Baishideng Publishing Group Inc, 7041 Koll Center Parkway, Suite 160, Pleasanton, CA 94566, USA
Share the Article
Lin X, Lin ZL, Li Y, Huang CD, Chen S, Shi DL, Chen XQ. Spatial macrophage signatures and prognostic relevance of NFKB1 in colorectal cancer. World J Clin Oncol 2026; 17(11): 125917 [DOI: 10.5306/wjco.125917]
Xin Lin, Zhen-Liang Lin, Yan Li, Cang-Dian Huang, Shuai Chen, Xian-Qiang Chen, Department of General Surgery (Emergency Surgery), Fujian Medical University Union Hospital, Fuzhou 350001, Fujian Province, China
Dong-Liang Shi, Department of Pathology, Fujian Medical University Union Hospital, Fuzhou 350001, Fujian Province, China
Co-corresponding authors: Dong-Liang Shi and Xian-Qiang Chen.
Author contributions: Lin X and Chen XQ designed the research study; Lin X and Lin ZL performed the bioinformatic analyses; Li Y and Huang CD collected and organized the data; Chen S provided methodological guidance; Lin X drafted the manuscript; Shi DL provided guidance on manuscript writing and performed English-language editing; Chen XQ critically revised the manuscript and supervised the study; Shi DL and Chen XQ played important and indispensable roles in the manuscript preparation as the co-corresponding authors; all the authors reviewed and approved the final manuscript.
AI contribution statement: During the preparation of this manuscript, the authors used ChatGPT (OpenAI) and Gemini (Google) solely to improve the English language, grammar, readability, and clarity of the manuscript. These tools were not used to generate research data, perform statistical analyses, interpret the results, formulate scientific conclusions, or select and verify references. All AI-assisted content was carefully reviewed and revised by the authors, who take full responsibility for the accuracy, integrity, originality, and final content of the manuscript.
Supported by Joint Funds for the Innovation of Science and Technology, Fujian Province, No. 2024Y9283; the Construction Project of Fujian Province Minimally Invasive Medical Center, No. (2021)76; and the Fujian Provincial Health Youth Research Project, No. 2022QNA020.
Conflict-of-interest statement: The authors declare that they have no conflicts of interest related to the publication of this manuscript.
Corresponding author: Xian-Qiang Chen, MD, Additional Professor, Department of General Surgery (Emergency Surgery), Fujian Medical University Union Hospital, No. 29 Xinquan Road, Fuzhou 350001, Fujian Province, China. cxq760818@163.com
Received: July 21, 2026 Revised: August 15, 2026 Accepted: October 28, 2026 Published online: November 24, 2026 Processing time: 120 Days and 11.1 Hours
Abstract
BACKGROUND
Tumor-associated macrophages (TAMs) exhibit substantial spatial and transcriptional heterogeneity in colorectal cancer (CRC). However, the prognostic relevance of spatial macrophage-associated transcriptional programs remains incompletely understood.
AIM
To characterize spatial macrophage-associated transcriptional programs and evaluate the prognostic relevance of NFKB1 in CRC.
METHODS
A publicly available conventional 10x Genomics Visium FFPE colorectal cancer specimen was integrated with The Cancer Genome Atlas colon adenocarcinoma cohort and external Gene Expression Omnibus bulk-transcriptomic cohorts. Spatial M1-like and M2-like macrophage-associated transcriptional-program activity was quantified using AUCell, whereas the corresponding bulk-cohort signature scores were calculated using single-sample gene set enrichment analysis. Candidate genes were prioritized using least absolute shrinkage and selection operator Cox regression and other survival-learning methods. Subsequently, the associations of NFKB1 with survival, clinicopathological characteristics, immune-related transcriptional features, representative protein staining, pathway enrichment, and predicted drug responses were evaluated.
RESULTS
Spatial distribution analysis revealed distinct structural characteristics of different compartments within the tumor immune system. The activity of an M2-like macrophage-associated signature was greater in tumor-core, invasive-front, and stroma-enriched regions, whereas the activity of an M1-like macrophage-associated signature was higher at tumor-peripheral regions and in areas at tumor-normal interfaces. Higher M1-like scores were correlated with early tumor stage and better overall survival (OS). The assessed prognostic models showed moderate discrimination (the best mean time-dependent area under the curve was 0.650). NFKB1 was prioritized as a candidate gene associated with the M1-like signature. In selected external cohorts, higher NFKB1 expression was associated with improved OS and correlated with transcriptional features characteristic of immune-active states. Differences in predicted drug response values were observed between the NFKB1-high and NFKB1-low groups.
CONCLUSION
M1- and M2-like macrophage-associated transcriptional programs showed distinct spatial distributions in colorectal cancer. Higher M1-like signature activity was preferentially observed in tumor-peripheral regions and was associated with immune-active features and more favorable survival. NFKB1 expression was also associated with the M1-like program and favorable outcomes in selected external cohorts. These findings support NFKB1 as a candidate bulk-transcriptomic biomarker.
Core Tip: This study integrated spatial and bulk-transcriptomic data to characterize macrophage-associated transcriptional programs in colorectal cancer. M1-like activity was preferentially observed at tumor-peripheral regions and was associated with immune-active features and favorable outcomes. NFKB1 was subsequently prioritized as a candidate bulk-transcriptomic biomarker. The findings describe spatial and prognostic associations.
