Machine Learning–Based Classification of Cancer Using Promoter-Associated 5-Hydroxymethylcytosine Signatures in Cell-Free DNA

Technical Summary: Machine Learning–Based Classification of Cancer Using Promoter-Associated 5-Hydroxymethylcytosine Signatures in Cell-Free DNA

Problem Statement
Cancer remains a leading cause of global mortality, with early detection being critical for patient survival. Conventional diagnostic modalities, including radiographic imaging and tissue biopsies, are often invasive, costly, and lack sensitivity for early-stage lesions. While circulating tumor DNA (ctDNA) analysis via liquid biopsy offers a non-invasive alternative, mutation-based approaches are limited by low variant allele frequencies in early disease. Epigenetic modifications, specifically 5-hydroxymethylcytosine (5hmC), offer a promising complementary signal. 5hmC is a dynamic, tissue-specific epigenetic mark associated with active transcription and gene regulation. However, the high dimensionality and complexity of genome-wide 5hmC datasets pose significant analytical challenges for conventional statistical methods. There is a need for reproducible, machine learning-based frameworks that can integrate epigenetic feature engineering with predictive modeling to distinguish cancer from healthy states across multiple tumor types (pan-cancer) using cell-free DNA (cfDNA).

Methodology
The study utilized a unified cohort of 804 cfDNA samples (458 cancer, 346 healthy controls) sourced from the Cell-Free Epigenome Atlas (CFEA), covering seven tumor types (esophageal, colon, lung, gastric, thyroid, pancreatic, and liver). All samples were profiled using 5hmC-Seal-seq.

1. Data Preprocessing and Rasterization: Raw BED files were converted into dense, one-dimensional signal arrays at a 1 kb resolution across the 23 standard chromosomes (hg19). Signals were normalized per chromosome to a [0, 1] range.
2. Feature Extraction: Instead of random genomic windows, features were anchored to Transcription Start Sites (TSS). Using Ensembl GRCh37 annotations, a representative TSS was selected for each gene. A fixed window of ±5 kb (11 bins of 1 kb) was centered on each TSS, and the mean normalized 5hmC signal was calculated, resulting in an initial feature matrix of 804 samples × 55,246 TSS-associated features.
3. Feature Spaces: Three distinct feature spaces were constructed and evaluated:
   • Baseline Space: The top 2,000 features with the highest cross-sample variance (label-independent), without biological restrictions.
   • Technical-Cleanup Space: The top 2,000 variance-selected features after excluding TSSs overlapping genomic blacklists, structural RNAs, or LOC-designated features.
   • Cancer-Gated Space: A restricted set of 605 features derived from the technical cleanup, limited only to genes present in established oncogenic signaling pathways (Sanchez-Vega et al.) and the MSK-IMPACT468 clinical panel.
4. Dimensionality Reduction and Modeling:
   • Random Forest (RF): Trained on the full 2,000-dimensional standardized matrix (baseline) and the 605-dimensional cancer-gated matrix.
   • Gradient Boosting (GB) and Logistic Regression (LR): Trained on a 100-component PCA-reduced representation (explaining 61.3% of variance) to handle correlated features.
5. Evaluation: Models were assessed using stratified five-fold cross-validation and an independent hold-out test set (201 samples). Performance metrics included ROC-AUC, accuracy, precision, recall, F1-score, and confusion matrix analysis. Feature importance was analyzed to identify biologically relevant loci.

Key Contributions

• Pan-Cancer Framework: The study demonstrates a unified analytical pipeline capable of distinguishing cancer from healthy controls across seven diverse tumor types using a single feature extraction strategy (TSS-centered).
• Comparative Model Evaluation: It provides a controlled comparison of Random Forest, Gradient Boosting, and Logistic Regression on the same curated cfDNA 5hmC dataset, moving beyond single-algorithm studies.
• Biological Interpretability: The study integrates model interpretability as a primary output, linking predictive features to specific biological pathways (e.g., PI3K–AKT–mTOR, RAS/MAPK, TGF-β, Wnt, MYC) and demonstrating that strong predictive signals exist both within and outside canonical cancer gene sets.
• Reproducible Workflow: The author provides a fully documented, end-to-end computational pipeline from raw BED files to classifier evaluation, addressing reproducibility gaps in the field.

Results

• Classification Performance: All three models demonstrated strong discriminative ability. Random Forest achieved the highest average cross-validation ROC-AUC (0.869 ± 0.025) and the highest hold-out ROC-AUC (0.884) with 77.6% accuracy. Gradient Boosting showed the highest sensitivity (87.7% on hold-out) but lower specificity, while Logistic Regression showed the highest specificity (75.9%) but lower sensitivity.
• Feature Space Analysis: Restricting the feature space to established cancer genes (Cancer-Gated) resulted in only a marginal performance drop (ROC-AUC 0.844 vs. 0.869 for baseline), indicating that while canonical genes are informative, the broader epigenomic landscape contains significant predictive signal.
• Feature Importance: In the baseline model, the top 20 features included non-coding RNAs and pseudogenes, none of which were in the predefined cancer-gene set. However, in the cancer-gated analysis, top features included NEGR1, RHEB, TAOK2, RASA3, SPRED1, SMAD2, AKT2, FZD8, and MYCL.
• Biological Correlation: The highest-ranked cancer-gated features showed enriched 5hmC coverage in cancer samples compared to healthy controls (e.g., NEGR1 showed a 1.62× enrichment). These genes are involved in critical oncogenic pathways, including cell adhesion, mTOR regulation, RAS/MAPK signaling, TGF-β signaling, Wnt signaling, and MYC transcriptional regulation.
• Robustness: Sensitivity analyses removing blood-lineage marker genes or dominant axes of variation from healthy controls showed that the classification signal persisted, suggesting the predictive power is not solely an artifact of blood-cell composition.

Significance and Claims
The paper claims that promoter-associated 5hmC signatures in cfDNA contain sufficient multidimensional information to support machine learning-based pan-cancer classification. The study establishes that:

1. Feasibility: TSS-centered feature engineering combined with supervised learning can effectively distinguish cancer from healthy samples across heterogeneous tumor types.
2. Signal Distribution: The discriminatory signal is distributed across coordinated promoter-associated 5hmC patterns involving multiple regulatory pathways, rather than being confined to a single universal biomarker or a narrow set of canonical cancer genes.
3. Biological Grounding: The predictive features align with known oncogenic biology, providing a biologically interpretable framework for liquid biopsy diagnostics.
4. Clinical Context: The author explicitly states that these findings represent a "proof-of-concept" and demonstrate "classification feasibility." They do not claim the model is ready for clinical diagnosis, noting that false positives and negatives remain and that further validation in independent, prospective cohorts is required before clinical translation. The work serves as a foundation for developing scalable, non-invasive approaches for future precision oncology applications.