Testing cellular provenance in the cancer extracellular matrix

1 Introduction

The extracellular matrix (ECM) is a dynamic component of tumor biology rather than an inert scaffold. Its proteins and regulators provide structural continuity, present ligands, store growth factors, transmit force, and guide vascular, immune, and tumor-cell movement. Matrisome definitions and ECM-enrichment proteomics made this complexity experimentally tractable (Naba et al., 2012). Imaging and mechanical studies subsequently linked matrix organization and mechanics to tumor initiation, invasion, metastasis, and progression (Provenzano et al., 2008; Levental et al., 2009; Acerbi et al., 2015; Fan et al., 2024). These advances moved the field beyond a descriptive view of desmoplasia.

Bulk assays report composition and abundance, whereas spatial measurements add location. Neither identifies the cells that synthesized and deposited protein that persists at a site. More generally, mRNA abundance is an imperfect predictor of protein abundance because mRNA turnover, translational efficiency, and protein degradation introduce additional regulatory layers (Vogel and Marcotte, 2012; Liu et al., 2016). For ECM proteins, secretion, extracellular processing, insolubilization, and retention add further separation between transcript abundance and mature matrix deposition. A cell assigned a fibronectin 1 (FN1) or collagen type I alpha 1 chain (COL1A1) transcript by single-cell RNA sequencing is therefore a candidate producer, not necessarily the source of mature deposited protein. Protein detection establishes composition, not origin.

The disconnect between expression and matrix history is biologically consequential. A shared gene symbol can conceal alternative transcripts, modified proteoforms, oligomers, or networks. Conversely, identical products can occupy different neighborhoods, encounter different proteases and forces, or engage different receptors. Total abundance therefore cannot capture architecture, mechanics, ligand accessibility, or local responder programs. Proteomic and temporal matrisome studies show that ECM protein abundance can diverge from corresponding transcriptional states and change dynamically during tumor progression (Naba et al., 2014; Papanicolaou et al., 2022; Pereira et al., 2024).

Single-cell and spatial transcriptomics identify candidate producer states. ECM-enrichment proteomics and species-resolved mass spectrometry provide protein- or compartment-level evidence, whereas imaging resolves deposits and mechanics. Because each method leaves part of the chain unresolved, credible source claims require complementary measurements. This is especially important for broad lineages, blood-borne proteins, and matrix that outlives its producers.

Current evidence permits three interpretations. Producer identity may create a persistent molecular, co-secretory, spatial, or assembly difference. Alternatively, identity may only track abundance, timing, location, cell state, or later remodeling. A third possibility is that products become interchangeable once molecular, spatial, mechanical, and receptor states are matched. Collagen I, tenascin C, and fibronectin are informative comparisons because each has multiple sources, context-dependent phenotypes, and unresolved cross-source equivalence. Examples from the literature map onto these three models: PDAC collagen I provides a producer-linked proteoform example, tenascin C illustrates stage-dependent source replacement, and fibronectin illustrates source redundancy and assembly dependence (Chen et al., 2022; Oskarsson et al., 2011; O'Connell et al., 2011; Lugano et al., 2018; Malik et al., 2010).

Testing cellular provenance therefore requires a causal rather than annotational standard. The question is whether producer identity or state creates a persistent functional difference after molecular and matrix context are accounted for. This distinction matters for metastasis, biomarkers, and therapeutic compensation, while leaving provenance-associated and provenance-independent outcomes possible. Figure 1 summarizes the conceptual framework and competing causal models.

2 Producer identity, matrix history, and matrix state

2.1 From expression to remodeling

A gene-expressing cell contains a detectable transcript. A protein-producing cell translates the relevant polypeptide, a protein-secreting cell exports it, and a matrix-depositing cell contributes it to an insoluble or functionally organized extracellular compartment. These categories often overlap, but none guarantees the next. Secreted procollagen can be degraded before fibril incorporation, soluble fibronectin can circulate without local fibrillogenesis, and a basement-membrane component can move before deposition. A provenance statement should therefore specify whether the evidence concerns RNA, intracellular or soluble protein, nascent secretion, or deposited ECM.

Matrix assembly is a distinct role. Fibronectin fibrillogenesis requires integrin-mediated traction and partner interactions, so the cell that organizes a fiber need not be its sole fibronectin producer. In glioma, CD93-MMRN2 signaling controlled β1-integrin activation and vascular fibronectin fibrillogenesis without changing FN1 transcription (Lugano et al., 2018), separating matrix assembly from gene expression. Cancer-associated fibroblasts (CAFs) can align fibronectin into migration-supporting tracks through contractility and receptor signaling (Attieh et al., 2017; Erdogan et al., 2017), while carcinoma cells can migrate collectively along fibronectin-rich interfaces (Gopal et al., 2017). Calling all of these cells “the source” would obscure the distinction between supplying material and organizing it.

Remodeling is likewise distinct. LOX-family enzymes can crosslink collagen made elsewhere, proteases can expose cryptic sites or release matrikines, and cells can orient pre-existing fibers or modify proteoglycans. In pancreatic ductal adenocarcinoma (PDAC), collagenolysis-dependent discoidin domain receptor 1 (DDR1) signaling shows that receptor output can depend strongly on the processed state of collagen (Su et al., 2022). Collagen abundance alone may therefore be insufficient to predict function. A rigorous report should name the producer, depositor, assembler, and remodeler only to the extent that each role is directly supported.

2.2 Provenance and matrix context

“Fibroblast-derived collagen” is often too coarse to be reproducible, but greater precision requires distinct terms rather than a broader definition of provenance. Cellular provenance specifies producer identity and state. Production context specifies when and where synthesis occurred, including the disease or treatment stage. Deposition context specifies the final microregion, neighboring cells, available receptors, and local protease and mechanical environment. Post-secretory history includes proteolysis, crosslinking, traction, alignment, assembly, degradation, and incorporation into networks. Final matrix state describes the proteoform, architecture, fiber orientation, porosity, stiffness, viscoelasticity, ligand accessibility, age, and receptor-readable configuration encountered by responder cells.

These categories can covary but are not synonymous. A myofibroblastic CAF and an inflammatory fibroblast can occupy distinct production contexts despite sharing a broad fibroblast identity (Öhlund et al., 2017). Endothelial fibronectin produced during angiogenic sprouting can enter a perivascular network (Lugano et al., 2018; Murphy et al., 2015), whereas hepatocyte-derived plasma fibronectin reaches tumors through the circulation and vascular leakage (Malik et al., 2010; von Au et al., 2013). Single-cell data may justify the statement “transcriptionally attributed to a myofibroblast state.” Species-specific peptides support “human tumor-cell versus mouse host contribution,” whereas a conditional allele supports a lineage requirement if deposition is measured. Claims should not exceed the resolution of the assay.

2.3 Gene, proteoform, and network identity

The phrase “the same ECM molecule” conceals several levels of identity. Products can share a gene identity yet differ in alternative splicing, promoter use, signal peptide, glycosylation, hydroxylation, proteolytic cleavage, disulfide pattern, or crosslinking. Fibronectin containing extra domain A or B is not molecularly identical to a plasma isoform. Collagen I α1 homotrimers are not the same proteoform as α1/α2 heterotrimers. A cleaved collagen fragment that engages a receptor is not equivalent to intact fibrillar collagen. At the next level, an identical purified polypeptide in solution, adsorbed to plastic, assembled into a fiber, or incorporated into a crosslinked network represents a different supramolecular state.

