data sets.

Introduction

Untargeted metabolomics employing liquid chromatography–mass spectrometry (LC-MS) has emerged as a powerful approach for comprehensive profiling of endogenous small-molecule metabolites in biological systems. − Unlike targeted metabolomics, which measures a predefined set of compounds, untargeted metabolomics interrogates thousands of features across diverse chemical classes, providing an unbiased snapshot of metabolic biochemistry. This broad analytical scope enables discovery of metabolites linked to disease pathogenesis, cellular adaptation, and phenotypic differences across biological conditions. , Complementary to metabolomics, lipidomics, the systematic analysis of cellular lipids, captures the complete landscape of lipid species and classes that structure membranes, serve as signaling molecules, and integrate energy metabolism. Together, metabolomics and lipidomics constitute the metabolome, encompassing the complete set of metabolic intermediates and end products that directly reflect phenotypic biochemistry across physiological and pathological states.

The primary challenge in translating untargeted metabolomics and lipidomics data into biological insights does not lie in feature detection, but in transforming detected features into biochemically and biologically meaningful annotations. Once raw analytical data are converted into a feature-by-sample matrix, a standard output from processing platforms such as Thermo Compound Discoverer and LipidSearch, researchers must integrate these features with curated biochemical knowledge to extract systemic understanding. This integration step, which maps individual metabolites to canonical pathways and reconstructs system-level metabolic organization, has become essential for connecting metabolic changes to disease mechanisms and identifying therapeutic targets.

The past two decades have witnessed the development of several comprehensive, publicly accessible knowledge bases that catalog metabolites, their biochemical transformations, and their participation in metabolic pathways. The Human Metabolome Database (HMDB), first released in 2007, currently contains over 200,000 metabolite entries with associated protein linkages, disease associations, and biofluid distributions. , Other databases such as KEGG (Kyoto Encyclopedia of Genes and Genomes) integrate pathway information across organisms, linking metabolites to reactions, enzymes, and higher-order metabolic processes through directed network representations. For lipidomic analyses, LIPID MAPS serves as a comprehensive repository of lipid structures, classifications, and biological roles, enabling annotation of diverse lipid species and their biochemical pathways. Additional specialized resources include PathBank, which integrates metabolic, signaling, and disease pathways, Reactome, which focuses on human molecular reactions and pathways, WikiPathways, a community-curated biological pathway resource, and SMPDB (Small Molecule Pathway Database), each providing complementary perspectives on metabolic organization and regulation.

Despite the abundance of these resources, translating annotated metabolites into actionable biological interpretation remains challenging. A critical limitation is that existing pathway databases, while comprehensive, often contain incomplete metabolite-pathway associations due to evolving biochemical knowledge and variable curation standards. Furthermore, lipid annotations, particularly those generated from untargeted lipidomics, frequently lack fully resolved structural identifiers in public databases, limiting their integration into traditional pathway frameworks.

To bridge the gap between individual metabolite annotations and systems-level biological understanding, researchers employ pathway enrichment analysis, a computational approach that maps metabolite changes onto known biochemical pathways and identifies those enriched with statistically significant metabolites. , Classical enrichment analysis employs overrepresentation testing using Fisher’s exact test, which evaluates the statistical probability that the overlap between a set of altered metabolites and a pathway is due to chance rather than genuine biological association. − The strength of Fisher’s test lies in its ability to account for both pathway size and the proportion of metabolites within a pathway that are observed to change, reducing bias toward large pathway gene sets. Following initial enrichment identification, pathway directionality scoring, which predicts whether a pathway is activated or inhibited based on the magnitude and direction of metabolite fold changes, enhances biological interpretability by indicating the likely regulatory state of affected processes.

Beyond enrichment analysis, network-based approaches to metabolic interpretation have proven effective for visualizing and analyzing multiscale relationships among metabolites, pathways, enzymes, and disease associations. − Metabolic network visualization, traditionally conducted through tools such as Cytoscape, an open-source platform for complex network analysis and visualization, enables exploration of metabolite-pathway connectivity, identification of functionally coherent metabolic modules, and integration of external data sets, for example, fold changes and biomarker scores as node properties. , Network topology analysis, quantifying properties such as degree centrality, betweenness centrality, and community structure, allows identification of hub metabolites or pathways that occupy central positions in the metabolic network and may serve as regulatory checkpoints. ,

The growing recognition that metabolic reprogramming represents a fundamental feature of disease pathophysiology has driven the development of integrated bioinformatics workflows that link raw LC-MS data to pathway- and network-level biological interpretation. Across diverse disease contexts, including cancer, neurodegenerative disorders, metabolic diseases, and traumatic injury, alterations in amino acid metabolism, lipid remodeling, and energy and nucleotide pathways have been associated with disease progression and phenotypic severity. − For example, in breast cancer, metabolic rewiring involving amino acid metabolism, particularly glutamine, serine, and glycine, lipid biosynthesis, and nucleotide metabolism has been associated with aggressive phenotypes and metastatic potential. Effective tools for mapping these metabolic changes onto curated pathway resources enable rapid hypothesis generation and mechanistic insight into tumor-specific metabolic dependencies.

Despite the availability of comprehensive pathway databases and diverse analytical approaches, current bioinformatics workflows typically require multiple software platforms and manual data integration steps, introducing friction and potential error propagation. A unified platform that integrates metabolite annotation, pathway mapping, enrichment analysis, and network visualization, while supporting flexible configuration of pathway universes, statistical thresholds, and visualization parameters, would substantially reduce analytical burden and enable more reproducible, publication-ready metabolomics studies.

MetaboGraph addresses these gaps by providing a standalone, modular Python-based framework designed specifically for untargeted metabolomics and lipidomics data annotation, pathway enrichment, and network analysis. The software integrates major public biochemical resources, including HMDB, KEGG, LIPID MAPS, PubChem, PathBank, WikiPathways, Reactome, and SMPDB, into locally optimized databases, while KEGG and Reactome are accessed through their respective application programming interfaces (APIs), enabling rapid, reproducible querying. The platform supports configurable pathway universe definition, multiple enrichment and directionality scoring strategies, including Fisher’s exact test and a MetaboGraph-defined Impact-Weighted Pathway Analysis (IWPA), and flexible visualization of metabolite-pathway-enzyme-disease interaction networks through both interactive and static rendering engines.

By consolidating annotation, enrichment, and network analysis into a single analytical environment with comprehensive logging and parameter tracking, MetaboGraph streamlines the interpretation of large-scale metabolomics and lipidomics data sets while maintaining full transparency and reproducibility.

Materials and Methods

Materials

HTB131 and HTB22 cancer cell lines, fetal bovine serum (FBS) were purchased from the American Type Culture Collection (ATCC), Manassas, VA. Eagle’s minimum essential medium (EMEM), Roswell Park Memorial Institute 1640 medium (RPMI), 1X media, trypsin-EDTA, penicillin-streptomycin solution, and phosphate-buffered saline 1X (PBS) were purchased from Corning, Manassas, VA. Sodium deoxycholate (SDC), ammonium bicarbonate (ABC), were purchased from Sigma-Aldrich Co. St. Louis, USA. High-performance liquid chromatography (HPLC)-grade water and HPLC-grade methanol, HPLC grade acetonitrile (ACN), HPLC grade dichloromethane, HPLC grade isopropanol, and HPLC grade formic acid (FA) were obtained from Thermo Fisher Scientific Meridian, Rockford, USA.

Cell Culture

HTB131 was maintained in RPMI-1640 media, and HTB22 was maintained in EMEM. All media were prepared with 10% FBS and 2% penicillin-streptomycin solution. Each cell line was cultured at 37 °C, 5% CO2, and a stable humidity level in their respective media. Cells were harvested at 90% confluency and washed in PBS three times.

