ssNetShift: single-sample metabolic network rewiring reveals hidden prognostic subtypes beyond clinical staging in gastric cancer

Abstract

Current metabolomics approaches predominantly rely on group-level differential abundance screening. While effective for identifying biomarkers, this paradigm often overlooks upstream regulatory hubs and fails to resolve the inter-individual heterogeneity inherent in complex diseases. To bridge this gap, we present ssNetShift, a single-sample network framework that identifies personalized topological driver metabolites by quantifying topological rewiring rather than static concentration deviations. By integrating linear interpolation-based network estimation with an extended neighbor-shift metric, ssNetShift systematically characterizes how specific metabolites alter their connectivity and centrality within individual patient networks. We applied ssNetShift to a multicohort gastric cancer dataset comprising 389 patients and 313 controls. Benchmarking analyses demonstrated that ssNetShift consistently outperformed conventional approaches: unlike group-level methods (e.g. NetShift), it recovered survival-associated driver metabolites masked by population averaging; unlike single-sample abundance methods (e.g. personalized perturbation profiles), it prioritized silent drivers, metabolites with stable abundance but drastic topological reorganization, thereby capturing system-level dysregulation. Crucially, ssNetShift revealed hidden prognostic subtypes within the same clinical stage, separating patients with identical tumor-node-metastasis (TNM) staging into distinct risk classes characterized by specific metabolic wiring patterns (e.g. nucleotide and tryptophan hubs) and significantly divergent survival outcomes. Collectively, ssNetShift provides a risk stratification dimension orthogonal to traditional staging, offering a robust tool for uncovering mechanistic drivers and refining prognostic resolution in heterogeneous malignancies.

Introduction

Metabolomics provides a high-resolution snapshot of cellular physiology, offering unique insights into disease mechanisms by profiling small-molecule dynamics [1, 2]. The current mainstream metabolomics analyses rely on group-level differential metabolite screening strategies, such as multivariate statistical analysis (e.g. OPLS-DA [3]), univariate methods (e.g. t-test [4]), and pathway enrichment techniques (e.g. MSEA) [5]. These approaches provide insights into pathological processes by comparing the metabolic profiles between disease groups and healthy controls [6, 7]. However, these conventional strategies suffer from two fundamental limitations. First, they assume population homogeneity, treating inter-individual variation as statistical noise rather than biological signal [8, 9]. This is particularly problematic in heterogeneous malignancies like gastric cancer (GC), where patients with identical clinical characteristics often exhibit divergent metabolic trajectories. Second, they treat metabolites as isolated variables [10], ignoring the synergistic interactions and regulatory networks that govern cellular metabolism. Consequently, abundance-based methods often identify downstream effectors, accumulated byproducts of dysregulation, while failing to pinpoint the upstream topological drivers that are central to these systemic changes [11].

In recent years, network topology analysis [12] has been introduced into the field of metabolomics to enhance the capability of system-level mechanism analysis. For example, metabolite module identification based on weighted gene co-expression networks (WGCNA [13]) and the construction of regulatory networks through information-theoretic methods (e.g. ARACNE [14]) have significantly strengthened the global correlation analysis of metabolic pathways. The heterogeneous directed graph model iMSEA [15], proposed by Wang et al., integrates metabolite interaction networks and pathway topology, showcasing unique advantages in analyzing drug combination mechanisms. However, these result-oriented methods, while capable of revealing which changes occurred, struggle to trace back to key topological drivers associated with the changes, which is the core issue. Notably, the NetShift method [16] have advanced the field by identifying driver metabolites based on topological differences between disease and healthy networks rather than concentration changes alone. However, NetShift remains constrained by a group-comparison framework. By aggregating samples into “Case” versus “Control” networks, it inherently averages out patient-specific signals, masking the unique wiring patterns that drive disease progression in individual patients [17]. This limitation is critical in the era of precision medicine, where the goal is to resolve intra-stage heterogeneity, explaining why patients at the same pathological stage display markedly different survival outcomes.

The rise of single-sample analysis (SSA) techniques has attempted to address this by quantifying individual-level deviations [18, 19], such as single-sample gene set enrichment analysis (ssGSEA [20], ssClustPA [21]), and the metabolite abnormality scoring method called personalized perturbation profiles (PEEPs) [22]. Yet, these methods largely remain tethered to the abundance-first perspective. They excel at detecting outliers (metabolites with extreme concentration Z-scores) but are blind to silent drivers, key regulatory hubs that maintain homeostatic concentrations despite undergoing drastic changes in their connectivity and network role [23, 24]. A metabolite may shift from a peripheral node to a central bottleneck in a patient’s network without a significant change in abundance; such topological rewiring is invisible to univariate outlier detection but represents a critical layer of metabolic dysregulation.

