Urine cfDNA EM-seq Ensemble Detects Bladder Cancer (AUC 0.932)

Investigators describe a urine cfDNA EM-seq ensemble assay that pairs whole-genome enzymatic methylation sequencing (EM-seq) with an XGBoost-based approach to classify bladder cancer from urine samples, using patients with benign disease (no history of malignant disease) as controls without cancer.

In the authors’ independent test-set evaluation, the final ensemble operating point was reported at 91.9% sensitivity with 80% specificity, alongside an AUC of 0.932 for bladder cancer detection and an AUC of 0.928 for non–muscle invasive bladder cancer (NMIBC). The report then outlines cohort assembly, methylation and copy-number feature generation and combination, and comparisons of urine results with plasma and mutation-based testing within the same dataset.

The authors describe profiling 143 urine samples in total, comprising 68 urothelial bladder cancer cases and 75 controls with benign disease (treated as healthy controls without cancer in the study). For marker selection, they compared 14 bladder cancer tissue samples with urine from 14 healthy individuals to define 113,052 methylation marker regions. Remaining urine samples were allocated to model development and evaluation, with 31 cancer and 41 control urines used for training/validation and 37 cancer and 20 control urines used as an independent test set, as reported. In this workflow, marker discovery from tissue–healthy urine comparisons preceded model fitting and performance assessment in a held-out test set.

For the assay workflow, the authors report extracting urinary cfDNA from urine supernatants (with urine collected pre-operatively and processed after sequential centrifugation to remove cellular material), then preparing whole-genome EM-seq libraries using a commercial enzymatic conversion kit and sequencing on an Illumina platform with paired-end reads. They describe read processing with adapter trimming, alignment to hg19, duplicate marking, and methylation-context extraction, and they report excluding samples with cytosine conversion efficiency below 97%.

For methylation quantification, they used 300 bp genomic windows (removing ENCODE blacklist regions) and applied coverage filtering before computing average methylation fraction; for CNV, they used ichorCNA on 1 Mbp bins and also used its tumor-fraction estimates as features. Modeling was organized as separate XGBoost models for methylation markers and CNV profiles, followed by an ensemble that evaluated combinations of candidate features and carried forward a final score for comparative analyses reported in the Results.

In matched biofluid analyses, the authors report EM-seq data from 41 matched urine–plasma pairs collected from bladder cancer patients to compare signal strength across sample types. In that paired comparison, urine showed higher estimated tumor fractions than matched plasma, and urine also showed higher concordance with bladder tissue methylation profiles when cosine similarity was computed against an average tissue reference profile. They also describe stage-linked changes in urine global methylation in contrast to more stable plasma profiles across the cancer groups shown. Within their dataset, these observations were presented as evidence that urine captured stronger and more tissue-representative tumor-derived signal than plasma.

Against mutation-based testing, the paper reports that the independent test set was restricted to samples with confirmed mutation status to enable direct comparison with a prior urine mutation panel approach. In that setting, the authors report the ensemble identifying 4 of 7 cancers that were mutation-negative by the comparator assay. They also present a decision curve analysis comparing the ensemble alone, mutation testing alone, and a combined “either-positive” strategy, describing higher net benefit for the combined approach across a broad range of threshold probabilities in their analysis.

In conclusion, the authors state that these results suggest complementary roles for methylation-based ensemble modeling and mutation detection.

Key Takeaways:

• In the independent test set, the authors report 91.9% sensitivity at 80% specificity (AUC 0.932 overall; AUC 0.928 for NMIBC detection).
• In matched pairs, urine was reported to carry higher tumor-fraction estimates and closer tissue-methylation similarity than plasma within the same patients.
• The study reports detection of 4 of 7 mutation-negative cases and reports higher net benefit in decision-curve analysis when an either-positive strategy (mutation or ensemble positive) was used.