Spatial membrane lipid remodelling encodes epithelial lineage and state in human colon cancer Authors: Ramon M Rodriguez, Karim Pérez-Romero, Albert Maimó-Barceló, Teresa Ximelis, Cristina Huergo-Baños, Marcos A Martínez, Juan José Segura-Sampedro, Rafael Morales-Soriano, Cristina Pineño-Flores, Álvaro García-Granero, Sebastián Jerí, Ana M Aransay, José A Fernández, Gwendolyn Barceló-Coblijn Published: 2026 ABSTRACT SummaryBackgroundCancer progression is driven by spatially organised intratumoural heterogeneity, yet the molecular basis of this organisation remains incompletely understood. While spatial transcriptomics has advanced rapidly, the spatial architecture and functional relevance of tumour lipid metabolism remain poorly characterised. We aimed to resolve whether membrane phospholipid remodelling is spatially organised across epithelial cell states in human colon cancer.MethodsWe developed an integrative spatial multi-omic strategy linking mass spectrometry imaging to transcriptome-defined cellular states on consecutive tissue sections from primary and intraperitoneal metastatic human colon cancer specimens. Epithelial lipid phenotypes were characterised, validated in an independent cohort, and linked to transcriptional programmes.FindingsWe identified four spatially organised epithelial lipid phenotypes defined by membrane phospholipid composition. These phenotypes aligned with transcriptional programmes reflecting epithelial lineage and differentiation state, and were organised along a differentiation axis marked by coordinated remodelling of arachidonic acid–, oleic acid–, and linoleic acid–containing phospholipids. Secretory-associated states, enriched in metastatic lesions, showed increased DHA-containing phospholipids and activation of peroxisomal pathways. Lipid phenotype assignments were reproducible across an independent cohort.InterpretationMembrane phospholipid composition encodes biologically meaningful epithelial cell states within human tumours, establishing lipid remodelling as a structured and spatially organised dimension of tumour heterogeneity.FundingThis study was supported by the ISCIII (PI19/00002, PI24/00313), the Health Research Institute of the Balearic Islands (IMP22/04), the Government of the Balearic Islands (AP_2021_001), the Basque Government (IT1491-22), the European Regional Development Fund, and the Scientific Foundation, Spanish Association Against Cancer (INVES222995RODR).Keywords: Spatial lipidomics, Mass spectrometry imaging, Colorectal cancer, Intratumoural heterogeneity, Phospholipid remodelling Research in context Evidence before this study It is increasingly clear that the phospholipid composition of the lipid membrane is associated with cell identity and function. In the context of colon epithelium, mass spectrometry imaging has shown that the balance between arachidonic acid– and oleic acid–containing phosphatidylinositols and plasmalogen species tracks the process of maturation along the crypt axis. In pathological conditions, such as chronic inflammation and early malignant transformation, this natural gradient along the colonic crypt is disrupted. Moreover, bulk lipidomic studies have identified tumour-associated lipid alterations and implicated fatty-acid metabolism in disease progression. However, bulk analysis is limited due to the masking of cellular heterogeneity inherent to solid tumours. Therefore, it is currently unknown how membrane phospholipid composition is spatially organised in colon cancer and whether lipidomic heterogeneity is associated with defined epithelial cell states. Added value of this study Here we utilise high-resolution mass spectrometry imaging and spatially resolved transcriptomics on consecutive colon cancer tissue sections. This methodology allowed us to identify four spatially organised epithelial lipid phenotypes in human colon cancer that recapitulate the epithelial differentiation hierarchy, spanning stem-like, transit-amplifying, absorptive, and secretory states. Interestingly, these phenotypes are organised along a differentiation axis defined by coordinated remodelling of arachidonic acid–, oleic acid–, linoleic acid–, and docosahexaenoic acid–containing phospholipids. In the particular case of intraperitoneal metastases, we observed a selective activation of a peroxisomal–DHA–plasmalogen programme, suggesting a link between tumour dissemination and membrane composition. These findings establish that intratumoural lipid heterogeneity is a biologically interpretable dimension of epithelial cell identity. Implications of all the available evidence Our findings indicate that membrane phospholipid composition is heterogeneous in colon cancer. Therefore, spatially resolved lipidomics emerge as a new molecular dimension of analysis to characterise tumour epithelial composition, complementing transcriptomic and proteomic approaches. The peroxisomal–DHA–plasmalogen axis identified in metastasis-enriched states represents a candidate mechanism underlying epithelial plasticity during colorectal cancer dissemination and warrants functional investigation. Future studies should determine whether epithelial lipid phenotypes carry independent prognostic value and whether targeting specific fatty-acid remodelling enzymes can modulate tumour cell plasticity in preclinical models. Introduction Traditional bulk genomic or metabolomic analyses hide cellular heterogeneity in solid tumours, where tumour cells differ in morphology and molecular profiles, making spatial context essential to preserve the tissue location of distinct molecular states.1, 2, 3 Among the molecular layers accessible to spatial analysis, membrane lipids are especially well suited to this purpose. Lipids are among the most abundant ionisable molecules in the cell, which allows them to be imaged directly from tissue by mass spectrometry imaging (MSI) without amplification or labelling.4,5 Just as importantly, membrane phospholipid composition is actively remodelled as cells change state. Using MSI in healthy human colon mucosa, we previously demonstrated that this remodelling tracks epithelial differentiation along the crypt axis with precision.6, 7, 8 In particular, phosphatidylinositols (PI) and phosphatidylethanolamine (PE) plasmalogens undergo pronounced and precise fatty-acid remodelling, with stem and progenitor compartments enriched in arachidonic acid (AA)–containing species that decline progressively during differentiation and are replaced by monounsaturated species, predominantly oleic acid (OA)–containing phospholipids. In contrast, phosphatidylcholine and most PE species remain comparatively stable. These observations indicate that membrane lipid profile encodes differentiation state and becomes disrupted during chronic inflammation and malignant transformation. Consistent with this principle, we previously showed that distinct immune cell populations within intestinal tissues also exhibit characteristic phospholipid compositions associated with activation and functional specialisation.9 Increasing evidence from spatial and developmental lipidomics further supports this concept, showing that coordinated lipid remodelling accompanies lineage specification and tissue development, with distinct cell types acquiring conserved lipid signatures during differentiation.10, 11, 12 These findings