Machine learning algorithms develop a tumor-educated platelets-related gene signature to predict colorectal cancer prognosis and therapy response Authors: Yueqiong Lao, Jie Yang, Minli Hu, Jing Li, Weikang Xu, Jiahui Xu, Huan Wang, Hechenhao Jiang, Zhihao Pei, Xinyi Qiu, Kunyuan Wang, Xuan Li, Hui Yang Published: 2026 ABSTRACT SummaryTumor-educated platelets (TEPs) have recently emerged as an important component of liquid biopsy, yet the clinical relevance in colorectal cancer (CRC) remains unclear. Here, we employed 10 machine learning algorithms to develop a stable, accurate TEP-related gene signature (TEPGS) to explore its links to tumor-associated macrophages (TAMs) and spatial platelet abundance. TEPGS correlated strongly with poor prognosis and outperformed 71 published gene signatures in predicting CRC overall survival. Multi-omics analysis displayed that high TEPGs were marked by increased TP53 mutations, copy number alterations, diminished immune features, enrichment of pro-tumor SPP1+/FCN1+ TAMs, and elevated spatial platelet abundance. Patients with high TEPGS exhibited resistance to immunotherapy but responded to a BRAF V600E inhibitor, while TEPGS showed tentative value for predicting cetuximab response and preliminary utility for bevacizumab. Functional assays confirmed ARPC1B as an oncogene. Our findings establish TEPGS as a valuable biomarker for prognostic stratification and tailored therapy selection in CRC. Introduction Colorectal cancer (CRC) ranks in third place in terms of incidence and remains the second leading cause of cancer-related death worldwide,1 with more than 1.9 million new cases and 903,859 deaths occurring in 2022. Despite compelling evidence that cancer screening reduces CRC-related mortality2 and significant advances in CRC diagnosis and treatment, the 5-year overall survival (OS) for metastatic CRC (mCRC) remains alarmingly low at approximately 14%.3 Chemotherapy, targeted therapy, and immunotherapy are the cornerstones of CRC treatment. In recent years, technologies such as whole-genome sequencing, bulk and single-cell RNA sequencing (scRNA-seq), spatial transcriptome sequencing, and transgenic mouse models have deepened our understanding of the molecular mechanisms of CRC and facilitated the development of targeted therapy and immunotherapy.4 Targeted therapy may avoid chemotherapy-related cytotoxicity, but its efficacy is often hampered by intrinsic or acquired resistance mechanisms.5 Currently, CRC exhibits particularly low responsiveness to immunotherapy, with only approximately 15% of patients (those with mismatch repair deficiency (dMMR)/microsatellite instability-high (MSI-H) tumors) demonstrating clinical benefit from such treatment.6,7 Although multiple predictive biomarkers, including programmed death-ligand 1 (PD-L1) expression, tumor-infiltrating lymphocytes, T cell receptor clonality, tumor mutation burden (TMB), neoantigen burden (TNB), and dMMR/MSI-H status, have demonstrated association with immunotherapy efficacy, their clinical utility remains limited by spatiotemporal heterogeneity, moderate accuracy, and small percentage populations.8,9,10 Therefore, the development of both more sensitive prognostic indicators and more robust predictive biomarkers represents an urgent need. Platelets are circulating anucleated cell fragments that are primarily involved in hemostasis and wound healing. Platelets play a crucial role in tumor development, metastasis, and drug resistance.11,12 Platelet-released adenine nucleotides reportedly induce endothelial barrier breakdown and promote cancer cell transendothelial migration and extravasation.13 In circulation, platelets protect tumor cells against attack by natural killer cells by transferring normal major histocompatibility class I molecules to the surface of tumor cells.14 In addition, platelets help circulating tumor cells establish a premetastatic niche and influence the sensitivity of patients with cancer to chemotherapy and targeted therapies.15 For instance, platelets increase the survival of colon cancer cells and increase their resistance to 5-fluorouracil and paclitaxel.16 Platelet counts and the mRNA profiles of tumor-educated platelets (TEPs) have emerged as potential cancer classifiers.17,18,19,20 Poor progression-free survival and/or OS are closely linked to elevated platelet counts in various malignancies, including lung, gastric, breast, ovarian, and colorectal cancers (CRC).11 In cancer diagnosis, TEPs-derived RNA panels can accurately discriminate patients with cancer from healthy controls.17,18 Although TEPs-derived RNA profiles have advanced in terms of blood-based detection and tumor origin site identification, their potential to predict mortality risk in patients with cancer remains unclear. TEPs and macrophages are recognized as important parts of the tumor microenvironment (TME), which consists of not only malignant cells but also other cell types, such as immune cells, endothelial cells, fibroblasts, and extracellular matrix.21 Advances in scRNA-seq technology have facilitated the identification of cell subtypes and functional status within tumors. A series of subpopulations of macrophages separated by scRNA-seq have been found to exert distinct immune functions, such as SPP1+ tumor-associated macrophages (TAMs), FCN1+ TAMs, and SLC40A1+ TAMs.22,23 Thus, exploring the relationship between TEPs and different TAM subtypes will provide a viewpoint on TEPs' functions. Constructing tailored and accurate survival prediction models based on transcriptome sequencing data for patients with CRC would facilitate a more objective risk classification and assessment in clinical settings and avoid under- or over-treatment. Currently, multiple prognostic factors such as age, gender, tumor-node-metastasis (TNM) stage, TMB, MSI, and TP53, KRAS, and BRAF mutation status are commonly applied in clinical practice. Additionally, various liquid biopsy components, including circulating tumor cells, circulating tumor DNAs, long non-coding RNAs, circular RNAs, and PIWI-interacting RNAs, are associated with CRC prognosis.24 However, as the latest component of liquid biopsies, the clinical significance of TEPs for personalized prognosis prediction and treatment in CRC has not been thoroughly addressed. In the present study, we comprehensively analyzed the clinical value of TEPs-derived RNA expression profiles in patients with CRC by integrating multi-omics data and a machine learning approach. Based on bulk RNA sequencing data, we first developed and verified a promising TEP-related gene signature (TEPGS) from a combination of 101 machine learning algorithms in 1598 patients with CRC from seven independent cohorts. To facilitate the clinical utility of the TEPGS, we evaluated its ability to predict the prognosis of CRC in the training and validation cohorts by constructing TEPGS-integrated nomograms. Mechanically, we revealed a robust correlation between TEPGS and mutational landscape, CNVs, immune landscape, and immunotherapy response of CRC. Notably, we applied scRNA-seq analysis to pinpoint a strong association between high TEPGS scores and SPP1+/FCN1+ TAMs, indicating a correlation between TEPGS and TAMs remodeling. Moreover, we examined the response of high- and low- TEPGS groups to chemotherapy and identified eleven potential chemotherapeutic compounds for patients with high TEPGS. The targeted therapy, including bevacizumab and cetuximab, response of TEPGS in 115 patients with mCRC from 4 independent datasets was further investigated. Finally, we demonstrated that ARPC1B knockdown inhibited cell proliferation in vitro and in vivo. The purpose of our study was to offer a useful tool for prognosis stratification and personalized therapeutic strategy of patients with CRC. Results Identification of TEPs-related genes and functional annotation We identified TEPs-related DEGs in blood samples from the GEO: GSE68086 dataset (patients with CRC vs. healthy donors). A total of 2306 TEPs-related DEGs were screened, including 1084 upregulated and 1222 downregulated genes with the false discovery rate threshold <0.05 and |log2 fold change| ≥ 0.5 (Figure 1A). Using the same screening criteria, we identified DEGs in TCGA-CRC samples between tumor and adjacent normal tissues (Figure 1B). Among them, 158 DEGs were significantly upregulated in both blood and tissue samples, whereas 212 DEGs were significantly downregulated (Figures 1C and 1D).Figure 1Identification of TEP-related DEGs in CRC(A) Volcano map of DEGs between cancer patients’ TEPs and healthy donors’ platelets from the GEO: GSE68086 dataset.(B) Volcano map of DEGs between tumor samples and adjacent normal tissues from the TCGA-CRC dataset.(C) Venn diagram shows the intersection of upregulated DEGs in the GEO: GSE68086 and TCGA-CRC datasets.(D) Venn diagram shows the intersection of downregulated DEGs in the GEO: GSE68086 and TCGA-CRC datasets.(E) KEGG analysis of 158 upregulated DEGs in two cohorts.(F) KEGG analysis of 212 downregulated DEGs in two cohorts. Identification of TEP-related DEGs in CRC (A) Volcano map of DEGs between cancer patients’ TEPs and healthy donors’ platelets from the GEO: GSE68086 dataset. (B) Volcano map of DEGs between tumor samples and adjacent normal tissues from the TCGA-CRC dataset. (C) Venn diagram shows the intersection of upregulated DEGs in the GEO: GSE68086 and TCGA-CRC datasets. (D) Venn diagram shows the intersection of downregulated DEGs in the GEO: GSE68086 and TCGA-CRC datasets. (E) KEGG analysis of 158 upregulated DEGs in two cohorts. (F) KEGG analysis of 212 downregulated DEGs in two cohorts. Furthermore, Kyoto Encyclopedia of Genes and Genomes (KEGG) analysis was utilized to investigate the biological roles associated with the DEGs connected to TEPs. The upregulated genes were mainly enriched in carbon metabolism, amino acid biosynthesis, pentose phosphate pathway, glycolysis/gluconeogenesis, and glutathione metabolism (Figure 1E). The downregulated genes were mainly related to the hematopoietic cell lineage, cell adhesion molecules, Th1 and Th2 cell differentiation, Th17 cell differentiation, and the T cell receptor signaling pathway (Figure 1F). These findings imply that TEPs-related DEGs are crucial to the CRC immune microenvironment. Construction and validation of a TEPGS Based on the bulk RNA expression profiles of 370 TEPs-related genes in the TCGA-CRC cohort, 44 prognostic genes were identified using univariate Cox regression analysis. Subsequently, these 44 TEPs-related genes were used to develop a consensus TEPGS in the TCGA-CRC cohort and five internal validation datasets (GEO: GSE17536, GEO: GSE17537, GEO: GSE29621, GEO: GSE38832, and GEO: GSE39582). Using a 10-fold cross-validation structure, 101 combinations of 10 machine learning algorithms—CoxBoost, Enet, GBM, Lasso, plsRcox, Ridge, RSF, stepwise Cox, SuperPC, and survival-SVM—were combined to generate the most powerful TEPGS with the highest C-index (Figure 2A; Table S1). The combination of CoxBoost and stepwise Cox (forward direction) was finally determined to be the optimal model with the highest average C-index (0.641), in which 12 TEPs-related genes (ARPC1B, OR2B6, TIMP1, TMEM40, RGS2, ERAP1, SRI, MMP1, FCER1A, ICAM3, MRPL37, and KIF13B) were selected (Figures 2A and 2B). Based on the weighted regression coefficients and mRNA expression levels of the 12 TEPs-related genes, we obtained the TEPGS score of each patient (Figure 2B). To further substantiate the stability and robustness of this model and the derived 12-gene signature, we performed 100 rounds of bootstrapping. The CoxBoost and forward stepwise Cox regression model consistently achieved the highest C-index across bootstrap iterations, verifying its superior and stable predictive performance relative to other 100 combinations (Figure S1A). Among the 44 candidate genes, the 12 selected genes defining our signature ranked top in terms of selection frequency across all 100 bootstrap resamples (Figure S1B). This selection frequency attests to the robustness of these genes.Figure 2A TEPGS developed using machine learning algorithms(A) The C-indices of 101 combinations of 10 machine learning algorithms in five internal validation cohorts.