The relevant experimental comparison must therefore be specified. A study may show functional non-equivalence between products with the same gene identity but different proteoforms. This supports a source-linked mechanism when producer biology generates the proteoform difference, but not an effect independent of proteoform. In PDAC, malignant cells predominantly produced collagen I α1 homotrimers, whereas fibroblasts produced conventional α1/α2 heterotrimers. The purified forms differed in integrin-dependent signaling (Chen et al., 2022). These results support a producer-mediated determinant mechanism through proteoform identity, not the idea that molecularly matched collagen differs merely because another cell produced it. If differences persist after measured states are matched, an unmeasured molecular or structural mediator should be sought instead of invoking an abstract “source memory.”

2.4 Multicellular matrix assembly

An individual protein can sometimes be assigned to one producer, but an ECM fiber or microdomain is usually a collective product. One cell may supply fibronectin, another collagen, a third LOX, a fourth a protease inhibitor, and a fifth the integrin traction that organizes the network. CAF-derived extracellular-vesicle LOX (Liu et al., 2023), fibroblast-derived LOX in cholangiocarcinoma (Lewinska et al., 2024), and macrophage-derived OSM induction of lysyl oxidase-like 2 (LOXL2) in PDAC cells (Alonso-Nocelo et al., 2023) illustrate how multiple cell types can directly or indirectly regulate collagen remodeling. Network-level provenance should therefore allow source convergence, with multiple producers contributing the same or different components to a common matrix microregion.

Basement membranes and vascular niches are especially complex because endothelial cells, pericytes, fibroblasts, and tumor cells may supply different chains while remodeling the same barrier. Describing a vascular basement membrane as “endothelial-derived” is therefore rarely adequate. The source of a chain, the cell that assembles the network, and the cell that breaches it during intravasation may all differ.

2.5 Local, systemic, and historical sources

Not all tumor-associated matrix originates locally. Hepatocyte-derived plasma fibronectin, liver-derived fibrinogen, platelet cargo, and circulating matrix regulators can enter tumors or metastatic sites. Genetic loss of plasma fibronectin reduced lung metastasis in a host model, whereas circulating fibronectin has shown context-dependent effects on tumor growth (Malik et al., 2010; von Au et al., 2013). Fibrinogen-deficient mice established causal contributions of a systemic coagulation-associated matrix protein to spontaneous and hematogenous metastasis, including protection of circulating tumor cells from immune elimination (Palumbo et al., 2002; Palumbo et al., 2000; Palumbo et al., 2005). These proteins should not automatically be classified as stable structural tumor ECM, but they show that a local cell map can miss a biologically important systemic source.

Platelets add another layer. Tumor-cell collagen and heat shock protein 47 (HSP47)-dependent platelet interactions, together with platelet-specific growth-factor programs affecting vessel integrity, can alter metastatic transit without identifying a local matrix-producing cell (Xiong et al., 2020; Zhang et al., 2020). In such settings, provenance must distinguish the distant organ that synthesized a circulating protein, the blood cell that stored or released it, and the tumor microregion in which it acted.

Historical sources create a different mismatch. ECM proteins can outlast the producing cell or state. Tissue proteome-turnover measurements show that protein half-lives vary over orders of magnitude, providing a physical basis for matrix legacy (Li et al., 2025). After therapy, involution, or cell-state transition, current populations may not resemble those that deposited the matrix. Aged and irradiated tissue matrices, together with obesity- and diet-induced remodeling, illustrate how earlier tissue states can generate persistent matrix-associated or microenvironmental legacy effects (Yang et al., 2024; Conner et al., 2024; Zhu et al., 2024; Pascual et al., 2021). Source convergence refers to several producers contributing to one microdomain. Source relay separates an initiating signal from a downstream matrix producer, whereas source replacement occurs when another lineage restores a product after the first producer is lost. Matrix legacy denotes persistent material, not a mystical memory of cellular identity.

3 Evidence for ECM production and causality

No single method captures every step from transcription to long-lived matrix function. Technologies should therefore be compared along five independent evidence dimensions: (1) source attribution, (2) protein deposition, (3) spatiotemporal localization, (4) source-specific causality, and (5) a mechanism of functional non-equivalence. A method can be strong in one dimension and weak in another. Spatial imaging may define final architecture with high resolution while revealing little about the original producer. Conditional genetics can identify a lineage requirement while leaving proteoform identity and spatial rescue unresolved. Table 1 summarizes representative studies that provide protein-level or causal evidence for tumor ECM provenance.

TABLE 1. Representative studies providing protein-level or causal evidence for tumor ECM provenance.

3.1 Single-cell and spatial transcriptomics

Single-cell RNA sequencing resolves candidate producer states, co-expression programs, and disease-associated transitions. Spatial transcriptomics adds neighborhood information and can place an FN1-high fibroblast state near an invasive edge or an endothelial ECM program near a vessel. Representative atlases reveal stromal heterogeneity and spatially restricted matrix programs that bulk expression obscures (Goveia et al., 2020; Wu et al., 2020; Jain et al., 2023; Grünwald et al., 2021; Risom et al., 2022; Heiser et al., 2023; Bansaccal et al., 2023).

However, the direct measurement is RNA. Dropout, ambient RNA, dissociation bias, and uncertain spatial deconvolution complicate even transcriptional attribution. More fundamentally, mRNA turnover, translational efficiency, protein stability, secretion, post-translational modification, extracellular processing, and matrix retention separate transcript abundance from mature ECM deposition (Vogel and Marcotte, 2012; Liu et al., 2016). RNA modifications such as N6-methyladenosine can further regulate transcript stability and translation (Zaccara et al., 2019). Examples from ECM biology make this distinction more concrete. Endothelial CD93 perturbation alters vascular fibronectin fibrillogenesis without changing FN1 transcription (Lugano et al., 2018), while PDAC collagen phenotypes vary with producer-specific processing and proteoform state (Tian et al., 2021; Chen et al., 2022). Spatial RNA location therefore need not match protein deposition. The appropriate conclusion is “a candidate producer state at this time and place.” Claims of mature ECM provenance require protein-level corroboration.

3.2 Species-resolved mass spectrometry and ECM-enrichment proteomics

Human–mouse xenografts provide a natural labeling experiment. Species-divergent peptides can assign ECM proteins to human tumor cells or mouse host cells after decellularization or biochemical ECM enrichment. This strategy demonstrated contributions from both compartments and identified tumor-cell-derived sushi, nidogen and EGF-like domains 1 (SNED1) and latent transforming growth factor beta binding protein 3 (LTBP3) in metastatic mammary carcinoma xenografts (Naba et al., 2014). Related analyses have characterized tumor education of the stromal proteome, treatment responses in pancreatic patient-derived xenografts, and species-deconvolved reference libraries (Wang et al., 2017; Wang et al., 2023; Krasny et al., 2020).

The main strength of this method is protein-level attribution, but its resolution is often binary: human tumor versus mouse host. It cannot distinguish a mouse fibroblast from a macrophage, endothelial cell, pericyte, or hepatocyte. Shared peptides may be unassignable or require probabilistic treatment. Sequence homology and database choice affect sensitivity, and tumor-host interactions cross species. Mouse stromal contamination can also distort apparent human tumor proteomes if deconvolution is not explicit (Shi et al., 2023). ECM enrichment introduces additional recovery bias because heavily crosslinked, insoluble, glycosylated, and low-abundance proteins are not extracted equally. The strongest conclusion is therefore often that deposited protein contains species-specific evidence from one compartment, not that a particular cell type produced every copy.

