Prostate cancer is the most commonly diagnosed cancer in men (29% of male cancers) and the second leading cause of cancer death in the US. While most cases grow slowly, approximately 20% of patients develop aggressive, lethal forms of the disease. There is an urgent need for biomarkers that can identify high-risk patients and guide treatment decisions more precisely.
Tumors do not exist in isolation -- they are embedded in a complex tumor microenvironment (TME) consisting of immune cells, blood vessels, connective tissue, and signaling molecules. Among the immune cells, tumor-associated macrophages (TAMs) play a particularly important role. They can be co-opted by the tumor to suppress immune responses, promote blood vessel growth, and stimulate cancer cell proliferation and invasion.
Macrophages exist on a spectrum between two functional states. M1 macrophages are pro-inflammatory and attack tumors. M2 macrophages are anti-inflammatory and typically support tissue repair -- but in cancer, they are hijacked to protect the tumor from immune attack. In most cancers, TAMs predominantly adopt an M2-like phenotype, and higher M2 infiltration correlates with more advanced disease and worse prognosis.
Despite extensive study, the specific subtypes of TAMs present in prostate cancer and their precise relationships with patient outcomes remained poorly characterized. This study uses single-cell RNA sequencing -- a technology that can analyze gene activity in individual cells -- combined with machine learning to map and classify TAM subtypes and build a prognostic tool from the findings.
The researchers downloaded single-cell RNA sequencing (scRNA-seq) data from 4 prostate cancer tissue samples and 4 adjacent benign tissue samples (dataset GSE193337). After quality filtering, 64,567 individual cells were analyzed. Using well-established macrophage marker genes (CD68, CD86, and CD163), the researchers identified macrophage populations and then performed sub-clustering to find distinct TAM subgroups.
The analysis revealed three distinct TAM clusters (TAM_0, TAM_1, TAM_2) with different gene expression profiles and different proportions of malignant versus normal cells. From these clusters, 161 differentially expressed genes (DEGs) were identified as potential TAM feature genes -- genes that distinguish one TAM cluster from another and may drive their different behaviors.
To select the most prognostically important genes from the 161 candidates, the researchers applied 10 different machine learning algorithms -- including LASSO, Ridge regression, Random Survival Forest, CoxBoost, Elastic Net, and support vector machine variants -- creating 101 algorithm combinations. These were trained on the TCGA-PRAD dataset (494 prostate cancer patients) and validated on five independent datasets totaling 1,027 patients.
The best-performing algorithm combination (LASSO + Random Forest) selected 9 TAM feature genes that were most predictive of disease progression. A risk score formula was then constructed using these 9 genes weighted by their Cox regression coefficients, allowing each patient to be assigned a numerical risk score that predicts cancer progression.
Among the three identified TAM clusters, TAM_2 stood out as the most tumor-promoting. Unlike TAM_0 and TAM_1, which had a higher proportion of normal cells, TAM_2 had a 65% malignant cell composition. In the TCGA-PRAD dataset, TAM_2 gene expression scores were significantly higher in tumor tissue than in normal tissue -- the opposite pattern from TAM_0 and TAM_1.
Survival analysis confirmed the clinical significance: patients with high TAM_2 scores had significantly worse disease-free survival than those with low TAM_2 scores. Pathway analysis showed that TAM_2 feature genes were enriched in biological processes related to ferroptosis (iron-dependent cell death), apoptosis regulation, and lysosomal transport -- processes that, when dysregulated, can protect cancer cells from dying and promote tumor growth.
The TAM_2 cluster's characteristics -- its predominance in malignant tissue, its association with worse survival, and its activation of pathways that support tumor progression -- suggest it most closely resembles the traditional M2-type immunosuppressive macrophage that promotes tumor growth and immune evasion.
The other clusters had different roles: TAM_0 was also associated with prognosis (high TAM_0 scores correlated with worse outcomes), while TAM_1 showed no significant correlation with prostate cancer progression, suggesting it may represent a more neutral macrophage population in this context.
The machine learning pipeline identified 9 TAM feature genes: ANGPT1, AOX1, APOE, C7, CLIC6, CREB3L4, MBNL2, SPDEF, and ST6GALNAC4. Six of these (ANGPT1, AOX1, C7, CLIC6, MBNL2, SPDEF) were classified as protective genes (higher expression associated with better outcomes), while three (APOE, CREB3L4, ST6GALNAC4) were identified as risk genes (higher expression associated with worse outcomes).
