The Aliquot

Every year, more than 35,000 American men die of prostate cancer, most of them not from the tumor that was first detected in their prostate, but from a transformed, therapy-resistant descendant that has shed its original identity and acquired the properties of a stem cell. A new study from Roswell Park Comprehensive Cancer Center now offers a quantitative measure of that transformation, a score derived from a tumor's gene expression profile that rises in lockstep with disease severity and, in patients with metastatic disease, predicts survival with a hazard ratio of 4.7.

The work, led by Xiaozhuo Liu and Dean Tang, draws on 3,102 RNA-sequencing samples and 84,081 microarray profiles spread across 26 distinct cohorts, covering every clinically meaningful stage of prostate cancer from normal glandular tissue to metastatic castration-resistant disease. The central finding is both conceptually clean and clinically provocative: as prostate cancer progresses, it becomes measurably more stem-like, and that stem-like quality is the single best molecular predictor of how badly things will go.

The Problem With Gleason

For decades, the Gleason score has been the clinical gold standard for grading prostate cancer. A pathologist examines the architectural pattern of prostatic glands under a microscope and assigns a number that reflects how far the tissue has drifted from normal. The system works reasonably well for primary tumors, but it breaks down precisely where it is needed most: in treatment-failed tumors and distant metastases, which have largely lost the glandular structures the score depends on.

The deeper limitation is conceptual. Gleason captures a snapshot of differentiation at one moment, in one biopsy, assessed by one pathologist. It does not quantify the underlying biological process driving that loss of differentiation, and it cannot be applied uniformly across the full arc of disease progression. What the field has lacked is a molecular ruler that works at every stage, from the first biopsy to the final metastatic lesion.

The Roswell Park team's approach draws on a concept from developmental biology: that cancer progression recapitulates, in distorted form, the process by which stem cells give rise to differentiated tissues. As tumors become more aggressive, they shed the gene expression programs of mature, specialized cells and reactivate programs associated with embryonic and adult stem cells. This 'stemness' is not merely a metaphor. It can be measured directly from a tumor's transcriptome using a Stemness Index (mRNAsi), a score that computes the Spearman correlation between a sample's gene expression profile and a stem cell-like reference signature.

Building the Ruler

Two complementary scores anchor the analysis. The Stemness Index measures oncogenic dedifferentiation: higher scores indicate a more primitive, stem-like transcriptional state. The canonical Androgen Receptor Activity score (AR-A) measures the opposite: how actively the tumor is running the normal differentiation program of prostatic epithelium, quantified as the combined Z-score of a validated 10-gene signature of known AR targets.

A third score, the MYC signaling activity (MYC-sig), was calculated using the same Z-score approach applied to a validated set of MYC target genes. Finally, the team derived a novel 12-gene 'PCa-Stem signature' by identifying genes consistently upregulated (fold change of 2 or greater, false discovery rate below 0.05) in the top third of Stemness-ranked tumors in both primary prostate cancer (TCGA-PRAD) and metastatic castration-resistant prostate cancer (SU2C 2019) cohorts simultaneously. The overlap of those two gene lists produced the signature.

Survival analyses used Kaplan-Meier estimation with log-rank testing, and multivariable Cox proportional hazards models adjusted for age, Gleason score, and tumor stage. Monotonic trends across ordered disease stages were assessed with the Jonckheere-Terpstra trend test. Genomic associations between Stemness and specific alterations, including MYC amplification and RB1 deletion, were drawn from cBioPortal data and tested with Fisher's exact test.

A Tale of Two Trajectories

The first finding is counterintuitive. In matched pairs of primary tumor and adjacent benign tissue from three independent cohorts (TCGA-PRAD, Wyatt 2014, Long 2020), both AR-A and Stemness are elevated in the tumor compared to normal tissue. Early prostate cancer initiation is characterized by a concordant rise in both scores. This makes biological sense: the androgen receptor is the master regulator of prostatic epithelial differentiation, and its activation in early tumors appears to drive not only luminal differentiation but also, paradoxically, a stem-like transcriptional state.

That concordance does not last. As primary tumors progress spontaneously to higher Gleason scores, the two trajectories split. Stemness continues to climb. AR-A begins to fall. The Jonckheere-Terpstra trend test confirms both trends are monotonic and statistically robust across the Gleason spectrum. The positive correlation between the two scores, which sits at a Pearson's r of 0.32 in primary tumors from the Bolis 2021 integrated cohort, weakens progressively with each increment in tumor grade and disappears entirely in the highest-grade primary cancers.