In complex multicellular tissues, source-resolved proteomics should therefore be interpreted as incomplete, abundance-dependent sampling rather than an exhaustive census of the matrix. ECM-enrichment protocols do not achieve complete biochemical separation from intracellular proteins, and different extraction strategies recover distinct subsets of the matrisome (Krasny et al., 2016; McCabe et al., 2021). Residual cytoplasmic or membrane-associated proteins may therefore complicate interpretation of an apparent matrix proteome. ECM extraction is also inherently selective: highly insoluble and crosslinked material can resist conventional extraction, leading to under-representation of structural ECM components (Barrett et al., 2017; McCabe et al., 2021). Abundant or readily ionized peptides are also more likely to be identified than low-abundance species. PTM-rich ECM proteins add another analytical challenge because many matrix proteins are heavily modified, and site-specific glycosylation or other modifications may escape standard bottom-up workflows (Naba, 2023); metabolic-labeling strategies have been developed to improve access to otherwise difficult-to-detect O-linked glycan modifications (Boyce et al., 2011). For multimeric ECM proteins, peptide intensities from individual chains should not be assumed to reconstruct native complex stoichiometry. Accordingly, absence of a source-specific peptide does not establish absence of production, and relative peptide abundance should not be converted directly into absolute producer contribution. In complex multicellular models, proteomic source attribution is most convincing when cross-validated with cell-selective labeling, spatial imaging, or lineage-specific perturbation.

No single analytical platform resolves all of these biases. Combining complementary methods can, however, reduce several of them. Fractionated extraction workflows that analyze cellular and ECM fractions separately can improve matrisome coverage and make cross-compartment carryover more explicit (McCabe et al., 2021). Cell-type-selective metabolic labeling can provide positive evidence of nascent synthesis from defined producers, whereas appropriately designed proximity-labeling approaches can strengthen source assignment; chemoselective enrichment of newly deposited ECM can improve detection of low-abundance nascent matrix proteins (Wei et al., 2021; Kim et al., 2021; Ling et al., 2025). Spatial extracellular proteomics can then map deposition context, and targeted LC–MS/MS with stable-isotope internal standards can provide calibrated quantification of prioritized ECM proteins or proteoforms (Macdonald et al., 2024; Rosin et al., 2024). When paired with source-specific perturbation and rescue, these methods can help identify ECM components whose cellular origin is well supported and whose functional relevance can be tested directly.

Conventional ECM proteomics remains informative without xenograft deconvolution. It identifies deposited matrisome proteins and can track temporal reprogramming or metastasis-associated changes (Papanicolaou et al., 2022; Jailkhani et al., 2023). These data should be paired with cell-resolved hypothesis generation or labeling rather than used to assign each protein retrospectively to the cell type with the highest RNA.

3.3 Cell-selective metabolic labeling and protein tags

Cell-type-selective proteomics can extend source attribution beyond the species level. Mutant aminoacyl-tRNA synthetases enable designated cells to incorporate noncanonical amino acids into nascent proteins for bioorthogonal enrichment (Alvarez-Castelao et al., 2017; Wei et al., 2021). Related secretome-labeling approaches can identify lineage-resolved extracellular proteins in tumors and circulation (Kim et al., 2021; Swietlik et al., 2023). These methods measure synthesis or release more directly than RNA and can define temporal windows.

Recent engineering approaches have expanded nascent-matrix analysis beyond bulk secretome measurements. In ex vivo lung tissue, local matrix manipulation combined with lineage-traced models enabled spatial analysis of nascent protein deposition and remodeling under controlled mechanical conditions (Ahmed et al., 2025). A glycosylation-enabled chemoselective workflow enriched newly synthesized ECM in bioengineered tumor tissues and revealed remodeling events that were under-represented by bulk proteomics (Ling et al., 2025). During murine forelimb development, three-dimensional mapping resolved the timing and architecture of ECM deposition at the developing muscle-tendon interface (Leng et al., 2025). These approaches improve temporal or spatial resolution of newly deposited matrix and could be combined with cell-selective labeling to strengthen producer assignment. By themselves, however, they do not yet provide complete cell-of-origin attribution in complex tumors.

For ECM, the analytical bottleneck lies downstream. A labeled nascent protein recovered from soluble fluid is not necessarily a deposited matrix protein. Labels or tags may alter secretion, multimerization, cleavage, or receptor binding. Low incorporation can be difficult to detect in slowly turning-over ECM, and enrichment methods optimized for soluble proteomes may miss insoluble, crosslinked material. Proximity labeling identifies residence near the enzyme rather than unambiguously identifying synthesis. Decisive validation requires recovery of labeled, mature protein from an ECM-enriched or spatially imaged compartment, with controls for labeling efficiency, protein assembly, and background transfer.

3.4 Lineage tracing and conditional genetic manipulation

Lineage tracing records the history of cells carrying a recombinase-defined label. It can reveal state changes in perivascular cells during pre-metastatic niche formation (Murgai et al., 2017). Conventional reporters do not usually tag every protein secreted by that lineage, and a labeled cell adjacent to a fiber is not proof that it produced the fiber. Long-lived ECM may also persist after the lineage disappears. Lineage history and protein history should therefore not be treated as synonymous.

Conditional deletion of an ECM gene provides causal leverage. If deletion in a defined producer reduces deposited protein and changes a tumor phenotype, it supports a requirement for that source in the tested model. In PDAC, alpha-smooth muscle actin (αSMA)-lineage Col1a1 deletion reduced fibrillar collagen and stiffness and accelerated tumor progression, whereas cancer-cell-selective collagen manipulations produced different phenotypes (Chen et al., 2021; Tian et al., 2021). Such studies provide substantially stronger evidence than transcriptional attribution. However, they are not complete provenance experiments because ECM-gene deletion can alter the producer’s adhesion, survival, migration, differentiation, stress response, and broader co-secretome. Recombination may also be incomplete or broader than the intended state. Protein deposition, producer abundance, and producer state must therefore be measured after deletion.

A conditional phenotype also cannot establish intrinsic non-equivalence. If one source is deleted while another remains in a different spatial compartment, the phenotype may reflect location or amount. The next steps are same-source rescue, cross-source rescue, abundance matching, and mechanistic comparison. Genetic epistasis downstream of a receptor is informative, but it does not substitute for replacing the missing matrix product.

3.5 Spatial proteomics, imaging, and structural analysis

Multiplexed immunofluorescence, imaging mass spectrometry, spatial extracellular proteomics, and ECM-targeted mass-spectrometric imaging can map proteins to microregions obscured by bulk proteomics (Macdonald et al., 2024; Boyle et al., 2020; Rujchanarong et al., 2021). Antibodies that recognize splice variants or neoepitopes can add proteoform information, although their specificity must be demonstrated in the relevant fixation and tissue context. Imaging mass spectrometry avoids antibody dependence but may sacrifice molecular completeness or cellular resolution.

Second-harmonic generation (SHG) and related nonlinear microscopy visualize fibrillar collagen without an exogenous label, including in living tumors (Bakker et al., 2022). SHG studies have defined collagen reorganization at tumor–stroma interfaces, prognostic fiber alignment, and local stiffening associated with collective invasion (Provenzano et al., 2006; Conklin et al., 2011; Koorman et al., 2022). Super-resolution imaging, electron microscopy, atomic-force microscopy, rheology, and tension probes further characterize fiber diameter, pore size, ligand exposure, stiffness, and viscoelasticity. These measurements define the final matrix state but do not identify the original producer. Fluorescence colocalization between a cell marker and a fiber is also insufficient because secreted proteins can move and persist.