Metabolite and Lipid Extraction

Cell pellets were transferred to screw-cap tubes containing beads and resuspended in 100 μL of 50 mM ABC buffer, and 100 μL of 5% SDC. Cells were lysed by bead-based homogenization at 4 °C, followed by sonication on ice for 60 min to ensure complete disruption. Lysates were centrifuged at 14,800 × g for 10 min, and the clarified supernatants were transferred to fresh tubes. The samples were dried under vacuum, resuspended in ABC buffer and aliquoted into 5 tubes for extraction.

Metabolites and lipids were extracted using a modified biphasic solvent protocol adapted from previously reported methods. Briefly, 200 μL of dichloromethane/methanol (1:2, v/v) was added to each 100 μL sample, vortexed for 30 s, and incubated at room temperature for 60 min. Subsequently, 75 μL of dichloromethane was added and vortexed, followed by the addition of 75 μL of ice-cold water and further vortexed. Phase separation was achieved by centrifugation at 5000 rpm for 15 min. The upper aqueous phase containing polar metabolites and the lower organic phase containing lipids were carefully collected into separate tubes. Extracts were dried under vacuum and reconstituted in methanol:water (1:1, v/v) for metabolomics analysis and a solution of acetonitrile (ACN), isopropanol and water (40:30:30, v/v) for lipidomics analysis prior to LC-MS/MS acquisition.

Liquid Chromatography (LC) and Mass Spectrometry Conditions

Chromatographic separation was carried out using an Acquity UPLC HSS T3 column (100 Å, 2.1 × 100 mm; Waters, Ireland) and BEH C8 column on a Vanquish UHPLC system (Thermo Scientific, San Jose, CA, USA) coupled to a Q Exactive HF Orbitrap mass spectrometer for metabolites and lipids separations, respectively. Mobile phase A for metabolomics runs consists of water containing 0.1% formic acid, while mobile phase B comprises methanol with 0.1% formic acid. The gradient began at 0.5% mobile phase B, increased linearly to 50% over 5.5 min, ramped to 98% B within 0.5 min, and held for 6 min. The column was then re-equilibrated at 0.5% B for 2 min. Also, mobile phase A for lipidomics runs consists of ACN and water (6:4 v/v/) containing 10 mM ammonium formate and 0.1% formic acid, while mobile phase B comprised of isopropanol, ACN and water (88:10:2 v/v/) containing 10 mM ammonium formate and 0.1% formic acid. The gradient started at 35% mobile phase B, increased linearly to 60% at 4 min over 8 min, increased to 85% for 9 min, ramped to 100% B for 3 min. Thereafter, column was re-equilibrated at 35% B for 4 min.

Mass spectrometric (MS) analysis was performed using an electrospray ionization (ESI) source operating in both positive and negative ionization modes, for both omics runs. In positive ionization mode, the spray voltage was set to 3.5 kV with a transfer tube temperature of 300 °C, whereas negative ionization mode employed a spray voltage of 2.8 kV and a transfer tube temperature of 262.5 °C. Full-scan MS metabolomics data were acquired using the Orbitrap analyzer over an m/z range of 75–750 at a resolution of 120,000. The automatic gain control (AGC) target was set to 3.0 × 10⁶ with a maximum injection time of 200 ms (positive ionization mode) and AGC target of 1.0 × 10⁶ with a maximum injection time of 100 ms (negative ionization mode). For lipidomics full-scan MS data were acquired over an m/z range of 250–1200 at a resolution of 120,000, AGC target of 1.0 × 10⁶ with a maximum injection time of 100 ms for both modes. An exclusion list comprising the 100 most intense background ions detected in blank runs from 0 to 6 min and 0 to 8 min were incorporated into the acquisition method, for metabolomics and lipidomics runs, respectively.

Data-dependent MS/MS acquisition was conducted using a 3 s duty cycle, selecting the five (metabolomics) and 20 (lipidomics) most intense precursor ions for higher-energy collisional dissociation (HCD). Fragmentation was performed using stepped normalized collision energies (metabolomics: 20%, 40%, and 60%; lipidomics: 25%, and 30%) in both ionization modes. Precursor isolation was achieved using a quadrupole with an isolation window of 4 m/z and a dynamic exclusion duration of 10 s in positive ionization mode, while negative ionization mode employed a 1 m/z isolation window with a 6 s dynamic exclusion. MS/MS spectra were acquired at resolutions of 30,000 (positive ionization mode) and 15,000 (negative ionization mode), with a maximum injection time of 50 ms (positive ionization mode) 40 ms (negative ionization mode). For lipidomics on the other hand, precursor isolation was achieved with an isolation window of 1 m/z and a dynamic exclusion duration of 8 s in both modes. MS/MS spectra were acquired at resolutions of 30,000 with a maximum injection time of 75 ms.

Data Analysis

Raw LC-MS/MS data from metabolomics and lipidomics analysis were processed with Compound Discoverer 3.4 and LipidSearch 5.1 respectively for feature detection, alignment, and quantification. Initial metabolite and lipid feature tables were exported for downstream analysis. All subsequent data cleaning, annotation, pathway enrichment, and network-based interpretation were performed using MetaboGraph, a standalone Python-based workflow developed for integrated metabolomics and lipidomics analysis. Statistical and exploratory analyses, including Principal Component Analysis (PCA), volcano plots, Venn diagrams, heatmaps, and boxplots, were performed using OmicsVisStat (v1.0), a Python (version 3.13)-based platform. The underlying scripts and analysis workflows are publicly available via Zenodo.

Within MetaboGraph, feature tables were subjected to quality-control filtering, including removal of poorly detected features and retention of compounds quantified in a minimum proportion of samples per group. Metabolite and lipid features were systematically annotated and harmonized to database-compatible identifiers using curated references from HMDB, KEGG, PubChem, and LipidMaps. This identifier harmonization enabled consistent pathway mapping across multiple pathway resources, including HMDB, KEGG, PathBank, WikiPathways, Reactome, and SMPDB, and facilitated interoperability with external pathway analysis platforms such as QIAGEN Ingenuity Pathway Analysis (IPA). Pathway enrichment analysis was performed using Fisher’s over-representation analysis (ORA) implemented in MetaboGraph, with Homo sapiens specified as the reference organism. Direction-aware pathway inference was achieved by integrating metabolite- and lipid-level fold changes with pathway membership to estimate relative pathway activation or inhibition. For benchmarking and contextual interpretation, curated metabolite identifiers results generated by MetaboGraph were exported for pathway analysis using QIAGEN IPA and MetaboAnalyst 6.0. Overlapping pathways across platforms were examined to assess concordance and highlight complementary strengths. All analyses were performed using reproducible, scriptable workflows within MetaboGraph. The software MetaboGraph (v1.0.1), together with its documentation and example workflows, is publicly available via GitHub (https://github.com/odaramola92/MetaboGraph) and archived on Zenodo with a persistent DOI to ensure transparency and reproducibility.

MetaboGraph Implementation

MetaboGraph is a standalone Python framework developed for metabolomics and lipidomics data processing, statistical analysis, pathway enrichment, and network-based biological interpretation. The software follows a modular workflow that integrates data input, preprocessing, annotation, enrichment, and network construction to enable end-to-end analysis from processed feature tables generated by untargeted small-molecule workflows. MetaboGraph incorporates an integrated annotation engine that indexes major biochemical databases, including HMDB, LIPID MAPS, PubChem, KEGG, PathBank, WikiPathways, Reactome, SMPDB, and Metabolika. Metabolites and lipids are systematically queried against these resources to retrieve standardized database-compatible identifiers, which are then used for pathway mapping. Pathway enrichment is performed using Fisher’s ORA, with the enrichment universe defined as the full set of annotated metabolites and lipids present in the uploaded data set. Pathway outputs include metabolite coverage, enrichment statistics, and direction-aware Z-scores derived from aggregate feature-level behavior rather than uniform pathway regulation.