(B) Lollipop diagram displays the coefficients of the 12 TEPs-related genes obtained using the CoxBoost and Stepwise Cox (forward direction) algorithm.(C–H) Kaplan–Meier survival curves compare OS between patients with CRC with high and low TEPGS scores in TCGA-CRC (C), GEO: GSE17536 (D), GEO: GSE17537 (E), GEO: GSE29621 (F), GEO: GSE38832 (G), and GEO: GSE39582 (H). A TEPGS developed using machine learning algorithms (A) The C-indices of 101 combinations of 10 machine learning algorithms in five internal validation cohorts. (B) Lollipop diagram displays the coefficients of the 12 TEPs-related genes obtained using the CoxBoost and Stepwise Cox (forward direction) algorithm. (C–H) Kaplan–Meier survival curves compare OS between patients with CRC with high and low TEPGS scores in TCGA-CRC (C), GEO: GSE17536 (D), GEO: GSE17537 (E), GEO: GSE29621 (F), GEO: GSE38832 (G), and GEO: GSE39582 (H). To investigate the TEPGS’s clinical applicability, we divided the CRC samples into high- and low-risk groups using the median (Figures 2C–2H). The Kaplan-Meier curve revealed that the TEPGS was significantly related to OS, and that patients in the high-risk group had a shorter OS than those in the low-risk group in the training cohort (TCGA-CRC, n = 570, p < 0.0001) and the five internal validation cohorts GEO: GSE17536 (n = 177, p = 0.022), GEO: GSE17537 (n = 55, p = 0.0067), GEO: GSE29621 (n = 65, p = 0.033), GEO: GSE38832 (n = 122, p = 0.0062), and GEO: GSE39582 (n = 551, p < 0.0001; Figures 2C–2H). Collinearity analysis revealed VIF values <5 between TEPGS and clinical/tumor covariates (age, gender, TNM stage, TMB, TNB, TP53, KRAS, BRAF), excluding significant collinearity. Comparison of the baseline model and TEPGS-integrated model demonstrated that TEPGS conferred significant incremental value in four cohorts (TCGA-CRC, GEO: GSE17536, GEO: GSE38832, GEO: GSE39582), as evidenced by NRI ≥0.1, IDI >0, and p < 0.05 (Table S2). No such incremental benefit was observed in GEO: GSE17537 and GEO: GSE29621, likely due to their small sample size (n < 100) and the 5-year endpoint exceeding the median follow-up time of these cohorts. We then employed multivariate Cox regression analysis to assess the prognostic value of the TEPGS. The analysis reveals that TEPGS is an independent risk factor across all cohorts after correcting for clinical and tumor features, including age, gender, TNM stage, TMB, TNB, TP53, KRAS, and BRAF. The hazard ratios (HRs) and 95% confidence intervals (CIs) were shown in Figures S2A–S2F (TCGA-CRC: HR = 2.653, 95% CI = 1.588–4.430; GEO: GSE17536: HR = 1.809, 95% CI = 1.107–2.956; GEO: GSE17537: HR = 3.024, 95% CI = 1.032–8.864; GEO: GSE29621: HR = 3.217, 95% CI = 1.296–7.977; GEO: GSE38832: HR = 2.430, 95% CI = 1.156–5.108; GEO: GSE39582: HR = 1.937, 95% CI = 1.278–2.936). Results from the meta-cohort, consisting of five internal validation cohorts after batch correction, also confirmed the harmful effect of TEPGS on OS (HR = 1.720, 95% CI = 1.375–2.152; Figure S2G). Taken together, these observations demonstrate that the TEPGS functions as an independent predictive biomarker for OS of patients with CRC. Predictive performance of the TEPGS and nomogram construction To investigate the performance of the TEPGS score for prognostic prediction, we calculated the time-dependent AUC and the C-index across all cohorts. The AUC values for 1-, 3-, and 5-year OS in the TCGA-CRC cohort were 0.696, 0.726, and 0.683, respectively (Figure 3A). The results for the five internal validation cohorts were as follows (Figures 3B–3F): GEO: GSE17536 (0.720, 0.658, and 0.601), GEO: GSE17537 (0.643, 0.763, and 0.685), GEO: GSE29621 (0.806, 0.717, and 0.618), GEO: GSE38832 (0.663, 0.671, and 0.604), and GEO: GSE39582 (0.623, 0.646, and 0.611). The C-index (95% CI) was 0.702 (0.651–0.753), 0.622 (0.559–0.686), 0.670 (0.555–0.786), 0.643 (0.525–0.761), 0.663 (0.540–0.786), and 0.608 (0.564–0.652) in the six cohorts, respectively (Figure 3G; Table S3). Altogether, the results from the ROC and C-index analyses reveal that the TEPGS score exhibits acceptable discriminative ability for predicting the OS of patients with CRC.Figure 3Predictive performance of the TEPGS(A–F) Time-dependent ROC curves show 1-, 2-, 3-, 4-, and 5-year OS in TCGA-CRC (A), GEO: GSE17536 (B), GEO: GSE17537 (C), GEO: GSE29621 (D), GEO: GSE38832 (E), and GEO: GSE39582 (F).(G) C-index of the TEPGS across all cohorts.(H and I) Comparison of C-index between the TEPGS and clinical characteristics in TCGA-CRC (n = 570, H) and GEO: GSE39582 (n = 551, I). Data in (G–I) are presented as mean ±95% confidence interval (95% CI). The p value of a two-sided Z score test represents the significance. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, and ∗∗∗∗p < 0.0001. Predictive performance of the TEPGS (A–F) Time-dependent ROC curves show 1-, 2-, 3-, 4-, and 5-year OS in TCGA-CRC (A), GEO: GSE17536 (B), GEO: GSE17537 (C), GEO: GSE29621 (D), GEO: GSE38832 (E), and GEO: GSE39582 (F). (G) C-index of the TEPGS across all cohorts. (H and I) Comparison of C-index between the TEPGS and clinical characteristics in TCGA-CRC (n = 570, H) and GEO: GSE39582 (n = 551, I). Data in (G–I) are presented as mean ±95% confidence interval (95% CI). The p value of a two-sided Z score test represents the significance. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, and ∗∗∗∗p < 0.0001. We compared the prognostic potential of the TEPGS scores with a variety of clinical characteristics, which have been shown to be closely related to the prognosis of patients with CRC. These characteristics include age, gender, pT, pN, pM, TNM stage, gene mutation status (TP53, KRAS, and BRAF), TMB, and TNB. The TEPGS score displayed higher predictive power than most clinical characteristics in TCGA-CRC and GEO: GSE39582 cohorts (Figures 3H and 3I) and non-inferior accuracy in the GEO: GSE17536, GEO: GSE17537, GEO: GSE29621, and GEO: GSE38832 cohorts (Figures S3A–S3D). Compared with 71 published mRNA gene signatures (Figure S3E; Table S4), TEPGS ranked second and fifth in the TCGA-CRC cohort and meta-cohort, respectively, and had a higher HR and C-index than most other signatures in five internal validation cohorts. Most of the gene signatures were not consistently reliable over several cohorts. For example, the Dai-FIG cohort’s gene signature was ranked first in the TCGA-CRC cohort but lower in the meta-cohort. Similarly, the Lin-JTM cohort’s gene signature scored top in the meta-cohort but only 27th in the TCGA-CRC cohort. Overall, our findings show that TEPGS performs better than published signatures. A TCGA-CRC nomogram was further developed based on three independent prognostic predictors (age, stage, and TEPGS) identified via univariate and multivariate Cox regression analyses to accurately predict 1-, 3-, and 5-year survival rates of patients with CRC (Figure S4A). The calibration plots showed that the predicted OS was comparable to the actual OS (Figures S4B–S4D). Furthermore, decision curve analysis revealed a higher net benefit for the nomogram compared with age, stage, and TEPGS (Figure S4E). These findings indicate that the nomogram model possesses superior predictive performance. Verification of the TEPGS model in an external independent validation cohort The external independent validation cohort GEO: GSE12945 was used to verify the performance of our TEPGS model. Using the median as the cutoff value, we divided the patients with CRC into TEPGS-low and -high groups. Consistent with our prior results, Kaplan-Meier survival analysis revealed that patients with high TEPGS exhibited shorter OS than those with low TEPGS (n = 58, p = 0.032, Figure 4A). The time-dependent ROC analysis indicated that the AUC values for 1-, 3-, and 5-year OS in the GEO: GSE12945 cohort were 0.713, 0.839, and 0.817, respectively (Figure 4B). The C-index (95% CI) of TEPGS was 0.798 (0.673–0.923), which had non-inferior accuracy compared to common clinical features, including age, gender, pM, and TNM stage (Figure 4C). Unfortunately, no significant incremental prognostic benefit of TEPGS was detected in the GEO: GSE12945 cohort (Table S2). Nevertheless, multivariate Cox regression analysis confirmed that TEPGS remained an independent prognostic factor in this cohort (log2HR = 2.257, 95% CI = 0.08–4.43; Figure S2H), after adjustment for clinical covariates. In addition, a GEO: GSE12945 cohort nomogram was established based on four prognostic predictors in Figure 4D. The calibration plots for 1-, 3-, and 5-year OS showed that the predicted OS was comparable to the actual OS (Figures 4E–4G). Furthermore, the result of the decision curve analysis indicated that our nomogram was of greater clinical utility (Figure 4H). Together, the results from the external cohort are consistent with our findings in the training and internal validation cohorts, further verifying the robustness of the TEPGS model.Figure 4Verification of the TEPGS model in the GEO: GSE12945 cohort(A) Kaplan–Meier survival curve compares OS between patients with CRC with high and low TEPGS scores (n = 58).(B) Time-dependent ROC curves show 1-, 2-, 3-, 4-, and 5-year OS.(C) Comparison of the C-index between the TEPGS and clinical characteristics. Data are presented as mean ±95% confidence interval (95% CI). The p value of a two-sided Z score test represents the significance. ∗p < 0.05 and ∗∗p < 0.01.(D) Nomogram based on four prognostic factors.(E–G) The calibration curve shows the accuracy of the nomogram for predicting 1-, 3-, and 5-year OS.(H) Decision curve analysis of OS for the predicted nomogram model. Verification of the TEPGS model in the GEO: GSE12945 cohort (A) Kaplan–Meier survival curve compares OS between patients with CRC with high and low TEPGS scores (n = 58). (B) Time-dependent ROC curves show 1-, 2-, 3-, 4-, and 5-year OS. (C) Comparison of the C-index between the TEPGS and clinical characteristics. Data are presented as mean ±95% confidence interval (95% CI). The p value of a two-sided Z score test represents the significance. ∗p < 0.05 and ∗∗p < 0.01. (D) Nomogram based on four prognostic factors. (E–G) The calibration curve shows the accuracy of the nomogram for predicting 1-, 3-, and 5-year OS. (H) Decision curve analysis of OS for the predicted nomogram model. TEPGS-related somatic mutational landscape and CNVs To explore the mechanism underlying poor survival, we characterized the somatic mutational landscapes of 20 FMGs previously reported in CRC,25,26,27,28 in the TEPGS-low and -high groups (Figures 5A and 5B). The mutation frequencies of TP53, SYNE2, and ZEP36L2 were significantly higher in the TEPGS-high group than in the TEPGS-low group (Figure 5C). In contrast, the mutation frequencies in SOX9, UNC13C, TCF7L2, and B2M were significantly lower (Figure 5C). Using GISTIC2.0, we further examined the relationship between TEPGS scores and CNVs. The jagged layout of CNV distributions across the genome indicated that amplification and deletion increased in the TEPGS-high group (Figures 5D and 5E). Additionally, the aneuploidy score and gene-level CNV in TEPGS-high tumors were considerably elevated compared to those in TEPGS-low tumors (Figures 5F and 5G). Interestingly, analyses of copy number amplification and deletion at the gene level yielded similar results (Figures 5H and 5I, both p < 0.001). In conclusion, these results reveal heterogeneity between TEPGS-high and -low tumors due to somatic mutations and CNVs, which may be connected to the unfavorable prognosis in the TEPGS-high group.Figure 5Comparison of FMGs and CNVs in different TEPGS groups of TCGA-CRC(A and B) The mutational landscapes of 20 FMGs in high TEPGS samples (A) and low TEPGS samples (B). The middle matrix displays the somatic mutations by gene (row) and sample (column). The top tracks show the clinicopathological characteristics, such as age, gender, TNM stage, survival status, and tumor mutation burden. The histogram on the right displays the number of somatic mutations accumulated in CRC samples in each gene.