This distinction has practical consequences. Tumor-cell DDR1 can promote collagen-fiber alignment and immune exclusion (Sun et al., 2021), establishing an organizer role without identifying the collagen-producing cell. CAF discoidin domain receptor 2 (DDR2) and collagen density affect stiffness, metastasis, and T-cell behavior through organization and signaling (Bayer et al., 2019; Kuczek et al., 2019). A collagen-rich image is therefore not a source map, and increased collagen content does not specify a particular fiber geometry or mechanical state.

3.6 Decellularized ECM as a functional comparison platform for add-back and source exchange

Decellularized matrices preserve complex cell-produced material while removing living producers. They are therefore best viewed as functional comparison platforms for testing how a pre-existing matrix state affects responder cells, rather than as direct source-attribution methods. Native or tumor-derived decellularized ECM usually integrates products from multiple cell types and retains differences in composition, architecture, stiffness, viscoelasticity, bound factors, and preparation history. Studies using primary liver cancer ECM, ovarian metastatic ECM, and decellularized organ matrices illustrate the value of this approach for preserving biologically relevant matrix cues (van Tienderen et al., 2023; Puttock et al., 2023; Pospelov et al., 2023). However, a phenotype observed after reseeding cannot be assigned to one producer or one protein unless source composition and the physical matrix state are independently controlled. Detergents, enzymes, osmotic treatments, and agitation can remove glycosaminoglycans and soluble factors, expose cryptic epitopes, disrupt basement membranes, alter fiber architecture and mechanics, or leave residual DNA and reagents (Crapo et al., 2011). If matrices from two sources require different processing or start from different cell densities, observed phenotypes may reflect technical or biomechanical differences rather than provenance.

An ideal source comparison combines source-specific knockout with direct verification that deposited target protein decreases. Same-source rescue tests whether replacing the product reverses the phenotype, whereas cross-source rescue asks whether the alternative producer’s product can substitute. Both rescues should match deposited abundance, spatial microregion, timing, fiber architecture, and bulk mechanics. Otherwise, failure to rescue cannot be attributed to producer identity. Reciprocal exchange is stronger than a one-way test, and purified-protein add-back is informative only if the relevant native proteoform and assembly state are restored.

Decellularized ECM is therefore most informative for provenance questions when combined with defined producer systems, source-specific perturbation, or matched add-back and exchange experiments, with composition and mechanics measured in parallel.

3.7 Evidence dimensions and causal standards

Evidence should be reported as a profile rather than a single prestige hierarchy. Transcriptional attribution is strong for cell state and timing but weak for deposition. Species-resolved proteomics is strong for protein source but weak for host-cell identity. Conditional genetics can provide strong source-specific causality while remaining weak on mechanism. Imaging is strong for architecture but weak for producer history. Table 2 makes these differences explicit across five evidence dimensions.

TABLE 2. Evidence-capability matrix for resolving ECM production.

A strong provenance claim should answer six questions. First, which cell produced and secreted the target product? Second, was that product directly detected in deposited ECM? Third, when and where was it produced and deposited? Fourth, does source-specific intervention alter deposition and phenotype without materially changing the producer population or state? Fifth, can a matched same-source product rescue the phenotype? Sixth, can a product from another source rescue under matched molecular, spatial, and mechanical conditions? If not, what proteoform, assembly, or receptor mechanism explains the failure? Most current studies answer the first three; fewer reach the fourth, and very few complete the fifth and sixth. Table 3 summarizes minimum experimental standards for progressively stronger provenance claims, while Figure 2 integrates the evidence dimensions and causal workflow.

TABLE 3. Minimum experimental standards for causal provenance claims.

4 Molecular and structural mediators

4.1 Alternative splicing and proteoforms

Different producer states can use different promoters and splice programs. Cellular fibronectin commonly includes extra domains A and B (EDA and EDB), whereas plasma fibronectin lacks these segments. EDA-containing fibronectin has been localized to tumor vasculature and genetically interrogated in angiogenesis models (Rybak et al., 2007; Astrof et al., 2004). Osteoblast-derived EDA-fibronectin can shape antitumor immunity in bone, and CAF-associated EDA programs have been linked to therapy resistance (Rossnagl et al., 2016; Dong et al., 2024). Recent splice manipulation further indicates that an FN1 isoform can create an invasion-permissive matrix (Franco-Valls et al., 2023). Together, these studies establish that transcript identity can matter. They support provenance dependence only when a producer state is shown to generate the isoform and when isoform, rather than producer abundance or location, explains the phenotype.

Tenascin C also has multiple splice forms, and laminin function depends on chain combinations. Expression of TNC or a laminin gene in two cell types therefore does not establish identical products. The relevant comparison should proceed from gene identity to transcript, proteoform, and supramolecular state. Source-linked splicing is a plausible determinant mechanism, but cell-resolved isoform deposition remains poorly resolved for most tumor ECM proteins.

4.2 Modification and proteolysis

Producer cells differ in glycosyltransferases, prolyl and lysyl hydroxylases, sulfation enzymes, chaperones, and redox systems. These differences can alter secretion, fibril stability, receptor binding, and protease susceptibility. P4HA1-dependent collagen prolyl hydroxylation has been linked to HIF-1α stabilization and TNBC chemoresistance (Xiong et al., 2018). PLOD2-linked succinate metabolism contributes to epithelial–mesenchymal plasticity and cancer cell stemness (Tong et al., 2023). HSP47-dependent collagen handling has been linked to metastatic progression (Wang et al., 2024; Ling et al., 2026), while proteome-wide analyses document altered proline hydroxylation in PDAC (Wu et al., 2025). Cell-autonomous prostate cancer and CAF collagen-processing studies further show that COL1A1 abundance alone does not define collagen state (Ippolito et al., 2024; Cencioni et al., 2025). Citrullination differences between breast tumors and adjacent tissue provide another modification landscape whose producer attribution remains unresolved (Tillotson et al., 2023).

Post-secretory proteolysis can override source. Endostatin and tumstatin, generated from collagens XVIII and IV, have receptor activities distinct from their parent networks (O'Reilly et al., 1997; Sudhakar et al., 2003). Collagen cleavage can expose DDR1-readable states, and proteases can release matrikines whose effects are not predictable from intact-protein abundance (Su et al., 2022). Evidence that modification and cleavage alter ECM function is strong, but evidence that two producer types create different deposited modifications under abundance- and space-matched conditions remains much rarer.

4.3 Secretion and co-secretomes

The producer’s endoplasmic-reticulum capacity, chaperones, glycosylation machinery, proteases, inhibitors, and vesicle pathways determine what is secreted and what accompanies it. HSP47-dependent collagen handling by tumor cells or in tumor–platelet interactions shows that chaperone programs can alter collagen secretion and downstream behavior (Xiong et al., 2020; Wang et al., 2024). Nicotinamide N-methyltransferase (NNMT)-driven CAF metabolic remodeling alters a broader stromal program rather than a single ECM protein (Eckert et al., 2019). Extracellular-vesicle-associated LOX and exosomal tenascin C further show that secretion route can alter localization and molecular context (Liu et al., 2023; Mirzaei et al., 2018).