Citation: Lin X, Lin ZL, Li Y, Huang CD, Chen S, Shi DL, Chen XQ. Spatial macrophage signatures and prognostic relevance of NFKB1 in colorectal cancer. World J Clin Oncol 2026; 17(11): 125917
Colorectal cancer (CRC) ranks as the third most common cancer type globally and is the second leading cause of cancer-related deaths[1]. The tumor microenvironment (TME) of CRC is enriched in immunocytes, among which tumor-associated macrophages (TAMs) are the major component[2]. Recently, TAMs have been shown to encompass heterogeneous and context-dependent transcriptional states that cannot be fully captured by a binary M1/M2 classification. Nevertheless, M1- and M2-like gene signatures remain useful descriptive frameworks for comparing inflammatory and immunoregulatory macrophage-associated programs[3]. M1-like TAMs are mainly activated via the LPS/TLR4 and IFN-γ-JAK/STAT1 pathways, resulting in the secretion of inflammatory cytokines such as tumor necrosis factor (TNF) and interleukin (IL)-12, which enhance antitumor immune responses[4]. In contrast, M2-like TAMs are typically polarized via the activation of the IL-4/IL-13-STAT6 axis. IL-10/transforming growth factor-beta (TGF-β)-mediated immunosuppression and the hypoxic TME, which increase HIF-1α expression and activate the pro-angiogenic VEGF pathway, are the primary pathways involved in the activation of M2-like TAMs in CRC. Additionally, TAMs can alter the extracellular matrix and inhibit the function of effector T cells[5]. TAMs have also been confirmed to function as key regulators of the immunosuppressive TME in various solid malignancies[6,7]. In CRC, macrophage-associated programs are spatially heterogeneous, with immunoregulatory and tissue remodeling states frequently enriched in hypoxic regions and at the invasive front.
To demonstrate the association between the location and the spatial function of TAMs in CRC, spatial transcriptomics has been widely used to map the distribution of different TAM phenotypes across spatial niches. Recent high-definition spatial transcriptomic profiling of CRC has demonstrated spatially organized macrophage states across distinct tumor niches[7]. Single-cell and spatial transcriptomic studies of multiple tumor types have established distinct functionally and spatially defined TAM states, such as SPP1high macrophages and SELENOPhigh macrophages[8]. SPP1high macrophages preferentially aggregate around hypoxic tumor stromal niches[9] and co-localize with FAP+ fibroblasts[10], which create desmoplastic and T-cell-excluded regions. These spatial interactions have been associated with SPP1-receptor signaling and inflammatory or profibrotic pathways, including those involving TGF-β and IL-1 signaling, thereby contributing to extracellular matrix remodeling and immune exclusion[10,11]. Across different tumor types, SELENOPhigh macrophages have been associated with comparatively immune-permissive tissue states, although their functions appear to vary according to tumor context[8,9,12]. Evidence from ovarian cancer suggests that hypoxia may promote the transition of SELENOPhigh macrophages toward an SPP1high state through VEGFA–EPHB2 signaling, although whether this mechanism also operates in CRC remains unknown[13]. In this study, integrating spatial transcriptional patterns with bulk tumor expression and survival data may help identify candidate genes associated with macrophage-related programs and clinical outcomes[14].
Biologically, NFKB1 encodes the p105 precursor, which is proteolytically processed into the p50 subunit of the NF-κB transcription factor family. Through this role, NFKB1 links inflammatory signaling with immune regulation and cell-survival programs[4,15,16]. However, its potential function in macrophage polarization in CRC remains uncertain. Previous studies have linked p50 activity to immunoregulatory macrophage states and unfavorable outcomes in CRC[17]. Moreover, tumors with a large number of TAMs that highly express p50 at the invasive front tend to have a poor prognosis[18]. These results suggest that the role of NFKB1 in CRC strongly depends on the specific polarization phenotype of TAMs. Moreover, it still remains unclear whether NF-κB can promote M1-type antitumor activity.
Here, we combined conventional spatial transcriptomic data from a publicly available CRC specimen with bulk-transcriptomic cohorts to characterize the spatial distribution of M1- and M2-like macrophage-associated transcriptional programs in colorectal cancer. Subsequently, machine learning and survival analyses revealed genes associated with the M1-like signature and clinical outcomes. NFKB1 was selected as a focus for additional analysis because of its relatively stable model coefficients and well-established role in immune regulation. We also explored its associations with immune-related transcriptional features, clinicopathological characteristics, survival outcomes, and predicted drug responses. This study was not intended to prove macrophage-specific causality but rather to identify spatial and transcriptomic associations.
A CRC transcriptome resource was assembled from the Gene Expression Omnibus (GEO) and The Cancer Genome Atlas (TCGA). The TCGA-colon adenocarcinoma (TCGA-COAD) cohort served as the principal model-development cohort. Model discrimination was explored across 14 dataset-endpoint combinations comprising TCGA-CRC overall survival (OS), disease-specific survival (DSS), and progression-free interval (PFI) and 11 external GEO cohort-endpoint combinations. Separately, four GEO cohorts-GSE17537, GSE29621, GSE71187, and GSE72970-were selected for cohort-specific visualization and evaluation of the prognostic association of NFKB1. The complete dataset-endpoint list, sample numbers, endpoint definitions, and model-performance estimates are provided in Supplementary Tables 3 and 5. RNA-Seq by Expectation Maximization-normalized RNA-sequencing data for TCGA-COAD were downloaded from the UCSC Xena browser and log2(x+1)-transformed. Primary TCGA-COAD tumor samples with complete overall-survival data were retained for the cohort-specific NFKB1 analysis, yielding 425 eligible cases from an initial 512. Separately, the original prognostic-model performance export labelled “TCGA-CRC” contained 588 OS-evaluable, 566 DSS-evaluable, and 586 PFI-evaluable cases. These model-performance objects were analyzed according to their original dataset labels and were not treated as identical to the 425-case TCGA-COAD NFKB1 analysis. Raw CEL files from Affymetrix platforms (GSE17537, n = 55; GSE29621, n = 65; GSE72970, n = 124) were processed with the robust multi-array average algorithm implemented in the oligo package (v1.66.0). Probes mapping to identical gene symbols were collapsed by retaining the probe with the highest mean expression value. GSE71187 (n = 52, Agilent platform) was background-corrected and quantile-normalized using limma (v3.58.1). Relapse-free survival was used for GSE17537, whereas OS was used for GSE29621, GSE71187, and GSE72970 in the cohort-specific NFKB1 survival analyses. Endpoints with different definitions were analyzed and reported separately and were not treated as interchangeable. Intercohort heterogeneity was assessed by principal component analysis. Because each validation cohort was analyzed independently rather than being pooled, cross-cohort batch correction was not applied; annotated processing batches within individual cohorts were corrected for using the ComBat function of the sva package (v3.50.0). Clinical variables, including OS status and time, tumor–node–metastasis (TNM) stage, age, and sex, were extracted and harmonized across cohorts under a common schema.
