Key molecular networks underlying retinol's biphasic effects on cell proliferation through multi-omics integrative analysis using genome-scale metabolic modeling Authors: Peishuang He, Guangming Xiang, Yizhen Yan, Jiangming Zhong, Yuting Liang, Peng Shu, Hongzhong Lu Published: 2027 ABSTRACT Retinol is widely used in skin anti-aging, yet its effects exhibit significant concentration dependence. However, the underlying biphasic mechanism, that is low concentrations promote cell proliferation while high concentrations are inhibitory, remains incompletely understood. Here, we employed multi-omics analysis and genome-scale metabolic models (GEMs) to systematically infer the key molecular networks underlying the distinct effects of retinol concentrations on human foreskin fibroblasts (HFFs). Our results suggest that low retinol concentrations appear to promote proliferation by activating the canonical retinoic acid signaling pathway and inducing sophisticated metabolic reprogramming. This reprogramming appears to involve prioritizing the coenzyme NADPH for retinol processing and antioxidant defense, which is associated with a compensatory suppression of NADPH-consuming cholesterol biosynthesis. This metabolic shift may foster a favorable intracellular environment for growth. Conversely, high concentrations are linked to a multi-system injury cascade. The observed key feature likely involves significant oxidative stress, associated with the buildup of toxic metabolic intermediates and a pro-inflammatory lipid storm marked by elevated leukotrienes. These stress responses align with the signatures of ferroptosis, a form of programmed cell death characterized by glutathione (GSH) defense system collapse and downregulation of the key regulator GPX4. Our findings suggest a potential mechanistic transition from a state of metabolic optimization to the activation of stress-induced injury pathways in response to varying retinol concentrations. This study provides a systems-level framework for understanding retinol's biphasic effects and highlights key regulatory nodes (e.g., DHRS3, GCLM) as potential targets for future research. These insights could contribute to developing safer and more effective retinol-based formulations in cosmetic and dermatological applications. Introduction Skin aging is a complex biological process driven by intrinsic factors (e.g., genetics) and extrinsic stressors (e.g., UV exposure). Its core pathophysiology involves the progressive deterioration of dermal structure and function [[1], [2], [3]]. Dermal fibroblasts, the predominant cell type, maintain skin mechanical strength and elasticity by regulating the dynamic balance of collagen, elastin, and other extracellular matrix (ECM) components. However, aging fibroblasts exhibit a marked decrease in proliferative capacity, biosynthetic activity, and mechanosensitivity [4,5]. This impairment stems from the dysregulation of key signaling networks, such as the inhibition of the YAP/TAZ pathway and the aberrant activation of the CCN1/TGF-β/AP-1 signaling axis, which together accelerate ECM degradation and the emergence of skin aging phenotypes [6,7]. Accumulating evidence implicates oxidative stress, mitochondrial dysfunction, and aberrant lipid metabolism as fundamental drivers of skin cell dysfunction and structural degradation [8]. Among these, ferroptosis – an iron-dependent programmed cell death driven by lipid peroxidation, is gaining increasing attention for its role in skin pathologies, fibroblast dysfunction, and oxidative stress-associated aging [9]. Hallmarks include diminished or reduced activity of glutathione peroxidase 4 (GPX4), glutathione (GSH) depletion, and toxic lipid peroxide accumulation, culminating in catastrophic membrane damage [10]. Indeed, emerging studies link ferroptosis during the skin aging process with collagen degradation, fibroblast death, and compromised ECM stability. For instance, critical environmental stressors such as ultraviolet radiation (UVR) have been shown to significantly promote the activation of ferroptosis in fibroblasts, thereby exacerbating their functional decline and dermal structural damage [11]. Retinol, a key active form of vitamin A, functions as a fundamental signaling molecule across the lifespan, precisely modulating gene expression to orchestrate key processes from embryonic patterning to adult tissue homeostasis [12,13]. It is widely recognized as the "gold standard" anti-aging agent in dermatology [14], stimulating fibroblast proliferation at low concentrations, promoting ECM repair via collagen upregulation and suppression of matrix metalloproteinases (MMPs) and CCN1 [15,16]. However, clinical utility is constrained by a critical paradox: its biological effects are highly concentration-dependent. Beyond a critical threshold, retinol shifts from a repair agent to a stressor, inducing potent inflammatory responses, cell cycle arrest, and apoptosis [[17], [18], [19]]. This stark phenotypic transition from a pro-proliferative to a pro-apoptotic state exemplifies the biphasic effect of retinol. To date, mechanistic studies on retinol were mainly focused on several isolated pathways [[20], [21], [22]], thus failing to explain this critical concentration dependency from a systemic perspective. This gap, to some extent, impedes the advancement of safer and more potent retinol-based interventions. To overcome the above bottleneck, the global analysis of cellular metabolism using systems biology procedures becomes increasingly important, among which the multi-omics integrative analysis is pivotal in profiling cellular responses to stimuli through simultaneous profiling of gene expression (transcriptomics), protein abundance (proteomics), and metabolite levels (metabolomics) [23,24]. Moreover, incorporating these multi-dimensional datasets into constraint-based genome-scale metabolic models (GEMs) could facilitate the predictive simulations that link molecular alterations to observable cellular outcomes [25]. This is particularly useful for the quantitative analysis of metabolic flux reprogramming, providing deeper insight into the core principles underlying cell fate decisions [26,27]. Indeed, such integrated approaches have already displayed immense potential in unraveling complex mechanisms in conditions like diabetes and stem cell aging [28,29]. In this study, we propose a core hypothesis: the biphasic effects of retinol—which promote cell proliferation at low concentrations but inhibit it at high ones—are driven by two distinct molecular programs. These different retinol concentration levels could exert hierarchical influences on cellular metabolism at the transcriptomic, proteomic, and metabolomic levels. To test this, we integrated transcriptomic, proteomic, and metabolomic analysis based on a genome-scale metabolic model, Human1 [30], to systematically dissect the biphasic regulatory mechanism of retinol on human foreskin fibroblasts (HFFs) proliferation. Through the integrative analysis, this study aimed to elucidate the clues of retinol's biphasic effects at both pro-proliferative and anti-proliferative concentrations, and identify the key regulatory nodes and metabolic pathways underlying the biphasic effects. The findings from this study will provide crucial molecular insights to maximize the anti-aging benefits of retinol while circumventing its potential side effects when used on human body. Results Phenotypic and molecular characterization of the biphasic regulation of HFFs proliferation by retinol To systematically characterize the concentration-dependent effects of retinol on HFF proliferation and identify optimal concentrations for subsequent multi-omics analyses, we performed a series of phenotypic and molecular assessments. First, the cell viability assays revealed that retinol exerts a pronounced biphasic, concentration-dependent regulation on HFF proliferation (Fig. 1A). At a low concentration (≤10 μg/mL), retinol significantly promoted HFF proliferation, peaking at 5 μg/mL (P < 0.001). Conversely, at a high concentration (≥30 μg/mL), retinol strongly inhibited HFF proliferation (P < 0.001).Fig. 1Biphasic regulation of HFF