This mechanism is plausible but experimentally difficult to isolate. Deleting a single ECM gene may rewire the broader secretory state, while conditioned medium transfers multiple factors. A phenotype attributed to “fibroblast-derived protein X″ could instead reflect protease balance, growth factors, vesicles, or assembly partners. Matched recombinant-protein rescue can test molecular sufficiency but may fail because it omits the native co-secretome. Conversely, rescue with whole conditioned medium cannot isolate the target. Fractionation, proteoform-resolved analysis, and receptor epistasis are therefore needed to narrow causality.

4.4 Supramolecular assembly

Source-linked co-secretion can shape stoichiometry, fiber diameter, pore size, network continuity, and growth-factor retention. Fibronectin is particularly instructive because endothelial, fibroblast, tumor-cell, and plasma sources may enter different assembly routes. CD93/MMRN2 perturbation disrupts fibrillogenesis without lowering FN1 transcription (Lugano et al., 2018), while an engineered non-fibrillar or arginine-glycine-aspartate (RGD)-deficient fibronectin state can alter metastatic behavior (Urtreger et al., 1998). Inhibiting fibronectin accumulation can change tumor growth, but such interventions target the final network rather than a single producer (Ghura et al., 2021).

SNED1 provides an emerging example of matrix assembly. Recent work indicates that fibrillar SNED1 assembly requires fibronectin and collagen I, while motif perturbation changes architecture (Leverton et al., 2026; Pally et al., 2026). Unless direct binding is demonstrated, these observations support dependence, colocalization, or candidate co-assembly rather than a defined binding interaction. They also do not establish that SNED1 produced by two cell types would be functionally non-equivalent.

4.5 Spatial mechanics and receptor reading

The same proteoform can have different effects at the tumor core, invasive front, vessel wall, immune-exclusion boundary, nerve, or premetastatic niche because local receptors, forces, proteases, and responder cells differ. Nerve-associated ECM and perivascular tenascin C illustrate this context dependence (Di Chiaro et al., 2024; Hongu et al., 2022). Crosslinking inhibition can improve T-cell migration and checkpoint response (Nicolas-Boluda et al., 2021). Matrix viscoelasticity can also promote liver cancer before overt cirrhotic stiffness develops (Fan et al., 2024).

A related example comes from our recent study of lung adenocarcinoma, which integrated single-cell and spatial transcriptomic data with ligand–receptor communication analysis and identified a potential SLC2A1–CAF–collagen signaling axis. POSTN + CAFs showed the strongest inferred contribution to collagen-mediated signaling toward SLC2A1+ malignant epithelial cells (Song et al., 2026). Although this study did not directly resolve collagen provenance or matrix mechanics, it illustrates how a malignant-cell state can be linked to a specific stromal collagen-signaling context.

These findings show that final matrix state and receptor context can explain an apparent producer effect. If fibroblast collagen is peritumoral and cancer-cell collagen is pericellular, a phenotype difference may arise from micrometer-scale location. If one niche expresses integrin α3β1 and another expresses leukocyte-associated immunoglobulin-like receptor 1 (LAIR1) or DDR1, the same material may be read differently. Producer identity is causal only insofar as it generates a persistent, measurable mediator state. Once molecular, spatial, mechanical, and receptor-readable states are genuinely matched, residual differences should prompt a search for an unmeasured mediator rather than attribution to provenance alone.

5 Rewiring ECM production during cancer progression

5.1 Early lesions

Tumorigenesis does not simply amplify a fixed matrix program. It changes which cells synthesize particular products, the states those cells occupy, and when and where their products are deposited. Resident fibroblasts, epithelial or malignant cells, endothelial cells, pericytes, macrophages, and recruited mesenchymal populations can enter or leave the producer pool. Temporal breast and pancreatic matrisome studies show staged protein changes rather than a single monotonic desmoplastic response (Papanicolaou et al., 2022; Pereira et al., 2024). Spatial studies likewise identify regional ECM competence for tumor initiation and progressive stromal changes during invasion (Risom et al., 2022; Bansaccal et al., 2023).

This process constitutes rewiring of ECM production. Producer identity or state may change, and production or deposition context may change even when the dominant producer does not. Early malignant cells can acquire a cell-autonomous ECM program while resident fibroblasts transition into contractile, inflammatory, or stress-associated states. In other settings, a pre-existing organ matrix determines which transformed cells can expand. Bulk tumor-versus-normal measurements cannot distinguish these routes. Longitudinal cell-state maps, pulse labeling, and deposited-protein measurements are needed to determine whether a new producer has emerged or an existing matrix is being remodeled.

5.2 Tumor core and invasive front

At an invasive edge, three questions should be separated: who supplies the major proteins, who organizes the tracks, and where invasive cells read them. CAF-led studies support fibroblast fibronectin assembly (Attieh et al., 2017), whereas tumor cells can organize autocrine matrices or produce their own collagen (Tian et al., 2021; Shinde et al., 2018). Collagen alignment at the tumor boundary can predict aggressive behavior even when producer identity is unknown (Provenzano et al., 2006; Conklin et al., 2011). In some tumors, deposition context and architecture may therefore explain more than producer identity.

Tumor-cell ECM is not uniformly pro-invasive. Autocrine fibronectin restrained breast-cancer metastasis (Shinde et al., 2018). By contrast, tumor-cell collagen I homotrimers promoted PDAC (Chen et al., 2022), while collagen III supported dormancy or tumor restriction in other settings (Di Martino et al., 2022; Stewart et al., 2024). These studies involve different products, proteoforms, stages, and endpoints and therefore do not support a general rule for “tumor-derived matrix.” The invasive front also illustrates role separation. CAFs may secrete collagen, tumor cells may align existing fibers through DDR1 signaling, and macrophage- or fibroblast-derived LOX-family enzymes may crosslink the network. SHG can define the resulting architecture, but a producer claim requires a cell-selective protein label or allele.

5.3 Perivascular niches and intravasation

Tumor cells, CAFs, endothelial cells, and pericytes converge at blood vessels. Endothelial CD93–MMRN2 signaling organizes perivascular fibronectin fibrils, and endothelial fibronectin can support tumor-cell adhesion in defined models (Lugano et al., 2018; Van Passen et al., 2025). However, endothelial fibronectin or its canonical receptors were dispensable for tumor angiogenesis in other conditional models (Murphy et al., 2015). This negative result is mechanistically informative because plasma fibronectin, stromal production, or alternative matrices may compensate. A universal requirement for the endothelial source is therefore untenable.

Perivascular CAF stress programs can drive angiogenesis, and tumor–stroma boundary organization can influence therapeutic response (Verginadis et al., 2022; Feng et al., 2024). Cells that synthesize a vascular basement-membrane chain need not be the same cells that assemble it or proteolyze it during intravasation. Perivascular staining therefore does not establish endothelial synthesis. A decisive design would separately pulse-label nascent protein in endothelial and perivascular lineages, image mature deposition over time, and perturb assembly or proteolysis independently.

5.4 Premetastatic source relay

Premetastatic niches make the distinction between signaling source and matrix producer especially clear. Primary tumor cells can send exosomes or cytokines, whereas resident cells in the liver, lung, bone, or omentum synthesize local ECM. Pancreatic-cancer exosomes initiate a liver cascade involving Kupffer and stellate cells and a fibronectin-rich niche (Costa-Silva et al., 2015). A widely cited report linked exosomal integrins to organotropic targeting, although the article now carries an expression of concern (Hoshino et al., 2015; Hoshino et al., 2026). Melanoma exosomes have also been reported to educate bone-marrow progenitors (Peinado et al., 2012). These studies identify upstream relay signals but do not assign production of the distant matrix to the tumor cell.