NFKB1 immunohistochemistry images were retrieved from the HPA (version 23.0) using antibody HPA027305. Three colorectal cancer specimens and three normal colon specimens from different patients were included in an exploratory semiquantitative image analysis. The corresponding HPA record identifiers were 1887511, 1887519, and 1887529 for colorectal cancer and 6636830, 6636814, and 6636822 for normal colon tissue. For each image, three nonoverlapping 100 × 100-pixel regions of interest were evaluated using fixed color-threshold parameters. The 3,3′-diaminobenzidine (DAB)-positive area percentage, integrated DAB optical density, and mean DAB optical density were recorded. The mean of the three regions of interest was calculated for each image and used as the independent patient-level statistical value. Comparisons between the two unmatched groups were performed using two-sided Welch’s t-tests. Because only three independent specimens were available per group, this analysis was considered exploratory rather than an independent clinical protein-validation cohort.
Spatial characterization of macrophage polarization
A publicly available human CRC spatial transcriptomic specimen generated using the conventional 10x Genomics Visium Spatial Gene Expression platform for formalin-fixed paraffin-embedded tissue was used for the analyses presented in Figure 1A-C. The specimen consisted of a 5-μm section from a colorectal cancer tissue block obtained from BioIVT Asterand Human Tissue Specimens (block 1281585B) and was profiled on slide V10L13-021, capture area B1. A total of 2660 spatial spots were detected, with a median of 7438 detected genes and 25312 unique molecular identifier counts per spot and a sequencing depth of approximately 77288 reads per spot. Processed expression matrices and corresponding spatial coordinates generated by the 10x Genomics workflow were used for downstream analysis. Because conventional Visium spots may contain transcripts from multiple adjacent cells, the resulting signature scores and cell-type enrichment scores were interpreted as composite spatial transcriptional signals rather than single-cell measurements or direct cell counts.
Figure 1 Spatial transcriptomics reveals distinct regions enriched for M1-like and M2-like signatures and their clinical relevance in colorectal cancer.
A: Spatial feature plots showing AUCell-based M1-like and M2-like macrophage-associated transcriptional-program scores across a conventional 10x Genomics Visium FFPE colorectal cancer (CRC) specimen from tissue block 1281585B, slide V10L13-021, capture area B1. Each point represents one spatial spot rather than an individual cell; B: Comparison of mean AUCell signature scores for the M1-like and M2-like macrophage-associated transcriptional programs across malignant, mixed, and normal regions of specimen CRC2; P values for regional comparisons are shown. AUCell scores represent gene-set activity and are distinct from receiver-operating-characteristic area under the curve (AUC); C: Correlation-based spatial association matrix and network visualization illustrating the association between macrophage signature activity (AUC) and inferred abundances of major tumor microenvironment cell types (e.g., CD4⁺ T cells, CD8⁺ T cells, NK cells, B cells, plasma cells, macrophages, DCs, neutrophils, endothelial cells, fibroblasts, epithelial cells, and tumor cells). The colors indicate positive/negative correlations, and the line thickness reflects the correlation strength; D: Violin/box plots of single-sample gene set enrichment analysis (ssGSEA) scores for M1 and M2 signatures in early-stage (I–II) vs advanced-stage (III-IV) CRC; E: Kaplan–Meier analysis of overall survival stratified by high vs low M1 or M2 ssGSEA scores; group sizes and log-rank P values are indicated. CRC: Colorectal cancer; ssGSEA: Single-sample gene set enrichment analysis; TME: Tumor microenvironment.
M1- and M2-like macrophage-associated transcriptional programs were evaluated using predefined gene sets[4,19-35]. Supplementary Table 1 provides the gene symbols and aliases, signature assignments, functional annotations, evidence interpretations, species-related cautions, and inclusion status for human CRC scoring. The supporting literature is cited in the Methods section[4,19-35]. Mouse-specific markers, including CHI3L3 and RETNLA, were retained in the table for source transparency but excluded from human CRC scoring. Spatial M1-like and M2-like macrophage-associated transcriptional-program activity was quantified using the AUCell package. For each spatial spot, detected genes were ranked according to their expression, and an area-under-the-recovery-curve score was calculated to represent the relative enrichment of the predefined signature genes. The AUCell score was interpreted as transcriptional-program activity rather than macrophage abundance. The resulting scores were projected onto the original spatial coordinates using SpatialFeaturePlot in Seurat, without interpreting individual spots as single cells. The resulting scores were projected onto the original spatial coordinates without additional spatial smoothing, imputation, or neighborhood-based averaging. Spatial cell-type enrichment scores were generated using the Cottrazm package. A CRC-specific single-cell RNA-sequencing reference was assembled from publicly available samples of the corresponding cancer type. After quality control, the mean expression of the top 25 cell-type-specific genes was used to construct a reference signature matrix. The get enrichment matrix and enrichment analysis functions were then used to generate cell-type enrichment scores for each spatial spot. Spatial tissue regions were classified according to the resulting malignant-cell enrichment score. Spots with a malignant-cell score of 1 were defined as malignant, those with a score of 0 as nonmalignant, and those with intermediate scores as mixed. These enrichment scores were used for exploratory regional classification and were not interpreted as exact cell proportions.
Differences in gene expression levels or macrophage-associated transcriptional program scores between spatial regions were explored using two-sided Wilcoxon rank-sum tests, with Benjamini-Hochberg correction for multiple comparisons where applicable. Associations among gene expression, transcriptional program scores, and externally derived cell-type proportion estimates were assessed using Pearson or Spearman correlation and interpreted as spatial covariation. The spatial data source, scoring procedures, and existing regional comparison results are summarized in Supplementary Table 6.