proliferation by retinol and ECM-related gene expression. (A) The cell viability assay displayed the effects of different concentrations of retinol on HFF proliferation. (B) Representative microscopy images illustrating the morphology and density changes of HFF cells following treatment with different concentrations of retinol. Scale bar = 1000 μm. (C) Quantitative analysis of absolute adherent cell density across the WT, A5, and A30 groups. Cell counts were measured within multiple randomly selected, fixed-size regions of interest (800 × 800 pixels) per field of view. (D) qPCR analysis of the relative expression levels of genes related to collagen synthesis (COL1A1), collagen degradation (MMP1), and upstream signaling (TGFB2) after 24 h of treatment. WT: untreated control group; A5: 5 μg/mL retinol treatment group; A30: 30 μg/mL retinol treatment group. Data are presented as mean ± SEM. ∗P < 0.05, ∗∗P < 0.01, ∗∗∗P < 0.001, ns = not significant.Fig. 1 Biphasic regulation of HFF proliferation by retinol and ECM-related gene expression. (A) The cell viability assay displayed the effects of different concentrations of retinol on HFF proliferation. (B) Representative microscopy images illustrating the morphology and density changes of HFF cells following treatment with different concentrations of retinol. Scale bar = 1000 μm. (C) Quantitative analysis of absolute adherent cell density across the WT, A5, and A30 groups. Cell counts were measured within multiple randomly selected, fixed-size regions of interest (800 × 800 pixels) per field of view. (D) qPCR analysis of the relative expression levels of genes related to collagen synthesis (COL1A1), collagen degradation (MMP1), and upstream signaling (TGFB2) after 24 h of treatment. WT: untreated control group; A5: 5 μg/mL retinol treatment group; A30: 30 μg/mL retinol treatment group. Data are presented as mean ± SEM. ∗P < 0.05, ∗∗P < 0.01, ∗∗∗P < 0.001, ns = not significant. These quantitative viability findings were directly corroborated by morphological observations (Fig. 1B) and quantitative image analysis of absolute cell density (Fig. 1C). Specifically, absolute cell counting within fixed regions of interest (ROIs) revealed that the low-concentration retinol (A5) significantly increased the adherent cell density compared to the WT group (P < 0.05), consistent with its pro-proliferative effect. In contrast, the high-concentration group (A30) exhibited a drastic reduction in absolute cell counts (P < 0.01), indicating massive cell detachment and cell death. Notably, morphometric analysis revealed no significant differences in the single-cell area or aspect ratio among the surviving adherent cells across all groups (Supplementary Fig. S1). To explore the molecular basis for this biphasic effect, we further assessed the expression patterns of core genes related to extracellular matrix (ECM) dynamics by qPCR (Fig. 1D). Consistently, the results demonstrated clear concentration-dependent regulation: the pro-proliferative concentration (5 μg/mL) significantly upregulated the expression of COL1A1 (collagen synthesis) and TGFB2 (upstream signaling). In contrast, the inhibitory concentration (30 μg/mL) significantly downregulated COL1A1 and TGFB2, while dramatically upregulating MMP1 (collagen degradation). Accordingly, 5 μg/mL was selected as the representative low-concentration group, and 30 μg/mL as the representative high concentration group for subsequent analyses to elucidate the molecular networks associated with this biphasic phenotypic transition. Concentration-dependent reprogramming of HFF transcriptional networks by retinol To investigate the concentration-dependent effects of retinol on the transcriptional landscape of HFF cells, we treated samples with three concentrations: untreated control group (WT), low concentration retinol (5 μg/mL, A5), and high concentration retinol (30 μg/mL, A30), followed by RNA sequencing for each group. The t-SNE cluster analysis confirmed high reproducibility within each group, and showed that the samples from different groups were separated into three distinct clusters (Fig. 2A). Differential expression analysis showed that, compared with the WT group, we identified 1348 significantly differentially expressed genes (DEGs) (680 down-regulated), whereas the A30 group induced a more extensive transcriptional reprogramming with 3135 DEGs (1729 down-regulated) (Fig. 2B and C). Among these, 508 genes were commonly altered in both treatment groups (Fig. 2B), suggesting that the addition of retinol has a broad effect on gene expression.Fig. 2Concentration-dependent effects of retinol on HFF transcriptome. (A) t-SNE clustering analysis of HFF transcriptomes after 24-h treatment with different retinol concentrations. Each point represents an independent biological replicate. Colors represent different treatment groups, demonstrating good reproducibility within groups and clear separation of transcriptional profiles. (B) Venn diagram presents the overlap and unique DEGs between A5 vs WT and A30 vs WT comparisons. (C) The volcano plot of DEGs for each concentration compared with the WT. Red dots indicate significantly upregulated genes, and blue dots indicate significantly downregulated genes. The selection criteria were |log2(Fold change)| ≥ 0.585 (corresponding to fold change ≥1.5 or ≤ 0.667) and an unadjusted P value < 0.05. (D) Bar plot of GO functional enrichment for DEGs from both the A5 vs. WT and A30 vs. WT comparisons. The length of each bar represents the number of enriched genes, and the color intensity reflects the enrichment significance. (E) Bubble plot of KEGG pathway enrichment analysis for DEGs from both the A5 vs. WT and A30 vs. WT comparisons. The bubble size represents the number of enriched genes, and the color intensity reflects the enrichment significance.Fig. 2 Concentration-dependent effects of retinol on HFF transcriptome. (A) t-SNE clustering analysis of HFF transcriptomes after 24-h treatment with different retinol concentrations. Each point represents an independent biological replicate. Colors represent different treatment groups, demonstrating good reproducibility within groups and clear separation of transcriptional profiles. (B) Venn diagram presents the overlap and unique DEGs between A5 vs WT and A30 vs WT comparisons. (C) The volcano plot of DEGs for each concentration compared with the WT. Red dots indicate significantly upregulated genes, and blue dots indicate significantly downregulated genes. The selection criteria were |log2(Fold change)| ≥ 0.585 (corresponding to fold change ≥1.5 or ≤ 0.667) and an unadjusted P value < 0.05. (D) Bar plot of GO functional enrichment for DEGs from both the A5 vs. WT and A30 vs. WT comparisons. The length of each bar represents the number of enriched genes, and the color intensity reflects the enrichment significance. (E) Bubble plot of KEGG pathway enrichment analysis for DEGs from both the A5 vs. WT and A30 vs. WT comparisons. The bubble size represents the number of enriched genes, and the color intensity reflects the enrichment significance. To functionally interpret these transcriptional changes, we first performed Gene Ontology (GO) enrichment analysis (Fig. 2D). We found that, under retinol treatment, several key biological processes are consistently enriched, including cell cycle, cell adhesion, and extracellular matrix (ECM)-related terms. Specifically, in the A5 group, it is observed that pathways related to Wnt signaling, cell growth, and ECM were significantly upregulated, whereas calcium ion transport, the TGF-β signaling pathway, and the ERK1 and ERK2 cascade were suppressed. This modulation of ECM-related genes aligns with prior in vitro evidence demonstrating the retinoid-dependent regulation of fibroblast matrix remodeling [31]. Conversely, in the A30 group, most of growth-related pathways, including Wnt-protein binding, cell growth, ECM-related terms were significantly suppressed, while genes related to phagocytosis and ferroptosis were strongly upregulated, suggesting a concurrent suppression of growth-related signaling and activation of oxidative stress responses at high concentrations [32]. To pinpoint