Perivascular-cell plasticity can contribute to pre-metastatic niche formation (Murgai et al., 2017), whereas a collagen-rich omentum can support integrin α2-mediated peritoneal metastasis (Huang et al., 2020). Other niche-remodeling routes include diet- and platelet-associated effects (Hergueta-Redondo et al., 2025), chronic-stress-associated neutrophil remodeling (He et al., 2024), and hypoxia-induced LOX-dependent bone-marrow-cell recruitment (Erler et al., 2009). The classical vascular endothelial growth factor receptor 1 (VEGFR1)-positive-cell/fibronectin-rich niche is compatible with several producer assignments and cannot be reduced retrospectively to one source without protein-level evidence (Kaplan et al., 2005). Periostin is better viewed within this relay framework than as a separate provenance case. Host periostin supports metastatic colonization (Malanchi et al., 2012; Fukuda et al., 2015; Wang et al., 2016; Gillot et al., 2022; Null et al., 2023), whereas tumor-cell periostin has been linked to a cytokine program that maintains breast cancer stemness (Lambert et al., 2016). Together, these studies support stage- and organ-dependent source convergence rather than cross-source non-equivalence.

5.5 Extravasation dormancy and recurrence

Disseminated cells can exploit an existing matrix, remodel it, or synthesize a new pericellular niche. Tumor-cell tenascin C supports early lung colonization, whereas S100 calcium-binding protein A4 (S100A4)-positive stromal cells provide tenascin C during metastatic outgrowth (Oskarsson et al., 2011; O'Connell et al., 2011). This sequence suggests source relay or replacement but does not establish intrinsic differences between tumor- and stromal-derived proteins. Section 6.2 examines the comparative evidence.

Dormancy adds temporal ambiguity. Perivascular and cell-matrix niches can regulate quiescence, survival, and recurrence (Ghajar et al., 2013; Ohta et al., 2022; Sreekumar et al., 2024). Tumor-cell-organized fibronectin and collagen III have been implicated in dormant states (Di Martino et al., 2022; Barney et al., 2020). During reactivation, an ECM program may be newly synthesized, inherited, or exposed by inflammation. Chemotherapy-associated neutrophil extracellular traps can awaken dormant cells through matrix-remodeling proteases, illustrating a post-secretory trigger rather than a change in the original producer (He et al., 2025).

5.6 Therapy-induced replacement and legacy

Chemotherapy, radiotherapy, antiangiogenic therapy, immunotherapy, and stromal interventions alter the survival and state of candidate producers. Radiotherapy can reorganize brain-tumor microenvironments (Watson et al., 2024), while collagen I–CXCL1 signaling can contribute to radioresistance in esophageal cancer (Yang et al., 2023). TNC has also been implicated in radiotherapy responses in head-and-neck and oral squamous carcinoma models (Loustau et al., 2026; Spenlé et al., 2021). TGF-β-associated immune exclusion, CAF states, and tumor-stroma boundaries influence checkpoint response (Feng et al., 2024; Mariathasan et al., 2018; Kieffer et al., 2020). Together, these studies support treatment-associated rewiring, but a post-treatment cell map cannot retrospectively identify which cells deposited persistent ECM.

Source replacement should be anticipated when one producer is inhibited. A surviving fibroblast state, macrophage, endothelial cell, malignant subclone, or systemic source may restore the same gene-identity product. Recent matrix- and fibrosis-directed interventions provide opportunities to measure such compensation (Tang et al., 2025; Khan et al., 2025; Parikainen et al., 2026; Wang et al., 2025). The restored material may occupy a different microregion or have a different proteoform. Post-treatment analysis should identify replacement producers, quantify deposited protein, map its architecture and location, and test whether the replacement matrix restores the phenotype.

Matrix legacy can persist after source loss. Treatment may clear malignant producers while leaving crosslinked collagen, basement membranes, matrix-bound vesicles, and mechanically stable tracks. Tissue injury can likewise generate material that outlives the acute inflammatory population. Producer-directed and matrix-state-directed interventions therefore operate on different timescales. Early failure of producer inhibition may reflect legacy matrix rather than absence of producer-mediated biology. Figure 3 summarizes these temporal changes in ECM provenance.

6 Comparative case studies

6.1 Collagen I and proteoform identity

αSMA-positive myofibroblasts supply much of the fibrillar collagen in PDAC, but malignant cells also produce a small, biologically consequential fraction. Species-specific peptide analysis, ECM enrichment, lineage-directed alleles, SHG, stiffness measurements, and procollagen-processing assays have been used to separate these contributions (Chen et al., 2022; Chen et al., 2021; Tian et al., 2021). These designs go beyond RNA attribution, although none provides a fully matched reciprocal in vivo source exchange.

Deleting Col1a1 in αSMA-positive myofibroblasts reduced fibrillar collagen and stiffness yet accelerated PDAC, accompanied by immunosuppressive myeloid changes (Chen et al., 2021). A related model supported tumor-restrictive type I collagen despite tumor-promoting effects of the broader fibroblast compartment (Bhattacharjee et al., 2021). Malignant-cell-selective studies produced apparently discordant outcomes. Bone morphogenetic protein 1 (BMP1)-dependent processing and deposition of tumor-cell collagen restrained PDAC in one study (Tian et al., 2021), whereas cancer-cell collagen I α1 homotrimers promoted growth and immune evasion in another (Chen et al., 2022). Differences in model, processing, proteoform, and endpoint preclude a simple “fibroblast good, tumor bad” interpretation.

Proteoform-resolved rescue in the homotrimer study most closely approaches a producer-mediated determinant model. PDAC cells suppressed collagen type I alpha 2 chain (COL1A2) and produced α1 homotrimers, whereas fibroblasts produced conventional α1/α2 heterotrimers. Cancer-cell Col1a1 deletion delayed disease despite maintained or increased stromal collagen. Purified homotrimer rescued spheroid behavior more effectively than heterotrimer and activated α3β1-integrin, focal adhesion kinase (FAK), protein kinase B (AKT), and extracellular signal-regulated kinase (ERK) signaling (Chen et al., 2022). These data support functional non-equivalence between distinct proteoforms generated by different producer programs, not an independent record of producer identity in chemically and structurally identical collagen.

Total abundance, homotrimer fraction, modification, cleavage, crosslinking, alignment, and receptor expression remain alternative explanations. HSP47-dependent folding and LOX-mediated crosslinking can alter collagen maturation and organization (Wang et al., 2024; Saatci et al., 2020). LAIR1 can read collagen-associated immune states (Peng et al., 2020), while postpartum remodeling shows that time and space can dominate a collagen phenotype (Lyons et al., 2011).

The collagen literature therefore provides the strongest current example of a producer-mediated effect acting through proteoform identity. Fibroblast- and tumor-cell collagen programs can be required for distinct PDAC phenotypes, and α1 homotrimers and α1/α2 heterotrimers are functionally different. Whether collagen remains non-equivalent after abundance, location, architecture, modification, and receptor access are matched is unknown. Resolving this question requires reciprocal in vivo deposition of characterized proteoforms in the same microregion with matched mechanics and receptor availability.