(C) Heatmap shows the differential frequencies of 20 FMGs between high and low TEPGS groups (Fisher’s exact test, ∗p < 0.05).(D and E) Frequencies of CNVs detected in high TEPGS samples (D) and low TEPGS samples (E), with gains in red and losses in blue.(F and G) Boxplots compare aneuploidy score (F) and gene-level CNV (G) between patients with high and low TEPGS.(H and I) Boxplots compare copy number amplification (H) and copy number deletion (I) at the gene level between patients with high and low TEPGS. In boxplots (F–I) center line indicates the median, the bounds of the box indicate the 25th and 75th percentiles, and whiskers indicate the minimal and maximal values. The error bars indicate 95% CI. The significance of the Wilcoxon rank-sum test is indicated by the p value. Comparison of FMGs and CNVs in different TEPGS groups of TCGA-CRC (A and B) The mutational landscapes of 20 FMGs in high TEPGS samples (A) and low TEPGS samples (B). The middle matrix displays the somatic mutations by gene (row) and sample (column). The top tracks show the clinicopathological characteristics, such as age, gender, TNM stage, survival status, and tumor mutation burden. The histogram on the right displays the number of somatic mutations accumulated in CRC samples in each gene. (C) Heatmap shows the differential frequencies of 20 FMGs between high and low TEPGS groups (Fisher’s exact test, ∗p < 0.05). (D and E) Frequencies of CNVs detected in high TEPGS samples (D) and low TEPGS samples (E), with gains in red and losses in blue. (F and G) Boxplots compare aneuploidy score (F) and gene-level CNV (G) between patients with high and low TEPGS. (H and I) Boxplots compare copy number amplification (H) and copy number deletion (I) at the gene level between patients with high and low TEPGS. In boxplots (F–I) center line indicates the median, the bounds of the box indicate the 25th and 75th percentiles, and whiskers indicate the minimal and maximal values. The error bars indicate 95% CI. The significance of the Wilcoxon rank-sum test is indicated by the p value. Immune-related characteristics and immunotherapy response in high- and low- TEPGS groups To explore the relationship between TEPGS score and the tumor immune microenvironment, we estimated the fraction of 28 immune cell types and calculated the expression levels of 28 immune modulators in CRC using bulk RNA-seq data. Based on the ssGSEA algorithm, we found that activated CD8 T cells, activated CD4 T cells, type 17 T helper cells (Th17), type 2 T helper cells (Th2), activated B cells, immature B cells, memory B cells, eosinophils, and neutrophils significantly decreased in the TEPGS-high group (Figure 6A), whereas the proportion of CD56dim natural killer cells was increased in the TEPGS-high group (Figure 6A). Notably, several immune infiltration status indicators, including the ImmuneScore and TMEscore, declined in the TEPGS-high group, whereas the StromaScore was moderately elevated using the xCell and TMEscore algorithms (Figures 6B–6D). In parallel, most immune checkpoint molecules showed a substantial negative correlation with the TEPGS score, including BTLA, NCR3, CD40LG, TMIGD2, TNFRSF9, TIGIT, ICOS, PDCD1LG2, CD274, CTLA4, CD27, IDO1, and HHLA2 (Figure 6E, p < 0.05). These results indicate that immunosuppression is more prevalent among individuals with high TEPGS scores.Figure 6Immune-related characteristics and immunotherapy response comparison in the TCGA-CRC dataset(A) Comparison of immune cell infiltration level estimated by ssGSEA between patients with high and low TEPGS scores. Act: activated; Tmem: memory T cell; Tcm: central memory T cell; Th: T helper cell; Tfh: follicular helper T cell; γδ T: gamma delta T cell; Treg: regulatory T cell; NK: natural killer cell; DC: dendritic cell; pDC: plasmacytoid dendritic cell; iDC: immature dendritic cell; Mφ: macrophage; MDSC: myeloid derived suppressor cell; NKT: natural killer T cell; IFN-γ: interferon gamma; IPS: immunophenoscore; CYT: cytotoxic activity; GEP: T cell-inflamed gene expression profile; TIDE: tumor immune dysfunction and exclusion.(B and C) Boxplots comparing the StromaScore (B) and ImmuneScore (C) estimated using the xCell algorithm.(D) Boxplot shows TMEscore in high and low TEPGS groups.(E) Boxplots display the difference in immune checkpoints in high and low TEPGS groups.(F–I) Boxplots compare the IFN-γ score (F), IPS (G), CYT score (H), and GEP score (I) between patients with high and low TEPGS.(J) TIDE score comparison between different TEPGS groups.(K) The fraction of patients in the high and low TEPGS groups that respond differently to immunotherapy (Fisher’s exact test). In boxplots (A–J) center line indicates the median, the bounds of the box indicate the 25th and 75th percentiles, and whiskers indicate the minimal and maximal values. The error bars indicate 95% CI. The significance of the Wilcoxon rank-sum test is indicated by the p value. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, and ∗∗∗∗p < 0.0001. Immune-related characteristics and immunotherapy response comparison in the TCGA-CRC dataset (A) Comparison of immune cell infiltration level estimated by ssGSEA between patients with high and low TEPGS scores. Act: activated; Tmem: memory T cell; Tcm: central memory T cell; Th: T helper cell; Tfh: follicular helper T cell; γδ T: gamma delta T cell; Treg: regulatory T cell; NK: natural killer cell; DC: dendritic cell; pDC: plasmacytoid dendritic cell; iDC: immature dendritic cell; Mφ: macrophage; MDSC: myeloid derived suppressor cell; NKT: natural killer T cell; IFN-γ: interferon gamma; IPS: immunophenoscore; CYT: cytotoxic activity; GEP: T cell-inflamed gene expression profile; TIDE: tumor immune dysfunction and exclusion. (B and C) Boxplots comparing the StromaScore (B) and ImmuneScore (C) estimated using the xCell algorithm. (D) Boxplot shows TMEscore in high and low TEPGS groups. (E) Boxplots display the difference in immune checkpoints in high and low TEPGS groups. (F–I) Boxplots compare the IFN-γ score (F), IPS (G), CYT score (H), and GEP score (I) between patients with high and low TEPGS. (J) TIDE score comparison between different TEPGS groups. (K) The fraction of patients in the high and low TEPGS groups that respond differently to immunotherapy (Fisher’s exact test). In boxplots (A–J) center line indicates the median, the bounds of the box indicate the 25th and 75th percentiles, and whiskers indicate the minimal and maximal values. The error bars indicate 95% CI. The significance of the Wilcoxon rank-sum test is indicated by the p value. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, and ∗∗∗∗p < 0.0001. We then assessed whether the TEPGS scores were significantly associated with tumor immunotherapy in CRC. As shown in Figures 6F–6I, several immunotherapy predictors, such as IFN-γ score, IPS, CYT score, and GEP score, were significantly elevated in TEPGS-low tumors, which predicted a better immunotherapy response. To provide additional evidence of the utility of the TEPGS for predicting the outcome of immunotherapy for TCGA-CRC tumors, we computed the TIDE score. According to our findings, patients in the TEPGS-high group had a considerably higher TIDE score and a higher proportion of non-responders to anti-PD1 or anti-CTLA4 treatment, all of which are indicators of a less favorable outcome from immunotherapy (Figures 6J and 6K). Overall, these findings suggest that individuals with high TEPGS are immunosuppressed and respond poorly to immunotherapy. Characteristics of TAMs in the TME of CRC Furthermore, we employed scRNA-seq data (GEO: GSE166555) to investigate the association between TEPGS and the immune microenvironment. We retrieved cell type annotations from the TISCH database. A total of 36,822 cells including immune cells (21,796), malignant cells (8,036) and stromal cells (6,128) from 13 CRC tumor tissues were analyzed and 13 major cell types were identified: B cells (1,876), CD4Tconv cells (8,137), CD8T cells (1,881), DC cells (611), endothelial cells (537), epithelial cells (3,664), fibroblasts (1,118), malignant cells (8,036), mast cells (437), mono/macro cells (2,379), myofibroblasts (512), plasma cells (6,475) and Tprolif cells (297) (Figures 7A and 7B). Utilizing the previously described TEPGS scoring algorithm, we mapped the TEPGS score to each cell at single-cell resolution. Results indicated that fibroblasts, mast cells, and mono/macrophage cells exhibited the highest TEPGS scores overall (Figures 7C and 7D), and notably, the TEPGS score of mono/macrophage cells was significantly higher than that of fibroblasts and mast cells (p < 0.01; Figure 7E). The heatmap further revealed that fibroblasts, mast cells, and mono/macrophage cells were all enriched in the TEPGS-high group (ratio >1; Figure 7F), with mono/macrophage cells exhibiting the highest enrichment level (Ratio = 1.47). Given that mono/macrophage cells are the primary precursors of TAMs, we focused our investigation on TAM subpopulations to assess the influence of TEPGS. We re-clustered macrophages and identified 5 subtypes of TAMs (Figures 7G and 7H): SPP1+ TAMs marked by SPP1 and CYP27A1, SLC40A1+ TAMs marked by SLC40A1 and SELENOP, FCN1+ TAMs marked by FCN1 and VCAN, CD3D+ TAMs marked by CD3D, CD3E, and TRBC2, CD1C+ TAMs marked by CD1C and CD1E. Using the median value as the cutoff, 13 CRC samples were divided into TEPGS-high and -low groups. We observed the distribution patterns of TAM subtypes between TEPGS-high and -low groups. SLC40A1+ TAMs and CD1C+ TAMs were more dominant in the TEPGS-low group and the FCN1+ TAMs and SPP1+ TAMs were greater in the TEPGS-high group (Figure 7I).Figure 7Single-cell transcriptomic analysis deciphers that TEPGS contributes to tumor growth(A) UMAP plot of 36,822 cells from 13 CRC tumor tissues, displaying immune cells, malignant cells, and stromal cells.(B) UMAP plot shows 13 major cell types.(C and D) UMAP plot (C) and Violin plot (D) displaying the TEPGS score for the defined cell types.(E) Violin plot compares the TEPGS scores among fibroblasts, mast cells, and mono/macrophage cells.(F) Heatmap displays the relative enrichment of each cell type in individuals with high and low TEPGS.(G) UMAP plot shows the composition of tumor-associated macrophages (TAMs) colored by cluster.(H) Dot plot shows the expression levels of selected genes for TAM subtypes. Dot size indicates the fraction of cells expressing the markers, colored based on normalized expression levels.(I) Bar plots show the proportion of 5 TAM subtypes in TEPGS-high and -low groups.(J) Pseudotime trajectory plot shows the evolution of TAMs inferred by the Monocle2 package. Color key from dark to bright indicates cancer progression from the early to the late stage.(K) The developmental trajectory was plotted by subtypes of TAM. Each dot represents a cell, and the color represents cell subtypes.(L and M) The density plots show the distribution of 5 TAM subtypes and different TEPGS groups along the pseudotime trajectory.