The challenge of lung adenocarcinoma. Lung adenocarcinoma (LUAD) is the most common subtype of non-small cell lung cancer and remains one of the deadliest cancers worldwide. Despite advances in targeted therapy, patients frequently develop resistance, and many are diagnosed at advanced stages when curative treatment is no longer possible.
Cancer-associated fibroblasts (CAFs) are specialized cells within the tumor's surrounding tissue that receive signals from cancer cells, immune cells, and other nearby components. These signals direct CAFs to promote tumor growth, support cancer spread, and help tumors resist treatments - though some CAF subtypes may also have tumor-suppressive roles.
CAF heterogeneity is a key complication in understanding their role. CAFs can be divided into distinct subtypes including antigen-presenting CAFs (apCAFs), myofibroblastic CAFs (myCAFs), inflammatory CAFs (iCAFs), and a less-studied group called extracellular matrix CAFs (eCAFs). Advances in single-cell technology now allow researchers to distinguish these subtypes with unprecedented precision.
Extracellular matrix CAFs (eCAFs) have been identified in gastric cancer as having pro-invasive properties, and they appear to closely interact with immune cells, particularly a type of immune cell called SPP1+ macrophages. However, their role specifically in lung adenocarcinoma had not been thoroughly investigated before this study.
Single-cell RNA sequencing data from 15 lung adenocarcinoma tumor samples were drawn from the publicly available GSE131907 dataset, which contains 58 samples including primary tumors, lymph node samples, brain metastases, and normal tissue. This technology allows researchers to measure gene activity in thousands of individual cells simultaneously.
Bulk transcriptomic and clinical data were collected from three independent patient cohorts: 503 samples from the TCGA-LUAD database, 226 samples from GSE31210, and 398 samples from GSE72094. Using multiple cohorts helps ensure that findings are reproducible and not specific to one patient group.
Quality control and batch correction were rigorously applied to the single-cell data, filtering out low-quality cells and removing technical artifacts that can arise when samples are processed at different times. Tools including DoubletFinder (to remove duplicate cell readings) and Harmony (to harmonize data across samples) were used to ensure data reliability.
Cell type identification was performed by matching gene expression patterns to known cell type markers from published databases. This allowed the team to classify cells as epithelial, stromal, or immune, and then further subdivide them into specific populations such as the four CAF subtypes and four macrophage subtypes.
The Scissor algorithm was used to link individual cells in the single-cell data to patient survival outcomes recorded in the bulk transcriptomic datasets. This approach identifies which cell subpopulations are associated with poor or favorable survival without requiring direct measurement of survival in single-cell experiments.
CIBERSORTx, a computational deconvolution tool, was then used to estimate the proportions of eCAFs and SPP1+ macrophages in the bulk RNA datasets. Patients were split into high- and low-infiltration groups based on median cell abundance, and Kaplan-Meier survival curves were generated to compare outcomes between groups.
Pseudotime trajectory analysis was performed using the Slingshot algorithm to reconstruct the developmental history of CAF and macrophage populations. This approach orders cells along a timeline of maturation, revealing how one cell type can evolve into another as the tumor progresses.
Cell-to-cell communication analysis was conducted using the CellChat software package, which identifies pairs of proteins (ligands and receptors) through which different cell types signal to one another. This revealed how eCAFs and SPP1+ macrophages exchange signals and cooperate in the tumor microenvironment.
Survival association using Scissor analysis revealed that eCAFs made up 75.9% of the CAF subpopulation associated with poor patient outcomes, while iCAFs dominated the group associated with better outcomes. Among macrophages, SPP1+ macrophages represented 62.4% of the poor-prognosis-associated macrophage population.
Multi-cohort Kaplan-Meier survival analysis confirmed that patients with high eCAF infiltration had significantly shorter overall survival in all three cohorts: TCGA (p=0.0033), GSE31210 (p=0.034), and GSE72094 (p=0.012). Similarly, high SPP1+ macrophage infiltration was consistently associated with worse outcomes across all cohorts.
Stage and grade analysis showed that both eCAF and SPP1+ macrophage proportions increased as tumors advanced to later stages. Importantly, eCAF abundance also rose as tumor differentiation decreased - meaning eCAFs were most prevalent in the most aggressive, poorly differentiated tumors.
Racial background was assessed and found not to significantly affect the infiltration levels of these cell subsets, indicating that the prognostic associations observed are likely broadly applicable across diverse patient populations.
Pseudotime trajectory analysis of CAFs revealed two distinct differentiation paths. In one path (Lineage 2), eCAFs serve as the evolutionary endpoint, with cells progressing from myofibroblastic CAFs through intermediate states toward eCAFs. This positions eCAFs as the most mature and potentially most aggressive CAF form in the tumor.
Functional evolution along Lineage 2 showed a biological progression from muscle contraction-related activities early in development, shifting through immune regulation and Wnt signaling, and ultimately arriving at extracellular matrix remodeling. This suggests eCAFs develop specialized machinery for remodeling the tissue scaffold around the tumor.
Macrophage trajectories showed that RETN+ macrophages serve as a starting point, with one lineage progressing ultimately toward SPP1+ macrophages through intermediate FOLR2+APOC1+ macrophages. Functionally, this lineage involves a shift from cellular immunity and antigen presentation toward signaling regulation and immune cell recruitment.
The parallel trajectories of eCAFs and SPP1+ macrophages both converge toward more immunosuppressive and pro-tumorigenic endpoints, suggesting that tumor progression drives coordinated evolution of both the stromal and immune compartments simultaneously.