Network analysis is supported through compound-pathway bipartite graphs and pathway-projection networks, enabling assessment of pathway interconnectivity and upstream regulatory context. Network topology metrics, including degree, betweenness centrality, and PageRank, are computed to identify highly connected pathways and regulatory hubs. All analytical steps record parameter settings, database versions, and intermediate results to support reproducibility and transparent reporting. The complete analytical pipeline is summarized schematically in Figure .

1.

Overview of the MetaboGraph workflow for metabolomics and lipidomics data analysis.

Results and Discussion

Database Setup

MetaboGraph includes an integrated Database Setup interface that allows users to prepare all required biochemical resources prior to annotation and pathway analysis. The platform natively supports five major local databases: HMDB, LipidMaps, PathBank, SMPDB, and WikiPathways which are uploaded by the user in their original XML, SDF, CSV, or GPML formats. Through a unified graphical interface, these resources are automatically parsed, standardized, and compiled into optimized Feather tables, ensuring consistent identifiers and rapid downstream querying. Database availability is verified within the interface, and a species-specific background universe for Fisher’s exact pathway enrichment is generated to support statistically valid interpretation.

In addition to locally processed databases, MetaboGraph also integrates external biochemical resources accessed directly through their public APIs, including PubChem, KEGG, and Reactome. These online databases require no user setup and are queried dynamically during annotation and pathway retrieval, enabling complementary coverage beyond the locally compiled repositories. A full summary of libraries used, and their roles, is provided in Supplementary Table S1. A complete description of the database processing workflow, extraction logic, required files, supported formats, and integration with external APIs is provided in the Supporting Information.

Data Cleaning

MetaboGraph includes a flexible data cleaning workflow designed to standardize LC-MS metabolomics and lipidomics data sets prior to annotation and statistical analysis. The module supports input data generated from Compound Discoverer 3.4, LipidSearch 5.1, or any user-defined data sets that follow the general output of feature search such as provided in Supplementary Table S2. Because exported data sets frequently differ in column naming and formatting across instruments and laboratories, MetaboGraph incorporates a verification step that allows users to confirm the assignment of essential fields such as feature identifiers and sample abundance or peak-area columns.

Once column assignments are validated, the cleaning workflow performs a series of harmonization steps that include removal of empty or redundant fields, and standardization of textual identifiers. Features lacking measurable signal across all samples are removed to eliminate nondetected species and reduce statistical noise. For data sets in which multiple adducts or ion forms represent the same chemical entity, polarity-specific prioritization rules ensure that only the most representative ion species is retained for analysis. The resulting cleaned data set provides a consistent and analysis-ready foundation for subsequent annotation, statistics, enrichment, and network modules within MetaboGraph. A complete description of the data-cleaning algorithm, quality-control steps, and ion-selection criteria is provided in the Supporting Information.

ID Annotation

MetaboGraph performs metabolite identification using a structured name-guided workflow that integrates PubChem, KEGG, and HMDB. The annotation process begins with a priority search in PubChem, followed by KEGG retrieval when necessary, and a final offline HMDB cross-reference. Retrieved information such as PubChem CID, KEGG compound ID, HMDB ID, InChIKey, SMILES, and ChEBI identifiers are consolidated into a unified annotation record, with user-defined control over the preferred identifier hierarchy and handling of unmatched features.

As shown in Supplementary Figure S1, MetaboGraph provides an interactive metabolite ID annotation interface that supports both initial compound identification and post hoc refiltering without requiring re-execution of the full annotation pipeline. Users may import cleaned metabolomics data or previously annotated results and dynamically adjust identifier filters to refine annotation confidence. Multiple identifier sources (HMDB, PubChem, KEGG, ChEBI, CAS, SMILES, and InChIKey) can be selectively enabled or disabled, allowing flexible control over annotation stringency depending on study requirements. Real-time processing logs and execution controls provide transparent tracking of database queries and annotation progress, supporting reproducibility and scalability across large data sets. This modular design enables rapid iteration during exploratory analysis while preserving deterministic annotation logic prior to downstream pathway and network interpretation.

Lipid identification follows similar workflow using name- and class-guided annotation strategy. While lipid annotation can be performed on cleaned feature tables alone, the use of lipid data sets that include statistical results is recommended to restrict annotation to biologically relevant features. This targeted approach reduces unnecessary database queries, improves computational efficiency. A detailed description of the identifier hierarchy, cross-referencing rules, and lipid-matching logic is provided in the Supporting Information.

Pathway Annotation

Pathway annotation in MetaboGraph integrates metabolite- and lipid-level identifiers with multiple pathway knowledge bases to produce a unified representation of biochemical context. After ID assignment, each feature is mapped to curated pathway sets from HMDB, PathBank, SMPDB, WikiPathways, KEGG, and Reactome using a consolidated annotation engine. Pathway names undergo extensive normalization, duplicate merging, and exclusion of nonmammalian or nonspecific entries to ensure consistency across databases. Disease-associated pathways are distinguished from physiological metabolic pathways through rule-based suffix detection and keyword classification, allowing disease terms to be retained separately for downstream biological interpretation.

Mappings may occur at the species level (for metabolites) or at the class/species level (for lipids), and all assignments are consolidated into a harmonized pathway list per feature. Mapping completeness depends on the availability of corresponding entries in the underlying pathway databases, and the identifiers present in the input data set. MetaboGraph facilitates this process by harmonizing identifiers across multiple resources and reporting feature-to-pathway assignments, but it does not infer pathway membership when such information is not available in the referenced databases. The annotation engine automatically merges multiple pathway sources and resolves variations in naming and database formatting to produce a robust and reproducible annotation layer. The full logic, including normalization rules, filtering patterns, and disease-pathway classification procedures, is provided in the Supporting Information.

The overall pathway and network analysis workflow is summarized in Supplementary Figure S2, which illustrates the modular progression from pathway annotation to enrichment, prioritization, and network-based interpretation within the MetaboGraph interface. The figure highlights how annotated metabolites and lipids transition through pathway enrichment, ranking, and network construction, and how users navigate between pathway-level summaries, contributing compounds, regulatory context, and integrated network views within a unified analytical environment.

Pathway Ranking and Prioritization Strategy

Conventional pathway enrichment ranks pathways primarily by statistical significance or feature count, which may bias interpretation toward large or highly annotated pathways. To mitigate this limitation, MetaboGraph integrates network topology metrics that quantify pathway centrality and connectivity within the metabolite-pathway graph, providing complementary analytical evidence for pathway prioritization. To prioritize biologically meaningful pathways beyond statistical significance alone, MetaboGraph implements a multilayered ranking framework that integrates enrichment statistics, directional effect size, network topology, and multiomics support. This approach addresses known limitations of conventional ORA, which relies primarily on p-values and does not account for pathway activity direction, coordinated metabolite behavior, or network context.

Pathway “activation” or “inhibition” as designated in MetaboGraph refers to a directional bias in metabolite or lipid level abundance changes within a pathway and does not imply direct measurement of metabolic flux, enzymatic activity, or regulatory control. Direction-aware Z-scores summarize the aggregate sign distribution of fold changes among pathway-associated features and should be interpreted as indicators of coordinated pathway behavior rather than definitive mechanistic activation.

Primary Pathway Selection

Pathways are first filtered using a hierarchical selection strategy designed to exclude noninformative annotations while retaining biologically interpretable metabolic processes. Only pathways passing Fisher’s exact test (default p < 0.05, with optional BH correction) are considered. The ORA universe is defined as all features with valid p-values, while the significant set includes metabolites meeting both statistical significance and fold-change criteria.