Machine learning-based screening and prognostic model construction
Genes associated with the M1-like macrophage transcriptional program were first identified by Spearman correlation analysis. Differential expression of these candidate genes between TCGA colorectal cancer and normal tissues was subsequently evaluated using PanCanAtlas EBPlusPlus-adjusted RNASeqV2 expression data. Expression values were standardized within the cancer type using z scores, and values below -3 or above 3 were treated as outliers and excluded. Tumor and normal tissues were compared using two-sided Wilcoxon rank-sum tests. Genes associated with the M1-like program and showing differential expression between tumor and normal tissues were retained for prognostic-model evaluation. Prognostic modeling was subsequently performed in the TCGA-COAD cohort using LASSO-Cox, ridge-Cox, elastic-net Cox, CoxBoost and stepwise Cox regression. The penalized Cox models were implemented using the glmnet package. For elastic-net regression, the mixing parameter was set to α = 0.5. The penalty parameter λ was determined by 10-fold cross-validation on the basis of partial-likelihood deviance, and coefficients corresponding to lambda.min were retained. CoxBoost modeling was performed using likelihood-based boosting methods, with the penalty parameter optimized using optimCoxBoostPenalty and the number of boosting steps selected by cv. CoxBoost. Stepwise Cox regression was conducted using the Akaike information criterion with bidirectional variable selection.
For each model, an individual risk score was calculated as the weighted sum of gene expression values using the corresponding regression coefficients. Model discrimination was evaluated using Harrell’s concordance index and time-dependent receiver operating characteristic analysis. The TCGA-COAD cohort was used for model development, whereas the independent GEO cohorts were used for external performance evaluation.
Survival analyses
Survival analyses were performed using OS or recurrence-related endpoints, depending on the available annotations for each cohort. The prognostic association of NFKB1 expression was evaluated using Kaplan–Meier analysis and Cox proportional hazards regression. For exploratory Kaplan–Meier visualization, patients were stratified into NFKB1-high and NFKB1-low groups using the cohort-specific median expression value, and survival distributions were compared using the log-rank test. Continuous NFKB1 expression was used in the principal Cox regression analyses. Median dichotomization was not interpreted as a clinically validated biomarker cutoff. NFKB1 expression was also evaluated as a continuous variable by Cox regression. Univariable Cox models were used to estimate hazard ratios (HRs), 95%CIs, and P values. No missing values were imputed. Each cohort-specific analysis was restricted to cases with complete data for the corresponding endpoint and included covariates. Consequently, the number of evaluable cases and covariates varied across analyses. Cohort-specific sample counts, endpoint labels, event counts, and observed time summaries available from the existing exports are provided in Supplementary Table 3. The univariable and multivariable Cox regression results, including hazard ratios, 95%CI, and P values, are provided in Supplementary Table 4.
Immune landscape profiling and biological function analysis
For complementary group-based comparisons, patients were divided into NFKB1-high and NFKB1-low groups according to the median expression level, and differences in immune cell or immune signature scores were evaluated using two-sided Wilcoxon rank-sum tests with false discovery rate (FDR) adjustment where applicable. Immune-related transcriptional features were further characterized using curated sets of immunostimulatory genes, immunoinhibitory genes, chemokines, and human leukocyte antigen genes obtained from the Tumor and Immune System Interaction Database, together with published cancer-immunity-cycle and immunotherapy-response signatures. Cancer-immunity-cycle activity was evaluated using Tracking Tumor Immunophenotype-derived scores, whereas cytolytic activity, tertiary lymphoid structures (TLSs), IFN-γ, T-cell-inflamed, and chemokine-related signatures were assessed using EaSIeR-associated immune response metrics.
For KEGG gene set enrichment analysis, TCGA-CRC samples were ranked according to NFKB1 expression. Samples in the highest 30% and lowest 30% of NFKB1 expression were defined as the NFKB1-high and NFKB1-low groups, respectively. Differential expression between these groups was evaluated using the limma package, and all genes were ranked according to the resulting log2 fold change. Gene set enrichment analysis was subsequently performed using the GSEA function in clusterProfiler with KEGG gene sets. Normalized enrichment scores (NES), nominal P values, and Benjamini-Hochberg-adjusted P values were calculated. Positive NES indicated relative enrichment in the NFKB1-high group, whereas negative scores indicated relative enrichment in the NFKB1-low group.
Drug sensitivity prediction
Drug-response phenotypes were estimated using the `oncoPredict` package. Pharmacogenomic reference expression profiles and corresponding drug-response measurements from a cancer cell-line panel were used to train ridge regression models, which were subsequently applied to the TCGA-COAD expression matrix to generate sample-level predicted drug-response values. During model fitting, empirical-Bayes batch correction and phenotype power transformation were applied, and the lowest 20% of genes by expression variance were removed. Predicted drug-response values were interpreted as estimated half-maximal inhibitory concentrations (IC50) rather than observed clinical responses.
Associations between NFKB1 expression and predicted drug response were further evaluated using Spearman correlation analysis, with Benjamini-Hochberg correction applied across the tested drug associations. Where NFKB1-high and NFKB1-low groups were compared, between-group differences in estimated IC50 values were assessed using non-parametric tests.
Statistical analysis
All analyses were conducted using R version 4.3.2. Continuous variables were compared using the Wilcoxon rank-sum test or Student’s t test according to the distributional properties assessed using the Shapiro-Wilk test, and categorical variables were compared using the χ2 or Fisher exact test. Survival distributions were estimated using the Kaplan-Meier method and compared using the log-rank test. HRs with 95%CIs were obtained from Cox proportional hazards regression. For analyses involving multiple simultaneous comparisons, P values were adjusted by the Benjamini-Hochberg procedure. Two-sided P values less than 0.05 were considered statistically significant.
RESULTS
Spatial distribution of macrophage-associated transcriptional programs in CRC
A distinct spatial compartmentalization pattern of M1- and M2-like TAM transcriptional programs was identified (Figure 1A). M2-like TAM transcriptional programs (CD163high and MRC1high) were predominantly observed in the tumor core and invasive margin regions. In contrast, transcripts characteristic of M1-like TAM transcriptional programs, including IL1B, TNF, CXCL9, and NOS2, accumulated around the tumor edge as well as at the junction between the tumor and normal tissue (Figure 1B). To further illustrate the ecological features and relationships among immune components, a Spearman correlation matrix and network-based visualizations were used (Figure 1C). Specifically, M1-like macrophage-associated transcriptional programs showed positive spatial correlations with signatures of several lymphocyte populations, including CD8+ T cells, CD4+ T cells, and NK cells, suggesting that regions with higher M1-like signature activity were also characterized by more immune-active transcriptional features. In contrast, M2-like macrophage-associated transcriptional programs were preferentially observed in regions characterized by hypoxia- and stromal-associated features, including fibroblast-enriched areas. These spatial associations suggest distinct immune and stromal contexts for M1- and M2-like macrophage-associated programs but do not establish direct cellular interactions or immune recruitment.