6.2 Tenascin C and stage-dependent replacement

Breast cancer cells can synthesize TNC during early lung colonization, whereas S100A4-positive host stromal cells provide it during later metastatic outgrowth (Oskarsson et al., 2011; O'Connell et al., 2011). Perivascular or reprogrammed stromal cells contribute in other models (Hongu et al., 2022; Lee et al., 2022). Brain tumor-initiating cells can export exosome-associated TNC that suppresses T-cell activity (Mirzaei et al., 2018). TNC can also inhibit T-cell activation and protect cancer stem-like cells from immune surveillance (Jachetti et al., 2015). The evidence therefore supports multiple producers and stage-dependent replacement, while producer identities remain coupled to different locations, splice forms, and secretory environments.

Deligne and colleagues directly compared host- and tumor-derived TNC in syngeneic mouse tumor models. Host-derived TNC was associated with antitumoral macrophages, whereas tumor-derived TNC favored an immunosuppressive phenotype. TNC- or toll-like receptor 4 (TLR4)-targeting antibodies altered these states and reduced tumor growth and spread (Deligne et al., 2020). This provides unusually direct source-stratified immune evidence. However, the study did not exchange equal amounts of isoform-matched TNC between identical spatial compartments, and the manipulations could also alter tumor or host co-secretomes.

Taken together, the Oskarsson and O'Connell studies are compatible with a relay in which early malignant TNC production is replaced or supplemented by host production (Oskarsson et al., 2011; O'Connell et al., 2011). They do not compare the two products under matched conditions. Alternative explanations include temporal availability, perivascular versus pericellular deposition, splice variants, glycosylation, vesicle association, macrophage composition, and local TLR4 accessibility. Radiotherapy-associated tenascin C phenotypes further show that treatment can alter both production context and biological reading (Loustau et al., 2026; Spenlé et al., 2021).

The tenascin C evidence is best classified as source-associated immune biology with stage-dependent replacement. Tumor and host cells can both supply consequential tenascin C, and Deligne et al. reported opposing macrophage phenotypes in mouse models. Whether this difference persists after abundance, isoform, spatial deposition, and co-secretome are matched remains unresolved. A decisive reciprocal test would express the same characterized proteoform at the same site and level before macrophage and tumor assays.

6.3 Fibronectin and source redundancy

Fibronectin arises from several local and systemic compartments and is therefore particularly susceptible to source redundancy. Tumor-cell and plasma fibronectin have been manipulated directly (Malik et al., 2010; Shinde et al., 2018), while endothelial pathways can control fibril assembly (Lugano et al., 2018). This diversity makes fibronectin a stringent test of apparent source dependence and functional interchangeability.

Endothelial CD93 deficiency disrupted perivascular fibronectin fibrils without lowering FN1 transcription (Lugano et al., 2018), whereas CAFs align invasion-guiding tracks (Attieh et al., 2017). Tumor-cell-organized fibronectin has also been linked to dormancy (Barney et al., 2020). Plasma and cellular fibronectin differ in EDA/EDB splicing. Genetic deletion studies indicate that individual EIIIA or EIIIB segments can be dispensable for angiogenesis in some contexts (Astrof et al., 2004), while more recent splice manipulation supports isoform-specific effects on invasion-permissive matrix formation (Franco-Valls et al., 2023). These experiments compare molecular variants or assembly states rather than producer identity alone. Soluble, adsorbed, and fibrillar fibronectin likewise represent distinct receptor-readable states.

Plasma fibronectin promoted lung metastasis (Malik et al., 2010), whereas autocrine tumor-cell fibronectin restrained breast-cancer metastasis (Shinde et al., 2018). Endothelial assembly promoted glioma vascular organization (Lugano et al., 2018), but conditional loss of endothelial fibronectin or cognate integrins did not prevent angiogenesis in other models (Murphy et al., 2015). Macrophage-derived fibronectin can suppress antitumor immunity (Li et al., 2026), while fibronectin from other stromal populations can also modulate antitumor immune responses (Lubosch et al., 2025). These findings remain compatible with both provenance-associated and provenance-independent explanations: another source may replace the protein, or a shared fibronectin state may function equivalently once deposited.

Local and systemic fibronectin sources therefore contribute to distinct tumor phenotypes, while compensation and assembly differences remain plausible explanations for apparent source dependence. Evidence is strongest when lineage-specific perturbation is combined with deposited-fiber imaging, splice analysis, or systemic-source genetics. A reciprocal, compensation-aware source exchange that matches proteoform, abundance, location, assembly, mechanics, and receptor readouts remains lacking.

6.4 Emerging tumor-cell-derived matrix organizers: SNED1 and LTBP3

6.4.1 Direct evidence

Species-resolved ECM proteomics in human breast-cancer xenografts assigned SNED1 and LTBP3 peptides to the human tumor compartment and detected them preferentially in highly metastatic tumors (Naba et al., 2014). Tumor-cell shRNA against either gene reduced invasion and metastatic dissemination without a comparable reduction in primary-tumor growth in the tested system. Matrix architecture also changed, including tumor encapsulation and peripheral collagen accumulation. Follow-up work supported a role for LTBP3 in early metastatic dissemination (Deryugina et al., 2018). Together, these findings provide coarse protein-level source attribution and source-specific perturbation evidence.

Recent biochemical and cell-based studies have refined the mechanism of SNED1. SNED1 supports adhesion through RGD-binding integrins, and its fibrillar assembly depends on fibronectin and collagen I in tested matrix systems (Leverton et al., 2026; Pally et al., 2025). Mutation of a leucine-aspartate-valine (LDV) motif alters matrix architecture and cell proliferation (Pally et al., 2026). Domain analysis and predicted interaction networks provide additional hypotheses (Vallet et al., 2021).

6.4.2 Suggestive evidence

Colocalization or assembly dependence suggests that SNED1 participates in fibronectin- and collagen-containing networks. It does not establish direct SNED1–fibronectin binding or identify which cell supplies each partner in vivo. Likewise, an invasion or metastasis phenotype after tumor-cell knockdown is consistent with a tumor-cell matrix-organizer function but may also reflect altered tumor-cell state, adhesion, TGF-β handling, or co-secretome.

6.4.3 Unresolved

Neither SNED1 nor LTBP3 has undergone a protein-matched same-source rescue followed by reciprocal exchange between tumor and stromal producers in vivo. It remains unknown whether another producer can deposit a molecularly matched product at the same location or whether source-linked modifications exist. The functional equivalence of the resulting matrices is therefore unresolved. For SNED1, direct binding partners in native tumor ECM and the receptor consequences of specific assembled states require further validation. For LTBP3, structural ECM function must be separated from latent TGF-β regulation.

Evidence for SNED1 and LTBP3 therefore supports tumor-compartment protein assignment and tumor-cell-specific necessity in a breast-cancer xenograft model, with emerging receptor and assembly mechanisms for SNED1. Producer-specific functional non-equivalence remains untested. The strongest evidence combines species-specific peptide assignment with source-directed knockdown. A deposition-verified rescue and reciprocal tumor/stromal source swap that matches proteoform, abundance, location, and TGF-β context are still required.

7 Alternative explanations and translational hypotheses

7.1 Confounders and causal inference

An apparent source-associated phenotype should first trigger a confounder audit. Total deposited protein may differ because of producer number, per-cell output, secretion efficiency, or protein half-life. Production stage and deposition microregion may also differ, as may the target proteoform, co-secretome, fiber architecture, and mechanics. Conditional manipulation can alter producer survival or state, while plasma leakage or another systemic source can contribute additional material. Long-lived historical matrix may persist. Technical confounders include species-specific assembly artifacts in xenografts and unequal damage during decellularization. Finally, responder cells may express different receptors across the compared locations.

