Multimodal cell-free DNA whole-genome TAPS is sensitive and reveals specific cancer signals

Abstract

The analysis of circulating tumour DNA (ctDNA) through minimally invasive liquid biopsies is promising for early multi-cancer detection and monitoring minimal residual disease. Most existing methods focus on targeted deep sequencing, but few integrate multiple data modalities. Here, we develop a methodology for ctDNA detection using deep (80x) whole-genome TET-Assisted Pyridine Borane Sequencing (TAPS), a less destructive approach than bisulphite sequencing, which permits the simultaneous analysis of genomic and methylomic data. We conduct a diagnostic accuracy study across multiple cancer types in symptomatic patients, achieving 94.9% sensitivity and 88.8% specificity. Matched tumour biopsies are used for validation, not for guiding the analysis, imitating an early detection scenario. Furthermore, in silico validation demonstrates strong discrimination (86% AUC) at ctDNA fractions as low as 0.7%. Additionally, we successfully track tumour burden and ctDNA shedding from precancerous lesions post-treatment without requiring matched tumour biopsies. This pipeline is ready for further clinical evaluation to extend cancer screening and improve patient triage and monitoring.

Fig. 1. Overview of the study.

We conducted a diagnostic accuracy study using a) cancer cases (GEL) from a cohort of symptomatic patients referred for urgent investigation for a possible gynaecological, lower GI, upper GI or renal cancer, and b) non-cancer controls with non-specific symptoms that might have been due to cancer from a rapid diagnostic centre (SCAN) or from Cambridge Bioscience (CBS). After collection of plasma samples, we conducted whole-genome sequencing at 80x or higher using TET-Assisted Pyridine Borane Sequencing (TAPS), aligned the generated reads against the human genome (GRCh38), and conducted analyses of copy number aberrations, methylation modifications, and somatic point mutations and indels, which included efficient denoising using the non-cancer SCAN controls. By integrating the analyses from all three data modalities, we generated sample-specific scores for the quantification of plasma ctDNA burden, which was used for cancer detection and post-treatment disease tracking. Matched tumour biopsies, if available, were used for validation, not for guiding the analysis. GEL Genomics England, SCAN Suspected Cancer Pathway, CBS Cambridge Biosciences Human Blood Products Supply Service.

Fig. 2. Analysis of copy number aberrations (CNA).

A Coverage signal for patient GEL195 (colorectal cancer) before denoising (Ai), after denoising (Aii) and in the biopsy after denoising (Aiii). A non-cancer control is also shown for comparison (Aiv). The aggregated coverage signal in each chromosome arm in the plasma sample is compared against the corresponding arm in a cohort of non-cancer control plasma CBS samples in search of gains (red) or losses (green). Gains in chromosomes 7, 13, 16, 20 and losses in chromosomes 4 and 18 in the plasma sample of patient GEL195 reflect aberrations in the same chromosomes in the matched biopsy, although this has not been used for guiding the analysis. Bi Scores quantifying coverage imbalances in the chromosome arms of each cancer plasma sample compared to the non-cancer plasma CBS controls. In each sample, each circle corresponds to a different chromosome arm. Red circles indicate a gain or loss of chromosomal material. Bii Integrated CNA scores over all chromosome arms in each plasma sample. Red circles indicate the gain or loss of chromosomal material in the corresponding samples. 29 out of 61 cancer samples were correctly identified (sensitivity 47.5%). C Integrated CNA score against cancer stage and type (Ci) and monotonic increase of median integrated CNA score with cancer stage (Cii). D In silico assessment of CNA analysis performance at increasing ctDNA fractions. At each ctDNA fraction, we simulated 1000 non-cancer and 1000 cancer plasma samples using actual non-cancer and cancer plasma samples as templates (see ‘Methods’). The area under the receiver operating characteristic (ROC) curve (AUC) was 80% at ctDNA fraction 0.7%. CTRL CBS controls (n = 9 subjects), CRC colorectal (n = 36 subjects), OES oesophageal (n = 8 subjects), PNCR pancreatic (n = 6 subjects), RNL renal (n = 5 subjects), OVR ovarian (n = 4 subjects), BST breast (n = 2 subjects). For each boxplot in (Ci, D) the box bounds, and centre correspond to the 25th, 50th (median), and 75th percentiles of the data in each corresponding group, and the whiskers extend to 1.5 times the interquartile range (IQR) above and below the box bounds. Source data is provided as a source data file.