To bridge this gap, we propose ssNetShift, a single-sample network framework designed to identify individualized driver metabolites by modeling dynamic network rewiring. Integrating the LIONESS [25] equation for individual network estimation with an extended Neighbor Shift (eNESH) metric, ssNetShift quantifies how specific metabolites alter their local coordination and global centrality in each patient relative to a healthy reference. In this study, we applied ssNetShift to a large, multi-cohort GC metabolomics dataset. We demonstrate that ssNetShift (i) uncovers a distinct class of survival-agnostic driver metabolites that are topologically active but statistically normal in abundance, (ii) outperforms both group-level (NetShift) and abundance-based (PEEPs) methods in prognostic stratification, and (iii) reveals hidden metabolic subtypes within the same clinical stage. By shifting the focus from static levels to dynamic wiring, ssNetShift offers a novel dimension of molecular risk stratification orthogonal to conventional staging systems.

Materials and methods

GC dataset

The GC metabolomics dataset utilized in this study was downloaded from an open-access literature source [26], which is a multi-cohort investigation, comprising 389 pathologically confirmed GC patients and 313 non-GC controls. Participants were recruited from 3 independent cohorts: Cohort 1 includes 145 GC patients and 281 Non-GC controls; Cohort 2 includes 63 GC patients and 32 Non-GC controls; Cohort 3 includes 181 GC participants with additional prognosis data. All GC patients were clinically staged (I–IV, detailed in Table S1). Plasma samples were collected from preoperative GC patients and non-GC controls after an overnight fast.

The samples were analyzed using LC–MS/MS, with chromatographic separation on a C18 column as described in Ref 21. The mass spectrometer (AB QTRAP 6500+, SCIEX) operated in multiple reaction monitoring (MRM) mode, and targeted 258 endogenous water-soluble metabolites, of which 147 were measured. Quality control (QC) samples, made by pooling plasma from all individuals, were inserted every 10 test samples to monitor instrument stability and normalize data. The data had been normalized during preprocessing by the original authors using a QC-based ratio method to correct for batch and analytical variation.

Overview of the ssNetShift framework

The core objective of ssNetShift is to decipher metabolic dysregulation at the individual-sample level by reconstructing the specific wiring patterns of metabolite interactions. Unlike traditional differential analyses that focus on static abundance deviations, ssNetShift postulates that disease progression is strongly characterized by the rewiring of metabolite–metabolite interactions. We define a “driver” metabolite (defined strictly as a key topological hub in this context, rather than implying direct biological causality without further experimental validation) not merely as a fluctuating node, but as a control point that undergoes a fundamental role shift relative to a healthy reference network. Specifically, this shift is quantified across two topological dimensions: local neighborhood rewiring (representing a change in regulatory partners) and global centrality gain (indicating increased influence on the network structure).

Operationally, the ssNetShift framework proceeds in three sequential steps (Fig. 1) to capture these personalized systemic dysregulations.

Figure 1.

Workflow of ssNetShift algorithm. (A) Construction of a single-sample metabolic network relative to a healthy reference network based on the LIONESS framework. (B) Quantification of local neighborhood rewiring for each metabolite using the extended Neighbor Shift (eNESH) score by comparing first-order interactions between the single-sample and reference networks. (C) Identification of driver metabolites based on joint criteria of statistically significant eNESH scores and increased global network importance, measured by positive differential betweenness centrality (ΔB).

Single-sample network construction

To resolve individual-specific interactions from population-level data, we employed the [25] algorithm. This method mathematically extracts the contribution of a single sample from the aggregate correlation structure of a reference population. Let denote the Pearson correlation matrix derived from a reference group of samples. When a specific sample is added to this reference, the updated correlation matrix is denoted as . The single-sample correlation matrix for sample () is then estimated as follows:

(1)

To ensure topological comparability, we applied a rank-based fixed-sparsity thresholding strategy. Specifically, edge weights in the single-sample network (Network ) were standardized to z-scores using the pooled distribution of all samples. Network sparsity was controlled via a significance level (default 0.05). The reference network (Network ) then retained the top-ranked edges by absolute z-score to exactly match the average density of these sparsified disease networks (Fig. 1A). This ensures that observed topological differences arise from wiring patterns rather than density variations.

Quantification of topological rewiring (eNetShift)

We postulate that a driver metabolite undergoes a significant shift in its local interaction partners (Fig. 1B). To quantify this neighbor exchange, we utilize the extended Neighbor Shift (eNESH) score [16]. We divide the single-sample network and reference network into positive (+) and negative () subnetworks, denoted as and . The eNESH score for metabolite in sample is defined as:

(2)

where , , and represent the intersection and symmetric differences of the neighborhoods, quantified by their set cardinalities:

Here, denotes the set of neighbors of node in the positive subnetwork . The notation represents the number of elements of a given set, and denotes the set difference operation. A higher eNESH score indicates a more drastic reconfiguration of the metabolite’s local regulatory environment.