the specific molecular circuits driving these biological processes, we further performed KEGG pathway enrichment analysis. This analysis systematically revealed the core network underlying retinol's biphasic effects, with the most representative pathways illustrated in Fig. 2E. Across both concentrations, pathways such as ABC transporters and retinol metabolism were significantly enriched, reflecting the fundamental cellular response to retinol uptake and processing, and the alterations of key retinol metabolic genes, such as RARB, CYP26A1, and CRABP2, are consistent with established indicators of intracellular retinoic acid signaling activity [33]. In the A5 group, we found significant activation of key pathways like focal adhesion, cell adhesion molecules, and cytochrome P450 drug metabolism. This is consistent with the enhanced cell proliferation, adhesion, and antioxidant capacity observed in the above GO enrichment analysis. Concurrently, pathways involved in steroid biosynthesis and inflammation (such as the interleukin-17 pathway) were significantly downregulated. In contrast, in the A30 group, the enriched KEGG pathways varied dramatically. Key pathways involved in cell adhesion and growth regulation, such as focal adhesion and Hippo signaling, were significantly suppressed. Conversely, pathways related to programmed cell death—including ferroptosis, necroptosis, phagosome, and glutathione metabolism—were significantly upregulated, indicating that the cells were undergoing severe stress and damage. Additionally, growth and survival pathways like cell cycle, ECM-receptor interaction, and PI3K-Akt signaling were only enriched among downregulated genes, further supporting the inhibitory effect of high concentration retinol on cell proliferation. Proteomics reveals retinol-induced remodeling of cellular functional networks To dissect the biphasic effects of retinol at the protein level, we performed a data-independent acquisition (DIA)-based proteomic analysis on HFF cells under different treatment conditions. The overall proteomic profile showed a concentration-dependent response. As observed in the t-SNE analysis, the samples from different groups were distinctly separated (Fig. 3A), and the A30 group induced more extensive changes in protein abundance than the A5 group (Fig. 3B and C).Fig. 3Concentration-dependent effects of retinol on the HFF proteome. (A) t-SNE dimensionality reduction of all proteomic samples, showing clear clustering separation among the WT, A5, and A30 groups. (B) The Venn diagram presents the overlap and unique DEPs between the A5 vs. WT and A30 vs. WT comparisons. (C) The volcano plot of differentially expressed proteins (DEPs) for the A5 vs. WT and A30 vs. WT comparisons. Red dots indicate significantly upregulated proteins, and blue dots indicate significantly downregulated proteins. Selected DEPs satisfied |log2(Fold change)| ≥ 0.585 and a P value < 0.05. (D) The bar plot of GO functional enrichment for DEPs from both the A5 vs. WT and A30 vs. WT comparisons. (E) Bubble plot of KEGG pathway enrichment for DEPs from both A5 vs. WT and A30 vs. WT comparisons.Fig. 3 Concentration-dependent effects of retinol on the HFF proteome. (A) t-SNE dimensionality reduction of all proteomic samples, showing clear clustering separation among the WT, A5, and A30 groups. (B) The Venn diagram presents the overlap and unique DEPs between the A5 vs. WT and A30 vs. WT comparisons. (C) The volcano plot of differentially expressed proteins (DEPs) for the A5 vs. WT and A30 vs. WT comparisons. Red dots indicate significantly upregulated proteins, and blue dots indicate significantly downregulated proteins. Selected DEPs satisfied |log2(Fold change)| ≥ 0.585 and a P value < 0.05. (D) The bar plot of GO functional enrichment for DEPs from both the A5 vs. WT and A30 vs. WT comparisons. (E) Bubble plot of KEGG pathway enrichment for DEPs from both A5 vs. WT and A30 vs. WT comparisons. Consistent with RNA sequencing analysis, the functional enrichment analysis based on proteomics revealed that HFFs exhibited a characteristic pro-proliferative state in the A5 group. In detail, GO enrichment analysis showed that anabolic processes, such as DNA synthesis and RNA polymerization, were activated, while cellular responses to the external stimuli and phospholipid binding activity were suppressed. KEGG pathway analysis indicated that a lower concentration of retinol induced a cellular state characterized by activation of Wnt signaling and signaling pathways regulating pluripotency of stem cells, supporting the proliferative state of HFFs. Consistent with prior experimental evidence indicating retinol's broad involvement in Wnt/TGF-β signaling [34], our proteomic dataset identified consistent alterations in core regulatory nodes such as SMAD family member 3 (SMAD3) and dishevelled associated activator of morphogenesis 2 (DAAM2) across both the A5 and A30 groups. Additionally, retinol-induced expression patterns suggested the establishment of a progenitor-like state, marked by upregulation of the polycomb repressive complex (PRC) and downregulation of ATP-dependent chromatin remodeling pathway. This pattern resembles the epigenetic features of stem/progenitor cells and may contribute to the suppression of differentiation-specific gene expression [35]. This shift toward a progenitor-like phenotype was further accompanied by reduced cytoskeleton in muscle cells and metabolic reprogramming characterized by decreased lipid biosynthesis and enhanced mineral absorption pathway. However, the cells in A30 shifted toward a stress and damage state. GO enrichment analysis revealed that a wide range of fundamental biological processes were significantly suppressed, including the Wnt signaling pathway, DNA replication, and processes related to cell growth, while pathways regulating apoptotic signaling in specific contexts were prominently engaged (Fig. 3D). KEGG pathway enrichment analysis further revealed the specific pathways involved in these dramatic changes (Fig. 3E). Upregulated proteins were significantly enriched in pathways for ferroptosis, phagosome, Hippo signaling, and antigen processing and presentation. This suggests the initiation of programmed cell death (particularly ferroptosis), activation of stress/immune responses (via Hippo and antigen presentation), and involvement of phagosomal pathways (associated with damaged cell clearance) in cells exposed to high-concentration retinol. By contrast, downregulated proteins were significantly enriched in pathways for cell cycle, signaling pathways regulating pluripotency of stem cells, Wnt signaling, ATP-dependent chromatin remodeling, and cholesterol metabolism, a mechanism consistent with recent reports that extreme retinoid exposure inhibits Wnt activity and stemness [36]. Marked up-regulation of the inflammation-related protein C1S and the stress regulator RCAN1 further supported the cytotoxic effects of high-concentration retinol. Additionally, RAB33A, a GTPase involved in Golgi-endosome trafficking and autophagosome formation, exhibited a striking reversed expression trend between the A5 and A30 groups, highlighting a biphasic reprogramming of membrane dynamics and cellular stress responses [37]. Metabolomics reveals the final functional phenotypes underlying retinol's biphasic effects To investigate the metabolic basis underlying the retinol-induced phenotypic transitions, we performed untargeted metabolomic analysis. OPLS-DA demonstrated a clear separation among the three groups at the metabolite level (Fig. 4A). A total of 531 significantly altered metabolites were identified (Fig. 4B), with the most pronounced changes observed in the A30 group, indicating high-concentration of retinol has a substantially stronger impact on the cellular metabolism compared to low-concentration of retinol.Fig. 4Concentration-dependent modulation of HFF metabolism by retinol. (A) OPLS-DA score plot of the metabolome, showing clear separation among WT, A5, and A30 groups at the metabolite level. (B) Volcano plots of differential metabolites for A5 vs WT and A30 vs WT comparisons. Differential metabolites were jointly selected based