(N) Heatmap shows the dynamic transcriptional changes along the pseudotime trajectory. Single-cell transcriptomic analysis deciphers that TEPGS contributes to tumor growth (A) UMAP plot of 36,822 cells from 13 CRC tumor tissues, displaying immune cells, malignant cells, and stromal cells. (B) UMAP plot shows 13 major cell types. (C and D) UMAP plot (C) and Violin plot (D) displaying the TEPGS score for the defined cell types. (E) Violin plot compares the TEPGS scores among fibroblasts, mast cells, and mono/macrophage cells. (F) Heatmap displays the relative enrichment of each cell type in individuals with high and low TEPGS. (G) UMAP plot shows the composition of tumor-associated macrophages (TAMs) colored by cluster. (H) Dot plot shows the expression levels of selected genes for TAM subtypes. Dot size indicates the fraction of cells expressing the markers, colored based on normalized expression levels. (I) Bar plots show the proportion of 5 TAM subtypes in TEPGS-high and -low groups. (J) Pseudotime trajectory plot shows the evolution of TAMs inferred by the Monocle2 package. Color key from dark to bright indicates cancer progression from the early to the late stage. (K) The developmental trajectory was plotted by subtypes of TAM. Each dot represents a cell, and the color represents cell subtypes. (L and M) The density plots show the distribution of 5 TAM subtypes and different TEPGS groups along the pseudotime trajectory. (N) Heatmap shows the dynamic transcriptional changes along the pseudotime trajectory. Moreover, pseudotime trajectory analysis was performed to visualize the developmental trajectory and imitate the cell evolution of TAMs (Figure 7J). CD1C+ TAMs were positioned at the initial point of the trajectory, suggesting their potential role as precursor cells for other TAMs. The evolution trajectory originated from CD1C+ TAMs and developed into two distinct cell fates: one generating SLC40A1+ TAMs, while the other led to FCN1+ TAMs and SPP1+ TAMs (Figure 7K). Then, we performed branching expression analysis modeling (BEAMbranching expression analysis modeling) and GO analysis to elucidate the transcriptional changes along two cell fates. The transition of CD1C+ TAMs to SLC40A1+ TAMs was associated with immune-related pathways, such as complement activation, cell junction disassembly, and antigen processing and presentation (Figure 7N). The transition of CD1C+ TAMs to FCN1+ TAMs and SPP1+ TAMs was characterized by pro-tumor related pathways, including the regulation of apoptotic signaling pathway, positive regulation of epithelial to mesenchymal transition (EMT), regulation of fibroblast proliferation, and glycolytic process (Figure 7N). The density plot showed that the number of CD1C+ TAMs located in the early stage elevated in the TEPGS-low group, while the number of FCN1+ TAMs located in the late trajectory stage increased in the TEPGS-high group (Figures 7L and 7M). All in all, our findings strongly suggest that elevated TEPGS expression is correlated with a malignant TAM phenotype and associated with tumor progression. Spatial association between platelet and TEPGS To validate the relationship between platelet abundance and TEPGS expression in CRC at the spatial level, we analyzed the spatial transcriptomics dataset GEO: GSE226997 (n = 3 samples). Given the limited research characterizing platelet distribution at spatial resolution in tumors, we first curated a panel of 11 well-established platelet-specific marker genes (PF4, ITGA2B, GP1BA, ITGB3, GP9, SELP, VWF, GP5, GP6, P2RY1, P2RY12) and developed a platelet score based on published scRNA-seq and spatial transcriptomics studies.29,30 Among the 11 platelet markers, ITGB3 and VWF were enriched in CRC tissues (Figures 8A–8C). In the Ajou_Visium_P1 sample, spatial mapping of the platelet score demonstrated an enrichment of platelet-associated signals at the tumor margin (Figure 8D). Notably, these platelet-high regions exhibited elevated TEPGS scores, and quantitative analysis confirmed a positive correlation between platelet score and TEPGS (r = 0.268, p < 2.2e−16; Figure 8D). Similarly, a consistent but modest positive correlation was evident between the platelet score and TEPGS in the Ajou_Visium_P2 and Ajou_Visium_P3 samples (Figures 8E and 8F). Collectively, these spatial transcriptomic data provide evidence that high TEPGS tumors are spatially associated with increased platelet abundance in the CRC TME.Figure 8Spatial transcriptomic analysis reveals an association between platelet and TEPGS in CRC(A–C) Spatial expression patterns of ITGB3 and VWF genes in CRC spatial transcriptomic samples (Ajou_Visium_P1, Ajou_Visium_P2, Ajou_Visium_P3).(D–F) Spatial enrichment and correlation of platelet and TEPGS in Ajou_Visium_P1, Ajou_Visium_P2, and Ajou_Visium_P3 samples. Spatial transcriptomic analysis reveals an association between platelet and TEPGS in CRC (A–C) Spatial expression patterns of ITGB3 and VWF genes in CRC spatial transcriptomic samples (Ajou_Visium_P1, Ajou_Visium_P2, Ajou_Visium_P3). (D–F) Spatial enrichment and correlation of platelet and TEPGS in Ajou_Visium_P1, Ajou_Visium_P2, and Ajou_Visium_P3 samples. Drug susceptibility prediction and potential treatment options for patients with high TEPGS scores To obtain effective therapeutic agents for patients with CRC with high TEPGS scores, two methods were used to analyze the RNA expression data from TCGA-CRC and GEO: GSE39582 cohorts. Compounds were first retrieved from the CTRP2.0 and PRISM databases. Subsequently, the pRRophetic package was used to predict drug responses, estimating the AUC value for each compound in the specimen. Thirdly, compounds exhibiting a significant association between AUC values and TEPGS scores were screened using Spearman’s correlation analysis (absolute r > 0.2 and p < 0.05 for CTRP2.0 and PRISM). Compounds with high TEPGS scores and low AUC values were selected for differential drug response analysis (p < 0.05). Notably, four common clinical medications, paclitaxel, lapatinib, docetaxel, and fluorouracil, were positively associated with the TEPGS score and were demonstrated to be effective in the TEPGS-low group in both cohorts (Figures 9A, 9B, S5A, and S5B). Crucially, five CTRP2.0 compounds (BRD-K79669418, NSC23766, PLX-4720, BMS-536924, and TG100-115) and six PRISM compounds (vemurafenib, AG-14361, lorlatinib, regorafenib, INC-280, and simvastatin) were shown to have greater TEPGS scores and lower estimated AUC values (Figures 9C–9E, and S5C–S5E).Figure 9Chemotherapy susceptibility analysis identifies drugs that are effective in the high TEPGS group of the TCGA-CRC cohort(A) Spearman correlation between the TEPGS score and AUC value of nine compounds in the CTRP2.0 dataset.(B) Boxplots compare the AUC levels of four common clinic medications from CTRP2.0 between patients with high- and low-TEPGS.(C) Boxplots compare the AUC levels of five compounds from CTRP2.0 between patients with high- and low-TEPGS.(D) Spearman correlation between the TEPGS score and AUC value of six derivatives in the PRISM dataset.(E) Boxplots display AUC levels of six derivatives from PRISM between patients with high and low-TEPGS. In boxplots (B–C, E), the center line indicates the median, the bounds of the box indicate the 25th and 75th percentiles, and the whiskers indicate the minimum and maximum values. The error bars indicate 95% CI. The significance of the Wilcoxon rank-sum test is indicated by the p value. ∗∗p < 0.01, ∗∗∗p < 0.001, and ∗∗∗∗p < 0.0001. Chemotherapy susceptibility analysis identifies drugs that are effective in the high TEPGS group of the TCGA-CRC cohort (A) Spearman correlation between the TEPGS score and AUC value of nine compounds in the CTRP2.0 dataset. (B) Boxplots compare the AUC levels of four common clinic medications from CTRP2.0 between patients with high- and low-TEPGS. (C) Boxplots compare the AUC levels of five compounds from CTRP2.0 between patients with high- and low-TEPGS. (D) Spearman correlation between the TEPGS score and AUC value of six derivatives in the PRISM dataset. (E) Boxplots display AUC levels of six derivatives from PRISM between patients with high and low-TEPGS. In boxplots (B–C, E), the center line indicates the median, the bounds of the box indicate the 25th and 75th percentiles, and the whiskers indicate the minimum and maximum values. The error bars indicate 95% CI. The significance of the Wilcoxon rank-sum test is indicated by the p value. ∗∗p < 0.01, ∗∗∗p < 0.001, and ∗∗∗∗p < 0.0001. Predictive value of bevacizumab and cetuximab benefits Currently, the vascular endothelial growth factor (VEGF) inhibitor bevacizumab and the epidermal growth factor receptor (EGFR) inhibitor cetuximab represent the two main US Food and Drug Administration (FDA)-approved targeted therapies for the first-line standard of care for patients with mCRC.31,32,33 TEP-derived RNA profiles, an important liquid biopsy component, have made great progress in blood-based detection, but their predictive power in targeted therapies merits further investigations. TEPGS has been revealed to be useful for immunotherapy and chemotherapy response prediction, but its association with the benefits of targeted therapies warrants further exploration. To investigate the predictive value of TEPGS in CRC, three cohorts treated with bevacizumab were retrieved, which contained 21 responders and 26 non-responders. A consistent trend of higher TEPGS scores was observed in bevacizumab-sensitive patients compared with resistant patients in GEO: GSE19860 (p > 0.05), GEO: GSE19862 (p > 0.05), GEO: GSE72970 (p > 0.05), and this trend reached statistical significance in the combined meta-cohort (p < 0.05) (Figures 10A–10D). The AUC values of TEPGS for predicting benefit from bevacizumab in the four cohorts were 0.743, 0.673, 0.630, and 0.648, respectively, which were slightly lower than those reported by Liu et al. (Figures 10E–10H).34 Furthermore, we confirmed the clinical significance of TEPGS using a publicly available cohort (GEO: GSE5851) of 68 patients with mCRC treated with cetuximab monotherapy, of which 25 were responders, and 43 were non-responders. As indicated in Figure 10I, we discovered that TEPGS expression was higher in tumors susceptible to cetuximab than in resistant cases (p < 0.05). The ROC analysis demonstrated that TEPGS slightly differed between cetuximab responders and non-responders with an AUC of 0.638 (Figure 10J), a value that reflects modest but detectable predictive performance for this targeted therapy. Taken together, given its relatively weak significance and moderate discriminatory capacity, TEPGS shows tentative promise as a predictive biomarker for bevacizumab and cetuximab therapy responses.Figure 10Predictive value of bevacizumab and cetuximab benefits(A–D) Boxplots comparing TEPGS level between responders and non-responders of bevacizumab therapy in GEO: GSE19860 (n = 12) (A), GEO: GSE19862 (n = 14) (B), GEO: GSE72970 (n = 21) (C), and the combined meta-cohort (n = 47) (D). Statistic tests: Student’s t test.(E–H) ROC curves of TEPGS to predict the benefits of bevacizumab therapy in GEO: GSE19860 (E), GEO: GSE19862 (F), GEO: GSE72970 (G), and the combined meta-cohort (H).(I) Boxplot comparing TEPGS level between responders and non-responders of cetuximab therapy in GEO: GSE5851 (n = 68). Statistic test: Student’s t test. (J) ROC curve of TEPGS to predict the benefits of cetuximab therapy in GEO: GSE5851. In boxplots (A–D, I), the center line indicates the median, the bounds of the box indicate the 25th and 75th percentiles, and the whiskers indicate the minimal and maximal values. The error bars indicate 95% CI. The significance of Student’s t test is indicated by the p value. ns p > 0.05, ∗p < 0.05. Predictive value of bevacizumab and cetuximab benefits (A–D) Boxplots comparing TEPGS level between responders and non-responders of bevacizumab therapy in GEO: GSE19860 (n = 12) (A), GEO: GSE19862 (n = 14) (B), GEO: GSE72970 (n = 21) (C), and the combined meta-cohort (n = 47) (D). Statistic tests: Student’s t test. (E–H) ROC curves of TEPGS to predict the benefits of bevacizumab therapy in GEO: GSE19860 (E), GEO: GSE19862 (F), GEO: GSE72970 (G), and the combined meta-cohort (H). (I) Boxplot comparing TEPGS level between responders and non-responders of cetuximab therapy in GEO: GSE5851 (n = 68). Statistic test: Student’s t test. (J) ROC curve of TEPGS to predict the benefits of cetuximab therapy in GEO: GSE5851. In boxplots (A–D, I), the center line indicates the median, the bounds of the box indicate the 25th and 75th percentiles, and the whiskers indicate the minimal and maximal values. The error bars indicate 95% CI. The significance of Student’s t test is indicated by the p value. ns p > 0.05, ∗p < 0.05. ARPC1B promotes cell proliferation and migration ARPC1B serves as an important component of the TEPGS. In a pan-cancer expression analysis across 27 cancer types, we found that ARPC1B was significantly elevated in most cancerous tissues compared to that in normal tissues, including colon adenocarcinoma and rectum adenocarcinoma (Figure S6A). Notably, the expression of ARPC1B was significantly higher in the TEPGS-high group than in the TEPGS-low group in both TCGA-CRC and meta-data cohorts (Figures S6B and S6C). We divided the CRC samples into high- and low-ARPC1B groups using the optimal cutoff values that the R package survminer had identified. The cutoff value of ARPC1B was 7.087 in TCGA-CRC, 11.699 in GEO: GSE17536, 10.959 in GEO: GSE17537, 11.113 in GEO: GSE29621, 12.877 in GEO: GSE38832, and 10.743 in GEO: GSE39582, respectively. Patients with high ARPC1B levels exhibited shorter OS than those with low ARPC1B levels (Figures 11A and S6D–S6H).Figure 11Knockdown of ARPC1B inhibits cell proliferation and migration(A) Kaplan–Meier survival curve of OS between patients with CRC with high and low ARPC1B expression levels in the TCGA-CRC dataset.