Driver metabolites identification

To distinguish biologically meaningful signals from stochastic noise, we implemented a robust statistical evaluation. The raw score is standardized into a z-score relative to the background distribution of the healthy reference group:

(3)

where and are the mean and standard deviation of the eNESH score estimated from the reference population. Given the large sample size of our healthy reference cohort (n = 313), we assume the background distribution of eNESH scores approximates normality. Consequently, relying on asymptotic properties, this z-score is converted to a two-sided P-value to quantify the statistical extremity of the individual rewiring event, with statistical significance determined by a predefined level α (default α = 0.05).

However, local rewiring alone does not guarantee systemic impact. To capture the global influence of a rewired node, we further calculated the Differential Betweenness Centrality [16] of metabolite between the reference network () and the single sample network () as follows:

(4)

where represents the betweenness centrality, quantifying the fraction of shortest paths passing through node .

Joint criterion for driver definition

A metabolite is defined as a driver metabolite in sample if and only if it satisfies a dual constraint (Fig. 1C): (i) Significant Local Rewiring: the eNESH score is statistically significant (), indicating a fundamental change in interaction partners; (ii) Increased Global Centrality: the differential betweenness is positive indicating the node has ascended to a more central, regulatory position in the disease network.

Results and discussion

s‌sNetShift reveals individualized driver metabolites masked by group-averaged NetShift

To evaluate the performance of the ssNetShift method in identifying individualized driver metabolites with potential clinical relevance, we conducted a comprehensive analysis using all available samples, including 313 healthy controls and 389 GC patients. For each GC patient, we applied ssNetShift to reconstruct individual-specific networks and detect personalized driver metabolites reflecting patient-specific metabolic dysregulation. In parallel, we applied the original NetShift algorithm to the same dataset to explore group-level causal mechanisms and establish a comparative baseline. It is important to note that we purposefully evaluated NetShift on the pooled dataset rather than stratifying by clinical stage or cohort. Group-level network construction requires substantial sample sizes to avoid mathematically unstable networks and spurious associations. More fundamentally, comparing our single-sample approach against a pooled baseline directly highlights ssNetShift’s unique capacity to resolve intra-group heterogeneity that is inherently masked by population averaging.

As illustrated in Fig. 2A, NetShift identified several group-level driver metabolites (Table S2) with known associations to GC, such as S-adenosylmethionine (SAM) and acetylcarnitine, but its findings obscured the unique metabolic trajectories of individual patients. In contrast, ssNetShift revealed substantial metabolic heterogeneity at the single-sample level. By calculating the driver metabolite frequency (the proportion of GC samples in which a metabolite was classified as a driver), we observed that several biologically critical metabolites were highly recurrent in ssNetShift despite being completely missed by NetShift. For example, Neopterin (frequency = 0.15) and inosine (frequency = 0.17) were frequently highlighted by ssNetShift. However, neither reached statistical significance when evaluated by the NetShift permutation test (P > .1). Neopterin is a progression-associated immunometabolite reflecting enhanced immune activation [27], while inosine is a biomarker associated with metabolic reprogramming [28], exemplifying our method’s sensitivity to key drivers overlooked by group-level metrics.

Figure 2.

Comparison of driver metabolites identified by ssNetShift and NetShift. (A) Driver metabolite identification by ssNetShift and NetShift. (B) Venn diagram illustrating the overlap between the top 24 driver metabolites identified by ssNetShift and the 24 significant driver metabolites identified by NetShift. (C) Pathway analysis of driver metabolite sets. Pathways are ranked by enrichment significance (P-value), with a minimum hit count of 2.

To systematically compare the approaches and ensure a strictly size-matched evaluation, we first established the baseline using NetShift, which identified exactly 24 significant driver metabolites under its permutation test threshold (P < .1). To prevent any bias from unequal feature space sizes and strictly avoid cherry-picking, we correspondingly selected the exact same number of top-ranked features (Top-24) prioritized by ssNetShift identification frequency. We then matched these two evenly sized feature sets (Tables S2 and S3). Venn diagram analysis (Fig. 2B) highlighted the limited overlap between the two methods, sharing only nine metabolites (Table S4). Pathway overrepresentation analysis (MetaboAnalyst v6.0) indicated that both methods implicated established GC-related pathways, such as tryptophan, nicotinate, and purine metabolism (Fig. 2C), which have been previously linked to GC pathogenesis [29–31]. Detailed enrichment criteria and pathway coverage reporting are provided in Materials S5. Importantly, because our targeted metabolomics panel (147 metabolites) inherently limits pathway coverage, these results provide supportive functional context rather than definitive evidence of pathway-wide dysregulation. In the ssNetShift framework, biological relevance does not require the coordinated dysregulation of an entire pathway. Instead, we uniquely captured perturbations in pathways like valine, leucine, and isoleucine degradation, demonstrating that the topological rewiring of a small set of central driver nodes can functionally disrupt metabolic networks even when abundance-level changes are sparse. Such key driver perturbations may be sufficient to functionally disrupt pathway activity, particularly in processes related to tumor suppression and redox balance [29, 32].