on the OPLS-DA model (VIP ≥1), t-test (P value < 0.05), and ROC >0.8. (C) Classification of differential metabolites at the superclass level (based on HMDB), illustrating changes in major categories such as lipids, amino acids, and organic acids under different treatment conditions. (D) Boxplots showing the relative abundance of key differential metabolites (e.g., phosphoethanolamine, leukotriene C4). (E) Key metabolic pathway maps, revealing metabolic reprogramming features supporting cell proliferation in A5 vs WT and multiple perturbed pathways driving ferroptosis in A30 vs WT. The pathways highlighted in red font on the left represent key pathways of focus that are primarily discussed in the main text.Fig. 4 Concentration-dependent modulation of HFF metabolism by retinol. (A) OPLS-DA score plot of the metabolome, showing clear separation among WT, A5, and A30 groups at the metabolite level. (B) Volcano plots of differential metabolites for A5 vs WT and A30 vs WT comparisons. Differential metabolites were jointly selected based on the OPLS-DA model (VIP ≥1), t-test (P value < 0.05), and ROC >0.8. (C) Classification of differential metabolites at the superclass level (based on HMDB), illustrating changes in major categories such as lipids, amino acids, and organic acids under different treatment conditions. (D) Boxplots showing the relative abundance of key differential metabolites (e.g., phosphoethanolamine, leukotriene C4). (E) Key metabolic pathway maps, revealing metabolic reprogramming features supporting cell proliferation in A5 vs WT and multiple perturbed pathways driving ferroptosis in A30 vs WT. The pathways highlighted in red font on the left represent key pathways of focus that are primarily discussed in the main text. Consistent with the transcriptomic and proteomic findings, the metabolic profile of HFFs in the A5 group exhibited a distinct pro-proliferative state. Firstly, differential metabolite analysis showed that lipids and lipid-like metabolites had significantly increased abundance (Fig. 4C), which is consistent with the high demand for cell membrane components (such as phospholipids) during cell proliferation [38]. Meanwhile, organic acids and their derivatives were generally reduced (Fig. 4C), suggesting that the cells were in a metabolically healthy state characterized by high energy efficiency and minimal accumulation of metabolic byproducts [39]. To gain deeper mechanistic insights into this metabolic pattern, we performed KEGG pathway mapping analysis (Fig. 4E). The results revealed suppressed activity in pathways critical for membrane synthesis, including glycerophospholipid and sphingolipid metabolism. In particular, the core precursor phosphorylethanolamine displayed a significant decrease in concentration (Fig. 4D). Given the proliferative phenotype observed in the A5 group, the decrease in concentration of phosphorylethanolamine may be due to the rapid consumption of the precursor for membrane phospholipid synthesis [40]. Furthermore, enhanced activity in metabolic pathways associated with essential coenzyme biosynthesis was observed (Fig. 4E), including the vitamin B6 metabolism pathway and the pantothenate and coenzyme A biosynthesis pathway. A decreased concentration in some key precursors, such as pyridoxal and dephospho-coenzyme A, suggests a potential increase in their consumption during coenzyme synthesis, possibly in response to the increased biosynthetic demand. In contrast to the orderly remodeling observed in the A5 group, the A30 group was associated with widespread metabolic dysregulation. The most prominent feature of this disruption was the pronounced imbalance in lipid metabolism, which accounted for the highest proportion of both increased and decreased metabolites (Fig. 4C), suggesting widespread perturbation of lipid homeostasis. Notably, phosphorylethanolamine accumulated significantly in the A30 group (Fig. 4D), potentially indicating membrane lipid breakdown and degradation driven by cellular stress, rather than de novo synthesis. The KEGG mapper analysis (Fig. 4E) provided insights into the potential clues underlying these perturbations. On the one hand, the arachidonic acid metabolism pathway was activated, leading to a marked accumulation of the inflammatory mediator, leukotriene C4 [41]. On the other hand, pathways closely associated with oxidative stress, such as glutathione metabolism and bile secretion involved in detoxification [42], were also activated, suggesting the cells may have been mounting defense responses against severe oxidative stress. More importantly, these dysregulated metabolic pathways showed strong associations with the known pathways of ferroptosis [43], providing supportive metabolic evidence for the programmed cell death observed in the proteomic analysis. Multi-omics data drives the pinpointing of core toxicity pathways To systematically integrate multidimensional datasets and elucidate the molecular mechanisms underlying the biphasic effect of retinol, we performed a multi-omics integrative analysis. First, the global comparison of the transcriptome and proteome revealed that extensive post-transcriptional regulation is a key feature of retinol action. The nine-quadrant plot (Fig. 5A) showed that, under both treatment concentrations, a large number of gene-protein pairs deviated from the diagonal, suggesting the induction of complex post-transcriptional regulatory patterns by retinol treatment. In the A5 group, fewer gene–protein pairs exhibited consistent changes, and no statistically significant pathway enrichment was observed. Functional annotations of key molecules such as dehydrogenase/reductase 3 (DHRS3), G protein-coupled receptor class C group 5 member A (GPRC5A), inhibitor of DNA binding 1 (ID1), helicase, lymphoid specific (HELLS), and SMAD family member 3 (SMAD3) indicated their involvement in retinoic acid signaling, DNA replication, and differentiation inhibition. In contrast, the A30 group showed a significant enrichment of coordinately changed genes and proteins in several core pathways (Fig. 5B). The downregulated pathways were significantly enriched in cell-related growth, including cell cycle, cholesterol metabolism, and the Hippo signaling pathway. The upregulated pathways were enriched in ferroptosis (P = 6.60∗10−6) and phagosome (P = 2.05∗10−3).Fig. 5Multi-omics integration analysis reveals the core mechanism of ferroptosis induced by high concentration retinol. (A) Nine-quadrant correlation analysis of transcriptome and proteome data, illustrating the partial discordance between mRNA and protein abundance and suggesting widespread post-transcriptional regulation. (B) KEGG pathway enrichment analysis of gene–protein pairs exhibiting coordinated regulation in the high concentration comparison (A30 vs WT). (C) Network of multi-omics key features (genes, proteins, and metabolites) identified by the DIABLO model that distinguish the different treatment groups. (D) Integrated pathway enrichment analysis across transcriptomic, proteomic, and metabolomic data, highlighting pathways significantly perturbed under high concentration retinol.Fig. 5 Multi-omics integration analysis reveals the core mechanism of ferroptosis induced by high concentration retinol. (A) Nine-quadrant correlation analysis of transcriptome and proteome data, illustrating the partial discordance between mRNA and protein abundance and suggesting widespread post-transcriptional regulation. (B) KEGG pathway enrichment analysis of gene–protein pairs exhibiting coordinated regulation in the high concentration comparison (A30 vs WT). (C) Network of multi-omics key features (genes, proteins, and metabolites) identified by the DIABLO model that distinguish the different treatment groups. (D) Integrated pathway enrichment analysis across transcriptomic, proteomic, and metabolomic data, highlighting pathways significantly perturbed under high concentration retinol. To identify the core molecular network underlying retinol's biphasic effects, we applied the data integration analysis for biomarker discovery using latent cOmponents (DIABLO) framework [44] for multi-omics integration to discriminate