Pathways predicted as activated or inhibited based on direction-aware Z-scores are prioritized over pathways with neutral status, ensuring that biologically active pathways are never displaced by statistically significant but directionally ambiguous results. When a fixed number of pathways is selected (e.g., top 30), activated or inhibited pathways are always selected first, with neutral pathways included only if additional slots remain.

Default Mode (recommended) ranks pathways by integrating activation status, multiomics support, number of contributing metabolites, absolute Z-score magnitude, consensus rank, and statistical significance. This mode balances statistical rigor with biological effect size and cross-omics validation.

Network-Based Scoring Metrics

Network metrics are computed on a metabolite–pathway bipartite graph constructed from enrichment results. Metabolite nodes are weighted by signed statistical significance (sign­(log₂FC) × −log₁₀(p-value)), and pathway-level scores are derived using established graph algorithms.

A Random Walk with Restart (RWR) algorithm (restart probability α = 0.3) is used to compute a diffusion-based machine learning (ML) score, capturing how coordinated metabolite signals propagate through interconnected pathways. Additional network descriptors, including PageRank and betweenness centrality, quantify pathway importance and bridging roles within the network topology.

To integrate statistical and diffusion-based evidence, a consensus rank is calculated as the rounded average of pathway ranks based on absolute Z-score magnitude (|Z|, ranked in descending order) and ML score (ranked in descending order), such that the consensus rank reflects the midpoint between the two ranking positions. Lower consensus rank values indicate higher pathway importance. Notably, PageRank and betweenness centrality are not used in the consensus ranking but are reported as complementary descriptors of pathway positioning within the network.

Rationale and Reproducibility

This ranking framework combines orthogonal evidence sources, statistical enrichment, effect directionality, network topology, and multiomics support using deterministic and reproducible algorithms. While network-based metrics may favor highly connected pathways, the use of consensus ranking and status-based prioritization mitigates this bias by requiring agreement with statistical evidence. A complete description of universe construction, ORA implementation, Z-score formulation, machine-learning diffusion scoring, and network-ranking logic is provided in the Supporting Information.

Network Analysis

Network analysis in MetaboGraph translates differential metabolite changes into pathway-level and system-level biological relationships. The framework integrates differential expression metrics, pathway membership assignments, enzyme and transporter associations, and disease-linked annotations into a structured multilayer network. Pathway enrichment is evaluated using classical Fisher’s Exact Test applied to a species-specific curated metabolite universe compiled from all databases used. For each pathway, MetaboGraph constructs a 2 × 2 contingency table comparing significant versus nonsignificant metabolites within the curated universe and computes enrichment p-values using the hypergeometric formulation, followed by FDR adjustment. Pathway directionality is assessed using a direction-aware Z-score derived from the sign distribution of log₂ fold changes across all pathway-assigned metabolites; this directionality reflects aggregate metabolite behavior and does not imply uniform regulation of all pathway members.

As an alternative to enrichment-based scoring, MetaboGraph also offers an Integrated Weighted Pathway Activation (IWPA) score, which summarizes pathway activation strength using fold-change magnitude, direction, statistical significance, and effect size. IWPA operates independently of Fisher ORA and provides an optional ranking strategy for users who prefer magnitude-weighted activation measures rather than count-based enrichment.

Pathway-level statistics, including ORA p-values, direction-aware Z-scores, and optional IWPA scores, are integrated with pathway membership relationships to construct a multilayer pathway–metabolite network that captures coordinated metabolic behavior across biological processes. This interactive network representation is shown in Figure . Within the network, nodes correspond to metabolites, pathways, enzymes, transporters, and disease-associated annotations, while edges represent curated biochemical, regulatory, or annotation-based relationships. Significant metabolites are emphasized as primary focus nodes, with optional inclusion of nonsignificant pathway members to preserve pathway connectivity and maintain biological context.

2.

Interactive pathway-metabolite network visualization in MetaboGraph, showing pathway-metabolite relationships and regulatory directionality. The control panel (left) allows user-defined layout algorithms, network modes, focused filtering, and visualization scaling. The central network displays pathways, metabolites, and upstream regulators as nodes, with edges representing activating, inhibiting, or associative relationships. Pathway nodes are color-coded by inferred activity state (activated, inhibited, or neutral), while metabolite nodes are colored according to relative regulation (upregulated or downregulated). Directional connections integrate statistical enrichment and expression trends, enabling intuitive exploration of pathway coordination and systems-level metabolic remodeling.

MetaboGraph includes an integrated interactive network viewer built using Dash and Cytoscape.js, which serves as the primary interface for exploring pathway–metabolite relationships. The viewer supports dynamic zooming and panning, node-level interrogation, multiple layout algorithms, and direct export of high-resolution PNG or SVG figures within the analytical environment. Networks may alternatively be exported as GraphML files for further interactive visualization or manual refinement in Cytoscape. A detailed description of universe construction, Fisher ORA implementation, Z-score formulation, IWPA computation, and network-building logic is provided in the Supporting Information.

Multiomics Integration and Comparative Analysis

MetaboGraph supports both multiomics integration and cross-study comparative analysis within a unified analytical framework. Annotated metabolomics and lipidomics data sets are processed independently using the same pathway annotation and ORA-based enrichment pipeline, ensuring consistent pathway definitions and statistical criteria across data types. For multiomics analysis, users upload curated metabolite and lipid data sets through a dedicated interface, after which the results are automatically integrated and propagated to the network analysis module. This design enables joint visualization and interpretation of metabolites, lipids, pathways, and upstream regulators without requiring a separate analytical workflow. During integration, MetaboGraph computes a multiomics integration score that prioritizes pathways supported by balanced evidence across multiple molecular layers (e.g., metabolomics and lipidomics), thereby emphasizing cross-platform validation and biologically robust pathway signals.

In addition to multiomics integration, MetaboGraph includes a comparative analysis module for evaluating multiple data sets derived from the same omics platform. This functionality enables cross-study comparison of independent metabolomics/lipidomics experiments to identify shared and study-specific metabolic features and pathways. Comparative analysis operates on independently analyzed data sets and focuses on feature overlaps, pathway consistency, and directional agreement across studies, supporting replication assessment, meta-analysis, and biological validation.

A detailed description of the multiomics integration algorithm, comparative analysis workflow, ranking metrics, and visualization outputs is provided in the Supporting Information.

Case Study: Metabolomics and Lipidomics Profiling of MCF7 (HTB22) and MDA-MB-453 (HTB131) Cell Lines

Two breast cancer cell lines (HTB22, and HTB131) were lysed, and intracellular metabolites and lipids were extracted and subjected to LC-MS/MS-based untargeted metabolomics and lipidomics analysis. The generated Thermo raw files were analyzed using Compound Discoverer 3.4 and LipidSearch 5.1, respectively.

HTB22 is a luminal A breast cancer cell line expressing ER and PR but not HER2. It is generally less aggressive, more hormone-dependent, and exhibits slower growth and lower metastatic potential. , In contrast, HTB131, also a TNBC line, represents a basal-like subtype and is considered more aggressive, displaying elevated glycolytic activity, metabolic flexibility, and enhanced proliferation. , These intrinsic biological differences establish a gradient of tumor aggressiveness and metabolic phenotype, ranging from hormone-dependent (HTB22) to basal-like (HTB131), which provides a relevant framework for interpreting the observed metabolic rewiring across comparisons.

Metabolomics Analysis

For the metabolomics analysis, processed data sets from Compound Discoverer were analyzed using MetaboGraph, beginning with data cleaning and feature annotation. Both positive and negative ionization mode features were annotated using a comprehensive suite of chemical identifiers including HMDB, KEGG, PubChem, ChEBI, SMILES, CAS, InChI, InChIKey, and LipidMaps.