To ensure that the Top-24 drivers prioritized from the full dataset were robust biological signals and not cohort-specific artifacts (Materials S1), we performed a cross-cohort reproducibility analysis. Independent ranking across the three GC cohorts revealed 9 core drivers present in all cohorts (3/3) and 21 replicated in at least 2 cohorts (≥2/3) (Fig. S1). Furthermore, a dedicated retention analysis confirmed (Materials S2) that these globally prioritized Top-24 drivers remained highly active and top-ranked specifically within Cohort 3, the independent survival cohort (Fig. S2).

Building on this validated core, we evaluated the prognostic significance of the 24 candidates from each method using univariate Cox regression [33] on data from Cohort 3. Using a screening threshold (Wald test P < .10), ssNetShift yielded 7 survival-associated metabolites, whereas NetShift yielded 3 (Fig. 3A). For instance, NetShift identified uracil (log₁₀(HR) = 0.22, P < .1) as a risk-associated metabolite and inosine monophosphate (IMP) (log₁₀(HR) = −0.45, P < .1) as a protective factor. In contrast, ssNetShift highlighted a greater number of prognostically relevant metabolites, uncovering the potential protective effects of serotonin (log₁₀(HR) = −0.40, P < .1), phosphocreatine (log₁₀(HR) = −0.16, P < .1), and tryptamine (log₁₀(HR) = −0.43, P < .1).

Figure 3.

Prognostic performance comparison between ssNetShift and NetShift. (A) Distribution of log-transformed hazard ratios (log10 HR) derived from univariate Cox regression analysis for driver metabolites identified by ssNetShift and NetShift. P-values were obtained from two-sided Wald tests in univariate Cox proportional hazards regression. (B) Comparison of prognostic model performance measured by concordance index (C-index). Error bars indicate 95% confidence intervals obtained by bootstrap resampling (100 iterations). (C, D) Kaplan–Meier survival curves for patients stratified into high-risk and low-risk groups based on ssNetShift and NetShift derived metabolites.

We then constructed multivariable Cox proportional hazards models using an entirely unsupervised, survival-agnostic feature selection strategy to prevent data leakage and overfitting. Consistent with our size-matched comparative framework, the previously established set of 24 candidate drivers from each method was utilized as the initial input for the models. To rigorously estimate predictive accuracy and its uncertainty, we employed a stratified bootstrap procedure (100 iterations). In each iteration, samples were resampled with replacement independently within the survival and death groups to preserve clinical outcome proportions, and the concordance index (C-index) was calculated (Material S3). The ssNetShift-based model achieved a superior mean C-index (0.73, 95% CI: 0.68–0.78) compared to the NetShift-based model (0.70, 95% CI: 0.65–0.75; Fig. 3B), an improvement confirmed as statistically significant by a Wilcoxon signed-rank test across paired bootstrap iterations. For Kaplan–Meier analysis [34], we deliberately avoided post hoc optimal cutpoint hunting. Patients were stratified strictly based on the median cohort risk score derived from the Cox models. Under this conservative stratification, the ssNetShift risk score yielded a significant survival separation (Fig. 3C), whereas NetShift did not (Fig. 3D).

Finally, we systematically assessed the algorithmic robustness of ssNetShift. Recognizing that single-sample network inference via the LIONESS algorithm depends heavily on the reference group, we performed repeated random subsampling of the HC reference at 20%, 40%, 60%, and 80% fractions. Our results demonstrated the exceptional rank convergence and the Top-24 drivers remained highly stable and retained substantial overlap with the original findings even under reduced reference sizes (Figs S3 and S4). We also conducted a sensitivity analysis on network construction parameters, confirming that the identification of key driver metabolites is highly consistent across varying correlation significance thresholds (0.01, 0.05, 0.10) (Fig. S5). Detailed subsampling procedures and correlation-threshold sensitivity analyses are provided in Supplementary Materials S4.

In summary, ssNetShift leverages single-sample network topology and dynamic correlation features to overcome the limitations of conventional group-based approaches, offering a novel and robust framework to decipher metabolic heterogeneity and personalized prognosis in GC.