between the different treatment groups. This identified a set of "multi-omics signatures" that contributed most to distinguishing the different phenotypes across all omics layers (Fig. 5C). This multi-omics network contains key antioxidant proteins such as heme oxygenase 1 (HMOX1) and glutathione peroxidase 4 (GCLM), pro-inflammatory metabolites such as leukotriene C4, and γ-glutamylcysteine as the synthetic precursor of glutathione (GSH), suggesting a central role of oxidative stress. Additionally, the KEGG pathway enrichment analysis revealed distinct biological signatures for each treatment condition (Fig. 5D). In the A5 vs WT group, there were only minor cellular perturbations. Pathways such as retinol metabolism and sphingolipid signaling pathway primarily enriched at the metabolomic level (P < 0.05). In the A30 vs WT group, the result showed strong and significant pathway enrichment across multiple omics layers. Ferroptosis previously identified as significantly upregulated in the transcriptomic data (Fig. 2D and E), also showed significant enrichment at both proteomic (P < 0.01) and metabolomic levels (P < 0.05). Glutathione metabolism was enriched across three omics layers and reached significance at both the transcriptomic (P < 0.05) and metabolomic (P < 0.05) levels. Furthermore, other pathways, such as cysteine and methionine metabolism, one carbon pool by folate, and tryptophan metabolism, were also enriched across three omics layers but showed significant enrichment in only a single omics layer (i.e., achieving P < 0.05 in only one type of omics dataset). GEMs uncover metabolic flux remodeling: a model-based framework for retinol's biphasic effect To achieve a comprehensive understanding of cellular function beyond isolated omics layers, we constructed context-specific GEMs by integrating transcriptomic data with the human GEM - Human1 [30]. Structural differences among the models were evaluated using the ftINIT algorithm [45]. To identify key metabolic reactions, we performed a reporter metabolite analysis with Kiwi [46], which helped elucidate the relevant metabolites and their associated genes. Meanwhile, we also conducted flux sampling with RIPTiDe [47] to characterize the alterations in flux redistribution induced by retinol (Fig. 6A).Fig. 6Genome-scale metabolic models (GEMs) revealed the metabolic flux remodeling driving retinol's biphasic effect. (A) The workflow diagram for constructing context-specific GEMs based on the transcriptomic data. (B) Comparison of metabolic subsystems in different context-specific GEMs shows the difference in network structure, such as the specific activation of eicosanoid metabolism in the A30 group. (C) Key subsystems with their associated gene-sets identified by Kiwi [40]. (D) Key metabolic reactions with the highest contribution to discriminating different treatment groups, identified by Random Forest modeling based on flux sampling through Riptide [41]. The red text highlights Beta oxidation of fatty acid as the dominant high-contribution reaction within fatty acid oxidation in the A30 vs WT comparison.Fig. 6 Genome-scale metabolic models (GEMs) revealed the metabolic flux remodeling driving retinol's biphasic effect. (A) The workflow diagram for constructing context-specific GEMs based on the transcriptomic data. (B) Comparison of metabolic subsystems in different context-specific GEMs shows the difference in network structure, such as the specific activation of eicosanoid metabolism in the A30 group. (C) Key subsystems with their associated gene-sets identified by Kiwi [40]. (D) Key metabolic reactions with the highest contribution to discriminating different treatment groups, identified by Random Forest modeling based on flux sampling through Riptide [41]. The red text highlights Beta oxidation of fatty acid as the dominant high-contribution reaction within fatty acid oxidation in the A30 vs WT comparison. By comparing subsystems and reporter metabolites in context-specific models constructed by ftINIT, we identified distinct concentration-dependent metabolic patterns (Fig. 6B and C). In the A5 group, vitamin C metabolism, associated with enhanced antioxidant capacity, was activated. Reporter metabolite analysis identified cholesterol biosynthesis as a central module of the metabolic network, which was consistently suppressed across its three constituent pathways, accompanied by a reduction in their precursor metabolites. In contrast, in the A30 group, retinol induced a global reduction in subsystem activity, notably in biopterin metabolism, ubiquinone synthesis, and several amino acid metabolic pathways. Reporter metabolite analysis highlighted fatty acid oxidation, leukotriene metabolism, and cholesterol biosynthesis as central nodes in the metabolic network. Interestingly, subsystems associated with inflammation, such as eicosanoid metabolism [48], were specifically activated in A30, underscoring the contribution of inflammatory stress to high concentration retinol-induced metabolic dysregulation and aligning with the pathways described above. To characterize metabolic state transitions, flux sampling was performed with Riptide [47]. Principal Coordinate Analysis (PCoA) based on flux distributions clearly separated the models into three groups (Supplementary Fig. S2). To identify the most influential reactions between different states, we constructed a Random Forest model to evaluate the contribution of individual metabolic reactions (Fig. 6D). In A5 vs WT, reactions with the highest contribution were mainly related to efficient cellular signaling and mitochondrial β-oxidation of fatty acids, which are involved in energy production [40]. Specifically, “11-cis-retinol:NAD + oxidoreductase”, which encodes a reaction belonging to “retinol metabolism” and functionally linked with multiple antioxidant enzymes, emerged as a key reaction based on the above contribution analysis. This may hint that HFFs exposed to low-concentration retinol could respond normally to maintain homeostasis. In the A30 vs WT comparison, the set of high-contribution reactions identified by the Random Forest model shifted, becoming dominated by fatty acid β-oxidation and related reactions in lipid metabolism and oxidation-reduction, indicating that the metabolic shifts in the A30 group are centered on pathways linked to lipid peroxidation and oxidative stress. Integration of flux analysis with multi-omics data suggests a shift in fatty acid β-oxidation from mitochondria to peroxisomes in A30-treated cells, potentially as an adaptive response to oxidative stress. This pattern was consistent with the findings reported in oxidative stress models [49], and aligned with our multi-omics analysis. Discussion This study employed an integrative systems biology approach, combining multi-omics analysis with GEMs to elucidate retinol's biphasic effects on proliferation of HFFs. The results suggest that the concentration-dependent effects of retinol on HFF proliferation may be governed by a metabolic transition: at low concentrations, retinol appears to optimize the metabolic network to support signal transduction and promote proliferative capacity. In contrast, at high concentrations, it appears to provoke metabolic dysregulation and oxidative stress, potentially leading to ferroptosis. At low concentrations, retinol seems to promote cell proliferation by inducing an NADPH-centric metabolic reprogramming (Fig. 7). Driven by a robust activation of retinol-specific signaling (DHRS3, CRABP2), this reprogramming may involve a dual metabolic adjustment: on one hand, the cell likely enhances the NADPH regeneration by upregulating G6PD and PGD (Fig. 7A and B), the rate-limiting enzyme of the pentose phosphate pathway (PPP). This is further supported by the downregulation of the competing enzyme AKR1B1 and the high weight assigned to the pentose and glucuronate interconversions pathway in our metabolic model. On the other hand, to prevent the exhaustion of this critical resource, cells actively suppress major NADPH-consuming pathways. Specifically, both multi-omics profiling and targeted qPCR confirmed a dramatic transcriptional suppression of central