A confidence-based quality filtering step was then applied in MetaboGraph, where features were classified as high, medium, or low confidence, and only high- and medium-confidence features were retained. High-confidence annotations were defined by the presence of at least one validated database identifier (HMDB, KEGG, or LipidMaps) together with MS/MS spectral matching confirmed using the Compound Discoverer-exported MS/MS match and MS/MS purity scores (maximum score = 100), while low-confidence features not meeting these criteria were excluded from further analysis. Metabolite annotations retained for downstream analysis correspond to Level 2 (putatively annotated compounds) as defined by the Metabolomics Standards Initiative (MSI), based on accurate mass, database matching, and MS/MS spectral similarity without confirmation using authentic reference standards.

Before normalization, further filtering was applied to retain only metabolites quantified in at least four out of five replicates within each group. After this threshold was applied, 447 metabolites were retained for HTB22 and 440 for HTB131. The data were normalized to percentage relative abundance prior to differential analysis to account for differences in total abundance between samples and to emphasize compositional shifts in intracellular metabolism between the two cell lines. Relative abundance normalization was selected to emphasize intracellular compositional shifts. All downstream statistical analyses were performed on the normalized data.

Exploratory data analysis was first conducted to characterize metabolic differences between the two cell lines. A Venn diagram summarized the overlap and uniqueness of detected metabolites, whereas principal component analysis (PCA), an unsupervised multivariate method, was used to visualize global variation across samples. As seen in the Venn (Supplementary Figure S3A), we observed 432 metabolites shared across the two cell lines, with each cell line also exhibiting a distinct set of unique features. For example, HTB131 displayed 8 unique metabolites, while HTB22 had 15 unique metabolites. PCA was performed on the normalized data set and visualized in two dimensions with 95% confidence ellipsoids (Supplementary Figure S3B). The PCA plot demonstrated strong separation among both cell lines, confirming substantial variation in their global metabolic profiles and minimal intragroup variability.

Furthermore, group comparisons were conducted using Welch’s t test, and statistical significance was determined using BH correction with a false discovery rate (FDR) threshold of 0.05. This comparative analysis identified 305 metabolites that were significantly different between the two cell lines. Supplementary Table S3 provides the complete list of metabolites included in the statistical analysis, along with their corresponding database identifiers (HMDB, KEGG, or PubChem etc.), raw and adjusted p-values, log₂ fold changes, and additional statistical metrics used in downstream pathway and network analyses.

Volcano plots were then generated for each pairwise comparison using an adjusted p-value cutoff of 0.05 and a fold change threshold of ±2 (Supplementary Figures S3C). These plots revealed a wide dynamic range of 72 upregulated and 105 downregulated metabolites between the cell lines. To further illustrate feature-level differences, a heatmap of selected significant metabolites was generated to highlight relative abundance patterns across samples (Supplementary Figure S3D), and representative boxplots were used to visualize group-specific differences for top significant metabolites (Supplementary Figure S3E).

To place these findings in a biochemical context, pathway enrichment analysis was performed using MetaboGraph’s integrated pathway-mapping framework. Prior to enrichment, all metabolites were preannotated with database-compatible identifiers (e.g., HMDB, KEGG, and LipidMaps IDs), a requirement for accurate pathway assignment within MetaboGraph. The pathway module also relies on statistical outputs from differential analysis, specifically the metabolite p-values (or adjusted p-values) and fold change (or log₂ fold change), which together enable direction-aware pathway scoring.

Pathway annotation was performed by mapping annotated metabolites to biochemical pathway definitions using a curated multidatabase reference composed of HMDB, KEGG, Reactome, PathBank, SMPDB, WikiPathways, and Metabolika. All annotation and enrichment computations were parallelized across eight CPU workers to improve computational efficiency. Pathway enrichment analysis was conducted using Fisher’s ORA with Homo sapiens as the reference organism and BH correction applied for multiple testing. Pathways were required to contain a minimum of two annotated metabolites, and statistical significance was evaluated at adjusted p-value <0.05. Metabolite-level significance was defined using an adjusted p-value threshold of 0.05, without imposing a fold-change cutoff at the enrichment stage. For visualization and interpretability, pathway networks were constructed using top ranked pathways exhibiting nonzero directionality (|Z-score| > 0), with additional manual curation to retain biologically relevant pathways for focused visualization.

Based on functional relatedness and directional coherence, the enriched pathways were first organized into two major metabolic programs: (i) a central carbon and nitrogen redistribution module and (ii) a lipid remodeling and nucleotide metabolic reprogramming module. The overall distribution of enriched pathways and their metabolite-level contributions for these two modules are summarized in Supplementary Figure S4. The chord diagram (Supplementary Figure S4A-B) illustrate the relationships between significantly altered metabolites and their associated pathways within each module. Supplementary Figure S4A corresponds to the central carbon and nitrogen redistribution module, whereas Supplementary Figure S4B represents the lipid remodeling and nucleotide metabolic reprogramming module. These plots highlight the extent to which individual metabolites contribute to multiple pathways, reflecting the interconnected nature of the metabolic network within each functional program.

The global enrichment patterns for the central carbon and nitrogen module are summarized in Figure . Figure A shows horizontal bar chart of the enriched pathways of central carbon and nitrogen redistribution, where bar length indicates the number of significant metabolites mapped to each pathway (metabolite coverage) and bar color encodes pathway significance as −log₁₀(adjusted p-value). Following enrichment, pathway-level relationships were visualized using a pathway enrichment network (Figure B) in which node size reflects metabolite coverage and node color represents enrichment significance. Edges denote biochemical interconnectedness defined by shared or coannotated metabolites across pathways, highlighting tightly connected modules involving amino-acid metabolism, nitrogen handling, carbon metabolism, and energy-associated processes, independent of pathway directionality.

3.

Central carbon and nitrogen redistribution pathway enrichment and network integration. (A) ORA-based pathway enrichment of significantly altered metabolites, where bar length indicates metabolite coverage and bar color represents pathway significance (−log₁₀ adjusted p-value). (B) Pathway enrichment network showing biochemical relationships among enriched pathways. Node size reflects metabolite coverage and node color encodes enrichment significance; edges represent connectivity based on shared or coannotated metabolites, independent of directionality. (C) Direction-aware metabolite-pathway network showing central carbon and nitrogen redistribution module highlighting coordinated shifts in amino acid and nitrogen metabolism. (D) Direction-aware metabolite-pathway network showing a lipid remodeling and nucleotide metabolic reprogramming module representing metabolic reprogramming associated with altered lipid metabolism and biosynthetic demand. (E) Legend describing node and edge annotations used in direction-aware networks.

Using the direction-aware Z-score framework implemented in MetaboGraph, we next inferred pathway activation and inhibition by integrating metabolite-level fold changes with pathway membership structure. Enriched pathways were subsequently grouped into biologically coherent functional categories and annotated as Activated or Inhibited in HTB131 relative to HTB22. A summary of pathways associated with metastatic metabolic reprogramming, along with their corresponding rankings, is provided in Supplementary Table S1.

Direction-aware pathway-metabolite networks were then constructed to resolve coordinated biochemical behavior. The resulting network (Figure C) captures a central carbon and nitrogen metabolic program, representing a core metabolic hub of the cell. This network highlights coordinated activation of glycolysis-associated processes, the Warburg effect, and pyruvate metabolism, alongside extensive remodeling of amino acid metabolism, including phenylalanine, methionine, tyrosine, and branched-chain amino acid pathways.