Topology-based ssNetShift captures survival-relevant metabolic signals beyond abundance-based single-sample analysis

To systematically benchmark ssNetShift against widely used single-sample differential analysis, we conducted a multi-dimensional comparison with PEEPs [22], focusing on methodological consistency and biological relevance. We first evaluated the concordance between driver metabolites identified by ssNetShift and differential metabolites detected by PEEPs across all GC samples. Using matched per-sample hit counts to ensure a fair comparison, we evaluated the population-level recurrence of each metabolite (Fig. 4A, Table S5). Notably, metabolites such as lactate, alanine, guanosine monophosphate (GMP), and tryptamine exhibited the highest identification frequencies in PEEPs, with GMP and tryptamine also showing high recurrence in both methods. However, the biological interpretation of these high-frequency signals differs fundamentally between the two analytical dimensions. While PEEPs successfully capture the extreme abundance fluctuations of these metabolites, often reflecting the downstream accumulation of metabolic byproducts, it remains inherently blind to their relational context. ssNetShift, in contrast, adds a crucial mechanistic layer by shifting the focus from static concentration outliers to dynamic topological rewiring. It reveals that these shared high-frequency metabolites do not merely fluctuate in quantity; they undergo a profound functional role shift, physically transitioning from peripheral nodes in a healthy state to central regulatory hubs that actively coordinate the tumor network.

Figure 4.

Comparison between ssNetShift- and PEEPs-identified metabolic drivers. (A) Scatter plot comparing metabolite identification frequencies across GC samples between ssNetShift (y-axis) and PEEPs (x-axis). Each point represents one metabolite. (B) Number of metabolites significantly associated with overall survival identified by ssNetShift and PEEPs under different top-k thresholds. Survival associations were evaluated using univariate Cox regression (P < .1). (C) Reference and single-sample networks for tryptamine in a representative patient. (D) Z-score of tryptamine in this patient shows no significant abundance change.

Biologically, disease progression in GC is often driven by a collapse in metabolic coordination, where central metabolites act as regulatory bridges or choke points. While abundance-based outliers make excellent descriptive biomarkers, they are often downstream byproducts; conversely, topologically rewired hubs represent the core control points maintaining the disease state. To illustrate this distinction, we analyzed a representative case (Patient No. 319) focusing on tryptamine (Fig. 4C and D). Interestingly, while tryptamine frequently exhibits massive abundance fluctuations across the broader GC population (as effectively detected by PEEPs), in this specific patient, PEEP-based analysis did not detect significant differences in tryptamine abundance. However, ssNetShift revealed that tryptamine underwent a dramatic topological leap, shifting from a peripheral node in the healthy reference network to a highly central node in the tumor network, establishing new high-confidence coordination with nucleotide metabolism (e.g. GMP) and energy pathways. Functionally, the topological role shifts of these metabolites highlight system-level regulatory restructuring crucial for GC progression. For instance, the marked increase in connectivity of tryptophan and tryptamine with redox-related metabolites reflects the pathological overactivation of the IDO1/TDO2-Kynurenine-AhR signaling axis. In advanced GC, the increased integration of tryptophan into the kynurenine pathway does more than deplete local tryptophan; it generates immunosuppressive metabolites that activate the Aryl Hydrocarbon Receptor (AhR) in T-cells, leading to their exhaustion and facilitating tumor immune evasion [35]. The hub status of these metabolites in our single-sample networks directly reflects this orchestrated effort to restructure the metabolic microenvironment in favor of immune tolerance. Similarly, ssNetShift prioritized nucleotide intermediates (e.g. GMP, IMP) as central hubs connecting one-carbon metabolism and glycolytic pathways. This high topological centrality captures the metabolic coordination required to sustain rapid DNA/RNA synthesis and maintain genomic integrity under high replication stress. Mechanistically, the rewiring of GMP/IMP indicates an enhanced capacity for purine salvage and de novo synthesis, which are critical for GC cell survival and have been linked to resistance against antimetabolite chemotherapies (e.g. 5-Fluorouracil) [36]. By identifying these metabolites as network centers rather than simple abundance outliers, ssNetShift highlights them as the bottlenecks of tumor anabolic metabolism, providing a prioritized roadmap for developing combination therapies that target these specific metabolic vulnerabilities. This captures a survival-agnostic topological reorganization, providing mechanistic insight missed by static concentration metrics.