suggest that membrane lipid composition can function as an integral feature of cellular identity rather than a passive metabolic readout. Although MSI overcomes many limitations inherent to bulk lipidomic analyses, the major challenge now lies not in data acquisition but in the biological interpretation of spatial lipidomic patterns. Spatial lipidomic datasets frequently reveal specific lipid profiles across neighbouring cells or regions. Nonetheless, the functional significance of these differences remains unclear without appropriate biological context. Interpreting such lipid heterogeneity therefore requires integrative strategies that link molecular lipid signatures to defined cellular identities and states. In this study, we sought to explore whether membrane phospholipid composition is spatially organised within human colon cancer and whether lipidomic heterogeneity is associated with transcriptionally defined epithelial cell states, by developing an integrative MSI–spatial-transcriptomics workflow on consecutive tissue sections. Methods Human tissue samples Human tissue specimens were obtained from two independent cohorts recruited at the University Hospital Son Espases (Palma, Spain). The study was approved by the Ethics Research Committee of the Balearic Islands (IB4568/21PI), and all participants provided written informed consent. Clinical and pathological characteristics of all participants are summarised in Table 1. Participant sex was determined from clinical records at the time of surgery. The discovery cohort comprised 12 participants. Primary tumour samples were collected from 8 patients with histologically confirmed colon adenocarcinoma undergoing surgical resection, between December 2017 and November 2020. Peritoneal metastatic implants of colorectal origin were obtained from 4 additional patients during cytoreductive surgery before hyperthermic intraperitoneal chemotherapy (HIPEC), between July 2022 and August 2023. Metastatic samples were collected from visceral adipose tissue, comprising mesentery (n = 2) and omentum (n = 2). Unlike the primary tumours, peritoneal deposits are not TNM-staged and were annotated for age, sex and site only (Table 1). The validation cohort comprised 15 patients with histologically confirmed colon adenocarcinoma undergoing surgical resection, collected between December 2020 and April 2022. Inclusion and exclusion criteria were identical to those applied in the discovery cohort. Spatial transcriptomics was not performed on this cohort; specimens were processed for mass spectrometry imaging only. Table 1Clinicopathological characteristics of the study cohort, stratified by sex.CharacteristicMaleFemaleTotalPrimary colorectal tumours (n = 23) Patients, n81523 Age, years — median (range)68 (57–81)76 (57–90)75 (57–90) Tumour location Right-sided/proximal51015 (65.2%) Left-sided/distal358 (34.8%) pT stage T2123 (13.0%) T341317 (73.9%) T4a303 (13.0%) pN stage N04812 (52.2%) N14610 (43.5%) N2011 (4.3%) pM stage M071421 (91.3%) Mx112 (8.7%) Differentiation grade G1044 (17.4%) G27916 (69.6%) G2–G3101 (4.3%) G3022 (8.7%) Mismatch repair (MMR) status pMMR (proficient)71118 (78.3%) dMMR (MLH1/PMS2 loss)145 (21.7%)Peritoneal metastatic implants (n = 4) Patients, n314 Age, years — median (range)65 (54–68)7166.5 (54–71) Implant location Mesentery202 Omentum112Primary colorectal tumours (n = 23) and peritoneal metastatic implants (n = 4). Clinicopathological characteristics of the study cohort, stratified by sex. Primary colorectal tumours (n = 23) and peritoneal metastatic implants (n = 4). Tissue processing and slide preparation Snap-frozen tissue specimens were embedded without OCT and stored at −80 °C until sectioning. Tissue sections (10 μm thickness) were mounted onto Visium Spatial Gene Expression slides (10× Genomics). Spatial Tissue Optimisation was performed according to the manufacturer’s protocol (10× Genomics, CG000238 Rev B), and images were acquired using a Zeiss Axio Observer D1 microscope operated with Zen Blue software. Spatial transcriptomics: library generation, sequencing, and preprocessing Spatial transcriptomics (ST) libraries were prepared following the 10× Genomics Visium Spatial Gene Expression Reagent Kits User Guide (CG000239 Rev G) using the Visium Spatial Gene Expression Slide & Reagent Kit (PN-1000187) and the Dual Index Kit TT Set A (PN-1000215). Following permeabilisation, cDNA was amplified and then purified using SPRIselect beads (Beckman Coulter). cDNA was subjected to fragmentation, end repair, and A-tailing. A double size selection was performed using SPRIselect beads (Beckman Coulter, Cat No. B23318), after which adaptors were ligated. After PCR amplification, libraries were quantified using Qubit (Thermo Fisher Scientific) and sequenced on an Illumina NovaSeq 6000. Transcriptomic analyses of ST data ST data were processed using the Space Ranger pipeline (10× Genomics, SCR_023571). Graph-based clustering was performed using the default Louvain Modularity Optimisation (LMO) algorithm to generate an unsupervised segmentation of barcodes. This clustering structure was subsequently used as the spatial reference for guided segmentation of the corresponding MSI section. In one primary tumour sample (P6), a spatially localised region with suboptimal permeabilisation was excluded from downstream analyses. After MSI-based identification of epithelial lipid phenotypes (EpLPs), each ST cluster was mapped to its corresponding EpLP. All ST datasets were then aggregated using the spaceranger aggr pipeline with read-depth normalisation to generate a unified gene expression matrix. Differential gene expression between clusters assigned to each EpLP was performed on the aggregated dataset. Gene Set Enrichment Analysis (GSEA v4.3.2, SCR_003199) was used to assess pathway differences across EpLPs. Enrichment was tested against (i) the Hallmark gene sets from the Molecular Signatures Database (MSigDB) and (ii) transcriptional signatures derived from published single-cell RNA-sequencing studies of human colon epithelium.13 For each lineage signature, the top 100 marker genes were used. We performed reference-free deconvolution using Space Ranger’s latent Dirichlet allocation (LDA)–based spot deconvolution. Barcodes in which epithelial-related signatures accounted for >50% of the inferred composition were retained for downstream analysis. All gene-expression p-values were adjusted for multiple comparisons using the Benjamini–Hochberg false discovery rate (FDR). Mass spectrometry imaging (MSI) Sections were mounted onto indium tin oxide (ITO)–coated glass slides (Hudson Surface Technology, Cat No. PL-IE-000010-P100) and stored at −80 °C until analysis to maintain lipidomic integrity. Prior to analysis, 1,5-diaminonaphthalene (DAN; Sigma–Aldrich, Cat No.56451) was applied as matrix for negative-ion detection. MSI data were acquired in negative-ion mode over a mass range of m/z 300–1300 with a resolution of 10 μm using a timsTOF fleX (Bruker Daltonics), operated at a mass resolution of approximately 38,000 at m/z 480. MSI datasets were visualised and processed using SCiLS Lab software (Bruker). MSI data processing, peak annotation, and normalisation All MSI experiments were processed with SCiLS Lab Software (Bruker, SCR_014426), following standard acquisition parameters. Individual MSI datasets