In parallel, the network resolves a nitrogen redistribution and amino acid supply module, with predicted activation of glutamate metabolism, biosynthesis of amino acids, and 2-oxocarboxylic acid metabolism. These pathways converge on metabolites such as l-glutamine, l-glutamic acid, l-aspartic acid, and other amino acid intermediates, supporting nitrogen assimilation, biosynthetic demand, and anaplerotic replenishment of central carbon metabolism. Together, these coordinated changes reflect a shift toward high-throughput carbon flux and metabolic plasticity characteristics of aggressive cancer phenotypes. Propanoate-related metabolism also showed predicted activation, consistent with increased propionyl-CoA-linked anaplerotic input, which can arise from short-chain fatty acid-related chemistry as well as branched-chain amino acid and lipid metabolism. , Such amino acid-driven metabolic reprogramming programs have been widely reported in highly proliferative and invasive cancer states. ,,

Furthermore, we visualized the lipid remodeling and nucleotide metabolic reprogramming module in Figure D. Within this network, a prominent nucleotide and biosynthetic demand module is evident, characterized by pathways such as purine metabolism, pyrimidine metabolism, and nucleotide metabolism, which are interconnected with adenine-, guanine-, and cytidine-related metabolites. These pathways are supported by upstream regulatory enzymes such as PNP, NT5C3, APRT, and DCK, indicating enhanced nucleotide turnover and salvage activity consistent with increased proliferative and transcriptional demand in metastatic cells.

Concurrently, the network reveals a lipid remodeling module centered on fatty acid and membrane-associated metabolism. Predicted activation of fatty acid metabolism is supported by altered levels of fatty acids and their connections to lipid-modifying enzymes including FASN, CES1, BAAT, ACER1, and ACER2. These relationships reflect enhanced lipid turnover and membrane remodeling, processes widely reported to support cancer cell growth, membrane plasticity, and invasive behavior. , In contrast, pathways associated with the biosynthesis of unsaturated fatty acids and α-linolenic/linoleic acid metabolism were predicted to be inhibited, consistent with reduced levels of long-chain polyunsaturated fatty acids such as docosahexaenoic acid (DHA) and eicosapentaenoic acid (EPA). Importantly, these lipid and nucleotide modules are highly interconnected within the network, with shared metabolites and regulatory nodes linking membrane remodeling to nucleotide metabolism. Together, these features suggest a coordinated metabolic adaptation in which enhanced lipid turnover supports membrane plasticity, while increased nucleotide recycling sustains biosynthetic and transcriptional activity.

Together, this illustrates how MetaboGraph integrates enrichment statistics, pathway directionality, and network structure to resolve coordinated metabolic programs. The results reveal a coupled metabolic phenotype in HTB131 characterized by enhanced lipid remodeling and membrane restructuring linked to increased nucleotide turnover and redox cofactor demand, consistent with established metabolic adaptations associated with invasive cancer states. A unified legend describing node and edge annotations used across the direction-aware networks is provided in Figure E, defining pathway, metabolite, and upstream regulator nodes, as well as activating, inhibiting, and associative edges.

Lipidomics Analysis

For the lipidomics analysis, processed data sets from LipidSearch 5.1 were analyzed using MetaboGraph, following the same data-cleaning, normalization, and statistical workflow applied to the metabolomics data set. Unlike small molecules, many lipid species do not have complete structural identifiers available in public databases, and a substantial fraction lack fully resolved entries in LipidMaps. Therefore, lipid annotation in MetaboGraph relies on a combination of available identifiers (LipidMaps, HMDB, PubChem) and KEGG-derived lipid class identifiers, which enable pathway-level mapping even when full molecular IDs are unavailable. This class-based annotation strategy allows lipids to be incorporated into pathway enrichment despite incomplete database coverage.

Annotation confidence for lipids was controlled using MS/MS-based identification and LipidSearch confidence grading. Only lipid species assigned grade A–C by LipidSearch were retained for downstream analysis. These grades reflect increasing confidence in MS/MS spectral matching and structural assignment, ensuring that retained lipids are supported by reliable fragmentation evidence. This approach minimizes false-positive annotations while maintaining biologically relevant lipid coverage.

Prior to normalization, quality filtering was performed to retain only lipid species quantified in at least three out of five biological replicates per group. After filtering, 394 lipid features were retained for HTB22 and 421 for HTB131. Based on MSI guidelines, lipid annotations correspond to Level 2 (putatively annotated lipids) where MS/MS evidence supports structural assignment, and Level 3 (putatively characterized lipid classes) in cases where full structural resolution could not be confidently assigned based on the available MS/MS data and lipid annotation workflow.

The data were normalized to percentage relative abundance and used for all subsequent analyses. Exploratory data analysis was then conducted to characterize lipidomic differences between the two cell lines. A Venn diagram summarized the overlap and uniqueness of detected lipid species (Supplementary Figure S5A). As shown in the Venn, we observed 202 lipid species shared across HTB22 and HTB131, with each cell line also exhibiting a distinct set of unique lipids. Specifically, HTB131 displayed 219 unique lipid species, while HTB22 exhibited 192 unique lipid species. PCA was performed on the normalized lipidomics data set and visualized in three dimensions with 95% confidence ellipsoids (Supplementary Figure S5B). The PCA plot demonstrated clear separation between the two cell lines, confirming substantial differences in their global lipidomic profiles and minimal intragroup variability across biological replicates.

Furthermore, differential analysis was conducted using Welch’s t test with BH-adjusted significance at FDR < 0.05, identical to the metabolomics workflow. This analysis identified 106 significantly altered lipids between the two cell lines. Supplementary Table S4 provides the complete list of lipids included in the statistical analysis, along with their identifiers, p-values, adjusted p-values, log₂ fold changes, and supporting metrics used for downstream pathway and network modeling.

Volcano plots were then generated for each pairwise comparison using an adjusted p-value cutoff of 0.05 and a fold change threshold of ±2 (Supplementary Figure S5C). Using these thresholds, the volcano plot highlights 36 significantly upregulated and 25 significantly downregulated lipid species between the cell lines, with additional features falling below the fold-change threshold or not meeting significance criteria. These significantly altered lipids were further visualized using a hierarchical clustering heatmap (Supplementary Figure S5D), which revealed consistent abundance patterns across biological replicates and clear separation between the two cell lines. Also, Supplementary Figure S5E shows representative boxplots of the significant lipid species illustrating relative abundance differences between HTB22 and HTB131.

Pathway enrichment and network modeling were then conducted using the MetaboGraph ORA-based framework and direction-aware scoring system described for the metabolomics data set. Lipid-associated pathways and upstream regulators were mapped using the combined lipid identifier and class-based annotation strategy, producing a lipid-centric pathway network (Figure ). As shown in Figure A, the chord diagram illustrates relationships between significantly altered lipid species and enriched lipid metabolic pathways, revealing selective remodeling of membrane lipid metabolism in HTB131 relative to the less metastatic HTB22 cell line. Multiple phosphatidylcholine (PC) species contribute to shared pathway enrichment, with dominant associations observed for phospholipid, phosphatidylcholine, and phosphatidylethanolamine biosynthesis. This convergence indicates coordinated alterations in membrane phospholipid composition and turnover rather than isolated lipid changes. Pathways with fewer lipid connections suggest more limited or context-dependent contributions, consistent with selective pathway engagement during metastatic progression.

4.

Lipid-centric pathway enrichment and network integration. (A) Chord diagram showing associations between significantly altered lipid species and enriched lipid metabolic pathways. Ribbon connections represent lipid–pathway relationships. (B) Lipid–pathway interaction network integrating altered lipids with enriched pathways and associated enzymes. Lipid nodes are colored by abundance change (green is upregulation and red is downregulation), while pathways reflect overall directionality inferred from lipid changes. Red dashed edges indicate activation, blue dashed edges indicate inhibition, and gray dashed edges represent associations.

The lipid-centric network in Figure B shows how significantly altered lipid species connect the three enriched pathways: ether lipid metabolism, glycerophospholipid metabolism, and choline metabolism in cancer. These pathways were predicted to be inhibited, indicating coordinated suppression of lipid remodeling processes in HTB131 relative to HTB22.