cholesterol biosynthesis hubs, including HMGCR and LDLR (Fig. 7A and B). This dual metabolic mechanism—boosting NADPH synthesis while curtailing its consumption in lipid metabolism—may contributes to the maintenance of a robust NADPH pool [50], which in turn likely enhances both the antioxidant systems [51] and retinoic acid (RA) metabolism [52], ultimately likely contributing to a favorable redox microenvironment for cell proliferation (Fig. 7C).Fig. 7Transcriptional validation of key metabolic hubs and the proposed working model of retinol-induced biphasic effects in HFF cells. (A) Heatmaps illustrating the global transcriptomic trends of representative genes involved in retinol signaling, NADPH regeneration, lipid metabolism (top), antioxidant defense, and ferroptosis/lipid peroxidation (bottom) across the WT, A5, and A30 groups. (B) qPCR validation of core regulatory genes orchestrating the metabolic transition and ferroptotic crisis. Data are presented as mean ± SEM. Statistical significance was determined using one-way ANOVA (∗P < 0.05, ∗∗P < 0.01, ∗∗∗P < 0.001, ns = not significant. (C) Schematic summary of the two divergent cellular states suggested by our integrated multi-omics analysis. Left: The metabolic optimization state associated with low-concentration (A5) retinol. This pro-proliferative phenotype is characterized by the upregulation of retinol metabolism (DHRS3, CRABP2) and the pentose phosphate pathway (PPP) (e.g., G6PD, PGD), suggesting an enhancement of NADPH production. This is concurrent with the downregulation of NADPH-consuming pathways such as lipid metabolism (HMGCR, LDLR, SQLE) to conserve the NADPH pool. Right: The cytotoxic cascade state associated with high-concentration (A30) retinol. The damage phenotype is characterized by a multi-system failure signature, including severe oxidative stress (ROS) and the activation of the inflammatory pathway (e.g., leukotriene C4). The cells initiate a compensatory glutathione (GSH) self-rescue response (GCLM/GCLC, SLC7A11); however, the overwhelming lipid peroxidation (driven by ACSL4) and the critical decline of the primary defense enzyme GPX4 ultimately potentially precipitate ferroptosis. Solid red/blue arrows indicate an observed upregulation/downregulation of a gene, metabolite, or pathway in this study; dashed arrows represent proposed regulatory relationships strongly suggested by the integrated multi-omics data that require further direct functional validation. Abbreviations: G6PD, glucose-6-phosphate dehydrogenase; HMGCR, 3-hydroxy-3-methylglutaryl-CoA reductase; LDLR, low-density lipoprotein receptor; SQLE, squalene epoxidase; ROS, reactive oxygen species.Fig. 7 Transcriptional validation of key metabolic hubs and the proposed working model of retinol-induced biphasic effects in HFF cells. (A) Heatmaps illustrating the global transcriptomic trends of representative genes involved in retinol signaling, NADPH regeneration, lipid metabolism (top), antioxidant defense, and ferroptosis/lipid peroxidation (bottom) across the WT, A5, and A30 groups. (B) qPCR validation of core regulatory genes orchestrating the metabolic transition and ferroptotic crisis. Data are presented as mean ± SEM. Statistical significance was determined using one-way ANOVA (∗P < 0.05, ∗∗P < 0.01, ∗∗∗P < 0.001, ns = not significant. (C) Schematic summary of the two divergent cellular states suggested by our integrated multi-omics analysis. Left: The metabolic optimization state associated with low-concentration (A5) retinol. This pro-proliferative phenotype is characterized by the upregulation of retinol metabolism (DHRS3, CRABP2) and the pentose phosphate pathway (PPP) (e.g., G6PD, PGD), suggesting an enhancement of NADPH production. This is concurrent with the downregulation of NADPH-consuming pathways such as lipid metabolism (HMGCR, LDLR, SQLE) to conserve the NADPH pool. Right: The cytotoxic cascade state associated with high-concentration (A30) retinol. The damage phenotype is characterized by a multi-system failure signature, including severe oxidative stress (ROS) and the activation of the inflammatory pathway (e.g., leukotriene C4). The cells initiate a compensatory glutathione (GSH) self-rescue response (GCLM/GCLC, SLC7A11); however, the overwhelming lipid peroxidation (driven by ACSL4) and the critical decline of the primary defense enzyme GPX4 ultimately potentially precipitate ferroptosis. Solid red/blue arrows indicate an observed upregulation/downregulation of a gene, metabolite, or pathway in this study; dashed arrows represent proposed regulatory relationships strongly suggested by the integrated multi-omics data that require further direct functional validation. Abbreviations: G6PD, glucose-6-phosphate dehydrogenase; HMGCR, 3-hydroxy-3-methylglutaryl-CoA reductase; LDLR, low-density lipoprotein receptor; SQLE, squalene epoxidase; ROS, reactive oxygen species. In contrast, the cytotoxicity of high concentration retinol manifests as multi-systemic disruption potentially driven by metabolic chaos. Phenotypically, our quantitative morphometric analysis implies that this toxicity is characterized by rapid cell detachment rather than gradual morphological degradation. Interestingly, high-concentration retinol (30 μg/mL) caused a significant reduction in overall cell density, with no significant differences in the single-cell area or aspect ratio among the surviving adherent cells (Supplementary Fig. S1). We postulate that this phenomenon reflects a survivor bias in our adherent cell model. Cells undergoing severe retinol-induced toxicity likely round up and detach from the culture surface, escaping the focal plane of brightfield imaging. This rapid loss of cellular adhesion and subsequent rounding is a well-documented morphological hallmark of severe cytotoxicity and programmed cell death, including ferroptosis [53,54], and is consistent with classic cytotoxic phenotypes observed under high-dose retinol exposure in dermal fibroblasts [32]. Traditional explanations have often attributed this phenotype to nonspecific inflammation or general stress [55], whereas our findings suggest that the oxidative stress may play a central role in the ferroptosis associated with high-concentration retinol (Fig. 7C). In support of this notion, multi-omics and GEMs analyses (Fig. 6C) revealed significant activation of pathways associated with lipid damage and NADPH consumption, including fatty acid oxidation, glycerophospholipid metabolism, and eicosanoid metabolism. Notably, the activation within eicosanoid metabolism involved a metabolic diversion toward pro-inflammatory leukotrienes (LTC4), providing a potential molecular basis for retinol-induced inflammation [48]. This observed pathway hyperactivation likely imposes critical pressure on cellular redox homeostasis, leading to NADPH imbalance. Such an imbalance may be a factor triggering a critical cascade: concomitant with oxidative stress, cells appear to activate a self-rescue program, supported by the highly significant compensatory upregulation of GCLM and the cystine/glutamate antiporter SLC7A11 to boost GSH synthesis (Fig. 7A and B). However, the concurrent upregulation of ACSL4 likely drives the accumulation of toxic lipid peroxides, which, combined with the severe NADPH decline, severely impairs GSH regeneration. This potential bottleneck appears to undermine the cell's antioxidant defenses, correlating with the collapse of the core SLC7A11/GSH/GPX4 defense axis – the central ferroptosis defense axis(Fig. 7A and B). Consequently, dysfunction of the GSH/NADPH axis might sensitize cells to ferroptosis and exacerbate overall retinol-induced cytotoxicity. Our study also uncovers a potential link between epigenetics and aging. High-concentration retinol significantly downregulated KMT5A expression, an epigenetic regulator and ferroptosis-protective gene [56]. This points to an attractive hypothesis: retinol may initiate a deleterious feedback loop in which its induced inflammation suppresses KMT5A, which in turn increases ferroptosis susceptibility, leading to further cellular damage that perpetuates the inflammation. This perspective provides a novel molecular framework