The causal sequence should be explicit. First, verify the deposited target and match abundance, location, timing, and producer state. Next, perform same-source rescue and reciprocal cross-source rescue before assigning the phenotype to a proteoform, architectural, or receptor mechanism. Failure at any step does not make the study uninformative; it places the conclusion at an earlier evidence dimension.

Several negative or discordant findings illustrate this logic. Endothelial fibronectin can be dispensable in some angiogenesis models despite strong endothelial assembly evidence elsewhere (Lugano et al., 2018; Murphy et al., 2015). Myofibroblast collagen I restrains PDAC in a conditional model, whereas tumor-cell collagen I can restrain or promote disease depending on processing and proteoform (Chen et al., 2022; Chen et al., 2021; Tian et al., 2021). Periostin can originate from host or tumor cells (Malanchi et al., 2012; Lambert et al., 2016), and TNC production can shift from tumor cells to S100A4-positive stroma during metastatic outgrowth (Oskarsson et al., 2011; O'Connell et al., 2011). These observations argue against a universal ECM-source rule and favor hypotheses defined by model, stage, and molecular state. Figure 4 summarizes the major confounders and causal decision framework.

7.2 Biomarker interpretation

High FN1, COL1A1, or periostin (POSTN) can report very different processes. FN1 may indicate a malignant-cell adhesion program, CAF expansion, angiogenic endothelium, activated macrophages, tissue injury, or influx of plasma fibronectin. COL1A1 may reflect myofibroblast abundance, a tumor-cell collagen program, wound repair, crosslinked legacy matrix, or activation of a distant metastatic organ. POSTN may reflect a fibroblastic niche, a treatment-associated wound state, or a tumor-cell program. Bulk RNA and even bulk protein collapse these alternatives.

A provenance-aware biomarker should capture five attributes when feasible: predominant producing cell and state, spatial microregion, transcript or proteoform, structural state, and disease or treatment stage. Species-resolved proteomics, spatial extracellular proteomics, and integrated colorectal ECM-origin mapping show that partial versions of this profile are technically feasible (Naba et al., 2014; Macdonald et al., 2024; Lee et al., 2025). In patient samples, however, cell type cannot be inferred solely from the most correlated single-cell reference. Orthogonal protein evidence or validated cell-selective labels remain necessary.

Provenance can improve interpretation even when it is not itself causal. A fibroblast-origin collagen signal may be more specific for a stromal state than total collagen, while a tumor-origin collagen proteoform may identify a malignant biosynthetic program. Conversely, a circulating ECM fragment may primarily report remodeling rather than the producer of the parent matrix. Biomarker claims should therefore state which link in the causal chain the assay measures.

7.3 Therapeutic hypotheses

Resolving provenance changes what can be tested therapeutically. One strategy removes or reprograms a producer. FAP-directed cellular approaches can reduce fibrosis in PDAC models, but perturbing fibroblast-derived matrix can also remove tumor-restrictive functions (Chen et al., 2021; Wang et al., 2025). A second strategy targets cell-state-linked secretory programs, including HSP47-dependent collagen handling (Wang et al., 2024), NNMT-linked CAF metabolic remodeling (Eckert et al., 2019), and stress-induced matrix programs (Insua-Rodríguez et al., 2018). A third targets splicing, hydroxylation, proteolysis, assembly, or crosslinking. A fourth targets receptors that read a defined matrix state, such as integrin, DDR1, or LAIR1 pathways. The most rational target is the layer supported by the strongest causal evidence.

These implications are experimental design principles, not clinically validated provenance-selective interventions. Human spatial or outcome associations can support biomarker relevance but not therapeutic causality. Genetically engineered mouse models can support a producer requirement within that model. Cell culture and purified proteins can define receptor or proteoform mechanisms but may not reproduce native matrix. Crosslinking and fibronectin-matrix interventions have animal support in specified settings (Nicolas-Boluda et al., 2021; Khan et al., 2025). Such approaches should not be described as producer-selective therapies unless the affected producer and deposited product are demonstrated.

The producer-mediated determinant model predicts that targeting a source-linked proteoform or its biosynthetic enzyme may spare a functionally beneficial product from another source. The provenance-associated model predicts that targeting location, architecture, or receptor context will outperform producer ablation. The provenance-independent model predicts that another cell can replace the product, making source-directed treatment fragile. These are experimentally distinguishable hypotheses rather than claims of clinical readiness.

7.4 Therapeutic compensation

After a source-directed intervention, total-protein measurements alone are insufficient. Four outcomes should be distinguished. If the target remains low and the phenotype follows, the result supports source-specific necessity. If total protein recovers from a replacement producer and function also recovers, it supports interchangeability. If total protein recovers but deposition shifts and function changes, the result supports a spatial mediator. If total protein and location recover but function remains different, proteoform, co-secretome, assembly, or receptor differences should be investigated.

Therapy studies should prospectively identify substitute producers by longitudinal single-cell or spatial profiling and then verify their deposited product with proteomics, imaging, or cell-selective labeling. They should test whether total target levels recover, whether fibers or mechanics normalize, and whether the replacement product restores the relevant tumor phenotype. This design converts “stromal compensation” from a retrospective explanation into a testable source-replacement mechanism.

8 Discussion and perspectives

Producer identity is an upstream causal variable, but its functional effect must be mediated through persistent molecular, spatial, mechanical, or assembly differences. Published examples show why these layers must remain separate. PDAC collagen I links producer programs to proteoform-specific integrin signaling, whereas collagen alignment, stiffness, and receptor context can alter invasion or immune behavior even when the original producer remains unresolved (Chen et al., 2022; Sun et al., 2021; Koorman et al., 2022; Kuczek et al., 2019). Most cancer ECM studies establish transcriptional attribution, protein-level compartment assignment, or producer-specific necessity. Far fewer match abundance and deposition context, restore material from the original producer, or exchange it across producers. The remaining task is to identify the proteoform, architecture, or receptor mechanism that explains any persistent difference. Provenance-associated and provenance-independent outcomes are therefore as informative as producer-mediated determinants.

Several unresolved questions define the next phase of ECM provenance research. How often do distinct producers generate different deposited proteoforms from the same ECM gene? When producer identity, spatial location, and co-secretome covary, which has an independent causal effect? How long does production history persist after the producer changes or disappears? Can source-resolved, proteoform-aware ECM proteomics be implemented in patient tissue without a species boundary? Does therapy commonly induce source replacement or relay, and can probes or drugs recognize source-linked proteoforms or assembly states rather than total protein?

Resolving these questions will require complementary methods. Spatial transcriptomics can define candidate states, cell-selective or isotope-based labeling can assign synthesis, imaging can locate mature deposits, and engineered matrices or conditional genetics can test causality. The experimental unit should be a specified product from a specified producer state and production context, not a gene symbol attached to a broad lineage.

The practical value of this framework lies in falsifiability. A well-designed experiment should allow producer identity to lose explanatory power when products are interchangeable and should identify the measurable mediator when a producer-linked difference persists. Applied in this way, testing cellular provenance can sharpen explanations of metastasis, biomarker ambiguity, and therapeutic compensation.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: Alessio Biagioni, University of Florence, Italy

Reviewed by: Brahim Chaqour, University of Pennsylvania, United States

Zihan Ling, Northwestern University, United States