Using these 9 genes, a risk score was calculated for each patient. Patients in the high-risk group had significantly worse progression-free survival than low-risk patients in both the training dataset and internal validation dataset. This finding held in an external validation cohort (SUPRAD, 136 patients) as well, confirming the model's generalizability. The AUC values for predicting disease progression at 1, 3, and 5 years ranged from 0.66 to 0.71 in validation sets.
A nomogram was constructed combining the 9-gene risk score with clinical stage information. The nomogram achieved a C-index of 0.71 -- a measure of the model's ability to correctly rank patients by risk -- and showed excellent calibration, meaning its predicted probabilities of progression matched actual observed outcomes closely at 1, 3, and 5 years.
Patients in the high-risk group also had higher tumor mutational burden (TMB) -- a measure of how many mutations are present in the tumor. Combining the TAM risk score with TMB provided even finer prognostic stratification: patients in the high-risk, high-TMB group had the worst outcomes, while low-risk, low-TMB patients had the most favorable prognosis.
Among the 9 TAM feature genes, APOE (Apolipoprotein E) showed the strongest and most specific association with M2 macrophages. Typically known for its role in lipid transport and its connection to cardiovascular disease and Alzheimer's disease, APOE has recently emerged as a significant player in cancer biology.
In this study, APOE was found to be predominantly expressed in prostate cancer macrophages (confirmed across three single-cell datasets) and showed essentially no protein expression in normal prostate tissue (confirmed via immunohistochemistry). Expression was elevated in aggressive prostate cancer cell lines (DU145 and 22RV1) compared to normal prostate epithelial cells, verified by qPCR and western blot.
Mendelian randomization analysis -- a statistical technique that uses genetic variants to establish causal relationships -- provided evidence that APOE causally promotes M2 macrophage development. Higher APOE expression increased levels of macrophage colony-stimulating factor 1 and mannose receptor 1, both key drivers of M2 polarization, and reverse analysis confirmed that M2 macrophage factors also increase APOE levels.
When APOE was knocked down in DU145 prostate cancer cells using siRNA (gene silencing), cell migration was significantly reduced in both scratch assays and Transwell migration assays. This functional experiment demonstrates that APOE actively promotes cancer cell movement -- a prerequisite for metastasis -- and represents a potential therapeutic target for aggressive prostate cancer.
Analysis of immune cell infiltration patterns showed that high-risk patients (defined by the 9-gene TAM signature) had significantly higher stromal cell infiltration scores, indicating a more immunosuppressive tumor microenvironment. Genes including ANGPT1, APOE, C7, and CLIC6 showed significant positive correlations with stromal and immune scores, while CREB3L4, MBNL2, SPDEF, and ST6GALNAC displayed negative correlations.
The relationship between the TAM signature and immune cell populations has direct implications for immunotherapy response prediction. In cancers where M2 macrophage dominance suppresses immune function, the TME may be resistant to checkpoint inhibitor immunotherapies (such as PD-1/PD-L1 blockers). The TAM risk score could help identify patients unlikely to respond to immunotherapy, guiding treatment selection.
Conversely, the identification of M2 macrophages as a tumor-promoting force opens therapeutic avenues. Strategies to deplete TAMs or repolarize them from M2 to M1 phenotype could enhance anti-tumor immunity. Prior research has already shown that TAM depletion can improve the effectiveness of docetaxel chemotherapy in castration-resistant prostate cancer, suggesting combination approaches.
The nomogram combining the 9-gene signature with clinical stage represents a practical clinical tool: physicians can enter a patient's tumor stage and gene expression data (from a biopsy) to generate a quantitative probability of disease progression, enabling more informed decisions about treatment intensity and monitoring frequency.
This study makes several meaningful contributions to prostate cancer biology. It provides the first systematic characterization of TAM subtypes in prostate cancer using single-cell sequencing, identifies a tumor-promoting TAM cluster (TAM_2) resembling M2 macrophages, and uses machine learning to distill the most prognostically important genes from this population into a clinically applicable 9-gene signature.
The model was validated in six independent datasets spanning over 1,600 patients, substantially strengthening confidence in its generalizability compared to models validated in single cohorts. The external validation in the SUPRAD dataset confirmed that the risk signature performs consistently across populations with different clinical characteristics.
Key limitations include the retrospective nature of all data sources. Prospective multicenter validation is needed to confirm whether the prognostic signature performs equally well in real-world clinical settings, where patient selection and data collection differ from curated research databases.
The functional validation of APOE as a cancer-promoting gene through siRNA knockdown experiments points to a clear future research direction: developing drugs that target APOE signaling or the APOE-M2 macrophage axis in prostate cancer. The identification of specific pathway interactions (such as APOE's connections to SYNE1, KMT2C, and LRP1B) also provides leads for understanding resistance mechanisms and combination therapy strategies.