Several phosphatidylcholine (PC) species, including PC(14:0_16:0), PC(14:0_15:0), PC(16:0_16:0), PC(28:2), PC(30:2), PC(40:1), PC(40:5), and PC(42:2), together with LPC(16:0), form key connections across the lipid pathways. The network highlights lipid-modifying enzymes, including CHKA, PCYT1A, LPCAT2, LPCAT3, LYPLA1, and LYPLA2, linking altered lipid species to choline metabolism, phosphatidylcholine turnover, and lysophospholipid remodeling.

Consistent with emerging evidence of lipid metabolic reprogramming in cancer, particularly in aggressive phenotypes, these findings suggest that HTB131 cells exhibit a shift toward selective suppression of lipid remodeling pathways while maintaining essential lipid processes. Overall, this pattern reflects coordinated metabolic adaptation rather than isolated changes in individual lipid species.

Overall, the results illustrate a coordinated lipid remodeling program in HTB131 characterized by altered phospholipid composition and rewiring of membrane-associated metabolic processes. These features are consistent with selective lipid metabolic reprogramming and adaptive lipid utilization observed in invasive cancer phenotypes, − and demonstrate how MetaboGraph integrates enrichment, directionality, and network structure to resolve biologically coherent lipid metabolic programs.

Multiomics Analysis

For the multiomics analysis, the independently annotated metabolomics and lipidomics data sets from HTB22 and HTB131 cells were imported into the multiomics integration framework of MetaboGraph. The same pathway annotation, over-representation analysis, and direction-aware scoring settings described above were applied to both data types to ensure consistency in pathway definitions, statistical thresholds, and interpretive criteria. Pathways visualized in the multiomics network were prioritized based on their multiomics rank, indicating contributions from both metabolite-level and lipid-level features.

The integrated multiomics analysis (Figure A–B) reveals a coordinated convergence of metabolite and lipid signals on membrane-associated lipid metabolic pathways. As shown in the chord diagram (Figure A), glycerophospholipid metabolism, ether lipid metabolism, choline metabolism in cancer, and sphingolipid metabolism represent the dominant functional axes linking lipid species and metabolite nodes. These pathways are interconnected through shared contributions from phosphatidylcholine (PC) species, lysophosphatidylcholine (LPC), and sphingolipid-related metabolites, including ceramides, sphingomyelins, sphinganine, and D-sphingosine, together with choline-associated intermediates.

5.

Multiomics integration of metabolite and lipid alterations in HTB131 vs HTB22 cells. (A) Chord plot showing associations between significantly altered metabolites and lipids with enriched pathways. Connections highlight shared contributions from phosphatidylcholine (PC), lysophosphatidylcholine (LPC), choline-related metabolites, and sphingolipid-associated features across glycerophospholipid metabolism, ether lipid metabolism, choline metabolism in cancer, and sphingolipid metabolism. (B) Integrated network of metabolites, lipids, enzymes, and pathways. Circular nodes represent metabolites and lipids, triangular nodes represent enzymes, and diamond nodes represent pathways. Node colors indicate feature regulation (upregulated or downregulated), while pathway nodes reflect overall directionality inferred from aggregated molecular changes. Red dashed edges indicate activation, blue dashed edges indicate inhibition, and gray dashed edges represent associations. The network highlights coordinated interactions linking phospholipid and sphingolipid metabolism through shared metabolite and lipid components.

The network representation (Figure B) further resolves these relationships into interconnected modules centered on phospholipid and sphingolipid metabolism. A central phospholipid module is defined by strong connectivity among PC species (e.g., PC(14:0_16:0), PC(14:0_15:0), PC(16:0_16:0), PC(28:2), PC(30:2), PC(40:1), PC(40:5), and PC(42:2)) and LPC(16:0), linked through lipid remodeling enzymes including LPCAT2, LPCAT3, LYPLA1, LYPLA2, and PNPLA6, as well as choline metabolism enzymes such as CHKA and PCYT1A. PLD2 is also integrated within this network, connecting phosphatidylcholine turnover to broader pathway interactions, consistent with established roles of these enzymes in phospholipid remodeling and membrane lipid dynamics. ,

Metabolite nodes, including choline, glycerol-3-phosphate, 1-palmitoyl-sn-glycero-3-phosphocholine, and N-methylethanolaminium phosphate, form key bridges between lipid species and pathway-level annotations, linking choline metabolism in cancer with glycerophospholipid and phospholipid biosynthetic processes. In parallel, a sphingolipid-associated module is observed, where ceramides, sphingomyelins, sphinganine, and D-sphingosine connect to sphingolipid metabolism and intersect with glycerophospholipid metabolism through shared intermediates such as glycerol-3-phosphate.

Direction-aware scoring indicates coordinated pathway behavior across molecular layers, with glycerophospholipid metabolism, ether lipid metabolism, and choline metabolism in cancer showing inhibitory trends, while phospholipid biosynthesis exhibits relative activation. This pattern suggests selective lipid metabolic reprogramming rather than uniform pathway activation. Increased phospholipid biosynthesis is consistent with sustained membrane production demands observed in rapidly proliferating and adaptive cancer cells. , In contrast, inhibition of glycerophospholipid and choline-associated pathways reflects reduced lipid remodeling and pathway-specific suppression. Additionally, inhibition of sphingolipid metabolism, involving ceramide-related metabolites, is consistent with reduced pro-apoptotic signaling, a feature commonly associated with aggressive cancer phenotypes.

Overall, this multiomics integration highlights coordinated changes across phospholipid and sphingolipid pathways, where metabolite-level inputs and lipid-level outputs are unified within a single network framework. These results demonstrate how MetaboGraph integrates metabolomics and lipidomics data to resolve pathway connectivity and coordinated biochemical changes across molecular layers, providing biologically meaningful insights that can guide further investigation and hypothesis generation.

Comparison with Established Pathway Analysis Methods

To contextualize MetaboGraph’s pathway-level outputs relative to commonly used metabolomics tools, results were compared with MetaboAnalyst and IPA. These tools were selected as representative commercial (IPA) and open-source (MetaboAnalyst) approaches commonly used for metabolomics pathway interpretation.

Comparison with MetaboAnalyst

For comparison with MetaboAnalyst, we uploaded the metabolites along with their corresponding quantitative values, using HMDB identifiers as the primary mapping reference. The use of HMDB identifiers improved mapping efficiency compared to metabolite names alone, which often lead to reduced pathway coverage in MetaboAnalyst due to naming inconsistencies and ambiguity in compound nomenclature. Pathway enrichment was performed using the Homo sapiens KEGG pathway library (December 2024 version), which contains curated human metabolic pathways.

MetaboAnalyst pathway enrichment analysis (Figure A) identified several significant pathways that overlapped with those detected by MetaboGraph, including purine metabolism, pyrimidine metabolism, glycolysis-associated processes, and multiple fatty acid-related pathways. This agreement confirms that the major metabolic programs underlying the data set are consistently captured across analytical platforms. Notably, MetaboGraph further resolves the directionality of these pathways, revealing activation of purine metabolism, glutamate metabolism, and pyruvate-associated energy pathways, alongside strong inhibition of unsaturated fatty acid biosynthesis, linoleic/alpha-linolenic acid metabolism, and lipid signaling pathways. However, MetaboAnalyst is limited to enrichment-based metabolite–pathway associations and does not incorporate pathway directionality or metabolite fold-change weighting. As a result, while it identifies significantly associated pathways, it cannot distinguish whether these pathways are upregulated or downregulated between conditions. In contrast, MetaboGraph integrates fold-change-weighted metabolite evidence and directional biochemical relationships, enabling identification of coordinated metabolic shifts. This includes activation of amino acid-driven nitrogen metabolism, nucleotide biosynthesis demand, and glycolytic flux, together with suppression of lipid biosynthesis and signaling pathways. Importantly, MetaboGraph also reveals asymmetry within nucleotide metabolism, with purine metabolism activated while pyrimidine metabolism is inhibited, a feature not discernible from enrichment analysis alone.