Cox proportional hazards models constructed using top-k metabolites from each method demonstrated ssNetShift’s advantage in survival prediction (Fig. 4B). At k = 10, ssNetShift identified tryptophan (log₁₀(HR) = 0.284, P < .1), a metabolite associated with tumor immune suppression through the KYN–AhR axis [37], which was not selected by PEEPs. At k = 20, ssNetShift detected 10 significant survival metabolites (P < .1), including serotonin (log₁₀(HR) = −0.282), a protective metabolite linked to monoamine oxidase A-mediated apoptosis [38], whereas PEEPs identified only 4 metabolites. Pathway enrichment of these ssNetShift-derived survival metabolites highlighted Tryptophan metabolism ( = .002) and Glutamate metabolism (P = .068) (Fig. S6), both implicated in advanced GC aggressiveness [39, 40]. In contrast, PEEPs primarily enriched Pterine Biosynthesis (P = .082) and Amino Sugar Metabolism (P = .096), with limited hits for each pathway.

To rigorously evaluate the prognostic efficiency of these distinct feature sets at a standardized round-number threshold independent of the previous NetShift-matched analysis, we conducted a comparative multivariable survival analysis using the Top-20 prioritized metabolites from each method. Following backward stepwise Cox proportional hazards regression optimized via the Akaike Information Criterion (AIC), the resulting models were visualized via forest plots (Fig. S7). The ssNetShift-based model demonstrated a superior fit with a lower AIC (AIC = 340.8) compared to the PEEPs model (AIC = 344.6), indicating a more optimal balance between model complexity and goodness-of-fit. Furthermore, the final ssNetShift model retained a broader spectrum of seven robust systemic drivers, whereas the PEEPs-based model collapsed to only three abundance-based outliers (Fig. S7a and b). Stratification utilizing the Cox-derived risk score demonstrated a markedly clearer separation between survival trajectories for the ssNetShift model than the PEEPs equivalent (Fig. S7c and d).

To further elucidate the dynamic metabolic reprogramming characteristics during GC progression, we employed ssNetShift to calculate frequencies across stages I–IV and systematically validated their associations with disease progression and comparative analyses with PEEPs were also conducted. Based on tumor staging data from 389 GC patients, we computed the frequency of each metabolite identified as a driver across stages (I–IV), generating a four-dimensional frequency vector per metabolite (Table S6). Hierarchical cluster analysis revealed distinct clusters of metabolites with varying driver trajectories (Fig. S8). Cluster 2 (increasing trend) showed significant enrichment under ssNetShift (hypergeometric test, ), whereas PEEPs identified another cluster with a much weaker stage association () (Fig. 5A). Trend analysis further highlighted ssNetShift’s sensitivity: metabolites in Cluster 2 exhibited a 129% average frequency increase in late stages (III and IV), while PEEPs-derived metabolites showed only a 104% increase (Fig. 5B).

Figure 5.

Stage-associated driver frequency patterns and comparative network analysis of ssNetShift and PEEPs. (A) Hierarchical clustering of metabolites based on their ssNetShift driver-frequency profiles across tumor stages I–IV (n = 389). Each metabolite is represented by a four-dimensional vector corresponding to stages I–IV. (B) Stage-associated trends of average driver frequencies across three trajectory clusters (increasing, decreasing, and mixed). (C) Pathway enrichment analysis of stage-associated driver metabolites identified by ssNetShift and PEEPs. (D) Comparison of network centrality metrics (degree and betweenness) for metabolites uniquely identified by ssNetShift and PEEPs. (E, F) Metabolite–metabolite interaction networks identified by ssNetShift and PEEPs. Node size indicates degree and edges represent high-confidence metabolite–metabolite associations, highlighting high-degree and high-betweenness hub metabolites.

Pathway overrepresentation analysis (MetaboAnalyst v6.0) of stage-associated metabolites further highlighted method-specific biological insights. While both methods identified GC-related pathways such as tyrosine and tryptophan metabolism, which have been consistently linked to GC progression [39, 41], the underlying evidence differs. Notably, ssNetShift additionally highlighted glycine/serine metabolism and glutamate metabolism (Fig. 5C), pathways closely associated with advanced GC aggressiveness by supporting energy production, redox homeostasis, and therapeutic resistance [40, 42]. It is important to note that our targeted metabolomics panel (147 metabolites) inherently limits broad pathway coverage. Therefore, rather than claiming whole-pathway enrichment, ssNetShift identifies key driver perturbations. For instance, while PEEPs captures static concentration changes within tryptophan metabolism, ssNetShift highlighted a markedly more central network role for key metabolites in this pathway, indicating enhanced cross-pathway connectivity driven by topological reorganization. In this rewired network, perturbations at these central nodes can propagate to multiple neighboring metabolites and other interconnected pathways.