(B) ARPC1B mRNA expression level in platelets between patients with CRC and healthy donors in the GEO: GSE68086 dataset.(C) ARPC1B mRNA expression level in CRC tumors and adjacent normal tissues in the TCGA-CRC dataset.(D) Relative mRNA levels of ARPC1B in the four CRC cell lines compared with NCM460 cells.(E and F) RT-qPCR assay (E) and western blot assay (F) showing the knockdown efficiency of ARPC1B siRNA.(G and H) CCK-8 assays display the proliferative abilities of SW480 (G) and HCT116 (H) cells after ARPC1B knockdown.(I) Colony formation assays show the proliferative abilities of SW480 and HCT116 cells after ARPC1B knockdown.(J) Migration abilities of SW480 and HCT116 cells were evaluated using a transwell assay after ARPC1B knockdown. Scale bars, 200 μm (100×).(K and L) Migration ability of SW480 (K) and HCT116 (L) was assessed using wound healing assays after ARPC1B knockdown at the indicated times. Scale bars, 100 μm (100×).(M) Xenograft tumors in shARPC1B #1, #2, and shControl groups at the end of the experiment.(N and O) Weight of subcutaneous tumors (N) and tumor growth curves (O) among the three groups (N = 5 per group). All the experiments are repeated at least three times. Data are shown as mean ± SEM. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, and ∗∗∗∗p < 0.0001, and NS, not significant of Student’s t test or two-way analysis of variance. Knockdown of ARPC1B inhibits cell proliferation and migration (A) Kaplan–Meier survival curve of OS between patients with CRC with high and low ARPC1B expression levels in the TCGA-CRC dataset. (B) ARPC1B mRNA expression level in platelets between patients with CRC and healthy donors in the GEO: GSE68086 dataset. (C) ARPC1B mRNA expression level in CRC tumors and adjacent normal tissues in the TCGA-CRC dataset. (D) Relative mRNA levels of ARPC1B in the four CRC cell lines compared with NCM460 cells. (E and F) RT-qPCR assay (E) and western blot assay (F) showing the knockdown efficiency of ARPC1B siRNA. (G and H) CCK-8 assays display the proliferative abilities of SW480 (G) and HCT116 (H) cells after ARPC1B knockdown. (I) Colony formation assays show the proliferative abilities of SW480 and HCT116 cells after ARPC1B knockdown. (J) Migration abilities of SW480 and HCT116 cells were evaluated using a transwell assay after ARPC1B knockdown. Scale bars, 200 μm (100×). (K and L) Migration ability of SW480 (K) and HCT116 (L) was assessed using wound healing assays after ARPC1B knockdown at the indicated times. Scale bars, 100 μm (100×). (M) Xenograft tumors in shARPC1B #1, #2, and shControl groups at the end of the experiment. (N and O) Weight of subcutaneous tumors (N) and tumor growth curves (O) among the three groups (N = 5 per group). All the experiments are repeated at least three times. Data are shown as mean ± SEM. ∗p < 0.05, ∗∗p < 0.01, ∗∗∗p < 0.001, and ∗∗∗∗p < 0.0001, and NS, not significant of Student’s t test or two-way analysis of variance. We then investigated the role of ARPC1B in CRC. As shown in Figures 11B and 11C, ARPC1B was considerably higher in tumor-derived platelets than in normal platelets and in tumor samples than in adjacent normal tissues. We found that ARPC1B was significantly higher in four CRC cell lines (HCT116, Caco-2, SW48, and SW480) than in the normal colonic epithelial cell line NCM460 (Figure 11D). Then, we selected SW480 and HCT116 cells for experimental validation. We confirmed the successful knockdown of ARPC1B at the mRNA and protein levels in SW480 and HCT116 cells (Figures 11E and 11F). Subsequently, we performed CCK-8 and colony formation assays on SW480 and HCT116 cells transfected with siARPC1B or siNC and showed that ARPC1B knockdown significantly inhibited cell proliferation (Figures 11G–11I and S6I). Transwell and wound healing experiments demonstrated that the cell migration ability was significantly decreased after ARPC1B knockdown (Figures 11J–11L, and S6J–S6L). We finally examined whether ARPC1B was involved in CRC progression in vivo. BALB/c male nude mice were subcutaneously injected with SW480 cells with ARPC1B knockdown or not (shARPC1B #1, #2, and shControl). We found that ARPC1B knockdown significantly suppressed tumor weight and volume (Figures 11M–11O). Overall, our findings confirm the oncogenic role of ARPC1B in CRC. Discussion Emerging evidence has revealed a relationship between platelets and tumorigenesis as well as cancer progression. As reported, platelets derived from patients with cancer often presented distinct RNA profiles owing to a biological process called education, including the transfer and storage of tumor-associated factors, platelet RNA processing, and abnormal platelet production by megakaryocytes.35 Since Best et al. published the platelet mRNA profiles of several patients with cancer and healthy donors, the diagnostic performance of TEPs in multiple cancers has gradually been revealed.17,18,20 TEPs have emerged as prospective biomarkers for tumor diagnosis, screening, and treatment monitoring, followed by circulating tumor cells, circulating tumor DNAs, and exosomes in liquid biopsy.36 Advances in deep learning algorithms have the potential to improve the accuracy and effectiveness of CRC detection.37 For example, using a deep learning neural network model, Liu et al. developed a platelet RNA signature for ovarian cancer prognosis prediction.19 However, the predictive significance of TEPs in CRC has yet to be fully elucidated. Growing evidence supports that TEPs and tumor tissue transcriptome are functionally connected through three well-documented mechanistic pathways, as summarized by Roweth et al.35 Firstly, tumor cells secrete membrane vesicles which transfer mutant or tumor tissue-specific RNA transcripts into platelets. Representative examples included EGFRvIII derived from patients with glioma, PCA3 from patients with prostate cancer, and EML4-ALK rearrangements from patients with non-small cell lung carcinoma (NSCLC).38,39 Secondly, Calverley et al. and Best et al. both uncovered differential RNA-splicing events in TEPs.35,40,41 The aberrant splice profiles detected in TEPs most probably reflect tumor-specific signals that lead to distinct pre-mRNA processing. Thirdly, tumor-derived factors can also alter the development and function of marrow-resident megakaryocytes, yielding a distinct population of circulating platelets. Together, these three mechanisms forge TEPs whose transcriptomic signature mirrors the tumor tissue expression profile. Therefore, after obtaining differentially expressed genes from the TEP transcriptome, we conducted analyses such as machine learning on the tumor tissue transcriptome. In our research, we developed a TEPGS signature comprising twelve TEPs-related genes (ARPC1B, OR2B6, TIMP1, TMEM40, RGS2, ERAP1, SRI, MMP1, FCER1A, ICAM3, MRPL37, and KIF13B) to predict prognosis and personalized therapeutic responses in patients with CRC. To evaluate the clinical predictive value of TEPGS in CRC, we initially performed multivariate Cox regression analysis, confirming its independent prognostic significance after adjusting for age, gender, and TNM stage. Subsequently, ROC and C-index analyses in both training and validation cohorts consistently demonstrated the acceptable predictive accuracy and stability of the TEPGS signature. We then collected and compared 71 published mRNA-based prognostic signatures, finding that TEPGS maintained relatively superior accuracy in prognosis risk assessment in each cohort. Furthermore, a nomogram was constructed based on the TCGA-CRC or GEO: GSE12945 cohort to assess the effectiveness of TEPGS in predicting patients’ outcomes. The decision curve analyses verified that the TEPGS-based nomogram exhibited superiority over other features for prognostic prediction. Our findings suggested that TEPGS could be a potential biomarker for prognostic stratification in clinical practice. To investigate the underlying molecular foundation and mechanism of TEPGS in predicting poor prognosis in patients, we deciphered the genomic and transcriptomic landscapes in the TEPGS-low and -high groups. Our research showed that the TEPGS score was correlated positively with TP53, SYNE2, and ZEP36L2 mutations and negatively associated with SOX9, UNC13C, TCF7L2, and B2M mutations. To the best of our knowledge, TP53 is identified as the most frequently mutated gene exhibiting mutation rates exceeding 50% in CRC tumors. Survival analyses have shown that the prognosis of patients with cancer is significantly different between diverse TP53 mutant subtypes.42,43 Sex-determining region Y-box 9 (SOX9) is a frequently mutated transcription factor in CRC that is transcriptionally repressed by the Wnt signaling pathway and associated with a favorable prognosis in individuals with stage II CRC.44,45 TCF7L2, a member of the T cell factor/lymphoid enhancer binding factor (TCF/LEF) family that participates in the Wnt/β-Catenin pathway and exhibits a tumor-suppressive effect in CRC.46 In clinical settings, the association of BRAF, KRAS, and NRAS mutations with increased recurrence rates in advanced patients with CRC provides mechanistic insight into their aggressive clinical behavior and unfavorable outcomes.26 Owing to the poorly characterized prognostic significance of recurrently mutated genes (SYNE2, ZEP36L2, SOX9, UNC13C, TCF7L2, and B2M) in CRC, the molecular mechanisms underlying the poor prognosis observed in TEPGS-high patients warrant further investigation. Accumulating evidence from recent decades has established that CNVs constitute critical genomic drivers of CRC pathogenesis and significantly influence disease progression and survival outcomes.47 In this study, patients with high TEPGS scores experienced elevated genomic instability, characterized by increased amplification and depletion events compared to individuals with low TEPGS scores, which may mechanistically contribute to their adverse prognosis. The initiation and progression of cancer are critically associated with the TME. Multiple cellular components of the TME, including cancer-associated fibroblasts, TAMs, and myeloid-derived suppressor cells, are linked to prognosis and responsiveness to immunotherapies in CRC.4,48 Investigating the correlations between TEPGS and immune cell infiltration can reveal the complex interplay between tumor genetics and the immune microenvironment, thereby providing immunological evidence for the poor prognosis observed in patients with high TEPGS expression. As established in the literature, tumors can be categorized into three primary immune phenotypes based on the distribution and abundance of cytotoxic immune cells in the TME: immune-inflamed, immune-excluded, and immune-deserted.49 Among these, immune-excluded and immune-desert tumors, collectively termed “cold tumors,” are respectively defined by the presence of T cells without infiltration into the tumor or the absence of T cells, and do not respond to ICB.50,51 Major mechanisms underlying ICB resistance include impaired antigen presentation, disrupted interferon-gamma signaling, and a suppressive TME characterized by insufficient or dysfunctional T cell infiltration or high expression of additional checkpoint inhibitors.52,53 Consistent with this framework, our study demonstrated that patients with CRC with elevated TEPGS levels exhibited significantly reduced immunescore