Chemotherapy benefit is modest and toxic. Cisplatin-based neoadjuvant chemotherapy (NAC) is the standard of care for muscle-invasive bladder cancer (MIBC), but the overall pathologic response rate is only approximately 40%, meaning the majority of patients undergo intensive chemotherapy with little or no benefit while delaying potentially curative surgery.
Existing biomarker candidates for predicting NAC response include molecular subtypes, DNA damage repair deficiency, and mutations in genes such as ERCC2, ERBB2, and FGFR3. The SWOG S1314 trial was specifically designed to evaluate the COXEN tumor tissue gene-expression score, but it did not confirm treatment-specific predictive ability, underscoring the continuing unmet clinical need.
Cell-free DNA as a liquid biopsy platform. Cell-free DNA (cfDNA) circulating in plasma offers a minimally invasive source of cancer-related molecular information, including somatic mutations and epigenetic data. Prior work has linked cfDNA methylation to cancer detection and molecular subtyping, motivating its investigation as a predictor of chemotherapy response.
DNA methylation is tissue-specific and reflects both tumor cell and host cell contributions in plasma, potentially capturing the composite biology of tumor behavior and immune response. This makes cfDNA methylation a uniquely comprehensive analyte compared to tumor tissue profiling alone.
SWOG S1314 clinical framework. The study was an exploratory analysis nested within SWOG S1314, a prospective phase 2 randomized trial comparing two cisplatin-based regimens (gemcitabine plus cisplatin, and dose-dense MVAC) in MIBC patients with clinical stage cT2-T4aN0M0 disease and at least 5 mm of viable tumor.
Plasma cfDNA was collected at two time points: before initiation of NAC (baseline) and on cycle 2 day 1 (after the first chemotherapy cycle). A total of 72 evaluable patients with both plasma and cystectomy pathology data were included, with pathologic response defined as achieving pT1N0 or better at radical cystectomy.
Infinium MethylationEPIC array. cfDNA methylation was measured using the Illumina MethylationEPIC BeadChip array, which covers over 850,000 CpG sites. This was the first report using this platform specifically for cfDNA (rather than tumor tissue), with 20 ng cfDNA validated as the optimal loading amount for reliable array performance.
Differentially methylated loci (DMLs) between responders and nonresponders were identified, and a Random Forest machine learning classifier was trained using a resampling procedure with leave-one-out cross-validation. An Elastic Net classifier was also built as a sensitivity analysis to confirm robustness of findings.
Circulating bladder DNA fraction was estimated using tissue-specific methylation reference profiles via non-negative least squares regression, providing a surrogate measure of tumor-derived cfDNA without requiring deep mutational sequencing.
No single locus is predictive. Comparing cfDNA methylation between responders and nonresponders identified 23,799 differentially methylated loci at a p value cutoff of 0.05, but none remained significant after Benjamini-Hochberg multiple testing correction, indicating that no individual methylation site carries sufficient predictive signal on its own.
DMLs less methylated in nonresponders were preferentially located on CpG islands and clustered into differentially methylated regions. Using the top 500 such loci, t-SNE visualization showed clear spatial separation between responders and nonresponders, supporting the feasibility of a composite biomarker approach.
mR-score classifier. The Random Forest model produced a methylation-based response score (mR-score) with an ROC area under the curve (AUC) of 0.636 (95% CI 0.498-0.773) for distinguishing responders from nonresponders. Importantly, median mR-scores showed a progressive gradient: lowest in complete responders, intermediate in partial responders, and highest in nonresponders.
An Elastic Net classifier produced a nearly identical AUC of 0.639, and the two methods showed strong correlation (Spearman coefficient 0.77), confirming the robustness of the methylation-based predictive signal regardless of algorithm choice. Similar predictive performance was observed for both chemotherapy regimens tested in S1314.
Early on-treatment assessment. Among the 72 patients, 57 also provided plasma samples after the first cycle of NAC on cycle 2 day 1. The pretreatment Random Forest model was applied to these on-treatment cfDNA methylation profiles to generate an on-treatment mR-score, enabling assessment of whether treatment-induced methylation changes added predictive value.
On-treatment mR-scores performed slightly better than baseline scores, achieving an AUC of 0.720 (95% CI 0.582-0.857), suggesting that one cycle of chemotherapy may further sharpen the methylation-based differentiation between responders and nonresponders.
Stability of methylation signal. A significant positive correlation was found between pretreatment and on-treatment mR-scores (Spearman coefficient 0.39, p = 0.003), indicating that the underlying methylation landscape driving the predictive signal is stable across treatment time points, supporting clinical utility at multiple sampling windows.