Crucially, network centrality analysis demonstrated that metabolites unique to ssNetShift possessed significantly higher node degree and betweenness compared to PEEPs-derived features (Fig. 5D, Table S7). Utilizing evidence-based STITCH database interactions for network mapping, the ssNetShift network revealed a dense regulatory environment surrounding its drivers (scaled by degree centrality), contrasting sharply with the fragmented and sparse isolated nodes prioritized by PEEPs (Fig. 5E and F). Collectively, these findings demonstrate that ssNetShift captures coordinated network rewiring rather than isolated concentration shifts, providing a superior mechanistic framework for understanding GC progression.

s‌sNetShift-derived metabolic subtypes capture stage-independent prognostic heterogeneity

To evaluate whether ssNetShift provides molecular resolution beyond clinical descriptors (e.g. TNM stage), we quantified sample-to-sample diversity and performed unsupervised subtyping based on ssNetShift driver profiles [43]. We first measured inter-sample similarity using the Jaccard index computed from binarized per-sample events (driver metabolites for ssNetShift; differential metabolites for PEEPs). ssNetShift yielded a significantly lower median Jaccard index (0.10, IQR: 0.00–0.20) than PEEPs (0.17, IQR: 0.08–0.27; Mann–Whitney U test, < .0 × 10⁻⁶; Fig. 6A), indicating that topology-based drivers capture stronger patient-specificity than abundance outliers.

Figure 6.

ssNetShift-derived metabolic subtypes capture prognostic heterogeneity beyond clinical staging. (A) Distribution of pairwise Jaccard similarity indices across GC samples for ssNetShift versus PEEPs. (B–F) Kaplan–Meier overall survival curves for ssNetShift and PEEPs subtypes (with confidence bands). (G) Stage-stratified survival analysis within TNM stage III patients showing significant survival stratification between ssNetShift subtypes. (H) Example of subtype-specific network rewiring (acetoacetate), illustrating differential topological roles across subtypes.

We then clustered patients using ssNetShift driver-frequency profiles. Silhouette analysis [44] supported a three-subtype solution (S1–S3; silhouette = 0.24), which provided higher internal separation than the best PEEPs solution (k = 2; silhouette = 0.18). Subtype S1 (n = 60) demonstrated the poorest survival, whereas S3 (n = 108) exhibited a lower-risk profile. Crucially, the distribution of clinical stages across S1 and S3 was not significantly different (early I–II versus advanced III–IV; two-sided Fisher’s exact test = .42; Fig. S9), indicating that the ssNetShift-derived metabolic subtypes are not simple reflections of conventional TNM staging. The lack of direct concordance between these metabolic subtypes and traditional staging underscores that ssNetShift captures a dimension of molecular risk that is highly orthogonal to anatomical TNM classification.

Given the small sample size of subtype S2 (n = 13), which exhibited a distinct hyper-proliferative nucleotide-hub pattern, we explicitly treat S2 as hypothesis-generating. To ensure statistical rigor and address potential over-clustering concerns, we performed a conservative two-subtype (k = 2) robustness check. In this alternative model, the high-risk S1 subtype remained strictly stable as an unchanged “anchor” cluster, while S2 and S3 merged into a single lower-risk group. This confirms that the distinct metabolic entity represented by S1 is statistically robust and independent of the chosen clustering parameter (Fig. S10). Survival analysis further supported this clinical relevance. The concordance index (C-index) for ssNetShift clusters was 0.70 (95% CI: 0.62–0.77), outperforming PEEPs at 0.62 (95% CI: 0.56–0.69) (Fig. S10). Kaplan–Meier curves, generated with conservative median-splits and including confidence bands to transparently reflect uncertainty, showed significant survival stratification between S1 and S3 (log-rank < .05), whereas PEEPs-derived subtypes were not significant (P = .29; Fig. 6B–F).

To definitively demonstrate added value beyond clinical homogeneity, we analyzed survival trajectories within TNM stage III patients. ssNetShift successfully separated stage III patients into distinct risk groups based on their metabolic subtype (S1 versus S3) (Fig. 6G). Beyond mere prognostic stratification, ssNetShift facilitates deeper mechanistic interpretation by pinpointing subtype-specific rewiring patterns. For example, acetoacetate showed a markedly higher rewiring score in S3 than S1 among stage III patients (Mann–Whitney U test, = 2.04 × 10⁻¹⁰; Fig. 6H). Representative ego-networks from patients sharing the same clinical stage but belonging to different subtypes display fundamentally distinct local metabolic neighborhoods. Acetoacetate serves as a high-degree hub in one patient while remaining topologically peripheral in another, highlighting that subtype divergence is rooted in network topology rather than static concentration changes (Fig. S11).