and TMEscore, suggesting an immunosuppressive TME. Notably, our analysis revealed a significant inverse correlation between TEPGS levels and the expression of key immune checkpoint molecules (BTLA, TIGIT, ICOS, CD274, CTLA4, CD27, and IDO1) in the TME. Furthermore, high TEPGS scores were associated with decreased levels of well-validated immunotherapy response predictors (IFN-γ scores, IPS, CYT scores, and GEP scores) and concurrently elevated TIDE scores—a robust biomarker of immunotherapy resistance. Collectively, these findings confirm that high-TEPGS tumors exhibit a classic cold tumor phenotype, marked by low immune infiltration, diminished checkpoint molecule expression, and elevated TIDE scores. ICB resistance in high-TEPGS tumors is likely attributable to the lack of immune cell infiltration, rather than checkpoint-mediated suppression. Recent advancements in scRNA-seq technologies have rapidly improved the identification of diverse cellular populations and phenotypic states within tumors, enabling a more comprehensive understanding of tumor heterogeneity and evolution within the TME.54 In our research, ScRNA-seq data were used to investigate the heterogeneity of TAMs in CRC and their relationship to the TEPGS signature. We identified five distinct TAM subtypes with varied distribution patterns between TEPGS-high and TEPGS-low tumors. By integrating spatial transcriptomics with scRNA-seq, Ozato et al. revealed that undifferentiated CRC cells at the invasion front can reprogram macrophages into SPP1+ TAMs, promoting tumor immunosuppression, growth, differentiation, and EMT.55 Additionally, pan-cancer single-cell analysis has revealed that SPP1+ TAMs may participate in tumor angiogenesis by interacting with adjacent cancer-associated fibroblasts.56 SPP1+ TAMs were enriched in hypoxic and necrotic tumor regions and were related to worse OS in CRC.57 These findings collectively demonstrate that SPP1+ TAMs are associated with tumor malignancy and unfavorable prognosis. Herein, we found that the preferential enrichment of SPP1+ and FCN1+ TAMs in TEPGS-high tumors, together with their interaction with pro-tumorigenic pathways including EMT, fibroblast proliferation, and glycolysis, thereby linking a high TEPGS score to a pro-tumor role within the TME. From a spatial perspective, high TEPGS tumors are spatially correlated with increased platelet abundance within the CRC TME, providing a preliminary link between liquid biopsy-derived TEPs and solid tumor transcriptomics. Notably, these detected marker signals only imply potential platelet involvement and cannot confirm the actual presence of intact platelets. Moreover, the Visium platform lacks subcellular resolution for precise cellular-level profiling. Although a modest, statistically significant correlation was observed across the three samples, this association is weak (r = 0.2–0.3). Due to the large number of spatial data points in the analysis, statistical significance alone bears limited biological implication. Accordingly, this preliminary relationship requires further verification in larger multicenter clinical cohorts. Due to poor immunotherapy response in high-TEPGS cases, we obtained compounds from the CTRP2.0 and PRISM databases and screened for potential therapeutic agents. We examined the IC50 value of commonly used chemotherapy drugs and found that patients with CRC with high TEPGS were resistant to paclitaxel, lapatinib, docetaxel, and fluorouracil. Notably, patients with high-TEPGS CRC displayed increased drug sensitivity to 11 drugs, including BRD-K79669418, NSC23766, PLX-4720, BMS-536924, TG100-11, vemurafenib, AG-14361, lorlatinib, regorafenib, INC-280, and simvastatin. PLX-4720, a selective BRAF V600E inhibitor, was more effective in treating patients with CRC with high TEPGS. PLX-4720 effectively suppresses ERK phosphorylation, inducing cell-cycle arrest and apoptosis.58 BRAF V600E mutation drives tumor progression by enhancing the phosphorylation of MEK and ERK and constitutively activating the mitogen-activated protein kinase (MAPK) signaling pathway, thereby promoting cancer cell proliferation and survival.59,60 Given the clinical challenges in treating BRAF V600E-mutated mCRC, the development of effective targeted therapies remains an urgent research priority. In this context, the BRAF V600E inhibitor PLX-4720 represents a potential therapeutic strategy in CRC. In first-line settings, bevacizumab or cetuximab, in combination with fluorouracil, oxaliplatin, or irinotecan-based chemotherapy, can extend patients’ median survival compared with chemotherapy alone.32 A variety of single potential markers have been demonstrated to predict response to bevacizumab or cetuximab in patients with mCRC.61,62,63,64 The combination of multiple biomarkers can better match the sensitivity and specificity required to predict treatment response, as tumor heterogeneity and resistance are strongly linked.34 In the present study, we identified that patients with high TEPGS (a 12-gene signature) exhibited clinical benefit from cetuximab treatment, while the association between high TEPGS and bevacizumab benefit was only detectable in meta-analysis results, with no statistical significance observed in any of the three independent cohorts. These results suggest a tentative link between high TEPGS scores and potential sensitivity to bevacizumab, though this association was not statistically robust in individual cohorts due to limited sample size constraints. Notably, although the AUC values of TEPGS for predicting benefit from bevacizumab were slightly lower than those reported by Liu et al.,34 our study nevertheless confirmed the predictive potential of TEPGS for cetuximab benefit, which was not reported in several similar studies.34,65,66 However, the ROC curves for TEPGS-based response prediction are marginally above random, indicating modest discriminatory capacity for both targeted therapies. Taken together, these data suggest a tentative predictive value of TEPGS for cetuximab response and highlight that its predictive value for bevacizumab remains highly preliminary and shows limited stability. Future investigations will be conducted to verify the potential link between high TEPGS expression and bevacizumab/cetuximab clinical benefit in our proprietary patient cohort. The exploration of TEPs-related genes is an urgent research priority, as systematic studies on their biological roles in CRC remain lacking. ARPC1B, a member of TEPGS, encodes a subunit of the Arp2/3 complex and is associated with poor prognosis in patients with CRC. ARPC1B exhibited significantly elevated transcriptional levels in CRC compared to the normal group. In glioblastoma, macrophage-derived ARPC1B promotes cell invasion, migration, and EMT through the IFNγ-IRF2-ARPC1B signaling axis.67 In another study, ARPC1B was shown to protect IFI16 and HuR from TRIM21-mediated degradation in glioma stem cells, thereby maintaining their mesenchymal phenotype and conferring radiation resistance.68 Importantly, elevated ARPC1B expression was significantly associated with worse prognosis in patients with glioblastoma. In this study, we demonstrate that ARPC1B significantly promotes cell proliferation and migration in vitro and enhances tumor growth in vivo. However, the underlying mechanism requires further investigation. In summary, patients with CRC exhibiting high TEPGS demonstrated significantly poorer survival outcomes, characterized by reduced antitumor immune infiltration, downregulated immune checkpoint expression, enrichment of pro-tumor SPP1+/FCN1+ TAMs, spatially elevated platelet abundance, and increased resistance to both immunotherapy and conventional chemotherapy. These patients showed enhanced therapeutic response to a BRAF V600E inhibitor. Notably, TEPGS displayed tentative predictive utility for cetuximab response, with its predictive performance for bevacizumab remaining highly preliminary. Genomic alteration analyses were performed to elucidate underlying mechanisms. Notably, we preliminarily confirmed the carcinogenic effects of ARPC1B in vitro and vivo. Our findings are encouraging for the application of the TEPGS for prognostic stratification and selecting therapeutic strategies tailored to patients with CRC. Limitations of the study Although our results highlight the clinical significance of TEPGS in CRC, some limitations of the study should be considered. Firstly, the TEPGS was constructed and validated using publicly accessible retrospective datasets. In future studies, we will strive to establish prospective multicenter cohorts and our own in-house patient cohorts to further verify the generalizability and clinical applicability of the TEPGS. Secondly, the information on clinical and molecular features was insufficient in some of the public datasets, which may have obscured the potential connections between TEPGS and specific features. Thirdly, the associations between TEPGS and therapeutic benefit for bevacizumab and cetuximab were established in small cohorts and demonstrated modest predictive strength, with AUC values often only marginally above random. Consequently, validation in larger targeted therapy cohorts is imperative—both to define the unstable relationship with bevacizumab benefit and to rigorously evaluate its potential for predicting cetuximab response. Fourthly, the TEPGS signature was originally derived from bulk tumor transcriptomic profiles and subsequently projected onto single-cell datasets. This approach yields a single signature score per cell, which inherently introduces bias by masking cell state-specific expression patterns within heterogeneous populations. Consequently, definitive assignment of TEPGS components to specific cell states will require signature discovery and validation directly from purified single-cell or high-resolution spatial transcriptomic datasets. Fifthly, our drug sensitivity predictions via pRRophetic rely on gene expression-drug response correlations established in cancer cell line models, which fail to recapitulate the multifaceted influences of the TME. Cancer cell lines differ substantially from primary tumors in lineage composition, signaling context, and drug metabolism, which can lead to inaccurate estimates of drug efficacy in clinical settings. Future work will therefore prioritize the validation of candidate drug-signature interactions in preclinical models, including patient-derived organoids and xenograft models that preserve TME context. Lastly, more in vivo and in vitro experiments are required to determine the functions of the TEPGS genes, as their roles in CRC remain unclear. Resource availability Lead contact Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Hui Yang (yanghui@gzhmu.edu.cn). Materials availability This study did not generate new unique reagents, and all materials in this study are commercially available. Data and code availability •This article analyzes existing, publicly available data from studies by Smith et al.,69 Freeman et al.,70 Chen et al.,71 Tripathi et al.,72 Marisa et al.,73 Staub et al.,74 Del Rio et al.,75 Khambata-Ford et al.,76 Uhlitz et al.,77 Park et al.,78 and Best et al.17 GEO Accession numbers are listed in the key resources table.•Essential code for conducting machine learning-based integrative procedure is available on the GitHub website (https://github.com/YueqiongLao/TEPGS).•Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. •This article analyzes existing, publicly available data from studies by Smith et al.,69 Freeman et al.,70 Chen et al.,71 Tripathi et al.,72 Marisa et al.,73 Staub et al.,74 Del Rio et al.,75 Khambata-Ford et al.,76 Uhlitz et al.,77 Park et al.,78 and Best et al.17 GEO Accession numbers are listed in the key resources table. This article analyzes existing, publicly available data from studies by Smith et al.,69 Freeman et al.,70 Chen et al.,71 Tripathi et al.,72 Marisa et al.,73 Staub et al.,74 Del Rio et al.,75 Khambata-Ford et al.,76 Uhlitz et al.,77 Park et al.,78 and Best et al.17 GEO Accession numbers are listed in the key resources table. •Essential code for conducting machine learning-based integrative procedure is available on the GitHub website (https://github.com/YueqiongLao/TEPGS). Essential code for conducting machine learning-based integrative procedure is available on the GitHub website (https://github.com/YueqiongLao/TEPGS). •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Acknowledgments This study was supported by the General Project of 10.13039/501100003453Guangdong Natural Science Foundation (no. 2021A1515010795), the 10.13039/501100001809National Natural Science Foundation of China (nos. 82102813, 82100610, and 82403294), the 10.13039/501100010256Guangzhou Science and Technology Plan Project (nos. 202201020157, 2023A03J0416, and 2024A03J1003), the National Foundation of Postdoctoral Program of China (no. GZC20233187), and the Guangdong Province ordinary universities characteristic innovation project (no. 2024KTSCX125). Author contributions Y.L., writing – original draft, methodology, investigation, formal analysis, data curation, and conceptualization. J.Y., software, resources, methodology, data curation, formal analysis, and funding acquisition. M.H., data curation and validation. J.L., visualization and data curation. W.X., data curation and validation. J.X., conceptualization and funding acquisition. H.W., data curation and investigation. H.J., methodology and visualization. Z.P., data curation and methodology. X.Q., data curation and methodology. K.W., supervision and funding acquisition. X.L., conceptualization, project administration, writing – review and editing, and supervision. H.Y., writing – review and editing, supervision, project administration, funding acquisition, and conceptualization. Declaration of interests The authors declare no competing interests. STAR★Methods Key resources table REAGENT or RESOURCESOURCEIDENTIFIERAntibodiesAnti-β-ACTIN antibodyCell Signaling TechnologyCat#4967; RRID: AB_330288Anti-ARPC1B antibodyProteintechCat#67320-1-Ig; RRID: AB_2882580Biological samplesBALB/c male nude miceLaboratory Animal Center of Southern Medical UniversityN/AChemicals, peptides, and recombinant proteinsDMEM mediumGibcoCat#11965092McCoy’s 5A mediumGibcoCat#16600082Fetal bovine serumGibcoCat#A5256701TRIzolInvitrogenCat#15596018CNLipofectamine™ RNAiMAXInvitrogenCat#13778150Enhanced Chemiluminescence KitFdbio scienceCat#FD800Cell Counting Kit-8 (CCK-8)SelleckCat#B34304Phosphate buffered salineProcellCat#PB180327ParaformaldehydeSolarbioCat #P1110puromycinBeyotimeST551Crystal violetMedChemExpressCat#HY-B0324ACritical commercial assaysPrimeScript RT Reagent KitTakaraCat#RR037ATaq Pro Universal SYBR qPCR Master MixVazymeCat#Q712-02Deposited dataBulk RNA-seq data for CRCTCGAhttps://tcga-data.nci.nih.gov/tcgaBulk RNA-seq data for CRCSmith et al.69 2010GEO: GSE17536Bulk RNA-seq data for CRCFreeman et al.70 2012GEO: GSE17537Bulk RNA-seq data for CRCChen et al.71 2012GEO: GSE29621Bulk RNA-seq data for CRCTripathi et al.72 2014GEO: GSE38832Bulk RNA-seq data for CRCMarisa et al.73 2013GEO: GSE39582Bulk RNA-seq data for CRCStaub et al.74 2009GEO: GSE12945Bulk RNA-seq data for CRCDel Rio et al.75 2017GEO: GSE72970Bulk RNA-seq data for CRCKhambata-Ford et al.76 2007GEO: GSE5851Single-cell RNA-seq data for CRCUhlitz et al.77 2021GEO: GSE166555Visium Spatial transcriptomic data for CRCPark et al.78 2024GEO: GSE226997RNA-seq data of tumor-educated platelets for pan-cancerBest et al.17 2015GEO: GSE68086Machine learning-based codeGitHub URLhttps://github.com/YueqiongLao/TEPGSCopy number variation dataUCSC Xenahttps://xena.ucsc.edu/Experimental models: cell linesHuman: SW480ATCCCCL-228: RRID:CVCL_0546Human: HCT116ATCCCCL-247: RRID:CVCL_0291OligonucleotidesPrimers for GAPDH and ARPC1BThis paper.N/ASiRNA and shRNA targeting sequence for ARPC1BThis paper.N/ASoftware and algorithmsImageJ softwareImageJN/AR 4.2.0RStudioN/AGraphPad Prism v9.1.1 for MacGraphPad SoftwareN/ATCGAbiolinks (v2.24.3)R packageN/ADESeq2 (v1.28.1)R packageN/AMaftools (v2.4.12)R packageN/AComplexHeatmap (v2.4.3)R packageN/AGISTIC2.0Webtool softwareN/ApRRophetic (v0.5)R packageN/AGSVA (v1.44.5)R packageN/AIOBR (v0.99.9)R packageN/ATMEscore (v0.1.4)R packageN/ASeurat (v5.1.0)R packageN/AMonocle2 (v2.32.0)R packageN/AtimeROC (v0.4)R packageN/ASurvival (v3.5-5)R packageN/Arms (v6.7–0)R packageN/AOtherPolyvinylidene fluoride membraneMilliporeCat#ISEQ000103-well Ibidi silicone culture insertIbidiCat#80369 REAGENT or RESOURCESOURCEIDENTIFIERAntibodiesAnti-β-ACTIN antibodyCell Signaling TechnologyCat#4967; RRID: AB_330288Anti-ARPC1B antibodyProteintechCat#67320-1-Ig; RRID: AB_2882580Biological samplesBALB/c male nude miceLaboratory Animal Center of Southern Medical UniversityN/AChemicals, peptides, and recombinant proteinsDMEM mediumGibcoCat#11965092McCoy’s 5A mediumGibcoCat#16600082Fetal bovine serumGibcoCat#A5256701TRIzolInvitrogenCat#15596018CNLipofectamine™ RNAiMAXInvitrogenCat#13778150Enhanced Chemiluminescence KitFdbio scienceCat#FD800Cell Counting Kit-8 (CCK-8)SelleckCat#B34304Phosphate buffered salineProcellCat#PB180327ParaformaldehydeSolarbioCat #P1110puromycinBeyotimeST551Crystal violetMedChemExpressCat#HY-B0324ACritical commercial assaysPrimeScript RT Reagent KitTakaraCat#RR037ATaq Pro Universal SYBR qPCR Master MixVazymeCat#Q712-02Deposited dataBulk RNA-seq data for CRCTCGAhttps://tcga-data.nci.nih.gov/tcgaBulk RNA-seq data for CRCSmith et al.69 2010GEO: GSE17536Bulk RNA-seq data for CRCFreeman et al.70 2012GEO: GSE17537Bulk RNA-seq data for CRCChen et al.71 2012GEO: GSE29621Bulk RNA-seq data for CRCTripathi et al.72 2014GEO: GSE38832Bulk RNA-seq data for CRCMarisa et al.73 2013GEO: GSE39582Bulk RNA-seq data for CRCStaub et al.74 2009GEO: GSE12945Bulk RNA-seq data for CRCDel Rio et al.75 2017GEO: GSE72970Bulk RNA-seq data for CRCKhambata-Ford et al.76 2007GEO: GSE5851Single-cell RNA-seq data for CRCUhlitz et al.77 2021GEO: GSE166555Visium Spatial transcriptomic data for CRCPark et al.78 2024GEO: GSE226997RNA-seq data of tumor-educated platelets for pan-cancerBest et al.17 2015GEO: GSE68086Machine learning-based codeGitHub URLhttps://github.com/YueqiongLao/TEPGSCopy number variation dataUCSC Xenahttps://xena.ucsc.edu/Experimental models: cell linesHuman: SW480ATCCCCL-228: RRID:CVCL_0546Human: HCT116ATCCCCL-247: RRID:CVCL_0291OligonucleotidesPrimers for GAPDH and ARPC1BThis paper.N/ASiRNA and shRNA targeting sequence for ARPC1BThis paper.N/ASoftware and algorithmsImageJ softwareImageJN/AR 4.2.0RStudioN/AGraphPad Prism v9.1.1 for MacGraphPad SoftwareN/ATCGAbiolinks (v2.24.3)R packageN/ADESeq2 (v1.28.1)R packageN/AMaftools (v2.4.12)R packageN/AComplexHeatmap (v2.4.3)R packageN/AGISTIC2.0Webtool softwareN/ApRRophetic (v0.5)R packageN/AGSVA (v1.44.5)R packageN/AIOBR (v0.99.9)R packageN/ATMEscore (v0.1.4)R packageN/ASeurat (v5.1.0)R packageN/AMonocle2 (v2.32.0)R packageN/AtimeROC (v0.4)R packageN/ASurvival (v3.5-5)R packageN/Arms (v6.7–0)R packageN/AOtherPolyvinylidene fluoride membraneMilliporeCat#ISEQ000103-well Ibidi silicone culture insertIbidiCat#80369 Experimental model and study participant details Data collection and processing Transcriptome data and clinical information of patients with CRC were retrieved from The Cancer Genome Atlas (TCGA) and Gene Expression Omnibus (GEO). Seven independent datasets (n = 1598) were used to construct and validate our signature: (a) The training cohort was TCGA-CRC (n = 570); (b) The five internal validation cohorts were GEO: GSE17536 (n = 177), GEO: GSE17537 (n = 55), GEO: GSE29621 (n = 65), GEO: GSE38832 (n = 122), and GEO: GSE39582 (n = 551); and (c) The independent external validation cohort was GEO: GSE12945 (n = 58). Four independent datasets (n = 115) were used to assess the benefits of bevacizumab and cetuximab: GEO: GSE19860 (n = 12), GEO: GSE19862 (n = 14), GEO: GSE72970 (n = 21), and GEO: GSE5851 (n = 68). Somatic mutation and copy number variation (CNV) data were accessed from UCSC Xena (https://xena.ucsc.edu/) and obtained using the R package TCGAbiolinks (v2.24.3).79 Cell culture The human CRC cell lines SW480 and HCT116 were a gift from Prof. Side Liu’s research group of Southern Medical University, China, and purchased from iCell Bioscience Inc., China, respectively. All cell lines used in this study were authenticated by short tandem repeat (STR) profiling before use and confirmed to be free of cross-contamination. Prior to experiments, all cells were verified to be free of mycoplasma contamination using PCR. No mycoplasma contamination was detected throughout the study. Cells were cultured in Dulbecco’s modified Eagle's medium (DMEM) or McCoy’s 5A medium supplemented with 10% fetal bovine serum (FBS) at 37 °C in a humidified incubator with 5% CO2. Xenograft mouse model BALB/c male nude mice (5 weeks old, 18–20 g) were purchased from the Laboratory Animal Center of Southern Medical University (Guangzhou, China). Mice were maintained under specific pathogen-free conditions in the Second Affiliated Hospital of Guangzhou Medical University animal center and received food and water ad libitum. Fifteen mice were randomly divided into shControl, shARPC1B#1, and shARPC1B#2 groups (N = 5/group). The SW480 cells, with or without ARPC1B knockdown, were suspended in 100 μL sterile 1X phosphate-buffered saline and then injected subcutaneously (6 × 106 cells/mouse) into the flanks of mice. Tumor volume was monitored via caliper measurements every 3 days. Sixteen days after tumor cell implantation, mice were sacrificed through cervical dislocation after inhalation of isoflurane, and the tumors were removed. Tumor volume was calculated using the formula 0.5 × length × width.2 The animal experiments and procedures were approved by the Institutional Animal Care and Use Committee of the Second Affiliated Hospital of Guangzhou Medical University (approval number: B2024-115). Method details Screening for TEPs-related genes TEPs-related differentially expressed genes (DEGs) in blood samples between patients with CRC and healthy donors were retrieved from Best et al.17 The raw count data of the gene expression of 570 TCGA-CRC samples (tumor and adjacent normal tissues) were collected from TCGA. The R package DESeq2 (v1.28.1)80 was used to identify DEGs between tumor and normal tissues with an absolute log2-fold change ≥0.5 and adjusted p value <0.05 as thresholds. Finally, DEGs that were simultaneously upregulated or downregulated in two cohorts were selected. TEPGS establishment A total of 101 combinations of 10 machine learning algorithms were used to screen TEPGS with the highest Harrell’s concordance index (C-index) to establish a consensus TEPGS with high accuracy and stability. The C-index, first introduced by F. E. Harrell Jr. et al.,81 was a widely used metric for evaluating the discriminatory power of survival prediction models. In the present study, the C-index was applied to assess the ability of machine learning models to predict and differentiate between patients with differing survival outcomes. The C-index ranges from 0.5 (indicating no predictive ability, equivalent to random chance) to 1.0 (indicating perfect discrimination), with higher values reflecting better performance in distinguishing between patients with distinct overall survival. All C-index values, along with their standard errors and 95% confidence intervals, were calculated using the rcorr.cens function from the Hmisc R package. The 10 machine learning algorithms included CoxBoost, elastic network (Enet), generalised boosted regression modeling (GBM), least absolute shrinkage and selection operator (Lasso), partial least-squares regression for Cox (plsRcox), Ridge, random survival forest (RSF), stepwise Cox, supervised principal components (SuperPC), and survival support vector machine (survival-SVM). The TEPGS was constructed following a standardized three-step workflow34: (a) Univariate Cox regression analysis was conducted to determine predictive TEPs-related DEGs in the TCGA-CRC cohort. (b) Based on the 10-fold cross-validation approach, the prediction models in the TCGA-CRC cohort were fitted using 101 combinations. Each model was then evaluated in five independent internal