Tumor- and circulating-free DNA methylation identifies clinically relevant small cell lung cancer subtypes

Abstract

Small cell lung cancer (SCLC) is an aggressive malignancy composed of distinct transcriptional subtypes, but implementing subtyping in the clinic has remained challenging, particularly due to limited tissue availability. Given the known epigenetic regulation of critical SCLC transcriptional programs, we hypothesized that subtype-specific patterns of DNA methylation could be detected in tumor or blood from SCLC patients. Using genomic-wide reduced-representation bisulfite sequencing (RRBS) in two cohorts totaling 179 SCLC patients and using machine learning approaches, we report a highly accurate DNA methylation-based classifier (SCLC-DMC) that can distinguish SCLC subtypes. We further adjust the classifier for circulating-free DNA (cfDNA) to subtype SCLC from plasma. Using the cfDNA classifier (cfDMC), we demonstrate that SCLC phenotypes can evolve during disease progression, highlighting the need for longitudinal tracking of SCLC during clinical treatment. These data establish that tumor and cfDNA methylation can be used to identify SCLC subtypes and might guide precision SCLC therapy.

Keywords: DNA methylation; SCLC; biomarker; cfDNA; ctDNA; epigenetics; gene expression; liquid biopsy; lung cancer; subtyping.

Figure 1:. Detection and Classification of SCLC.

A Receiver operator characteristics (ROC) analysis of a DNA methylation-based test for the detection of SCLC from plasma. B Predictive models were generated to classify SCLC based on RNA-seq (Gene Ratio Classifier; GRC) and consensus of several combined predictive models is shown. A subtype was called when the consensus ≥ 0.5, else a sample was called equivocal. In addition, the expression of the three transcription factors ASCL1 (for SCLC-A), NEUROD1 (for SCLC-N) and POU2F3 (for SCLC-P) is shown normalized across the two cohorts. Furthermore, genes involved in neuroendocrine and non-neuroendocrine (Non-NE) as well as in tumor inflammation (TIS) and expression of HLA is shown. C Immune infiltration estimation using RNA-seq data (using the ESTIMATE algorithm). Boxplot shows the median as thick line, the box highlighting the first and third quartile with the whiskers highlighting 1.5x the interquartile range. D Characterization of SCLC consensus heterogeneity. The consensus agreement value for each subtype is plotted on the axis for each subtype by its consensus fraction of the respective subtype, demonstrating overlaps between SCLC subtypes. The line plot at the axis characterizes the distribution of subtypes across the axis. Wilcoxon test was used to compute p-values between groups. See also Figure S1, Table S1, and Data S1 and S2.

Figure 2:. Subtype-specific DNA methylation in SCLC.

A DNA methylation was assessed using reduced-representation bisulfite sequencing (RRBS) and DNA methylation was averaged per sample and subtype over 100kbp bins and the rolling average over 500 bins (= 50mbp) is highlighted in the c1 tumor samples. B-G Analysis of gene expression per SCLC subtype for DNA-methyltransferase 1 (DNMT1; B), DNA-methyltransferase 3A (DNMT3A; C) and 3B (DNMT3B; D), methionine adenosyltransferase 2A (MAT2A; E) and histone lysine methyltransferase (SUV39H1; F). G Overview of mechanism that links SUV39H1 expression with histone methylation. H Scheme highlighting the analysis and selection of DNA methylation sites associated with each of the SCLC subtypes using 100bp bins. By calculating the area under the curve by receiver operator characteristics (AUROC) we defined genomic region with high (AUC > 0.8) association with one the four respective subtypes. DNA methylation bins are shown related to their position within the genome for each chromosome for SCLC-A (I), SCLC-N (J), SCLC-P (K) and SCLC-I (L) and number of regions is stated for each subtype. Boxplot shows the median as thick line, the box highlighting the first and third quartile with the whiskers highlighting 1.5x the interquartile range. Wilcoxon test was used to compute p-values between groups. See also Figures S2-5 and Table S2.

Figure 3:. DNA methylation-based subtyping in SCLC.

A Scheme describing the process to develop the SCLC DNA methylation classifier (SCLC-DMC). Both cohorts were combined and the dataset was split in a training and a testing set and highly predictive DNA methylation sites were selected using area under the receiver operator characteristics curve (AUROC) to create predictive models using extreme gradient boosting with Dropouts multiple Additive Regression Trees (xGB-DART) with leave one out cross validation (LOOCV). For each subtype, 500 models were individually trained. Performance was assessed on the testing set. A cfDNA adjusted consensus classification approach (SCLC-cfDMC) was created using the same DNA methylation sites as used for the SCLC-DMC to predict subtypes in liquid biopsies. B Classification of SCLC tissue specimen using the SCLC-DMC approach. Prediction of subtype is shown in the training set, the independent testing set as well as in samples were classification by RNA (GRC) was not possible due to the absence of RNA-seq data (untested). The consensus in percentage of agreement between the models is shown. C Correlation of computed circulating tumor DNA (ctDNA) fraction by ultra-low pass whole genome sequencing (ULP-WGS) and a classifier based on seven methylation sites (Calculated Fraction [%]). D Differences in ctDNA fraction per DNA methylation were compared between samples analyzed at baseline prior to treatment and samples at tumor progression. E Differences in genome-wide DNA methylation between tumor tissue samples and matched baseline plasma samples were compared. DNA methylation was averaged per sample and subtype over 100kbp bins and changes between tumor DNA methylation and plasma DNA methylation were analyzed for each 100kb bin for each patient represented by a row in the heatmap across each chromosome as highlighted above. Furthermore, mean methylation per bin across the samples is highlighted in grey color above the heatmap together with the rolling average depicted by a black line. A histogram to the right highlights the distribution of differences for each bin across all samples. F The classification of SCLC subtypes using the SCLC-cfDMA approach is shown in plasma sample taken at baseline prior to treatment. Additionally, to the consensus, the classification based on the gene-ratio approach (GRC) as well as based on the tissue DMC approach is shown. Samples with GRC classification were included in the training cohort and inclusion for each sample is shown. G Classification of SCLC-subtypes using the SCLC-cfDMC approach is shown for samples with matched baseline plasma and plasma at progression. Boxplot shows the median as thick line, the box highlighting the first and third quartile with the whiskers highlighting 1.5x the interquartile range. Wilcoxon test was used to compute p-values between groups. See also Figures S6-9 and Table S3.

Figure 4:. Influence of SCLC subtyping methods on in vitro drug screening and clinical outcome.

Comparison of IC50 values for the A CDKi R-547 and the B AURKi CYC-116 between cell lines assigned to SCLC-A and SCLC-N using SCLC-DMC. C-D Clinical outcome depending on classification method used. Overall survival of SCLC patients stratified by classification using the SCLC-GRC (RNA-seq) and SCLC-DMC (DNA Methylation) method for C SCLC-A and D SCLC-N. Statistical significance is calculated using log-rank test. Cox-proportional hazard ratio is calculated and shown with 95% confidence interval. Boxplot shows the median as thick line, the box highlighting the first and third quartile with the whiskers highlighting 1.5x the interquartile range. Wilcoxon test was used to compute p-values between groups.