From Primary Tumor to Lethal Metastasis

The full picture emerges when the analysis is extended across the entire disease spectrum. Stemness rises continuously from normal prostate tissue through treatment-naive primary tumors, through tumors that have been exposed to androgen deprivation therapy, and into metastatic castration-resistant prostate cancer (mCRPC). The endpoint is striking: mCRPC carries the highest Stemness scores of any clinical sample type in the dataset, despite the fact that AR messenger RNA levels are actually elevated in mCRPC compared to primary tumors. The tumor has amplified the gene but lost the ability to use it for differentiation.

AR-A tells the opposite story. It peaks in primary tumors and reaches its nadir in mCRPC. The correlation between the two scores, positive at r = 0.32 in primary disease, flips to a negative r = -0.15 (p = 0.005) in mCRPC. This is not a subtle statistical signal. It represents a fundamental rewiring of the relationship between differentiation and stemness as the disease evolves.

The pattern holds across independent data sources. In the Decipher GRID database of 82,470 prospectively collected biopsy samples, Stemness increases monotonically across NCCN risk groups from very low to very high risk. In the Taylor 2010 microarray dataset, metastatic prostate cancer carries the highest Stemness of any clinical stage. In genetically engineered mouse models, the most aggressive triple-knockout tumors (lacking Pten, Rb1, and Trp53) show the highest Stemness and lowest AR-A, while the indolent single-knockout tumors show the reverse.

One finding warrants particular attention. Tumor-adjacent benign tissues, which are routinely used as 'normal' controls in cancer studies, carry substantially higher Stemness scores than truly normal prostate tissue from cancer-free organ donors in the GTEx database. These tissues are not normal. Their transcriptomes have shifted toward a cancer-like state, a finding consistent with prior observations that field cancerization extends beyond the visible tumor margin.

Predicting Who Dies

A biological score is only clinically useful if it predicts outcomes. High Stemness is consistently associated with worse overall survival across every cohort where survival data were available. In the Spratt 2017 cohort of 855 radical prostatectomy specimens, the log-rank p-value is 0.001. In the Tosoian 2020 cohort of 405 high-risk localized tumors, it is 0.014. In the CHAARTED 2021 cohort of 160 metastatic hormone-sensitive tumors, it is 0.006.

The 12-gene PCa-Stem signature performs even more sharply. In primary prostate cancer, high signature scores predict worse progression-free survival with a multivariable-adjusted hazard ratio of 1.636 (95% CI 1.014 to 2.64), independent of age, Gleason score, and stage. In mCRPC, the effect is larger: patients with high PCa-Stem signature scores have a multivariable-adjusted hazard ratio for overall survival of 4.712 (95% CI 2.011 to 11.04, p less than 0.001). That is a nearly fivefold difference in the risk of death, after accounting for standard clinical variables.

The 12 genes in the signature are predominantly involved in mitosis and cell-cycle progression, consistent with the finding that Stemness-high tumors are enriched for a 31-gene cell-cycle progression signature. Eleven of the 12 genes show statistically significant correlation with disease progression scores across the Bolis 2021 transcriptome atlas. The Human Protein Atlas independently identifies these same 11 genes as unfavorable prognostic markers in prostate cancer and several other malignancies, raising the possibility that the PCa-Stem signature may have prognostic utility beyond prostate cancer.

The Genomic Drivers of Stemness

High Stemness is not merely a transcriptional state. It is accompanied by measurable genomic instability. Stemness-high tumors carry a greater fraction of altered genome and a higher tumor mutation burden than Stemness-low tumors at both the primary and metastatic stages, despite no difference in patient age between the groups.

The specific genomic events associated with high Stemness differ by disease stage. In primary prostate cancer, MYC amplification is the most prominent genomic correlate of high Stemness. The prevalence of MYC amplification rises from 2% in low-grade (Gleason 6) primary tumors to 12% in high-grade (Gleason 9/10) tumors (p = 0.006). Stemness-high primary tumors are also enriched for PTEN deletion and FOXA1 mutation. In mCRPC, RB1 deletion is the genomic alteration most significantly associated with elevated Stemness. AR amplification, which occurs in 49% of mCRPC cases, is exclusively treatment-induced and absent from treatment-naive primary tumors.