6.

Pathway and network analysis using IPA and MetaboAnalyst. (A) Metabolite-pathway association network generated using MetaboAnalyst based on HMDB identifiers and KEGG (Homo sapiens) pathways. (B) Cancer-relevant regulatory network generated using Ingenuity Pathway Analysis (IPA) from the MetaboGraph-derived results table. Pathways were filtered using p < 0.05. Node and edge colors indicate predicted pathway directionality, with red representing activation and blue representing inhibition.

Taken together, the overlap between MetaboAnalyst and MetaboGraph supports the biological validity of the observed metabolic signatures, while demonstrating that MetaboGraph provides additional mechanistic insight by resolving pathway directionality and coordinated metabolic behavior.

Comparison with IPA

For IPA, we uploaded the MetaboGraph-derived results table containing preannotated metabolite identifiers and associated statistics (adjusted p-value, and log₂ fold change). IPA successfully mapped 199 metabolites with valid identifiers following filtering based on an adjusted p-value threshold (<0.05) and subsequently identified a set of cancer-relevant metabolic pathways.

Several pathways identified by IPA overlapped directly with those detected by MetaboGraph, including nucleotide biosynthesis, aspartate and asparagine metabolism, purine metabolism, pyrimidine metabolism, glycolysis/gluconeogenesis and related glucose-utilization pathways, and citrulline-associated nitrogen metabolism. This concordance indicates that both platforms consistently capture the core metabolic programs underlying the data set.

Importantly, IPA-derived pathway directionality (Figure B) was consistent with MetaboGraph for key biosynthetic and nitrogen-associated processes, with both approaches indicating activation of nucleotide biosynthesis and amino acid-linked metabolic pathways. In addition, IPA highlights connections between these pathways and functional outcomes such as cell viability and cell death of tumor cell lines through metabolite-level interactions involving intermediates such as adenosine, adenosine monophosphate (AMP), guanosine monophosphate (GMP), and l-glutamine. These observations reflect methodological differences between the platforms. IPA infers pathway activity based on curated upstream–downstream relationships and integrates metabolite changes into causal network predictions, linking metabolic alterations to phenotypic outcomes. In contrast, MetaboGraph derives pathway directionality directly from fold-change-weighted metabolite evidence and resolves coordinated behavior within metabolic pathways without incorporating phenotype-level predictions.

Taken together, the agreement between IPA and MetaboGraph supports the biological relevance of enhanced nucleotide biosynthesis and amino acid-associated metabolism in the data set, while highlighting that IPA provides phenotype-linked network context, whereas MetaboGraph offers direct resolution of metabolite-driven pathway behavior. In addition to pathway-level overlap, several metabolites that also contributed to MetaboGraph pathway interpretation, including AMP, GMP, adenosine, l-glutamine, and l-aspartic acid, were central to the IPA network structure. Other mapped metabolites visible in the IPA network, such as l-methionine, niacinamide, choline, lactate, arachidonic acid, palmitic acid, myristic acid, taurine, adenine, and 5′-methylthioadenosine, further connected sulfur amino acid metabolism, nucleotide biosynthesis, nicotinate metabolism, aspartate and asparagine metabolism, and neurotransmitter release cycle to predicted cellular outcomes. IPA also linked these metabolite-centered pathway relationships to functional annotations including cell viability, cell viability of tumor cell lines, and cell death of tumor cell lines. While MetaboGraph includes enzymes and transport processes as annotation-level network features to contextualize metabolite-pathway relationships, it does not currently perform causal upstream regulator prediction. In contrast, IPA places these metabolite changes within a curated causal framework that connects molecular patterns to predicted biological functions.

MetaboGraph and IPA are therefore best viewed as complementary tools. IPA contextualizes metabolite changes within a curated knowledge base that connects pathways to higher-order functional outcomes, whereas MetaboGraph provides a transparent, metabolite-centric workflow for identifier harmonization, direction-aware pathway scoring, and network construction tailored to untargeted metabolomics and lipidomics data sets. In practice, MetaboGraph addresses a major bottleneck in untargeted analysis by explicitly harmonizing metabolite identifiers across databases and retaining annotation provenance, thereby improving compatibility with downstream pathway and network analysis while preserving analytical transparency and reproducibility.

In summary, the overlap in enriched pathways identified by MetaboGraph, IPA, and MetaboAnalyst supports the biological relevance of the metabolic signatures observed in this study. Concordance with IPA for key biosynthetic and amino acid-associated pathways further supports the validity of MetaboGraph’s pathway-level interpretations. At the same time, MetaboGraph’s metabolite-centric, direction-aware framework enables resolution of coordinated metabolic behavior based directly on fold-change patterns, providing insight beyond enrichment-based associations. These results demonstrate that MetaboGraph serves as a complementary analytical approach, integrating pathway enrichment with direction-resolved, metabolite-driven network analysis to facilitate more detailed interpretation of metabolic reprogramming.

Conclusion

MetaboGraph provides a unified and reproducible framework for metabolomics and lipidomics analysis that bridges the gap between statistical feature detection and biologically meaningful pathway interpretation. By integrating metabolite-level statistics with curated biochemical knowledge, MetaboGraph enables direction-aware pathway enrichment and network modeling that capture coordinated metabolic behavior rather than isolated metabolite changes. Application of the platform to breast cancer cell-line data sets demonstrated its ability to recover biologically coherent metabolic programs consistent with known features of tumor aggressiveness, including amino acid rewiring, nucleotide biosynthesis, lipid remodeling, and altered energy metabolism. Although demonstrated here in a cancer context, MetaboGraph is agnostic to biological domain and broadly applicable to metabolomics and lipidomics data sets involving group-based comparisons across diverse biological systems. Compared with established tools such as IPA and MetaboAnalyst, MetaboGraph offers complementary strengths by emphasizing metabolite-centric, transparent pathway scoring and polarity-aware network construction. Together with its modular design and extensive database integration, MetaboGraph provides a flexible and accessible platform for pathway-level interpretation and hypothesis generation in system-level metabolic research.

The mass spectrometry data has been deposited to massive (https://massive.ucsd.edu/) with the Data set identifier “MSV000100627”.

The Supporting Information is available free of charge at https://pubs.acs.org/doi/10.1021/acs.analchem.6c01446.

• Exploratory metabolomics and lipidomics analyses, complete pathway analysis result tables, IPA benchmarking outputs, computational dependency summaries, and the MetaboGraph User’s Guide and implementation documentation (PDF)

• Data on metabolite positive and negative and lipid positive and negative (XLSX)

• Annotated data (XLSX)

• Annotated data (XLSX)

Conceptualization: O.D.; recourses, funding acquisition, project administration, and supervision: Y.M.; methodology, validation, formal analysis, investigation, writingoriginal draft preparation, and editing, visualization: O.D., J.N., O.O., M.F.; methodology, software, investigation, and visualization: O.D., M.F., Y.M.; writing and revision: O.D., J.N., Y.M.; software and visualization: O.D., O.O., J.N. All authors have read and agreed to the published version of the manuscript. All authors have approved the final version of the manuscript.

This work is supported by grants from the Robert A. Welch Foundation (No. D-0005) and the CH Foundation.

MetaboGraph (v1.0.1) is freely available as an open-source tool. The software, along with its documentation and example workflows, is accessible via GitHub (https://github.com/odaramola92/MetaboGraph) and archived on Zenodo with a persistent DOI to ensure transparency and reproducibility.

The authors declare no competing financial interest.