Page 115 - Read Online

Li et al. Cancer Drug Resist. 2025;8:31 Page 21 of 26

capability for 3- and 5-year survival, with AUC values significantly exceeding 0.85. However, the 1-year
survival AUC reached 1.0, a result primarily due to the extremely low number of events (deaths) within the
first year in the TCGA-PRAD cohort, leading to inflated performance metrics at this time point.
Consequently, the 1-year AUC should be interpreted with caution and not considered a reliable indicator of
model performance. The 3- and 5-year AUC values more accurately reflect the model’s true predictive
power. Future studies incorporating larger sample sizes and independent prospective cohorts are warranted
to further validate the model’s accuracy for short-term prediction. Additionally, the model predicted
differences in prognosis between high- and low-risk patients when used alone or in combination with the
TMB subgroups in survival analysis, confirming the critical role of immune features in tumor development.
Notably, primary PCa typically exhibits low TMB, which likely partially accounts for the lack of statistically
significant differences in our cohort. Collectively, these findings suggest that TMB alone has limited clinical
utility as a prognostic indicator within the context of primary PCa. To explore the underlying reasons for the
differences in prognosis, we analyzed the biological characteristics of the two groups. The high-risk group
had a significantly higher gene mutation rate, primarily centered on TP53. TP53 is a well-known tumor
suppressor gene that helps maintain genomic stability and cancer resistance . Mutations in TP53 result in
[49]
the loss of its tumor-suppressive function, increased cell resistance, and promotion of PCa progression .
[50]
Moreover, the high gene mutation rate in the high-risk group can generate more neoantigens, which
enhance antigen presentation and lead to increased active immune activity within the TME.

The prognostic model we established not only stratifies patients into high- and low-risk groups, but also
reveals subtype-specific drug tolerance patterns. Notably, the 10-gene signature demonstrates strong
predictive value for both prognosis and therapeutic response. For instance, high-risk patients identified by
the model exhibit enhanced energy metabolism and dysregulated immune signaling pathways, suggesting
they may be better suited for metabolic-targeted therapies or immunotherapies. This model provides a
practical tool for clinical risk stratification and personalized treatment selection, enabling clinicians to
identify patients who may benefit from immune checkpoint inhibitors, metabolic inhibitors, or combination
regimens. Furthermore, our study lays the foundation for translating this signature into a clinical assay to
guide precision therapy for advanced PCa. These results underscore the necessity of developing precision
treatment strategies, such as targeting metabolic vulnerabilities or tailoring regimens based on
subtype-specific drug sensitivities.

With the advancement of single-cell technology, the exploration of the TME has become more common,
helping to capture immune cell distribution within the tumor and to identify distinct cancer cell subgroups.
Using our model, we screened high-risk cancer cell subgroups with high immune activity and employed the
GSEA to explore their biological features. These results indicated that high-risk cancer cell subgroups could
increase glycolysis rate and promote ATP production. Therefore, we hypothesized that the higher the
proportion of high-risk cancer cell subgroups, the more vigorous their energy metabolism and the worse
patient prognosis. Subsequent survival analysis based on the proportion of high-risk cell subgroups
confirmed that patients with a higher proportion of these subgroups had a poorer prognosis. Although
leveraging high-risk subpopulation proportions at single-cell resolution achieved statistically significant
survival discrimination, the corresponding 1- and 5-year survival prediction AUC values (0.637 and 0.623,
respectively) were substantially lower than those of the bulk RNA-seq model (0.854 and 0.889, respectively).
This indicates considerable potential for enhancing the predictive capacity of single-cell-derived risk
subpopulation metrics, likely constrained by limited sample sizes, cellular heterogeneity, or technical noise
artifacts. Future investigations should integrate larger-scale single-cell datasets with multimodal data
integration strategies to optimize model performance.

108