Hydroxymethylation profile of cell-free DNA is a biomarker for early colorectal cancer

Abstract

Early detection of cancer will improve survival rates. The blood biomarker 5-hydroxymethylcytosine has been shown to discriminate cancer. In a large covariate-controlled study of over two thousand individual blood samples, we created, tested and explored the properties of a 5-hydroxymethylcytosine-based classifier to detect colorectal cancer (CRC). In an independent validation sample set, the classifier discriminated CRC samples from controls with an area under the receiver operating characteristic curve (AUC) of 90% (95% CI [87, 93]). Sensitivity was 55% at 95% specificity. Performance was similar for early stage 1 (AUC 89%; 95% CI [83, 94]) and late stage 4 CRC (AUC 94%; 95% CI [89, 98]). The classifier could detect CRC even when the proportion of tumor DNA in blood was undetectable by other methods. Expanding the classifier to include information about cell-free DNA fragment size and abundance across the genome led to gains in sensitivity (63% at 95% specificity), with similar overall performance (AUC 91%; 95% CI [89, 94]). We confirm that 5-hydroxymethylcytosine can be used to detect CRC, even in early-stage disease. Therefore, the inclusion of 5-hydroxymethylcytosine in multianalyte testing could improve sensitivity for the detection of early-stage cancer.

Figure 1

Flow chart of subjects included in the study. Control samples were made up of individuals who were CRC and adenomatous polyp negative (colonoscopy confirmed). A total of 8.3% of the control individuals were diagnosed with peptic ulcers, arthritis or COPD.

Figure 2

(A) Hydroxymethylome capture procedure. (B) 166 bp synthetic spike-in controls with 1, 3, 6 5hmC residues demonstrate that the hydroxymethylome enriches for 5hmC over controls containing 6 5mC residues and unmodified cytosines.

Figure 3

A two-dimensional representation of 5hmC quantified within gene enhancers over the training and validation samples displays evidence of clustering by disease status (CRC = green, control = orange), with little bias for gender (male = green, female = orange) or age (45–59, 60–69, 70–85 years) (t-SNE parameters: perplexity = 20, theta = 0.5).

Figure 4

(A) Classifier trained on 5hmC levels in enhancer regions shows equivalent performance on the training (dotted line) and validation sets (solid line) and high performance across all CRC stages versus controls, with AUCs ranging from 88.6 to 93.6%. Cut-of-values are reported in Supp. Table 1. (B) The IchorCNA tumor fraction was positively correlated with tumor stage in validation samples. The correlation with the CRC classifier score was lower in early-stage samples (stages 1 and 2), with higher p-values than in later-stage samples. (C) CRC classifier score on validation samples with ichorCNA values ≤ 3% tumor fraction, demonstrating that the 5hmC-based classifier maintains robust performance on samples with a low tumor fraction. The corresponding table presents the percentage of samples on either side of the classification threshold (0.5), demonstrating that the classifier performs similarly across CRC stages. (D) The top significant biological pathways identified that relate to the pathway relationship of genes regulated by enhancer features in the 5hmC classifier indicate a global immune response to tumorigenesis.

Figure 5

(A) Classifier trained using a DELFI-like approach demonstrates CRC stage-dependent performance versus controls in validation samples. Cut-of-values are reported in Supp. Table 1. (B) The classifier prediction probability shows strong concordance with the estimated tumor fraction (ichorCNA), particularly in late stages in both training and validation samples.

Figure 6

(A) Overall, the median performance estimate was higher for the 5hmC classifier than for the DELFI and NPS classifiers. (B–D) Median AUC and sensitivity at 95% specificity of 5hmC, DELFI-like fragmentomics approach and combined classifier. The 5hmC classifier performs better than the DELFI-like fragmentomics classifier in early CRC stages (1 & 2), while in late stages (3 & 4) 5hmC shows significant additivity at higher specificity (95% specificity).

Figure 7

(A, B) A classifier trained on 5hmC levels in enhancer regions maintains performance at early-stage cancer(CRC Stage vs AUC: Spearman’s rho = 0.80, p = 0.333) compared to a model trained on cfDNA fragment size and coverage (DELFI-like approach) (CRC Stage vs AUC: Spearman’s rho = 0.95, p = 0.05). (C) The 5hmC-based classifier performs comparably to reported classifiers for stage 1 CRC. To gain approximately comparable confidence intervals, 95% binomial confidence intervals were computed for all classifiers using publicly available information^,,,,. The CRC classifier from Putcha et al. contains both Stage 1 and Stage 2 samples.