validation cohorts. (c) The model with the highest average C-index across all internal validation cohorts was deemed most valuable. The incremental predictive value of TEPGS To ensure the robustness of multivariate Cox regression analyses, we assessed collinearity among variables using the Variance Inflation Factor (VIF) across all cohorts. A VIF value <5 was considered indicative of no significant collinearity, while VIF >10 suggested strong collinearity. For evaluating the incremental prognostic value of TEPGS beyond clinical/tumor variables, we constructed two models: (1) a baseline model incorporating clinical/tumor covariates (age, gender, TNM stage, TMB, TNB, TP53, KRAS, BRAF); and (2) a TEPGS-integrated model combining the baseline covariates with the TEPGS score. The Net Reclassification Index (NRI) and Integrated Discrimination Index (IDI) were applied to compare the performance of these two models at the 5-year follow-up endpoint.82,83,84 An NRI ≥0.1 with a p value <0.05 and an IDI >0 with a p value <0.05 were defined as statistically significant improvements in model performance. Comparison of somatic mutations and CNVs between TEPGS-high and -low groups Based on previous reports on CRC,25,26,27,28,85 we summarised several frequently mutated genes (FMGs). To identify differential genes with somatic mutations, we applied the mafCompare algorithm via the R package maftools (v2.4.12) to compare TEPGS-high and -low tumors, and then selected 20 FMGs. Subsequently, the mutation landscapes of these 20 genes were visualised using the maftools and ComplexHeatmap (v2.4.3) R packages. CNV data were analyzed using the GISTIC2.0 webtool in Genepattern. In a tumor genome, we considered a gene undergoing homozygous deletion, single copy deletion, low-level copy number amplification, or high-level copy number amplification if the corresponding GISTIC score = −2, −1, 1, or 2, respectively. Therefore, we defined gene level amplification as genes with a GISTIC score of 1 or 2 and gene level deletion as genes with a GISTIC score of −1 or −2. Chemotherapeutic response prediction Gene expression data (transcript per million) for human cancer cell lines (CCLs) were retrieved from the Broad Institute Cancer Cell Line Encyclopedia project (https://portals.broadinstitute.org/ccle/). The drug sensitivity data (area under the dose-response curve (AUC) measures) for 482 CCLs across 1492 compounds were downloaded from PRISM, and 835 CCLs across 481 compounds were available from the Cancer Therapeutics Response Portal (CTRP2.0) dataset. The AUC value revealed the relationship between the drug concentration and cell viability, and a low AUC value indicated high sensitivity to drug therapy. After excluding compounds with >20% missing data and imputing the missing AUC values using the k-nearest neighbor imputation method, we applied the calcPhenotype algorithm in the R package pRRophetic (v0.5) to estimate the drug response of each sample with the CCLs expression matrix, CCLs AUC values, and TCGA-CRC or GEO: GSE39582 expression matrix as inputs.86 Immune-related characteristics annotation Utilising the single-cell gene set enrichment analysis (ssGSEA) algorithm, which was a part of the R package GSVA (v1.44.5), immune cell fractions were calculated.87 The xCell algorithm was employed to calculate the ImmuneScore and StromaScore using the R package IOBR (v0.99.9).88 The TMEscore (v0.1.4) R package was applied to calculate the tumor microenvironment score, which was another method used for predicting immune cell infiltration.89 Interferon gamma (IFN-γ) score, immunophenoscore (IPS), cytotoxic activity (CYT) and T cell-inflamed gene expression profile (GEP) score were considered as superior predictors to immune checkpoint blockade (ICB).90,91 The IFN-γ score was inferred by six genes (IFNγ, CXCL9, CXCL10, IDO1, STAT1, and HLA-DRA) using the R package GSVA (v1.44.5).92 The IPS was calculated using the IOBR R package (v0.99.9). As we previously reported, the CYT score was determined by calculating the geometric mean of GZMA and PRF1.93 According to previous studies,90,91 the weighted sum of the normalised expression values for 18 T cell-inflammation prediction genes was used to calculate the GEP score. The effectiveness of anti-PD-1 and anti-CTLA4 therapy was assessed using the tumor immune dysfunction and exclusion (TIDE) score, which was produced from the TIDE website (http://tide.dfci.harvard.edu/).94 ScRNA-seq data analysis ScRNA-seq data and cell annotation file were downloaded from the TISCH database (GEO: GSE166555).95 Cells with 500–5000 genes, no more than 50000 total counts and 10% of mitochondrial gene counts were chosen for further analysis using the standard Seurat (v5.1.0) workflow.96 Gene counts were normalized and 2000 highly variable genes were selected. We then performed principal component analysis (PCA). Harmony was used to remove the batch effects. Next, cell clusters were identified using FindNeighbors and FindClusters methods, and cell type annotation was used as previously published.97 Marker genes for macrophage subtypes were identified using FindAllMarkers. Cell developmental trajectory analysis was conducted using Monocle2 (v2.32.0) with default parameters.98 To examine the enrichment of each cell type between TEPGS-high and -low groups, we used the following formula as previously mentioned99:Ratio=cik/∑j=1mcijCk/∑z=1mCz where cik represented the counts of a cell type i in one group (named group k), m represented total counts of group, and Ck represented the total counts of cells in group k. One cell type was assumed to be enriched in a group when Ratio > 1. Spatial transcriptomic analysis Spatial transcriptomic data of CRC tissues were retrieved from the GEO database under accession number GEO: GSE226997, which encompassed three independent tumor samples (Ajou_Visium_P1, Ajou_Visium_P2, Ajou_Visium_P3) profiled using the 10x Genomics Visium platform.78 All analyses were conducted in R (v4.2.0) using Seurat (v5.1.0). Briefly, Seurat objects were constructed for each sample with the CreateSeuratObject function, followed by data normalization using NormalizeData. Spatial expression of target genes was visualized via the SpatialFeaturePlot function. To quantify local platelet abundance, we calculated a platelet score for each Visium spot using Seurat’s AddModuleScore function, which aggregates the expression of a curated panel of platelet-specific marker genes. In parallel, the TEPGS score was computed for each spot using our previously established signature. The spatial association between platelet abundance and TEPGS expression was assessed via pearson correlation analysis, with results presented as scatterplots. RNA extraction and real-time quantitative PCR Total RNA extraction and reverse transcription were performed using TRIzol reagent and a PrimeScript RT Reagent Kit (RR037A, Takara, Japan). Taq Pro Universal SYBR qPCR Master Mix (Q712-02, Vazyme, China) was used to quantitate the mRNA expression in triplicate on a LightCycler 480 II Real-Time PCR System (Roche) according to the manufacturer’s instructions. The primers used for real-time quantitative PCR (RT-qPCR) were listed in Table S5. Cell transfection and transduction When the confluency of SW480 and HCT116 cells in the 6-well plate reached 50–70%, the cells were transfected with siARPC1B or control siRNA (siControl) using Lipofectamine RNAiMAX (#13778150, Invitrogen, USA) reagent according to the manufacturer’s instructions. The siRNAs targeting ARPC1B (Table S5) were purchased from Kidan Biosciences (Guangzhou, China). To obtain stable knockdown cells, two short hairpin RNAs (shRNA) targeting ARPC1B and a negative control (shControl) were synthesized and inserted into PLKO.1-U6-GFP-Puro lentiviral shRNA vector (Kidan Biosciences, Guangzhou, China) (Table S5). CRC cells were infected with lentiviruses production more than 12 h and were subsequently cultured in DMEM containing 10% FBS for 24 h, followed by selection with puromycin (10 μg/mL) for 2 weeks. Western blot analysis Total protein was separated by sodium dodecyl sulphate-polyacrylamide gel electrophoresis and transferred onto a polyvinylidene fluoride membrane (Millipore, USA). Antibodies against β-ACTIN (#4967) and ARPC1B (67320-1-Ig) were purchased from Cell Signaling Technology or Proteintech. The membrane was incubated overnight at 4 °C with primary antibody and 2 h with secondary antibody. The signal was detected using an Enhanced Chemiluminescence Kit (FD800, Fdbio science, China) and a UVITEC Mini HD9 System. The full images of the blots were provided in Figure S7. All blots shown in a figure were derived from the same experiment and processed in parallel. Cell viability and colony formation assays Cell viability was measured using a CCK-8 kit (#B34304, Selleck, USA), and colony formation ability was estimated by counting the number of colonies after staining with crystal violet. For the cell viability assay, cells were seeded 3000 cells/well in five replicates in 96-well plates. At the indicated times (24, 48, 72, 96, and 120 h), cells were incubated with 110 μL DMEM containing 10 μL CCK-8 for 2 h. The absorbance of the cells in each well was measured at 450 nm using a microplate reader (BioTek). For the colony formation assay, SW480 and HCT116 cells were seeded 3000 cells/well in 6-well plates. After incubation at 37 °C for 10 days, at room temperature, colonies were fixed with methanol for 30 min, stained with 0.5% crystal violet for 30 min and then washed using ddH2O. Cell migration assay Cell migration ability was evaluated using 8.0 mm pore polycarbonate membrane inserts in 24-well transwell plates (Corning, USA). Briefly, cells transfected with siARPC1B or siNC were digested and resuspended in DMEM or McCoy’s 5A medium containing 1% FBS and then seeded into the upper chamber. The lower chamber was filled with 600 μL DMEM or McCoy’s 5A medium containing 15% FBS. After incubation at 37 °C for 48 h, cells on the upper surface of the membrane were removed, and those on the bottom surface were fixed with 4% paraformaldehyde for 30 min and stained with 0.5% crystal violet for 30 min. Finally, five fields of view were randomly selected to count the number of transmembrane cells. Wound healing assay The wound healing assay was performed using a 3-well Ibidi silicone culture insert (cat. 80369; Germany) to assess cell migration. A total of 1×105 transfected cells in 100 μL DMEM or McCoy’s 5A medium containing 10% FBS were added to the cell-free gap of the insert. When the cells reached 100% confluence in the 6-well plates, the wound healing inserts were removed. The plates were washed twice with phosphate buffered saline and incubated for 48 h in DMEM or McCoy’s 5A medium containing 1% FBS. Images were captured at 0, 24, and 48 h, and the wound healing rate was analyzed via ImageJ software. Quantification and statistical analysis Statistical analyses were performed using R 4.2.0, Python v3.8.12, and GraphPad Prism v9.1.1. The Wilcoxon rank-sum test was used to compare non-normally distributed continuous variables between groups. Student’s t test was used to compare normally distributed continuous variables between groups. Categorical variables were compared using Fisher’s exact test. Spearman’s rank correlation analysis was used to measure the correlation between two continuous variables. The median of TEPGS was used to divide CRC patients into high- and low-risk groups. Cox regression and Kaplan–Meier analyses were performed using the survival package in R (v3.5-5). The time-dependent area under the receiver operating characteristic (ROC) curve (AUC) for TEPGS was determined using the timeROC R package (v0.4). The nomogram was plotted using the rms R package (v6.7–0). For functional assay results, Student’s t test was used to compare two groups and two-way analysis of variance was used to compare three independent groups with two factors. All data were presented as the mean ± SEM. Statistical significance was indicated as follows: NS, not significant; ∗p < 0.05; ∗∗p < 0.01; ∗∗∗p < 0.001, ∗∗∗∗p < 0.0001.