Non-invasive diagnosis of early-stage lung cancer using high-throughput targeted DNA methylation sequencing of circulating tumor DNA (ctDNA)

Abstract

Rational: LDCT screening can identify early-stage lung cancers yet introduces excessive false positives and it remains a great challenge to differentiate malignant tumors from benign solitary pulmonary nodules, which calls for better non-invasive diagnostic tools. Methods: We performed DNA methylation profiling by high throughput DNA bisulfite sequencing in tissue samples (nodule size < 3 cm in diameter) to learn methylation patterns that differentiate cancerous tumors from benign lesions. Then we filtered out methylation patterns exhibiting high background in circulating tumor DNA (ctDNA) and built an assay for plasma sample classification. Results: We first performed methylation profiling of 230 tissue samples to learn cancer-specific methylation patterns which achieved a sensitivity of 92.7% (88.3% - 97.1%) and a specificity of 92.8% (89.3% - 96.3%). These tissue-derived DNA methylation markers were further filtered using a training set of 66 plasma samples and 9 markers were selected to build a diagnostic prediction model. From an independent validation set of additional 66 plasma samples, this model obtained a sensitivity of 79.5% (63.5% - 90.7%) and a specificity of 85.2% (66.3% - 95.8%) for differentiating patients with malignant tumor (n = 39) from patients with benign lesions (n = 27). Additionally, when tested on gender and age matched asymptomatic normal individuals (n = 118), our model achieved a specificity of 93.2% (89.0% - 98.3%). Specifically, our assay is highly sensitive towards early-stage lung cancer, with a sensitivity of 75.0% (55.0%-90.0%) in 20 stage Ia lung cancer patients and 85.7% (57.1%-100.0%) in 7 stage Ib lung cancer patients. Conclusions: We have developed a novel sensitive blood based non-invasive diagnostic assay for detecting early stage lung cancer as well as differentiating lung cancers from benign pulmonary nodules.

Keywords: Early-stage lung cancer; circulating tumor DNA; high-throughput targeted DNA methylation sequencing.

Figure 1

The AnchorIRIS^TM assay and performance assessment. (A) Workflow of the ultra-sensitive AnchorIRIS^TM library preparation method. (B-C) A bake-off experiment comparing assay performance between the AnchorIRIS^TM assay and the SWIFT^® accel-NGS Methyl-seq^TM assay. The IRIS assay presents superior molecule conversion efficiency (C) with much higher average unique coverage for each input amount tested (B). (D and E) The sensitivity of the AnchorIRIS^TM assay was evaluated by diluting tumor gDNA into WBC gDNA, showing that significantly more informative co-methylated CpG regions above WBC background can be detected at dilutions ≥ 0.033% by Z-test (D). Dilutions higher than 10% (gray box) preserve a linear response of average co-methylation signal to the tumor fractions of input DNA (E).

Figure 2

Characterization of tissue level hypermethylation signatures of lung cancer. (A) Heatmap showing randomly selected 1000 hypermethylation regions for representative lung cancer and benign tissue samples. Methylation level of each region was calculated as co-methylated reads fraction. Samples are ordered from left to right by malignant/benign status (top color bar) and corresponding subtypes (second color bar). Subtypes from left to right are IA (n=33), MIA (n=19), AIS (n=8), FUN (n=11), INF (n=9), GRAN (n=4), TB (n=25), and HAM (n=21). Signal is shown in linear scale of color, with red indicating high methylation signal and green indicating low methylation signal. (B) A representative receiver operating curve (ROC) displays the tissue classification performance for distinguishing IA samples (n=65) against benign lesions (n=101) based on 10 bootstraps of 2-fold cross-validation of a regularized logistic regression. 95% confidence interval (CI) is shown in blue shade.

Figure 3

Lung cancer tissue co-methylation patterns can be captured in the cfDNA pool. Concordance of co-methylation between paired tissue (row) and plasma (column) samples is calculated using the percentage of reads sharing pre-defined co-methylation patterns and displayed in the heatmap. The highest similarity of a tissue sample to its matched plasma is shown in the diagonal of the heatmap, with ranking and Wilcoxon test p-values of each self-pair compared to the rest tissue-plasma pairs shown on the right. The smaller the rank (and p-value), the better the match of self-pair.

Figure 4

Cancer classification using plasma DNA. (A) Workflow chart of building a plasma level diagnostic prediction model. (B) Heatmap of the 9 hypermethylated markers used for the diagnostic prediction model in the training and independent test data sets. Methylation level of each marker was calculated as co-methylated reads fraction. (C and D) ROC curves plot the performance of plasma level classification with the 95% confidence interval (CI) of sensitivity in the training (C) and test (D) data sets. (E) Performance of Mayo model in our plasma cohort. P, partial solid nodule; S, solid nodule; G, ground-glass nodule.