We also evaluated the clinical implications by categorizing patients according to tumor stage and comparing their immune scores. In advanced-stage CRC (stages III-IV), M1-like signature scores were lower in stage III-IV disease than in stage I-II disease (P < 0.001; Figure 1D). Moreover, Kaplan-Meier analysis demonstrated that a lower M1 signature score was associated with poorer OS (P = 0.020; Figure 1E).
Identification of NFKB1 as a key prognostic biomarker by machine learning
Through LASSO-Cox regression analysis, we identified 14 candidate genes associated with M1-like TAM transcriptional programs (Figure 2A and B). Within the TCGA-COAD cohort, differential expression analysis revealed that the expression levels of seven genes, namely, CCL2, CCL5, CD86, NFKB1, NFKBIA, CYBB, and GBP2, were lower in tumor tissues than in their normal counterparts (Figure 2C). NFKB1 was consistently recognized across multiple machine learning models, with stable coefficient estimates. Given its well-established role in macrophage polarization and immune regulation, it was prioritized for inclusion in subsequent analyses. To construct a prognostic signature with better generalizability, we incorporated these candidate genes into a comprehensive machine learning framework and evaluated multiple modeling approaches, including Ridge regression, CoxBoost, Elastic Net, and the stepwise Cox model (Figure 2D). Among the evaluated algorithms, the backward and bidirectional stepwise Cox models achieved the highest overall mean of the 1-, 3-, and 5-year time-dependent AUC values across the evaluated dataset-endpoint combinations (overall mean AUC = 0.650). The cohort-specific mean AUC values for these models ranged from 0.566 to 0.754, with the highest value observed for GSE17537 relapse-free survival. These results indicate modest and heterogeneous discrimination rather than uniformly strong predictive performance (Figure 2D). Despite variations in the coefficient estimates derived from different algorithms, NFKB1 exhibited negative coefficients in multiple models (β = -0.326 to -0.344; Figure 2E).
Figure 2 Construction and evaluation of an M1-related prognostic model using LASSO-Cox and multialgorithm machine learning.
A: Tenfold cross-validation curve for LASSO Cox regression showing model deviance across log(λ), with the number of retained features indicated; B: LASSO coefficient profiles of candidate genes across log(λ); dashed vertical lines denote the selected λ values; C: Differential expression analysis of M1-like TAM-associated candidate genes between colorectal cancer tissues and normal tissues; D: Comparison of prognostic-model performance across 14 dataset-endpoint combinations using the mean of the 1-, 3-, and 5-year time-dependent area under the curve (AUC) values. The evaluated algorithms included backward, forward, and bidirectional stepwise Cox regression, CoxBoost, ridge Cox, elastic-net Cox, and LASSO-Cox. Because the endpoint definitions differed across datasets, each dataset-endpoint combination was evaluated separately. The complete AUC values are provided in Supplementary Table 5; E: Distribution of regression coefficients of candidate genes across multiple survival-learning models. bP < 0.01; dP < 0.0001. AUC: Area under the curve.
Cohort-specific prognostic associations of NFKB1 in independent cohorts
NFKB1 transcript expression was lower in CRC tissues than in adjacent normal tissues. Exploratory semiquantitative analysis of publicly available HPA immunohistochemistry images obtained using antibody HPA027305 showed lower NFKB1 staining measurements in CRC specimens than in normal colon specimens (Figure 3A and B). The patient-level means were lower in CRC than in normal colon specimens for DAB-positive area (CRC vs normal: 2.65% vs 73.90%, P = 0.016), mean DAB optical density (75.01 vs 120.85, P = 0.005), and integrated DAB optical density (20433.67 vs 899072.44, P = 0.023). Each group included three specimens from different patients, and the mean of three regions of interest per image was used as the patient-level statistical value. Subsequently, we appraised the prognostic implications of NFKB1 expression. In GSE29621, higher NFKB1 expression was associated with a lower hazard in both univariable analysis (HR = 0.530, 95%CI: 0.337-0.833, P = 0.006) and multivariable analysis incorporating the available clinicopathological variables (HR = 0.474, 95%CI: 0.300-0.750, P = 0.001). Similar protective associations were observed in GSE72970 in the univariable (HR = 0.779, 95%CI: 0.633-0.957, P = 0.017) and multivariable analyses (HR = 0.712, 95%CI: 0.556-0.910, P = 0.007). In TCGA-COAD, the association was borderline in univariable analysis (HR = 0.684, 95%CI: 0.465-1.004, P = 0.052) and was not statistically significant after multivariable adjustment (HR = 0.750, 95%CI: 0.481-1.168, P = 0.203) (Figure 3C). Therefore, the data support cohort-specific protective prognostic associations rather than uniform independent validation across all cohorts. Exploratory median-stratified Kaplan-Meier analyses showed less favorable relapse-free survival in the NFKB1-low group in GSE17537 (log-rank P = 0.025) and less favorable OS in the NFKB1-low groups in GSE29621 (log-rank P < 0.001), GSE71187 (log-rank P = 0.001), and GSE72970 (log-rank P = 0.018) (Figure 3D). Because the cohorts used different endpoints and median dichotomization, these analyses were interpreted separately and as exploratory visualizations.
Figure 3 Exploratory Human Protein Atlas staining differences and cohort-specific prognostic associations of NFKB1 expression in colorectal cancer.