were segmented using hierarchical clustering with bisecting k-means and Euclidean distance metrics. Tissue regions were initially separated into stromal and epithelial segments according to H.E. optical images co-registration and checking for ion distribution of previously described epithelial lipid markers: m/z 861.549 ([PI 36:2-H]-)m/z 863.565 ([PI 36:1-H]-) and m/z 857.518 ([PI 36:4-H]-).7,8 Epithelial regions were sub-segmented into two to four different clusters, based on ST data on the consecutive tissue section. A combined experiment in SCiLS Lab Software was created incorporating all individual MSI segmented datasets for combined feature extraction. Data were normalised to Total Ion Count (TIC), and Individual peaks were moved to local max for m/z alignment, with a minimal uncertainty interval width of 20. A list of the 523 features most abundant features was exported, and lipid annotations were performed using Matlab (Mathworks, SCR_001622). Mass accuracy was always better than 10 ppm and lipid species were filtered according to the commonly detected for adducts.14 Assignments were additionally cross-confirmed against an independent lipidomic analysis of colon tissue performed by an external reference laboratory (Institute of Clinical Chemistry, University of Regensburg), using high-resolution FTMS (Q Exactive hybrid quadrupole-Orbitrap, Thermo Fisher Scientific) and triple-quadrupole SRM (QTRAP 6500+, Sciex), which confirmed that the same species were detected by orthogonal instrumentation. A minor subset, in classes not interpreted at the acyl level, rested on accurate mass and adduct rules alone. For a subset of species, the fatty-acyl composition was already confirmed by MS/MS in human colon.6, 7, 8 Annotation confidence for all 99 species is provided in Supplementary Table S1. Relative abundance of each lipid species was calculated as the percentage of its intensity to the total intensity of all identified lipid species within the corresponding lipid family. Integration of MSI and ST data for guided spatial segmentation ST and MSI datasets were generated from consecutive tissue sections. ST-derived clustering structure was used as spatial reference for MSI segmentation. For each ST section, unsupervised graph-based clustering using the Louvain Modularity Optimisation (LMO) algorithm (Space Ranger) established the number of epithelial clusters present in that specimen. This cluster number was then used to define the segmentation granularity of the corresponding MSI section. Restricting segmentation to the epithelial compartment and enforcing matched cluster number yielded MSI partitions that closely resembled the spatial structure of the ST-derived epithelial domains. MSI clusters were paired to ST clusters by maximising their shared tissue area after geometric alignment of the two sections. When more than one pairing was plausible, assignments were guided by the joint evaluation of spatial proximity, continuity of epithelial domains, and preservation of tissue architecture. Because the two modalities derived from consecutive rather than identical sections, correspondence was assessed at the level of histological domains rather than individual cells. To quantify it, each epithelial ST spot was assigned the MSI cluster present beneath it after alignment and cross-tabulated against its transcriptomic cluster. Agreement between the two independently derived partitions was summarised as per cent spot concordance and the adjusted Rand index (ARI; chance-corrected and label-invariant, and therefore independent of the pairing rule), computed per sample and pooled across all epithelial spots (aricode, R, SCR_001905). Spots classified as epithelial in ST but not in MSI were tabulated separately as epithelial-mask disagreement. Across 5655 epithelial spots, the partitions showed concordant spatial organisation (pooled concordance 62.9%, ARI 0.507; mask agreement 76.2%; per-sample values in Supplementary Table S2). In one primary tumour sample (P8), ST data were lost due to tissue damage and could not be used to derive a transcriptionally informed segmentation. Nevertheless, P8 was retained in the analysis because MSI data quality and tissue morphology were preserved and was excluded from the concordance analysis. In this case, MSI epithelial segmentation was performed using three clusters. Lipidomic profiling and differential lipid abundance analyses Epithelial lipid profiles were first standardised (μ = 0, σ2 = 1). t-SNE was used to explore grouping patterns among epithelial segments and revealed an apparent structure of four groups. To define these groups objectively, k-means clustering was performed in R on all 99 annotated phospholipid species using the default Hartigan–Wong algorithm; k = 4 was consistent with the structure observed by t-SNE and supported by silhouette analysis. The four resulting clusters, each corresponding to a distinct lipid profile, were designated epithelial lipid phenotypes (EpLP1–EpLP4). To visualise the lipid features distinguishing these phenotypes, the 50 species ranked highest by ANOVA across clusters were displayed as heatmaps for hierarchical visualisation of lipid abundances. Lipid–gene association analyses A curated list of lipid-related genes was assembled by merging all lipid-associated categories from Gene Ontology, KEGG, Reactome, and WikiPathways databases. After intersecting this list with genes detected in the ST dataset and excluding genes with low average expression (<1 count per barcode), 237 genes were retained for downstream analysis. Pairwise correlations were computed between the expression levels of these 237 genes and the abundances of 99 annotated phospholipid species quantified across the 32 epithelial MSI segments. Pearson correlation coefficients were calculated using GraphPad Prism software (version 8.0.1, SCR_002798). Gene–lipid pairs with absolute correlation values > 0.6 were used to construct a bipartite correlation network. Network was built using Cytoscape (SCR_003032),15 representing genes and phospholipid species as nodes within the network and correlation as edges. The resulting network was exported to Gephi (SCR_004293; https://gephi.org/) and the final graph was generated using the Fruchterman–Reingold force-directed algorithm. For the peroxisomal gene analysis, a comprehensive list of peroxisome-associated genes was compiled beginning with KEGG pathway hsa04146 and complemented with Reactome (R-HSA-390918, R-HSA-389542, and R-HSA-9033241), WikiPathways (WP1941 and WP5275), GO Biological Process (GO:0007031), and transcriptional regulators or coactivators known to influence peroxisomal pathways. Cross-cohort projection and EpLP assignment An independent cohort of 15 primary colon tumour specimens was processed for MSI as described above. Because spatial transcriptomics data were not available for this cohort, epithelial MSI segments were defined using up to four clusters per sample, consistent with the range observed in the discovery cohort. EpLP assignment was performed in normalised, z-scored phospholipid space restricted to the 50 lipid species most discriminative of EpLP identity in the discovery cohort, as ranked by ANOVA. For each discovery EpLP, a reference centroid was computed as the mean z-scored profile across all constituent segments. Each validation segment