Functionally, summarizing subtype-wise driver frequencies revealed distinct metabolic vulnerabilities (Table S8). S1 was highly enriched in immune-associated metabolic intermediates (e.g. tryptophan, tryptamine, acetoacetate), primarily involving tryptophan metabolism and purine metabolism, which have been linked to immunosuppression in aggressive tumors [45, 46]. In contrast, the S3 profile (and preliminarily, the small-n S2 cluster) was largely nucleotide-driven, indicating heightened nucleic acid synthesis [47], ATP energy metabolism, and redox regulation [48], e.g. GMP, UMP, AMP, and inosine. While these subtype-specific patterns nominate distinct metabolic vulnerabilities, such as nucleotide metabolism targeting for S3 (and potentially S2) versus immune–tryptophan axis modulation for S1, we emphasize that the specific hyper-proliferative nucleotide-hub pattern observed in S2 must be interpreted with caution due to its limited sample size (n = 13), strictly requiring future validation in larger independent cohorts.

Collectively, our within-stage survival stratification and topology-driven mechanistic insights support ssNetShift as a powerful complementary layer of molecular stratification, providing prioritized therapeutic hypotheses beyond the resolution of traditional TNM staging.

Conclusion

In this study, we present ssNetShift, a topology-driven single-sample metabolomics framework that integrates LIONESS-based network estimation with an extended NEighbor SHift (eNESH) metric and differential betweenness to identify personalized driver metabolites. Across a multi-cohort GC dataset, ssNetShift revealed individual, survival-agnostic drivers and demonstrated superior prognostic stratification compared to group-level approaches (NetShift) and abundance-based single-sample methods (PEEPs). By prioritizing topological rewiring over static concentration changes, ssNetShift enabled clinically meaningful metabolic subtyping with within-stage (e.g. Stage III) prognostic separation, providing a molecular resolution orthogonal to conventional TNM staging.

However, several methodological and biological limitations should be explicitly acknowledged. First, the performance of LIONESS-based network estimation is inherently dependent on the quality, size, and composition of the healthy reference group. Although our extensive subsampling and correlation-threshold sensitivity analyses demonstrated high algorithmic robustness, single-sample edge estimates could still be susceptible to bias in smaller or less diverse cohorts. Second, the targeted nature of our metabolomics panel (147 metabolites) restricts comprehensive pathway coverage; thus, our pathway-level findings are intended for supportive biological interpretation, highlighting key driver perturbations, rather than serving as definitive claims of pathway-wide dysregulation. Third, the limited sample size of the specific S2 metabolic subtype (n = 13) necessitates cautious interpretation. Consequently, we currently treat the S2 findings as hypothesis-generating and have provided a conservative two-subtype (k = 2) robustness check to validate the statistical stability of our primary risk stratification. Finally, while we demonstrated strong cross-cohort reproducibility within this multi-cohort dataset, our study currently lacks fully independent external validation and wet-lab experimental confirmation. Future functional studies, such as cell culture experiments and metabolic flux assays, are required to mechanistically confirm the causal and regulatory roles of the identified network hubs (e.g. tryptamine, GMP, acetoacetate).

Overall, by bridging single-sample network topology with clinical outcomes, ssNetShift shifts the analytical focus from downstream byproducts to upstream regulatory control points. It offers a robust framework for individualized driver discovery, risk scoring, and subtype characterization in heterogeneous malignancies, ultimately providing a prioritized roadmap and highly specific hypotheses for follow-up experimental oncology and precision therapeutic interventions.

Key Points

• ssNetShift is a single-sample network framework that identifies individual-level driver metabolites by modeling dynamic metabolite correlations and network rewiring rather than static abundance changes.

• By incorporating network topology into single-sample analysis, ssNetShift overcomes the limitations of both group-level methods and univariate outlier approaches, which fail to capture metabolite–metabolite interactions.

• Applied to a gastric cancer cohort, ssNetShift identified a broader set of survival-associated driver metabolites and achieved improved prognostic stratification compared with existing methods.

• Importantly, ssNetShift enables clinically meaningful disease subtyping within the same pathological stage, revealing network-defined risk classes that are not captured by conventional staging and supporting more precise patient stratification.

Funding

The work was supported by the National Natural Science Foundation of China (82372087 and 82360363) and the Natural Science Foundation of Jiangxi province, China (20232BAB206136). K.K.C. was supported by the Malaysian Ministry of Higher Education under Fundamental Research Grant Scheme (FRGS/1/2020/WAB13/UTM/02/1). D.R. is supported by the National Cancer Institute of the National Institutes of Health, USA (P30CA015704).

Data availability

Python code of the ssNetShift algorithm is available at GitHub: https://github.com/BioNet-XMU/ssNetShift. To facilitate community adoption and reproducibility, the repository also includes a representative test dataset for quick reproduction and comprehensive documentation detailing the computational environment requirements.