The MYC story is developed in detail in the final section of the paper. MYC messenger RNA and MYC signaling activity both increase in primary tumors compared to matched benign tissue, continue to rise with Gleason grade, and reach their highest levels in mCRPC. Critically, neoadjuvant androgen deprivation therapy, which reduces both AR-A and Stemness in primary tumors, has no discernible effect on MYC activity. The Pearson correlation between MYC signaling activity and Stemness in mCRPC is 0.76, a tight positive relationship that holds across every stage of disease examined.

A Two-Phase Model of Stemness

The data support a two-phase model of how stemness is acquired and maintained during prostate cancer progression. In the first phase, covering early tumor initiation and low-grade primary disease, AR signaling drives both differentiation and stemness simultaneously. The two are positively correlated. Neoadjuvant hormone therapy, by suppressing AR-A, reduces Stemness as a consequence. Loss of PTEN and RB1, present in 9% and 13% of Gleason 6 tumors respectively, may contribute additional stemness at this stage.

In the second phase, covering high-grade primary tumors and all forms of advanced disease, the coupling between AR-A and Stemness breaks down. AR-A declines as tumors lose differentiated glandular architecture. Stemness continues to rise, now driven by accumulating genomic alterations, particularly MYC amplification and increased MYC signaling activity. Because MYC activity is insensitive to androgen deprivation, this second-phase stemness persists and grows even as AR-targeted therapies suppress canonical AR signaling. By the time a tumor reaches the mCRPC state, it has achieved maximum stemness through a mechanism that is largely independent of the pathway that all current standard-of-care therapies target.

What the Score Does Not Capture

The study's primary methodological limitation is one the authors acknowledge directly: the Stemness Index was not developed for prostate cancer. It was derived from a pan-cancer analysis using embryonic stem cell reference signatures, and its sensitivity for distinguishing subtle differences within a single disease stage has not been fully validated. Whether it can reliably separate, say, Gleason 7 favorable-intermediate from Gleason 7 unfavorable-intermediate tumors is an open question.

The bulk RNA-sequencing data used throughout the analysis presents a second constraint. A bulk transcriptome reflects the average gene expression of all cells in a biopsy, including stromal fibroblasts, immune infiltrates, and endothelial cells alongside the cancer epithelium. A high Stemness score in a bulk sample could reflect genuine dedifferentiation of cancer cells, or it could partly reflect a shift in the cellular composition of the tumor microenvironment toward more stromal or immune cell types that happen to express stem-like genes. Single-cell RNA-sequencing analysis, or the application of alternative stemness pipelines such as CytoTRACE that are designed to work at single-cell resolution, would help disentangle these contributions.

The role of noncanonical AR signaling in mCRPC also remains incompletely addressed. AR amplification occurs in 49% of mCRPC cases and AR mutation in 11%, both exclusively treatment-induced. These alterations drive castration-resistant AR activity through mechanisms distinct from the canonical differentiation program measured by the AR-A score. Whether this noncanonical AR activity contributes to the high Stemness of mCRPC, or whether it is orthogonal to it, is a question the current analysis cannot answer. Targeted inhibitor studies or AR-specific perturbation experiments in mCRPC models would be needed to resolve this.

A Ruler for Aggressiveness

The practical value of this work lies in its potential to stratify patients more precisely than current tools allow. The Gleason score and PSA level, the two workhorses of prostate cancer risk assessment, capture different aspects of tumor biology but neither directly measures the stem-like state that appears to drive lethality. A Stemness score, or the 12-gene PCa-Stem signature derived here, could in principle be applied to any tumor biopsy with gene expression data, including the microarray-based Decipher Genomic Classifier already in clinical use for 82,000 patients in the GRID database.

The mechanistic findings point toward a specific therapeutic vulnerability. If MYC amplification and signaling are the primary drivers of stemness in advanced prostate cancer, then MYC-targeting strategies, which have historically been difficult to develop but are now an active area of drug discovery, could in principle reduce stemness and restore sensitivity to AR-targeted therapies. The observation that neoadjuvant hormone therapy reduces Stemness in primary tumors but leaves MYC activity untouched suggests that combining AR pathway inhibition with MYC inhibition early in the disease course, before the second-phase genomic drivers take hold, might prevent the emergence of the high-stemness mCRPC state altogether.

That hypothesis remains to be tested. What this study provides is the quantitative framework to ask the question rigorously, and a set of validated scores that can track the answer across the full arc of disease.