A: Representative NFKB1 immunohistochemistry images retrieved from the Human Protein Atlas using antibody HPA027305. Three colorectal cancer specimens and three normal colon specimens from different patients are shown. The corresponding Human Protein Atlas record identifiers, patient characteristics, pathological annotations, staining intensity, staining quantity, staining location, and image names are provided in Supplementary Table 2; B: Exploratory semiquantitative comparisons of the 3,3′-Diaminobenzidine (DAB) positive area percentage, mean DAB optical density, and integrated DAB optical density between normal colon and colorectal cancer specimens. Three nonoverlapping regions of interest were evaluated per image, and their mean was used as the independent patient-level value. Each point represents one patient specimen (n = 3 per group). Comparisons were performed using two-sided Welch’s t-tests; C: Forest plot of univariable and multivariable Cox regression analyses evaluating the prognostic impact of NFKB1 together with clinicopathological variables (e.g., age, sex, stage/T/N/M) in The Cancer Genome Atlas and independent Gene Expression Omnibus validation cohorts; D: Exploratory Kaplan–Meier curves stratified by cohort-specific median NFKB1 expression. Relapse-free survival was analyzed in GSE17537, whereas overall survival was analyzed in GSE29621, GSE71187, and GSE72970. Group sizes and log-rank P values are shown. Median dichotomization was used for visualization and does not represent a clinically validated cutoff. aP < 0.05; bP < 0.01; TCGA: The Cancer Genome Atlas; HR: Hazard ratio. DAB: 3,3′-Diaminobenzidine.
Immune-related transcriptional features and computationally predicted drug-response differences associated with NFKB1 expression
To characterize the immune-related transcriptional features associated with NFKB1 expression, we performed Spearman correlation and immune signature analyses. Spearman correlation analysis revealed a positive correlation between NFKB1 expression and the expression of immune-stimulating molecules, specifically C10orf54, ENTPD1, and ICOSLG, as well as T-cell-recruiting chemokines, such as CCL2, CCL5, CXCL9, CXCL10, and CXCL12 (Figure 4A). Similarly, NFKB1 expression was associated with the expression of immune checkpoint genes, including CD274, LGALS9, and PDCD1, as well as HLA-DQA1, further supporting an association between NFKB1 expression and immune-related transcriptional features. Quantitative immune-signature analysis showed that the NFKB1-low group had lower T-cell-inflamed, tertiary lymphoid structure, interferon-γ, and cytolytic-activity scores than the NFKB1-high group (all P < 0.05; Figure 4B). These differences represent bulk-transcriptomic associations and do not establish an immunologically cold phenotype or a causal effect of NFKB1. KEGG-GSEA demonstrated distinct pathway patterns between the NFKB1 expression groups. ECM–receptor interaction (NES = 2.90, FDR-adjusted P = 5.43 × 10-9), B-cell receptor signaling (NES = 2.16, FDR-adjusted P = 4.28 × 10-5), and TGF-β signaling (NES = 2.05, FDR-adjusted P = 1.83 × 10-4) were relatively enriched in the NFKB1-high group. In contrast, autophagy–animal (NES = -1.76, FDR-adjusted P = 7.07 × 10-5), RIG-I-like receptor signaling (NES = -1.64, FDR-adjusted P = 2.75 × 10-4), and cytosolic DNA-sensing (NES = -1.55, FDR-adjusted P = 0.0078) were relatively enriched in the NFKB1-low group (Figure 4C). Cell-line-trained pharmacogenomic models were used to explore computationally predicted drug-response differences associated with NFKB1 expression. The NFKB1-low group showed different estimated IC50 values for selected compounds related to genome maintenance and MAPK/ERK signaling (Figure 4D). These estimates do not represent observed treatment responses in patients with CRC.
Figure 4 Immune-related transcriptional features, KEGG-GSEA patterns, and computationally predicted drug-response associations according to NFKB1 expression.
A: Heatmaps of differential immune-related molecules between NFKB1-low and NFKB1-high tumors, including immunostimulators, chemokines, and immunoinhibitors; significance is indicated by superscript lowercase letters; B: Distribution comparisons of immune activation signatures between groups, including the T-cell–inflamed score, tertiary lymphoid structure signature, interferon-γ signature, and cytolytic activity, are shown as density plots with embedded box plots and P values; C: KEGG gene set enrichment analysis comparing the highest 30% and lowest 30% of TCGA-CRC samples ranked by NFKB1 expression. All genes were ranked according to the limma-derived log2 fold change. Positive normalized enrichment scores indicate relative enrichment in the NFKB1-high group, whereas negative scores indicate relative enrichment in the NFKB1-low group. Benjamini-Hochberg-adjusted P values were used to control the false discovery rate (FDR). Complete results are provided in Supplementary Table 7; D: Predicted drug sensitivity and mechanism-of-action mapping based on the pharmacogenomic resource Genomics of Drug Sensitivity in Cancer. Candidate compounds associated with subgroup-specific vulnerability are linked to their putative target pathways; edges represent drug–pathway associations. All available GDSC1 and GDSC2 compound–NFKB1 correlation coefficients and nominal P values from the original significant-result exports are provided in Supplementary Table 8. The source exports did not retain an FDR-adjusted P value column. These results are computational associations and do not represent observed clinical responses. aP < 0.05; bP < 0.01; cP < 0.001. TLS: Tertiary lymphoid structure; IFN-γ: Interferon-γ; CYT: Cytolytic.
DISCUSSION
In CRC, TAMs demonstrate characteristic spatial distributions and polarized activation states, namely, M1-like TAM transcriptional programs and M2-like programs. Such classification may influence malignant behaviors within the TME[36,37]. Overall, M1-like macrophages in the classically activated state can enhance tumor cytotoxicity and immune-stimulatory functions. Conversely, M2-like TAM transcriptional programs in the alternatively activated state are generally linked to immune evasion, angiogenesis, and the dissemination of metastatic foci. Clinicopathological research has further indicated that the relative proportions of these subpopulations within tumors have prognostic implications. An elevated M1/M2 ratio within the tumor is associated with a more favorable prognosis in diverse cancers and with a stronger antitumor immune response and longer survival[2,38,39]. In contrast, a TME dominated by M2-like TAM transcriptional programs is associated with greater invasion and an inferior prognosis[40]. These observations support a prognostic association between macrophage-related transcriptional states and CRC outcomes, although they do not establish that M1-like or M2-like macrophage programs directly determine disease progression. Recent investigations in spatial transcriptomics have further elaborated on these perspectives by revealing the heterogeneity of TAMs at both the cellular and anatomical levels. Through in-depth mapping analyses of CRC specimens, transcriptionally unique TAM subpopulations situated within specific microenvironmental zones have been identified[7,12]. Spatial and single-cell analyses have revealed an invasive-front niche in which HLA-G⁺ tumor cells are closely associated with SPP1⁺ macrophages, suggesting the coordinated establishment of an immunosuppressive microenvironment[41].