The consistency of predictive performance across both chemotherapy regimens and both time points supports the biological interpretation that cfDNA methylation reflects intrinsic patient and tumor characteristics rather than being dependent on treatment-specific effects.
Estimating tumor-derived cfDNA without sequencing. Using tissue-specific methylation reference profiles, the fraction of circulating bladder DNA in total cfDNA was estimated as a surrogate for circulating tumor DNA. Complete responders had the lowest circulating bladder DNA fractions, followed by partial responders, with nonresponders showing the highest fractions.
As a standalone biomarker, circulating bladder DNA fraction yielded an AUC of 0.600 for predicting response, and an AUC of 0.616 for predicting complete response. Of responders, 84% had circulating bladder DNA below the threshold of 1.11%, compared with only 51% of nonresponders, giving this marker high sensitivity for identifying likely responders.
Independence from mR-score. Critically, circulating bladder DNA fraction and pretreatment mR-score showed no linear correlation (Spearman coefficient 0.073), confirming that these two metrics capture complementary and non-redundant aspects of bladder cancer biology in plasma and can be combined for additive predictive power.
Healthy donor samples had negligible circulating bladder DNA, validating the specificity of this methylation-based tissue-of-origin estimation approach for detecting tumor-associated signal in cancer patients.
Two-step risk stratification. A combined risk-stratification model was developed that first categorized patients using both the pretreatment mR-score and the circulating bladder DNA fraction, then applied the on-treatment mR-score to further stratify patients who fell into an intermediate-risk category after the initial assessment.
Using Youden's index as the cutoff criterion for all three measures, the combined model correctly classified 45 of 57 patients (79%) who had both pretreatment and on-treatment cfDNA samples, representing a meaningful improvement over either biomarker used alone.
Clinical impact estimation. In the proof-of-concept cohort, this approach would have identified 17 patients as unlikely to benefit from NAC before or after the first cycle (10 at baseline and 7 after cycle 1), at the cost of incorrectly withholding beneficial treatment from only 3 patients. For the remaining 17 patients identified as non-responders, upfront surgery or enrollment in clinical trials could have been offered instead of toxic chemotherapy.
The sequential design of the model is clinically rational: the majority of patients can be classified at baseline without needing to wait for on-treatment samples, while the on-treatment mR-score adds value specifically for those in the intermediate-risk zone where the baseline data are insufficient for a confident prediction.
Unique properties of methylation as an analyte. DNA methylation degrades slowly, is functionally meaningful, and reflects dynamic shifts in transcriptional regulation. Unlike somatic mutations, it captures both tumor cell and host cell (including immune cell) contributions to plasma cfDNA, potentially encoding information about immune response to chemotherapy in addition to tumor biology.
The MethylationEPIC array approach offers time efficiency, cost effectiveness, and a mature analysis workflow compared to bisulfite sequencing or methylation immunoprecipitation sequencing methods, enabling high-throughput profiling of clinical samples with high reproducibility at scale.
The mR-score likely captures host factors. Given that MIBC patients undergoing cystectomy have limited disease burden and low circulating bladder DNA fractions (0-3%), and that the mR-score did not correlate with circulating bladder DNA, it is likely that the majority of the DMLs driving the mR-score reflect host factors, particularly peripheral blood leukocyte methylation patterns related to immune status, rather than tumor-derived signal.
Important limitations include the small sample size requiring validation in larger prospective cohorts, the relatively low tumor-derived DNA fraction in locally advanced disease (compared to up to 80% in metastatic patients), and the potential confounding effect of transurethral resection performed before sample collection, which may have already eliminated residual disease in some patients.
The leukocyte-derived component of cfDNA methylation may actually represent a future strength of this approach, as immune cell methylation patterns could prove predictive of response to immunotherapy, which has become a standard treatment in metastatic urothelial carcinoma.
First use of MethylationEPIC array for cfDNA. This study establishes the feasibility of using the Infinium MethylationEPIC array to profile cfDNA methylation at scale in a multicenter cooperative group trial setting, demonstrating that high-quality methylation data can be obtained from clinical plasma samples.
The combination of the mR-score and circulating bladder DNA fraction into a risk-stratification model represents a novel liquid biopsy framework that correctly predicts NAC response in approximately 80% of patients, potentially enabling clinical decision-making about whether to proceed with NAC, go directly to surgery, or enroll in a clinical trial.
Broader applicability. The cfDNA methylation approach could extend beyond locally advanced disease to metastatic bladder cancer, where higher tumor DNA fractions would provide stronger tumor-specific signal, and where the immune cell methylation component might also predict response to checkpoint inhibitor immunotherapy.
Validation in independent prospective cohorts remains the critical next step, with the potential for this minimally invasive, blood-based biomarker approach to guide treatment selection and spare non-responders from the toxicity of ineffective chemotherapy.