Biochemical recurrence (BCR) refers to a rise in prostate-specific antigen (PSA) levels after a patient has undergone treatment for prostate cancer -- a sign that cancer cells may still be active despite apparent treatment success. Approximately 20-30% of prostate cancer patients will experience BCR after radical prostatectomy or radiation therapy.
The Gleason score, the standard system for grading prostate cancer aggressiveness based on the microscopic appearance of cancer cells, was designed to assess the initial cancer before treatment. It is less informative for predicting which treated patients will develop BCR, creating a critical surveillance gap.
RNA sequencing technologies have enabled researchers to study gene expression patterns across thousands of genes simultaneously, creating the possibility of identifying specific gene signatures -- small sets of genes whose expression levels together predict the likelihood of BCR. However, most published gene signatures have been too large, too complex, or too unreliable for routine clinical use.
This study introduces a machine learning pipeline based on data discretization -- a mathematical technique that converts continuous gene expression values into categorical intervals -- to identify a compact, interpretable 10-gene signature that predicts BCR within 24 months of completing treatment.
Discretization is a data preprocessing technique that converts continuous numerical values into a set of discrete categories called bins or intervals. For gene expression data, instead of using the exact measured RNA level (a continuous value), discretization assigns each measurement to a range -- for example, 'low' (0-0.2), 'medium-low' (0.2-0.4), and so on.
This transformation helps machine learning models in several ways. It reduces noise in genomic data by grouping similar expression values together rather than treating every small measurement difference as meaningful. It also reduces the degrees of freedom in the dataset, which helps prevent models from memorizing training data rather than learning general patterns.
A key advantage specific to this application is that discretization allows identifying not just which genes predict recurrence, but the specific expression range within which each gene's effect is most important. This gives each gene a defined 'zone of action' -- for example, a gene may only be predictive of recurrence when its expression falls between 0.2 and 0.4 but not at other levels.
Two types of discretization were compared: equal width binning (dividing the expression range into intervals of equal size) and equal frequency binning (dividing into intervals containing equal numbers of data points). These were tested across 8 different rules for determining the optimal number of intervals.
The study used publicly available RNA sequencing data from four prostate cancer datasets. The training cohort came from TCGA (The Cancer Genome Atlas) with 143 patients balanced between recurrent and non-recurrent cases. Two independent testing datasets -- from Memorial Sloan Kettering Cancer Center (MSKCC) and GSE54460 -- were used for model testing. A fourth dataset (GSE70769) containing 93 patients with PSA levels served as the final validation set.
After log2 normalization of the 20,000+ gene expression values and removing low-variance genes, 7,800 genes remained for analysis. Discretization was then applied using different binning strategies. The Correlation-based Feature Selection (CFS) algorithm ranked genes by their correlation with BCR status, retaining only those with a merit score above 0.60 to prevent overfitting with too many variables.
The final classification model was deliberately chosen to be logistic regression rather than more complex algorithms. While neural networks or decision trees could potentially achieve higher accuracy, logistic regression produces a directly interpretable equation showing the weight of each gene's contribution -- allowing clinicians to understand why the model predicts recurrence in any given patient.
The dataset imbalance problem -- many more non-recurrent than recurrent patients in public datasets -- was addressed by selecting a balanced training set with equal numbers of BCR and non-BCR patients matched for similar follow-up time, preventing the model from being biased toward predicting no recurrence by default.
The optimal model used five-bin equal width discretization (based on the Cencov rule for estimating bin count) and identified a signature of exactly 10 genes. This model achieved 82% accuracy on the testing datasets (MSKCC and GSE54460), correctly classifying 47 out of 51 BCR-positive patients in GSE54460. Performance held at 75% accuracy on the independent validation dataset GSE70769.
Equal width discretization consistently outperformed equal frequency discretization across all bin-count rules. Notably, it also outperformed no-discretization at all -- the same logistic regression approach without the discretization preprocessing step achieved only 75% accuracy with 19 genes, confirming that discretization both improves performance and reduces the number of variables needed.
The 10 identified genes were: ZFHX3, EMP2, ITPR1, NFIB, PCCA, RGS2, WDR5, SRGAP2, AIDA, and SLC25A30. In the gene signature equation, ZFHX3, NFIB, PCCA, and WDR5 are predominantly over-expressed in patients without BCR (protective), while reduced expression of SLC25A30 and ITPR1 and elevated AIDA are associated with recurrence risk.
To predict the timing of recurrence (BCR within 24 months), PSA levels were added to the gene signature model. This combined model achieved a precision of 71% for the testing dataset and 69% for the validation dataset in predicting early BCR, with a PSA level below 0.38 ng/mL identified as a key additional risk indicator for recurrence within two years.
A striking biological observation from the study is that the 10 identified genes share a common functional theme: most are directly or indirectly related to the extracellular matrix (ECM) -- the structural scaffold of proteins and molecules that surrounds cells in tissue, providing organization to the prostate gland and regulating cell behavior.
ZFHX3 acts as a tumor suppressor in prostate cancer -- its reduced expression disrupts the organization of prostate glands and the smooth muscle layer separating the cell layers. Greater ZFHX3 levels correlate with better survival in published studies. WDR5 interacts with androgen signaling to accelerate prostate cancer cell proliferation and is expressed at higher levels in cancer than in healthy prostate tissue.
ITPR1 is involved in controlling programmed cell death (apoptosis). Its depletion impairs apoptotic control, allowing cancer cells to survive longer -- a mechanism previously documented in colorectal and ovarian cancers that appears to apply to prostate cancer as well. NFIB is implicated in prostate gland organization, with its reduced expression associated with gland enlargement and ECM disruption.
The convergence of multiple independent studies on ECM-related genes in BCR prediction is biologically significant. The ECM provides structural context for prostate glands, regulates cell migration, and changes in ECM composition have been directly linked to tumor cell invasion and metastasis -- connecting recurrence risk to the same processes that drive cancer spread.
A key advantage of this study's approach is that the final 10-gene logistic regression model produces an explicit, transparent prediction equation that shows the contribution and expression range of each gene. This interpretability is clinically essential -- unlike black-box neural networks, clinicians can directly evaluate why the model assigns a particular recurrence risk to a given patient's gene expression profile.
The study acknowledges that the number and identity of genes in the signature depends on the discretization approach and bin number chosen. Too few bins oversimplify gene relationships and may miss important distinctions; too many bins add noise. The five-bin equal width approach struck the optimal balance for this dataset.
An important limitation is the relatively modest validation dataset size (93 patients) and the fact that all datasets were from existing public repositories rather than prospectively collected cohorts with standardized treatment protocols. BCR prediction is inherently complex because recurrence patterns differ substantially between patients treated with surgery versus radiation versus hormone therapy.
The authors note that the discretization methodology used here is directly applicable to other cancer types. The approach of converting continuous RNA expression data into interpretable interval-based signatures could be extended to any cancer where biochemical or molecular recurrence prediction is clinically needed -- making the methodology itself a transferable contribution beyond prostate cancer specifically.