NFKB1, a pivotal gene, integrates inflammatory signals with cell survival mechanisms. Aberrant regulation of NFKB1 is closely linked to the initiation and progression of CRC[42]. In our analysis, decreased NFKB1 expression was associated with an immunosuppressive TME characterized by decreased TLS abundance, cytolytic activity, and IFN-γ-related scores, together with reduced expression of T-cell-recruiting chemokines, including CXCL9 and CXCL10. These tissue-level associations do not establish an immunosuppressive mechanism or identify the cellular source of NFKB1 expression. These associations are consistent with the broader spatial and functional heterogeneity of TAMs described previously[43]. Cytokines, including TNF-α, IL-1, IL-6, and CXCL8, play a vital role in the recruitment and activation of leukocytes, which also maintain the inflammatory state and increase tumor survival capabilities[15,44]. An aberrant NFKB1 signaling pathway involving p50 can shift the myeloid cell program toward an M2-like TAM transcriptional program, thereby promoting a cytokine milieu associated with immune escape and rejection[4,45]. Spatial organization of the tumor stroma boundary has been linked to immune activation and immunotherapy efficacy in CRC[46]. In our dataset, higher NFKB1 expression was similarly associated with stronger immune-inflammatory and chemokine signatures. NFKB1 expression was lower in CRC tissues than in normal colon tissues, whereas relatively higher expression among patients with CRC was associated with more favorable outcomes in selected cohorts. These observations are not contradictory because they represent different comparisons. The tissue comparison evaluates the overall expression change associated with CRC, whereas the survival analyses evaluate interpatient variation within CRC. Thus, NFKB1 may be generally downregulated in CRC, while relative preservation of its expression within tumors may be associated with a more favorable clinical outcome. Overall, bulk-tumor NFKB1 expression was associated with immune-active transcriptional features, although these associations do not establish macrophage-specific NFKB1 activity or a direct regulatory mechanism. In this context, differences between high and low NFKB1 expression may reflect distinct immunological states. Higher NFKB1 expression was associated with chemokine activity and lymphocyte-related transcriptional signatures. These observations generate a hypothesis regarding potential immunotherapy relevance. Our integrated analysis prioritized NFKB1 as an exploratory bulk-transcriptomic candidate showing protective prognostic associations in selected cohorts. Together with previous evidence supporting therapeutic reprogramming of TAMs[3,6], these findings provide a rationale for investigating NFKB1-directed modulation of the CRC immune microenvironment.
Shifting TAMs from the M2 type to the M1 type can enhance the antitumor immune response and increase the efficacy of immunotherapy[47,48]. F1929-1458, a small-molecule NFKB1 agonist, suppressed CT26 tumor growth in a T-cell-dependent manner by activating NFKB1 signaling in tumor cells, promoting the release of immunogenic damage-associated signals, and enhancing dendritic cell activation[49]. In addition, altering innate signal transduction and reprogramming TAMs can reshape the myeloid microenvironment and thus inhibit M2 polarization[5,50]. Because the functions of M1-like TAM transcriptional programs are closely coupled to glycolytic metabolism, metabolic intervention may provide an additional strategy for promoting pro-inflammatory macrophage activity[51]. Whether this approach restores NFKB1-dependent functions remains unknown. Beyond metabolic reprogramming, the oncoPredict analysis showed lower estimated IC50 values for selected agents in the NFKB1-low group. The cell-line-trained pharmacogenomic analysis identified correlations between NFKB1 expression and estimated IC50 values for selected MAPK-related agents. Although the KRAS-MAPK pathway represents an established therapeutic axis in CRC[52], the specific relationship between NFKB1 expression and MAPK inhibitor sensitivity has not been experimentally demonstrated. Oncogenic RAS-MAPK signaling can promote an immunosuppressive microenvironment by restricting T-cell infiltration and recruiting suppressive myeloid populations[53]. Preclinical studies have indicated that MAPK inhibition can alter immune cell composition and activity[54]. Consistent with this rationale, early clinical evidence in BRAFV600E-mutant CRC has demonstrated the feasibility and antitumor activity of combined PD-1, BRAF, and MEK inhibition[55], although the benefit of such combinations remains molecularly and context-dependent[56]. Thus, the lower estimated IC50 values observed for selected MAPK-related agents in the NFKB1-low group may reflect an indirect association between MAPK/JNK and NF-κB-related transcriptional states. This interpretation remains a computational hypothesis and was not tested experimentally or clinically in the present study[15,57]. This hypothesis requires experimental validation in CRC cells and macrophage-containing model systems. Modulating the NFKB1/NF-κB signaling pathway can reshape the TME, increase dendritic cell activation and T-cell infiltration, and potentially even improve the efficacy of immune checkpoint blockade therapy[7,49]. Our study involved a comprehensive analysis of the macrophage polarization phenomenon in CRC. These findings support NFKB1 as an exploratory bulk-transcriptomic candidate associated with immune-related transcriptional features and cohort-specific survival outcomes. They do not establish that NFKB1 directly regulates antitumor immunity or is ready for clinical prognostic application.
We acknowledge several limitations in our study. The study design is inherently subject to selection bias. Validating our results across the four GEO datasets was also complicated by small sample sizes and platform differences. The spatial analysis was based on a single conventional Visium FFPE CRC specimen and therefore could not capture interpatient spatial heterogeneity or provide single-cell resolution. The main limitation, however, is the lack of single-cell sequencing and proteomics data. We had to rely on algorithms to estimate the M2 macrophage polarization phenotype instead of measuring it directly. These findings still require validation through experimental studies and prospective clinical trials.