was then assigned to the nearest centroid by Euclidean distance. Assignment confidence was quantified as (d2 − d1)/d2, where d1 and d2 denote the distances to the nearest and second nearest centroid respectively, with values approaching 1 indicating unambiguous assignment and values approaching 0 indicating proximity to the boundary between two EpLPs. Internal consistency of the centroid-based classification was confirmed by reassignment of the epithelial clusters in the discovery samples, which yielded 100% concordance with original EpLP labels. Statistical analysis Lipid-species abundances were compared across epithelial lipid phenotypes using the Kruskal–Wallis test followed by Dunn's test for multiple comparisons; non-parametric tests were chosen because abundance values were not assumed to be normally distributed. Unless otherwise stated, p values arising from multiple comparisons were adjusted using the Benjamini–Hochberg false discovery rate (FDR), with an adjusted p value < 0.01 considered significant. Test-specific details for differential gene expression, gene-set enrichment (GSEA) and gene–lipid correlation (Pearson) are given in the corresponding Methods subsections. Analyses were performed in R and GraphPad Prism 8.0.1 (GraphPad Software Inc., SCR_002798). No formal sample-size calculation was performed. As an exploratory spatial multi-omic study, cohort size was determined by the availability of suitable specimens (discovery cohort, n = 12; validation cohort, n = 15). Inclusion and exclusion criteria are described under “Human tissue samples”. Role of funders The funding sources had no role in the study design, data acquisition, data analysis, data interpretation, or the writing of the manuscript. Results Integrating ST and MSI data defines epithelial lipid phenotypes in colon cancer To investigate the potential of spatial lipidomics in colon cancer (CC), we performed spatial transcriptomic (ST) and mass spectrometry imaging (MSI) analyses on consecutive histological sections from seven primary colon tumours and four intraperitoneal metastatic implants (Fig. 1A). ST was initially attempted on all twelve samples. In one primary tumour (P8) the ST dataset failed quality control and was excluded from the transcriptomic analyses, but its high-quality MSI data were retained and segmented into three epithelial clusters on the basis of lipid profiles alone. Integrated spatial transcriptomics and mass spectrometry imaging workflow and segmentation of colon cancer samples. (A) Overview of the analytical strategy. Consecutive tissue sections from primary and metastatic colon tumours were analysed by spatial transcriptomics (ST) and mass spectrometry imaging (MSI). ST-derived transcriptional segments were used as spatial references to guide epithelial segmentation in the corresponding MSI sections. (B) ST segmentation across all 11 analysed sections. For each tissue section, the H&E image is shown alongside its ST segmentation. Segmentation of the ST data generated 67 segments across all samples that recapitulated the underlying histological architecture enabling reliable annotation of clusters (Fig. 1B and Supplementary Fig. S1). We identified 32 tumour epithelial segments in the 11 samples, ranging from two to four per histological section, and characterised by high expression of epithelial markers (EPCAM, CDH1, TFF1, MUC1) (Supplementary Fig. S1 and Table S3). Conversely, MSI provided a higher spatial resolution of 10 μm per pixel, allowing clear discrimination between epithelial and stromal regions. Because epithelial areas could be readily identified in the MSI data, we adjusted the segmentation granularity so that the number of epithelial clusters matched that observed in the corresponding ST section. In this way, ST-derived segments, which define biologically interpretable tissue domains based on gene expression, served as spatial references to guide the MSI segmentation strategy, resulting in a comparable segmentation structure (Fig. 2A and Supplementary Fig. S2). To quantify this correspondence, we compared the transcriptomic and lipidomic cluster assignments at the level of individual ST spots. Across epithelial spots from eleven samples, the two independently derived partitions were spatially concordant (62.9% spot concordance; adjusted Rand index 0.507), and the modalities agreed on epithelial identity for 76.2% of spots. In general, concordance was higher in primary tumours, including those resolved into multiple epithelial domains, and lower in most metastatic implants (Supplementary Table S2), consistent with the reduced lipid-phenotypic heterogeneity of these lesions. Spatially resolved lipidomic profiling defines four epithelial lipid phenotypes (EpLPs). (A) Representative samples of matched histological (H&E), spatial transcriptomic (ST) and MSI segmentation, illustrating the correspondence between transcriptional and lipidomic spatial domains. (B) t-SNE of all MSI-derived epithelial clusters. K-means clustering (k = 4) identified four epithelial lipid phenotypes (EpLP1–EpLP4), visualised by distinct colours in the t-SNE plot. For each EpLP, the normalised spatial intensity (0–1) of a representative lipid species is shown on separate t-SNE projections (EpLP1: [PI 36:2-H]-, m/z 861.5433; EpLP2: [PA 38:4-H]-, m/z 723.4965; EpLP3: [PI 36:4-H]-, m/z 857.5122; EpLP4: [PE_P 38:6-H]-, m/z 746.5075). (C) Heatmap of the top 50 differentially distributed lipid species in EpLPs. (D) Two representative tissue sections displaying the spatial distribution of EpLPs, each section containing two EpLPs. For each example, MSI maps illustrate the localised enrichment of two representative lipids with EpLP-specific spatial patterns ([PE 32:0-H]-, m/z 690.5033; [PE_P 36:5-H]-, m/z 720.4930; [PI 36:2-H]-, m/z 861.5433). For differential lipidomic analysis, 99 phospholipid species were annotated. Lipid profiles were then normalised so that each species was expressed as a percentage of the total within its corresponding phospholipid class, including phosphatidylethanolamines (PE), PE-plasmalogens (PE_P), phosphatidylinositols (PI), phosphatidylserines (PS), sphingomyelins (SM), sulfatides, phosphatidic acids (PA), and phosphatidylglycerols (PG) (Supplementary Table S4). t-SNE of all epithelial lipid segments revealed a clear clustering pattern indicative of four major groups (Fig. 2B). Unsupervised clustering confirmed this structure, and the number of clusters (k = 4) was chosen to best reflect the data points distribution observed in the t-SNE projection. This analysis defined four epithelial lipid phenotypes (EpLPs) based on phospholipid composition (Fig. 2C). EpLP1–EpLP4 comprised 11, 13, 5, and 6 epithelial segments, respectively (Supplementary Table S2), with three tumours contributing clusters to more than one phenotype. Therefore, these four EpLPs were not uniformly distributed within tumours but instead formed spatially organised epithelial domains (Fig. 2D). Individual sections contained one or two EpLPs in proximity, indicating tumour heterogeneity at the level of membrane lipid composition (Supplementary Fig. S3). EpLPs link lipid composition to transcriptional and lineage programmes We identified 1240 differentially expressed genes (DEGs) (FDR-adjusted p value < 0.01) across EpLPs (Supplementary Table S5). Gene Set Enrichment Analysis (GSEA) revealed differential enrichment patterns among the EpLPs involving key pathways in CC, including TGFβ and WNT/β-catenin signalling (upregulated in EpLP1 and EpLP2), MYC targets (upregulated in EpLP2 and EpLP3), and angiogenesis (upregulated in EpLP4) (Fig. 3A). Interestingly, several of the most differentially expressed genes corresponded to lineage-related markers (Fig. 3B). EpLP1 overexpressed genes associated with the enterocyte lineage (CLDN4 and MUC1), EpLP2 was enriched in genes linked to cell proliferation and division (MCM7 and MKI67), EpLP3 displayed higher expression of genes characteristic of immature colonocytes (LEFTY1 and OLFM4), while EpLP4 showed elevated levels of TFF1, TFF2 and AQP5, genes associated with goblet cells. These patterns suggested that each EpLP may correspond to a different stage or lineage within the epithelial differentiation hierarchy. To test this hypothesis, we performed GSEA using transcriptional signatures derived from scRNA-seq studies of the human colon epithelium, representing major epithelial lineages (Fig. 3C and Supplementary Table S6).13 EpLP1 and EpLP2 showed enrichment of the transient amplifying (TA) cell signature, mixed with enterocyte or goblet, respectively. In contrast, EpLP3 was enriched for the stem-cell signature, while EpLP4 displayed a bias towards all secretory lineages. Transcriptional programmes associated with epithelial lipid phenotypes (EpLPs). (A) Gene Set Enrichment Analysis (GSEA) of Hallmark pathways across the four EpLPs, shown as a bubble plot. For each EpLP, the five most positively and five most negatively enriched gene sets (ranked by normalised enrichment scores, NES) were selected, yielding 25 unique gene sets. Bubble colour represents the NES and bubble size the statistical significance (nominal p-value). (B) Violin plots displaying the expression of representative genes associated with colon epithelial lineage differentiation, maturity, or stemness, highlighting distinct transcriptional features of each EpLP (EpLP1, CLDN4, MUC1, MUC12, FOS; EpLP2, MCM7, MKI67, MUC2, AGR2; EpLP3, LEFTY1, OLFM4, EPHB2, DMBT1; EpLP4, TFF1, TFF2, AQP5, CLDN18). (C) Heatmap of NES values for epithelial lineage–related transcriptional signatures. NES values reflect the relative enrichment of each signature within an EpLP compared with the other EpLPs. (D) Relative abundance of assigned PI species and PE-plasmalogens (PE_P) across the four EpLPs, expressed as percentages within their respective lipid classes. Differences across EpLPs were assessed by Kruskal–Wallis test followed by Dunn's multiple-comparison correction. Error bars represent standard error (SE). Statistical significance is indicated as p < 0.05 (∗), p < 0.01 (∗∗), and p < 0.001 (∗∗∗). We next examined class-specific phospholipid profiles across the EpLPs (Fig. 3D and Supplementary Fig. S4). We observed differences in PI 38:4 abundance across the EpLPs. EpLP1 exhibited the lowest relative levels of PI 38:4 (22.7%), whereas EpLP2 and EpLP3 showed higher proportions (37.0% and 37.2%, respectively), with EpLP4 displaying intermediate values (29.2%). These differences suggest that EpLPs span a spectrum of epithelial differentiation states that is encoded at the level of membrane phospholipid composition. Consistent with this interpretation, EpLP1, characterised by the lowest PI 38:4 levels, showed the strongest enrichment for mature absorptive transcriptional programmes. In contrast, the elevated PI 38:4 levels observed in EpLP2 and EpLP3 align with their enrichment for TA and stem-like gene expression signatures, respectively. EpLP4 occupied an intermediate position, concordant with its bias towards secretory lineage programmes. Based on these findings, we focused our subsequent analyses on PI and PE_P classes, previously shown to undergo differentiation-associated remodelling. Within these classes, the most consistent and biologically informative changes involved phospholipid species containing arachidonic acid (AA), oleic acid (OA), linoleic acid (LA), and docosahexaenoic acid (DHA) (Fig. 3D and Supplementary Fig. S4). EpLP2, EpLP3, and EpLP4 all displayed an increased relative abundance of AA–containing species in PI and PE_P, and to a lesser extent within PE. At the level of the PI class, AA-containing species (38:4 and PI 38:5), were significantly more abundant in EpLP2 and EpLP3 (Fig. 3D). Given our previous observations in healthy human colon, elevated AA content suggested an undifferentiated state in these groups. In contrast, this relationship was not as clearly preserved within the PE_P and PE classes (Fig. 3D and Supplementary Fig. S4). Instead, several AA-containing species within these classes, including PE_P 36:4 and PE_P 38:5, as well as PE 38:4 and PE 38:5, were most significantly increased in EpLP4. In contrast, OA-containing species such as PI 34:1 and PI 36:1 were significantly reduced in EpLP2 and EpLP3 (Fig. 3D and Supplementary Fig. S4). This reciprocal regulation of AA- and OA-containing PI species reinforces the positioning of EpLPs along an epithelial differentiation axis. However, this pattern was not uniformly observed in other phospholipid classes. In addition, we observed a global decrease in linoleic acid (LA)–containing species across EpLP2, EpLP3, and EpLP4 relative to EpLP1, including 34:2 and 36:2. Given that LA is an essential fatty acid and a biosynthetic precursor of AA whose conversion is known to be activated in inflammatory contexts,16,17 this pattern is consistent with increased LA utilisation towards downstream polyunsaturated lipid species. In line with this interpretation, PI 38:3, which is compatible with incorporation of a 20:3 acyl chain generated by LA elongation, accumulated in EpLP2, EpLP3, and EpLP4 compared with EpLP1. Finally, we observed that DHA-containing species (38:6 and 40:6) were enriched in EpLP4 and, to a lesser extent, in EpLP3. This enrichment was most pronounced within the plasmalogen fraction, where DHA-containing species are more abundant, but similar trends were also observed in PI and PE. Epithelial lineage programmes differ across anatomical contexts Metastatic epithelial clusters were assigned to either EpLP2 or 4, which were characterised by a higher enrichment of the secretory transcriptional signature (Fig. 4A). To further characterise these epithelial transcriptional features, we applied agnostic (reference-free) deconvolution to our ST data. Among the inferred signatures, we selected those corresponding to colon epithelium and retained barcodes in which these signatures represented >50% of total cells. Accordingly, we assigned 3178 barcodes across all samples into 17 primary and 11 metastatic epithelial populations. Differential expression analysis between primary and metastatic cells identified 930 genes (log2 fold change > 1.5, FDR-adjusted p value < 0.01) (Fig. 4B and C, Supplementary Table S7). Among them, we found genes related to transcriptional regulation (PROX1, KLF5), WNT signalling (SFRP2, APCDD1, WIF1), the enteroendocrine lineage (PCSK1N), chemotaxis (CXCL2, CXCL3), and lipid metabolism and transport (APOC2, PLCG1, LPCAT1, HMGCR). In addition, we found a consistent downregulation of some stem-related genes in metastasis (OLFM4 and DMBT1). Using GSEA, we confirmed that metastatic cells showed an overexpression of secretory signatures (Fig. 4D and E). Among these, the enteroendocrine signature was the most clearly biased towards metastasis, with the neuroendocrine marker PCSK1N being the most significantly overexpressed gene. Transcriptomic differences between primary and metastatic epithelial populations identified by reference-free deconvolution. (A) t-SNE of all epithelial segments, coloured by anatomical origin (primary, red; metastatic, blue). (B) Heatmap of differentially expressed genes (DEGs) between primary and metastatic epithelial populations. (C) Boxplots showing the expression of representative DEGs (OLFM4, DMBT1, KLF5, PROX1), with primary epithelial populations in red and metastatic populations in blue. (D) Left panel shows a heatmap of normalised enrichment scores (NES) from GSEA of epithelial lineage–related transcriptional signatures, comparing each inferred epithelial population against all others within the data set. Right panel show NES values from a group-level comparison of all metastatic versus all primary epithelial populations. (E) Volcano plots displaying differential gene expression between metastatic and primary epithelium. Genes belonging to the stem-like signature (left) or neuroendocrine signature (right) are highlighted in light purple to illustrate their distribution within groups. Coordinated regulation of DHA-containing phospholipids and peroxisomal pathways in EpLP4 To study the relationship between lipid metabolism and transcriptional programmes, we analysed the differential expression of a curated set of 1150 lipid-related genes (Supplementary Table S8). After excluding genes with low average expression (<1 count per barcode), 237 genes were retained for correlation analysis. We then calculated the pairwise correlations between these genes and the 99 lipid species quantified across the 32 epithelial segments. Of the 23,463 possible gene–lipid interactions, 347 pairs showed strong correlations (|r| > 0.6). These associations were visualised as a bipartite lipid–gene correlation network, where nodes represented lipid species or lipid-related genes and edges denoted significant correlations (Fig. 5A). Lipid–gene correlation structure and differential expression of lipid-related genes across EpLPs. (A) Bipartite correlation network linking lipid species with lipid-related genes across epithelial clusters. Only lipid–gene pairs with strong correlations (|r| > 0.6, Pearson) are shown. Nodes are positioned using a Fruchterman–Reingold force-directed layout, which clusters highly connected nodes by minimising the overall energy of the graph. Edge colours indicate positive (red) or negative (blue) correlations, and node colour represents node centrality within the network. (B) Heatmap of the top 25 lipid-related genes differentially expressed among the four EpLPs. (C) Violin plots showing the expression of selected lipid-related genes with significant variability across EpLPs (AMACR, ACSL1, PEX19, CAT, GNPAT, ACAA1, PEX16, IDH1). Among the phospholipid species, PE_P 38:6, PE 40:6, and PI 40:6 were the most highly connected, sharing 16 associated genes (Fig. 5A) and suggesting the existence of a coordinated regulatory module. All three are DHA-containing phospholipids, and their shared gene set included several peroxisomal genes (ACAA1, AMACR). Therefore, we analysed the expression of 105 peroxisome-related genes across the four EpLPs (Fig. 5B and Supplementary Table S9). EpLP4 showed the highest number of upregulated peroxisomal genes (20 genes), followed by EpLP3 (12 genes), EpLP2 (9 genes), and EpLP1 (3 genes). The upregulated genes in EpLP4 included multiple aspects of peroxisomal function, including lipid and ether lipid metabolism (AMACR, GNPAT, ACAA1, ACSL1), peroxisome biogenesis (PEX19, PEX16), and reactive oxygen species detoxification (CAT) (Fig. 5C). Epithelial lipid phenotypes generalise across independent primary tumour cohorts Having fully characterised the transcriptional and lipidomic identity of each EpLP in the discovery cohort, we next asked whether these lipid-defined phenotypes are reproducible in independent tumour specimens. We projected an independent set of 15 primary colon tumours onto the EpLP lipid space defined in the discovery set. Because spatial transcriptomics data were not available for this cohort, epithelial MSI segments were assigned up to four clusters per sample, consistent with the range observed in the discovery cohort (Supplementary Table S10). Reference centroids were computed from discovery samples in normalised, z-scored phospholipid space, and each validation segment was assigned to its nearest centroid using Euclidean distance across the 50 lipid species most discriminative of EpLP identity in the discovery cohort. Assignment confidence was quantified using a silhouette-derived score reflecting the relative distance to the assigned versus the next nearest centroid (Supplementary Table S11). Projection of the validation cohort identified samples across all four EpLP classes (Fig. 6A). EpLP4 and EpLP1 showed the most coherent cross-cohort correspondence, with validation samples occupying the same lipid space regions as their discovery counterparts. EpLP2 and EpLP3 validation samples showed greater dispersion and mutual overlap, consistent with the biological proximity of these phenotypes, which share broadly similar phospholipid acyl-chain compositions and partially overlapping transcriptional programmes, positioning them along a continuum rather than at discrete poles in lipid space. The spatial distribution of EpLP assignments across the 15 validation tissue sections is shown in Fig. 6B. Together, these results indicate that EpLP reproducibility tracks biological distinctiveness, with EpLP4 showing the strongest cross-cohort generalisability, consistent with its unique lipidomic and transcriptional identity. Cross-cohort projection of epithelial lipid phenotypes. (A) t-SNE of discovery (filled circles) and validation (triangles) epithelial segments, coloured by EpLP assignment. Triangle size reflects assignment confidence. Validation segments were assigned by nearest-centroid classification in z-scored phospholipid space using the 50 most discriminative lipid species from the discovery cohort. (B) MSI segmentation maps of all 15 validation tissue sections coloured by EpLP assignment. Non-epithelial regions in grey. Discussion By integrating MSI with ST, our study reveals that polyunsaturated fatty acid (PUFA) remodelling in CC is not uniformly distributed but instead segregates according to tumour epithelial programmes. These findings indicate that lipid remodelling represents a structured feature of tumour heterogeneity rather than a metabolic hallmark shared by all cancer cells. This concept is consistent with emerging evidence that phospholipid acyl-chain composition can influence epithelial biology. In the intestinal epithelium, acyl-chain remodelling enzymes such as LPCAT3 regulate stem and progenitor function,18 while lysophospholipid acyltransferases modulate membrane properties required for epithelial homoeostasis in vivo.19 In addition, it has been described in cancer that elongation of phospholipid acyl chains through ELOVL5 supports proliferation, mitochondrial function, and metastatic progression in prostate cancer.20 Together, these observations support a model in which acyl-chain