for understanding both high-concentration retinol cytotoxicity and skin aging. We propose that this KMT5A feedback loop is one of the key mechanisms driving the toxic outcome in the high-concentration arm of this bifurcation of fate. Furthermore, our data revealed several other key regulatory molecules that collectively orchestrate this concentration-dependent decision. For instance, DHRS3 may serve as a key node in the retinol-induced metabolic cycle, connecting NADPH metabolism to signal transduction and antioxidant defenses [57]. GCLM, representing a critical node that gauges cellular stress and probably determines progression toward ferroptosis, may serve as a therapeutic target to alleviate retinol-induced stress [58]. Moreover, we observed significant inhibition of Hippo signaling pathway under high concentration of retinol, implying that down-regulation of YAP/TAZ activity may synergize with metabolic disorder to suppress cell proliferation [59]. Notably, this extreme concentration sensitivity appears to be a general principle in skin biology, aligning with recent findings in hair follicle stem cells [60]. The innovation of this study lies in its systematic paradigm, shifting the understanding of retinol's effects from limited pathway analysis to global metabolic profiling based on genome-scale modeling, which could potentially be applied to other skin cell contexts. As our findings are mainly based on omics integrative analysis, and lack direct targeted quantification of intracellular metabolites, further functional validations, such as experiments with ferroptosis inhibitors, gene knockdowns, and endocytic pathway interventions [61], or confirmation in more complex in vivo or 3D skin models [62,63], are needed. In sum, this study provides a systems-level insight into the potential molecular basis for retinol's biphasic effects. Our findings highlight a narrow concentration window where cellular responses seem to shift from adaptive metabolic optimization toward states associated with cytotoxic injury. We propose that these insights could serve as valuable references for the evidence-based determination of retinol concentrations in cosmetic formulations. Therefore, rather than following the conventional assumption that higher concentration equals greater efficacy, we suggest that future formulation strategies may benefit from incorporating data-driven, low-concentration approaches to optimize efficacy while mitigating potential risks— an approach potentially aligned with the latest regulatory guidance in the European Union [64]. Materials and methods Cell culture with retinol treatment Human foreskin fibroblasts (HFFs; Chinese Academy of Sciences Stem Cell Bank) were routinely cultured in DMEM supplemented with 10% fetal bovine serum (Gibco) and 1% penicillin-streptomycin (Gibco) at 37 °C in a humidified incubator with 5% CO2. The culture medium was replaced every 2-3 days, and cells were passaged or cryopreserved when reaching 80-90% confluence. Retinol (Sigma-Aldrich) was first dissolved in dimethyl sulfoxide (DMSO) to create a 25 mg/mL stock solution, which was stored at −20 °C in the dark. For experiments, the stock solution was diluted with complete culture medium to achieve the desired working concentrations (0, 5, 10, 25, 30, 60 μg/mL), the final DMSO concentration in all treatment groups was less than 0.5%, a concentration which does not significantly affect cell viability or gene expression in human cell lines, including dermal fibroblasts [65,66]. The selected DMSO concentration was also validated in preliminary experiments to confirm the absence of cytotoxic effects under our assay conditions. Cell proliferation assay HFF cells in the logarithmic growth phase were seeded into 96-well plates at a density of 4000 cells per well. After 24 h of incubation, the medium was replaced with fresh medium containing different concentrations of retinol. Each group was set up with three biological replicates and four technical replicates. After a further 24 h of incubation, 10 μL of CCK-8 reagent (Dojindo) was added to each well and incubated for 1.5 h in the incubator. The absorbance (OD) at 450 nm was measured using a microplate reader (BioTek, USA). Cell viability was calculated as: cell viability (%) = (A_experimental_group - A_blank)/(A_control_group - A_blank) ∗ 100%. Quantitative analysis of cell density and morphology To quantify cell density and overcome the limitations of automated segmentation in highly confluent brightfield images, a fixed region of interest (ROI) sampling method was employed using ImageJ software (NIH, Bethesda, MD, USA). For each treatment group, multiple independent 800 × 800 pixel ROIs were selected per field of view. The absolute number of adherent cells within each ROI was manually counted using the Multi-point tool to assess cell density. Furthermore, to evaluate morphological alterations at the single-cell level, 30 representative individual cells with distinct boundaries were randomly selected and manually traced across different fields of view for each group. Two key morphological parameters were calculated: (1) Cell area (μm2), to evaluate potential cell shrinkage, and (2) Aspect ratio (AR), defined as the ratio of the major axis to the minor axis of the fitted ellipse, to quantify the degree of cell elongation versus rounding. Real-time quantitative PCR analysis HFF cells were seeded in 6-well plates at an appropriate density. Once cell confluence reached 70%, the medium was replaced with culture medium containing either 5 μg/mL (low concentration) or 30 μg/mL (high concentration) of retinol. An untreated control group was also included. After 24 h of treatment, total RNA was extracted from the cells using Trizol reagent (Invitrogen). RNA quantity and purity were measured using a NanoDrop spectrophotometer. Next, 0.5 μg of total RNA was reverse-transcribed into cDNA using the All-in-One First-Strand kit. The qPCR reaction was performed on a qTOWER3 series real-time fluorescence quantitative gene amplification instrument, using 2 × Universal SYBR Green qPCR Premix. The reaction system had a total volume of 20 μL, and the protocol was as follows: an initial denaturation at 95 °C for 3 min, followed by 40 cycles of 95 °C for 15 s and 60 °C for 1 min. Each sample was run in three technical replicates. GAPDH was used as the internal reference gene, and the relative expression of the target genes was calculated using the 2−ΔΔCt method. The primer sequences used in this study are detailed in Supplementary Table S1. Multi-omics sample preparation and sequencing HFF cells were seeded into 10-cm culture dishes and grown until reaching approximately 70% confluence. Cells were then treated with 5 μg/mL retinol (low-concentration group, A5) or 30 μg/mL retinol (high-concentration group, A30). The control group (wild-type cell, WT) received an equal volume of culture medium containing DMSO. After 24 h of treatment, cells were harvested and washed twice with pre-cooled PBS. The cell pellets were snap-frozen in liquid nitrogen and stored at −80 °C. Samples for transcriptomics, proteomics, and metabolomics were all prepared following this protocol, with biological replicates (n = 3 or 4 per group), and were submitted to Novogene Co., Ltd. (Beijing, China) for subsequent sequencing and analysis. Transcriptomic data analysis RNA integrity was assessed using the Agilent 2100 Bioanalyzer system. mRNA was enriched from total RNA using poly-T oligo-attached magnetic beads and fragmented under elevated temperature. First- and second-strand cDNA were synthesized, followed by end repair, A-tailing, adaptor ligation, and PCR amplification. Libraries with insert sizes of approximately 370–420 bp were selected using the AMPure XP system and quality-validated prior to sequencing. Cluster generation was performed using the TruSeq PE Cluster Kit, and libraries were sequenced on the Illumina NovaSeq 6000 platform to generate 150 bp paired-end reads (PE150). Raw reads were processed using fastp to remove adaptor sequences, poly-N reads, and low-quality reads (Supplementary Table S2), and the clean reads were aligned to the human reference genome (GRCh38) using HISAT2 (v2.1.0). Gene expression was quantified as raw read counts using HTSeq (v0.6.0) and subsequently normalized to fragments per kilobase of transcript per million mapped reads (FPKM). Differentially expressed genes (DEGs) were identified with the R package limma [67], applying the thresholds |log2(Fold change)| ≥ 0.585 (corresponding to FC ≥ 1.5 or ≤ 0.667) and an unadjusted P value < 0.05. This threshold was selected to capture subtle but biologically significant metabolic reprogramming and coordinated expression shifts in metabolic enzymes, a standardized criterion widely accepted for studying mild physiological regulatory effects [68,69]. Proteomic data analysis Proteomic profiling was performed using a data-independent acquisition (DIA) strategy on a Thermo Orbitrap Astral mass spectrometer coupled to a Vanquish™ Neo UHPLC system. Proteins were extracted, trypsin-digested, desalted, and subjected to LC–MS/MS analysis following standard protocols. Raw data were processed using DIA-NN software for spectral library searching against the UniProt human protein database. Carbamidomethylation of cysteine was set as a fixed modification, while methionine oxidation and N-terminal acetylation were treated as variable modifications. Peptide and protein identifications were filtered at a false discovery rate (FDR) < 1%. Differentially expressed proteins (DEPs) were identified using the R package limma [67], with the same thresholds as applied for transcriptomic data (|log2(Fold change)| ≥ 0.585 and an unadjusted P value < 0.05), ensuring consistent identification of moderate but functional alterations across the multi-omics layers [68]. Metabolomic data analysis Untargeted metabolomic profiling was performed using liquid chromatography–mass spectrometry (LC–MS). Raw UHPLC–MS/MS data were processed using the XCMS package in R for peak detection, alignment, missing value imputation, and quantification. Metabolite identification was conducted by matching accurate mass (mass deviation ≤10 ppm) and MS/MS spectra against public databases, including KEGG, HMDB, and LIPID MAPS. Background ions were removed based on blank samples. Signal intensities were normalized using QC-based normalization, and metabolites with a coefficient of variation (CV) > 30% in QC samples were excluded to ensure data quality. The resulting relative peak areas were used for downstream statistical analyses. Orthogonal partial least squares discriminant analysis (OPLS-DA) was employed to evaluate group separation [70]. Differential metabolites were selected based on the criteria of P value < 0.05, VIP >1, and ROC >0.8. Pathway enrichment and network analysis Differentially expressed genes, proteins, and metabolites were subjected to Gene Ontology (GO) and Kyoto Encyclopedia of Genes and Genomes (KEGG) pathway enrichment analysis using the clusterProfiler R package [71]. Pathways categorized under 'Human Diseases' were excluded to focus the functional interpretation specifically on core metabolism and cellular processes. A P value < 0.05 was considered statistically significant for enrichment. Multi-omics integration and statistical analysis To integrate the multi-omics data, several methods were used. Transcriptome-proteome integration, mRNA-protein expression relationships were evaluated using nine-quadrant plots and Spearman correlation analysis. Three-omics integration. To extract combined molecular features discriminating different treatment groups, the machine learning framework DIABLO (Data integration analysis for biomarker discovery using latent cOmponents) [44,72] was employed to identify combined molecular features that contributed most to discriminating different treatment groups. Pathway integration, KEGG pathways served as the framework for mapping multi-omics data, and visualization was performed using tools such as Pathview [73]. Network construction, the basic protein-protein interaction (PPI) networks were constructed using the STRING database [74], and further correlation networks of key molecules were constructed using Cytoscape [75]. Genome-scale metabolic model (GEM) analysis Using Human1 model as a template, we integrated our transcriptomic data via the ftINIT algorithm to build context-specific HFF metabolic models in a MATLAB 2024b environment with the Gurobi solver (Gurobi Optimization, LLC) [30,45]. We then used the RAVEN Toolbox [76] to analyze the models' structure, including their subsystem coverage. Next, we performed GSA analysis using the piano R package [77], and predicted key reactions and metabolites with gene-set analysis by Kiwi [68]. Meanwhile, we employed the RIPTiDe algorithm in a Python environment to perform 1000 flux samplings for each model, predicting the metabolic flux distribution of each reaction [47]. The resulting flux data was then subjected to Principal Coordinate Analysis (PCoA) using the vegan R package [78] and the ape R package [79] Finally, we built a random forest model using the caret R package [80] to identify the key reaction fluxes that contributed most to discriminating between the different cellular states. Statistical analysis All statistical analyses and data visualizations were performed using R (v4.4.3), Python (v3.12), or GraphPad Prism (v9.0.0; La Jolla, CA, USA). A P value < 0.05 was considered statistically significant unless otherwise specified. For the multi-omics datasets, dimensionality reduction was performed using t-distributed Stochastic Neighbor Embedding (t-SNE) [81] for transcriptomic and proteomic data, and Principal Coordinate Analysis (PCoA) [78,79] based on Bray-Curtis distances for metabolic flux data. For metabolomic data, Orthogonal Partial Least Squares Discriminant Analysis (OPLS-DA) [70] was employed to assess group separation. Differential expression analysis for transcriptomic and proteomic data was conducted using the limma R package [67], which employs a moderated t-test. The criteria for significant differential expression were an absolute log2(Fold change) > 0.58 and a P value < 0.05. Differential metabolites were identified based on a Variable Importance in Projection (VIP) score >1 from the OPLS-DA model, a Receiver Operating Characteristic (ROC) curve area >0.8, and an unadjusted P value < 0.05 from a Student's t-test. Correlation analyses between molecular features were calculated using Spearman's rank correlation coefficient using the cor() function in base R. For cellular assay data, values are presented as the mean ± SEM (n = 3). Differences between two groups were assessed using a two-tailed Student's t-test assuming unequal variances. CRediT authorship contribution statement Peishuang He: Writing – review & editing, Writing – original draft, Visualization, Software, Methodology, Formal analysis, Data curation. Guangming Xiang: Writing – review & editing, Writing – original draft, Validation, Supervision, Methodology, Investigation, Data curation, Conceptualization. Yizhen Yan: Supervision, Resources, Project administration, Investigation, Data curation. Jiangming Zhong: Resources, Project administration, Investigation, Data curation. Yuting Liang: Resources, Project administration, Investigation, Formal analysis. Peng Shu: Supervision, Resources, Project administration, Funding acquisition, Formal analysis, Conceptualization. Hongzhong Lu: Writing – review & editing, Supervision, Software, Resources, Project administration, Funding acquisition, Data curation, Conceptualization. Code and data availability All data generated or analyzed during this study are included in this published article (and its supplementary information files). All codes used are available at GitHub repository (https://github.com/hongzhonglu/Human-GEM_Antiage). The Human1 model is available on GitHub (https://github.com/SysBioChalmers/Human-GEM). Declaration of competing interest The authors declare the following financial interests/personal relationships which may be considered as potential competing interests: Yizhen Yan, Jiangming Zhong, Yuting Liang and Peng Shu are currently employed by Shenzhen Hujia Technology Co., Ltd. The research project is funded by HBN Research Institute and Biological Laboratory.