CONCLUSION
This retrospective, hypothesis-generating study identified distinct spatial distributions of M1-like and M2-like macrophage-associated transcriptional programs in the analyzed CRC specimen. NFKB1 transcript expression and exploratory protein-staining measurements were lower in CRC tissues than in normal colon tissues. Nevertheless, among patients with CRC, relatively higher tumor NFKB1 expression was associated with immune-active transcriptional features and more favorable outcomes in selected cohorts. These findings support NFKB1 as an exploratory bulk-transcriptomic candidate with protective prognostic associations rather than a clinically validated or macrophage-specific biomarker.
ACKNOWLEDGEMENTS
The authors sincerely thank the Department of General Surgery (Emergency Surgery) at Fujian Medical University Union Hospital for its institutional and academic support throughout this study. The authors also gratefully acknowledge the funding agencies and grant recipients for providing financial support for this research.
Oliveira MF, Romero JP, Chung M, Williams SR, Gottscho AD, Gupta A, Pilipauskas SE, Mohabbat S, Raman N, Sukovich DJ, Patterson DM; Visium HD Development Team, Taylor SEB. High-definition spatial transcriptomic profiling of immune cell populations in colorectal cancer.Nat Genet. 2025;57:1512-1523.
[RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)][Cited by in RCA: 217][Reference Citation Analysis (0)]
Bill R, Wirapati P, Messemaker M, Roh W, Zitti B, Duval F, Kiss M, Park JC, Saal TM, Hoelzl J, Tarussio D, Benedetti F, Tissot S, Kandalaft L, Varrone M, Ciriello G, McKee TA, Monnier Y, Mermod M, Blaum EM, Gushterova I, Gonye ALK, Hacohen N, Getz G, Mempel TR, Klein AM, Weissleder R, Faquin WC, Sadow PM, Lin D, Pai SI, Sade-Feldman M, Pittet MJ. CXCL9:SPP1 macrophage polarity identifies a network of cellular programs that control human cancers.Science. 2023;381:515-524.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 119][Cited by in RCA: 568][Article Influence: 189.3][Reference Citation Analysis (2)]
Porta C, Ippolito A, Consonni FM, Carraro L, Celesti G, Correale C, Grizzi F, Pasqualini F, Tartari S, Rinaldi M, Bianchi P, Balzac F, Vetrano S, Turco E, Hirsch E, Laghi L, Sica A. Protumor Steering of Cancer Inflammation by p50 NF-κB Enhances Colorectal Cancer Progression.Cancer Immunol Res. 2018;6:578-593.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 38][Cited by in RCA: 50][Article Influence: 6.3][Reference Citation Analysis (2)]
Murray PJ, Allen JE, Biswas SK, Fisher EA, Gilroy DW, Goerdt S, Gordon S, Hamilton JA, Ivashkiv LB, Lawrence T, Locati M, Mantovani A, Martinez FO, Mege JL, Mosser DM, Natoli G, Saeij JP, Schultze JL, Shirey KA, Sica A, Suttles J, Udalova I, van Ginderachter JA, Vogel SN, Wynn TA. Macrophage activation and polarization: nomenclature and experimental guidelines.Immunity. 2014;41:14-20.
[RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)][Cited by in Crossref: 5296][Cited by in RCA: 5102][Article Influence: 425.2][Reference Citation Analysis (5)]
Xue J, Schmidt SV, Sander J, Draffehn A, Krebs W, Quester I, De Nardo D, Gohel TD, Emde M, Schmidleithner L, Ganesan H, Nino-Castro A, Mallmann MR, Labzin L, Theis H, Kraut M, Beyer M, Latz E, Freeman TC, Ulas T, Schultze JL. Transcriptome-based network analysis reveals a spectrum model of human macrophage activation.Immunity. 2014;40:274-288.
[RCA] [PubMed] [DOI] [Full Text][Cited by in Crossref: 2120][Cited by in RCA: 1749][Article Influence: 145.8][Reference Citation Analysis (0)]
Ozato Y, Kojima Y, Kobayashi Y, Hisamatsu Y, Toshima T, Yonemura Y, Masuda T, Kagawa K, Goto Y, Utou M, Fukunaga M, Gamachi A, Imamura K, Kuze Y, Zenkoh J, Suzuki A, Niida A, Hirose H, Hayashi S, Koseki J, Oki E, Fukuchi S, Murakami K, Tobo T, Nagayama S, Uemura M, Sakamoto T, Oshima M, Doki Y, Eguchi H, Mori M, Iwasaki T, Oda Y, Shibata T, Suzuki Y, Shimamura T, Mimori K. Spatial and single-cell transcriptomics decipher the cellular environment containing HLA-G+ cancer cells and SPP1+ macrophages in colorectal cancer.Cell Rep. 2023;42:111929.
[RCA] [PubMed] [DOI] [Full Text][Cited by in RCA: 107][Reference Citation Analysis (3)]
Tian J, Chen JH, Chao SX, Pelka K, Giannakis M, Hess J, Burke K, Jorgji V, Sindurakar P, Braverman J, Mehta A, Oka T, Huang M, Lieb D, Spurrell M, Allen JN, Abrams TA, Clark JW, Enzinger AC, Enzinger PC, Klempner SJ, McCleary NJ, Meyerhardt JA, Ryan DP, Yurgelun MB, Kanter K, Van Seventer EE, Baiev I, Chi G, Jarnagin J, Bradford WB, Wong E, Michel AG, Fetter IJ, Siravegna G, Gemma AJ, Sharpe A, Demehri S, Leary R, Campbell CD, Yilmaz O, Getz GA, Parikh AR, Hacohen N, Corcoran RB. Combined PD-1, BRAF and MEK inhibition in BRAF(V600E) colorectal cancer: a phase 2 trial.Nat Med. 2023;29:458-466.
[RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)][Cited by in Crossref: 199][Cited by in RCA: 177][Article Influence: 59.0][Reference Citation Analysis (1)]
P-Reviewer: Racz A, MD, PhD, Full Professor, Professor, Croatia; Yang Y, MD, PhD, Researcher, Visiting Professor, Germany S-Editor: Liu H L-Editor: A P-Editor: Lei YY