remodelling constitutes a functional layer of tumour cell specialisation that becomes apparent only when lipid composition is resolved in its spatial and cellular context. Our previous analysis demonstrated that the balance between OA- and AA-containing phospholipids tracks epithelial maturation.6,7 The re-emergence of this compositional signature across EpLPs indicates that tumour lipid heterogeneity reflects structured epithelial lineage states rather than stochastic metabolic variation. Although AA enrichment has been reported in tumours, our data indicate that AA incorporation is selectively partitioned across phospholipid classes according to epithelial state, concentrating in PI in EpLP2–3 and in PE_P in EpLP4. Concomitant depletion of LA-containing species alongside accumulation of LA-derived elongated and desaturated products further supports activation of a coordinated fatty-acid remodelling programme associated with less differentiated tumour populations. Together, these observations suggest that membrane phospholipid composition is coupled to epithelial differentiation state and may contribute to stabilising lineage-associated cellular programmes. A plausible mechanism linking AA-containing PI enrichment to epithelial state is phosphoinositide signalling, which is a central regulator of epithelial polarity, proliferation, and differentiation.21, 22, 23 Although phosphoinositide signalling is defined by phosphorylation of the myo-inositol headgroup, an emerging question is whether the acyl-chain composition of the precursor PI affects the signalling competence of phosphoinositide pools. Prior work supports a model in which the PI acyl-chain composition shapes the efficiency of phosphoinositide synthesis and thereby influences downstream signalling pathways. PI 38:4 (18:0/20:4), the dominant AA-containing PI species, represents the preferred substrate for phosphoinositide kinases and is therefore efficiently converted into phosphorylated signalling lipids.24 Also, activation of WNT signalling promotes phosphoinositide kinase activity and increases the production of PI-derived signalling intermediates required for pathway activation and proliferation.25,26 In this context, the abundance of PI 38:4 may influence the size of the kinase-accessible PI pool, thereby modulating the capacity of epithelial cells to sustain downstream signalling. In addition, phosphoinositides also function as membrane docking molecules for signalling effectors such as AKT, linking PI remodelling to pathways controlling proliferation, survival, and differentiation.27 Together, these observations suggest that remodelling the PI pool could contribute to tumour cell state by shaping signalling responsiveness. On the other hand, PUFA accumulation (AA and DHA) within PE_P occurred preferentially in EpLP4. This observation indicates that expansion of the tumour PUFA pool is not uniformly distributed across phospholipid classes but instead segregates according to tumour cell identity. Consistent with this lineage-restricted PUFA redistribution, our ST-lipidomic analysis shows a high correlation between DHA-containing phospholipids and peroxisomal enzymes. Because it is well known that key steps of DHA and plasmalogen synthesis take place in the peroxisome,28, 29, 30 this result suggests the presence of a coordinated peroxisome–DHA-plasmalogen programme in EpLP4. Indeed, increased rates of fatty acid oxidation due to peroxisomal activity have been linked to cancer progression31 and several peroxisomal enzymes, such as AMACR and ECI2, have also been implicated in colorectal cancer.32,33 Nonetheless, our data suggests that membrane remodelling may be playing a role in tumour biology as well. Plasmalogen and DHA content can modify the biophysical properties of membranes, such as curvature and fluidity, which are critical for membrane fusion, fission, and vesicle formation in secretory cells.34, 35, 36 Therefore, these lipid changes may be functionally linked to secretory activity in specific tumour epithelial populations. Such secretory features could reflect either the ontogeny of the tumour, preserving molecular traits of secretory epithelial lineages, or an adaptive reprogramming that confers selective advantages during tumour progression. In particular, enhanced secretory capacity may support processes relevant to migration, invasion, or resistance to oxidative stress, all of which rely on dynamic membrane remodelling and vesicle trafficking. Together, these observations point to a secretory-associated membrane lipid configuration as a coherent and biologically meaningful axis of tumour epithelial heterogeneity. In addition, metastatic cells exhibited loss of canonical intestinal stemness. Concomitantly, metastatic cells showed increased expression of PROX1, a transcription factor recently identified as a key regulator of epithelial plasticity during colorectal cancer progression.37 Moorman et al. demonstrated that PROX1 is induced in injury-associated, foetal progenitor–like states, where it acts to restrict non-intestinal lineage programmes, and that metastatic progression in distant organs is associated with progressive loss of PROX1-dependent lineage. In this context, elevated PROX1 expression in intraperitoneal metastases could be consistent with a plastic epithelial state, rather than maintenance of crypt-base intestinal stem cell identity. Finally, it is important to mention that primary and metastatic samples were not patient-matched. Nonetheless, the reproducibility of these features across independent intraperitoneal metastases supports a site-specific adaptation model in which peritoneal dissemination retains greater epithelial lineage continuity while adopting selected transcriptional programmes. This study has limitations that should be considered when interpreting our findings. The discovery cohort is modest in size and was recruited at a single institution, and the number of biological samples contributing to each phenotype is small. We addressed this by recovering all four phenotypes in an independent cohort of fifteen primary tumours, although confirmation in larger, multi-institutional series will be required to establish their generality. Because MSI and ST were acquired from consecutive rather than identical sections, the two modalities do not sample the same cells, and cross-modal correspondence could be evaluated only at the level of histological domains rather than individual cells. In addition, the primary and metastatic specimens were not patient-matched, so the anatomical comparisons reflect population-rather than individual-level differences. Finally, the present work is an observational, spatially resolved characterisation rather than a perturbation study. The candidate mechanisms it suggests (modulation of the phosphoinositide pool through AA-containing PI, and a coordinated peroxisome–DHA–plasmalogen programme in secretory and metastatic states) are hypotheses generated by the data that warrant dedicated functional investigation. Our study provides a generalisable spatial multi-omic workflow anchored in ST that enables the derivation of cell-state–resolved metabolic phenotypes from mass spectrometry imaging. Using this approach, we demonstrated that membrane phospholipid composition encodes structured epithelial cell states within human tumours, revealing membrane lipid remodelling as an integral and spatially organised dimension of tumour heterogeneity. Contributors R.M.R. and G.B.