Gene Set Variation Analysis (GSVA) compared pathway activity across CAF subtypes and found that eCAFs were particularly enriched in pathways related to angiogenesis, epithelial-mesenchymal transition, protein secretion, Notch signaling, and glycolysis - all hallmarks of aggressive tumor behavior.
SPP1+ macrophages showed enrichment in overlapping pathways including angiogenesis, epithelial-mesenchymal transition, inflammatory response, and Notch signaling, reinforcing the functional alignment between these two cell types in driving tumor progression.
Cell communication analysis using CellChat identified strong signaling between eCAFs and SPP1+ macrophages specifically through the COL1A1-CD44 and COL1A2-CD44 ligand-receptor pairs. COL1A1 and COL1A2 encode components of type I collagen, a major structural protein of the extracellular matrix, while CD44 is a receptor that mediates cell attachment and migration.
COLLAGEN and FN1 (fibronectin) signaling networks were also found to be highly active in eCAFs, particularly in their communication with epithelial cancer cells. Dense collagen is known to physically block immune cells from entering tumors, and eCAFs appear to be a primary source of this immunosuppressive matrix scaffold.
Feature selection involved a three-step filtering process: first identifying genes distinguishing CAFs from other cell types, then finding genes specifically elevated in eCAFs within the CAF population, and finally screening for genes upregulated in tumor versus normal tissue in the TCGA dataset. This process converged on 28 core eCAF-related gene signatures.
101 machine learning algorithm combinations were systematically tested using the Mime1 framework, a comprehensive toolkit that trains and evaluates a wide range of model types. Models were ranked by their mean C-index performance across all three patient cohorts, with the C-index measuring how well the model discriminates between patients who survive longer versus shorter.
The best-performing model was the StepCox[forward] + plsRcox combination, achieving C-index values of 0.65 in the TCGA training set, 0.75 in the GSE31210 validation set, and 0.63 in the GSE72094 validation set. The TCGA cohort was used for model training, while the other two cohorts served as independent validation sets.
Only genes with statistically significant associations in univariate survival analysis (Cox proportional hazards) were included in the model, ensuring that each gene in the signature contributed meaningful prognostic information independently before being combined into the final multi-gene model.
High-risk versus low-risk classification using the median risk score from the model consistently separated patients with significantly different survival outcomes. In the TCGA training set, the hazard ratio was 1.92 (95% CI: 1.43-2.58, p less than 0.001), meaning high-risk patients were nearly twice as likely to die than low-risk patients.
Even stronger separation was observed in the GSE31210 validation cohort with a hazard ratio of 7.41 (95% CI: 3.8-14.42, p less than 0.001), and the GSE72094 cohort also showed significant stratification with a hazard ratio of 1.81 (95% CI: 1.25-2.62, p=0.002). This consistency across independent datasets demonstrates the model's generalizability.
ROC curve analysis evaluated the model's ability to predict survival at 1, 3, and 5 years. In the GSE31210 validation set, area under the curve (AUC) values were particularly strong: 0.805 at 1 year, 0.771 at 3 years, and 0.78 at 5 years. AUC values above 0.7 are generally considered good predictive performance.
Meta-analysis combining results across all cohorts showed a pooled hazard ratio of 2.05 (Fixed Effect Model) to 2.66 (Random Effect Model), confirming that the model's risk stratification effect is stable and reproducible even when accounting for variability between datasets.
This study provides the first systematic characterization of eCAFs in lung adenocarcinoma, establishing their central role in driving aggressive tumor behavior through extracellular matrix remodeling, immune exclusion, and promotion of epithelial-mesenchymal transition - a process by which cancer cells acquire the ability to spread.
The eCAF-SPP1+ macrophage crosstalk identified here parallels findings in colorectal and gastric cancers, suggesting this stromal-immune partnership may be a conserved mechanism across cancer types. In colorectal cancer, CAFs and SPP1+ macrophages together create a fibrotic barrier that prevents immune cells from reaching tumor cells.
Dense collagen produced by eCAFs can physically restrict T cell movement into tumors, creating what researchers call an immune-excluded microenvironment. By identifying COL1A1-CD44 and COL1A2-CD44 as the key communication channels between eCAFs and SPP1+ macrophages, this study points toward specific molecular targets that could be disrupted to restore immune access to tumors.
The eCAF scoring signature (eCAFsRS) outperformed several established clinical biomarkers and immune checkpoints in predicting 1-, 3-, and 5-year survival, suggesting it could add value to existing prognostic tools used in clinical practice. However, the authors caution that further validation in prospective clinical trials incorporating full treatment and staging information is needed.
This study establishes eCAFs as key drivers of lung adenocarcinoma progression, linked to poor prognosis through their roles in extracellular matrix remodeling and immunosuppressive crosstalk with SPP1+ macrophages. The COL1A1/COL1A2-CD44 signaling axis represents a promising therapeutic target to disrupt this cooperation.
The 28-gene prognostic model provides a clinically useful tool for stratifying lung adenocarcinoma patients into high- and low-risk groups, with performance validated across three independent cohorts. Its incorporation of eCAF-specific biology represents an advance over conventional prognostic approaches.
Key limitations include the purely computational nature of the findings, which lack in vitro and in vivo experimental validation. The sample size could also be expanded, and the model requires validation in real-world clinical cohorts that include detailed treatment information.
Future research directions proposed by the authors include constructing animal models of LUAD, using longitudinal spatial transcriptomics to map microenvironment changes during treatment, and screening targeted drug combinations using patient-derived organoids. These experiments would validate the causal role of eCAFs and test whether disrupting eCAF-macrophage interactions can improve treatment outcomes.