-C. conceived and designed the study. R.M.R. performed the computational and statistical analyses, integrated ST and lipidomic datasets, prepared figures, and drafted the manuscript. K.P.-R. processed clinical samples and performed ST analysis. A.M.A. generated ST libraries and contributed to ST data analysis. T.X. and A.M.-B. performed mass spectrometry imaging (MSI) analyses. C.H.-B. performed MSI data acquisition and readout. J.A.F. contributed to MSI analyses and study design. M.A.M. evaluated histopathology and supported tissue annotation. J.J.S.-S., R.M.-S., S.J., C. P.-F. and A. G.-G selected patients and provided surgical specimens. R.M.R. and G.B.-C. wrote the manuscript with input from all authors and have accessed and verified the underlying data. All authors reviewed and approved the final manuscript. Data sharing statement All data supporting the findings of this study are available as described. Spatial transcriptomic data have been deposited at NCBI GEO database (accession number GSE325755). Processed lipidomic are presented in the Supplementary Tables. Mass spectrometry imaging raw data are not publicly deposited owing to their large file size and proprietary format but are available from the corresponding author upon reasonable request. All analyses were conducted using established, previously published methods implemented in standard, publicly available and appropriately cited software packages. Declaration of interests A.G.-G. serves as an external advisor to Cella Medical Solutions and received no compensation related to this manuscript. S.J.McF. has received a research grant from the Asociación Española de Coloproctología (AECP 2025) for work unrelated to this manuscript. All other authors declare no competing interests. Acknowledgements We gratefully acknowledge the nurses and physicians from the Departments of Gastroenterology, Pathological Anatomy, and Gastroenterology and General Surgery at the University Hospital de Son Espases (HUSE, Palma, Spain) for their valuable contributions to patient enrolment and sample collection. We also thank the Microscopy and Genomics and Bioinformatics Platforms at the Health Research Institute of the Balearic Islands for their commitment and technical support. We thank Javier Piérola Lopetegui for assisting with microscopy imaging and Aarne Fleischer for assisting with sequencing library preparation. The authors thank for technical and human support provided by SGIker of UPV/EHU and European funding (ERDF and ESF). During the preparation of this work the authors used Grammarly (Grammarly Inc.) in order to improve the readability and language of the manuscript. The authors have reviewed and confirmed the validity of the text and take full responsibility for the content of the publication. This study was supported in part by the Institute of Health Carlos III (PI19/00002, PI24/00313), the Health Research Institute of the Balearic Islands (IMP22/04), Conselleria de fons europeus, universitat i cultura, Government of the Balearic Islands (AP_2021_001) and the Basque Government (IT1491-22) and the EC (European Regional Development Fund, ERDF). RM R holds a postdoctoral contract supported by the Scientific Foundation, Spanish Association Against Cancer (INVES222995RODR). K P-R held a predoctoral contract of the Health Research Institute Carlos III (FI20/00180) and co-funded by ESF (European Social Fund, Program 2014–2020, “ESF invest in your future”). A M-B holds a contract of that is part of the grant (PTA2022-021759-I) funded by Ministerio de Ciencia, Innovación y Universidades (MCIU/AEI/10.13039/501100011033) and the ESF+ (European Social Fund+); TX holds a predoctoral contract from the Health Research Institute from the Balearic Islands funded by the Sustainable Tourism Tax Fund (ITS) of the Government of the Balearic Islands (ITS2023-057). Footnotes Supplementary data related to this article can be found at https://doi.org/10.1016/j.ebiom.2026.106473. Contributor Information Ramon M. Rodriguez, Email: ramonmaria.rodriguez@ssib.es. Gwendolyn Barceló-Coblijn, Email: gwendolyn.barcelo@idisba.es. Supplementary data Supplementary_materials Supplementary Table S1 Supplementary Table S1. Annotation confidence and MS/MS validation of the 99 phospholipid species. Supplementary Table S2 Supplementary Table S2. Per-sample MSI–ST spatial concordance results. Supplementary Table S3 Supplementary Table S3. Complete gene expression matrix of spatial transcriptomics clusters with histological annotations for all analysed tissue sections. Supplementary Table S4 Supplementary Table S4. MSI-derived lipid peak intensities and lipid class–normalised abundances of all annotated lipid species across all epithelial clusters. Supplementary Table S5 Supplementary Table S5. Differential gene expression analysis of all annotated genes across epithelial lipid phenotypes (EpLPs). Supplementary Table S6 Supplementary Table S6. Gene sets defining epithelial lineage transcriptional programmes in the human colon. Supplementary Table S7 Supplementary Table S7. Differential gene expression analysis between primary and metastatic epithelial populations identified by deconvolution analysis. Supplementary Table S8 Supplementary Table S8. Curated list of lipid-related genes used for lipid–gene correlation analyses. Supplementary Table S9 Supplementary Table S9. Curated list of peroxisome-associated genes used in this study. Supplementary Table S10 Supplementary Table S10. MSI segmentation data for the validation cohort. Supplementary Table S11 Supplementary Table S11. EpLP assignment and confidence scores for validation cohort segments. References Associated Data Supplementary Materials Supplementary_materials Supplementary Table S1 Supplementary Table S1. Annotation confidence and MS/MS validation of the 99 phospholipid species. Supplementary Table S2 Supplementary Table S2. Per-sample MSI–ST spatial concordance results. Supplementary Table S3 Supplementary Table S3. Complete gene expression matrix of spatial transcriptomics clusters with histological annotations for all analysed tissue sections. Supplementary Table S4 Supplementary Table S4. MSI-derived lipid peak intensities and lipid class–normalised abundances of all annotated lipid species across all epithelial clusters. Supplementary Table S5 Supplementary Table S5. Differential gene expression analysis of all annotated genes across epithelial lipid phenotypes (EpLPs). Supplementary Table S6 Supplementary Table S6. Gene sets defining epithelial lineage transcriptional programmes in the human colon. Supplementary Table S7 Supplementary Table S7. Differential gene expression analysis between primary and metastatic epithelial populations identified by deconvolution analysis. Supplementary Table S8 Supplementary Table S8. Curated list of lipid-related genes used for lipid–gene correlation analyses. Supplementary Table S9 Supplementary Table S9. Curated list of peroxisome-associated genes used in this study. Supplementary Table S10 Supplementary Table S10. MSI segmentation data for the validation cohort. Supplementary Table S11 Supplementary Table S11. EpLP assignment and confidence scores for validation cohort segments.