Fig. 3. Burden analysis of somatic single nucleotide variants (SNVs) and INDELs.

A Somatic mutation burden in different cancer types and in non-cancer CBS controls (Ai), distribution of mutation numbers across genes (Aii) and consequences of mutations (Aiii). In (Ai), each circle corresponds to a different plasma sample. Bi Scores quantifying mutation burden imbalances in the chromosome arms of each cancer plasma sample compared to the non-cancer plasma CBS controls. In each sample, each circle corresponds to a different chromosome arm. Red circles indicate a difference in the somatic mutations burden of the chromosome arm in relation to the same arm in the CBS controls. Bii Integrated somatic mutation scores over all chromosome arms in each plasma sample. A red circle indicates a higher mutation burden in the corresponding sample, when compared to the CBS controls. We identified correctly 32 out of 61 cancer plasma samples (sensitivity 52.5%). C Integrated somatic mutation scores against cancer stage and type (Ci) and moderate correlation between median integrated somatic mutation score and stage (Cii); Spearman’s r = 50%). D In silico validation of somatic mutation analysis at increasing ctDNA fractions. At each ctDNA fraction, we simulated 1000 controls and 1000 cancer plasma samples using actual non-cancer and cancer plasma samples as templates (see ‘Methods‘). The area under the receiver operating characteristic (ROC) curve (AUC) is 74% at ctDNA fraction 1%. CTRL CBS controls (n = 9 subjects); CRC colorectal (n = 36 subjects), OES oesophageal (n = 8 subjects), PNCR pancreatic (n = 6 subjects), RNL renal (n = 5 subjects), OVR ovarian (n = 4 subjects), BST breast (n = 2 subjects). For each boxplot in (Ai, Ci, D), the box bounds, and centre correspond to the 25th, 50th (median) and 75th percentiles of the data in each corresponding group, and the whiskers extend to 1.5 times the interquartile range (IQR) above and below the box bounds. Source data is provided as a source data file.

Fig. 4. Overview of methylation analysis.

Ai Scores quantifying imbalances in the methylation burden in any of 377 regions (extracted from TCGA; see ‘Methods’) in each cancer plasma sample compared to the non-cancer plasma CBS controls. Each circle corresponds to a different region and red circles indicate over-methylation of the corresponding regions between the cancer plasma and the CBS controls. Aii Integrated methylation scores over all regions in each plasma sample. A red circle indicates an over-methylated plasma sample, when compared to the CBS controls. We identified correctly 28 out of 61 cancer plasma samples, which corresponds to a 45.9% sensitivity. B Integrated methylation scores against cancer stage and type (Bi) and monotonic increase of median integrated methylation scores with cancer stage (Bii). C In silico validation of methylation analysis at increasing ctDNA fractions. At each ctDNA fraction, we simulated 1000 non-cancer and 1000 cancer plasma samples using actual non-cancer and cancer plasma samples as templates (see ‘Methods’). The area under the receiver operating characteristic (ROC) curve (AUC) is 87% at ctDNA fractions 0.9%. CTRL CBS controls (n = 9 subjects); CRC colorectal (n = 36 subjects), OES oesophageal (n = 8 subjects), PNCR pancreatic (n = 6 subjects), RNL renal (n = 5 subjects), OVR ovarian (n = 4 subjects), BST breast (n = 2 subjects). For each boxplot in (Bi, C), the box bounds, and centre correspond to the 25th, 50th (median) and 75th percentiles of the data in each corresponding group, and the whiskers extend to 1.5 times the interquartile range (IQR) above and below the box bounds. Source data is provided as a source data file.

Fig. 5. Integration of genomic data modalities for ctDNA detection.

A Multimodal scores for the quantification of plasma ctDNA generated from the integration of copy number aberrations, somatic SNVs and INDELs, and methylation signals in each plasma sample. A red circle indicates a higher ctDNA burden in comparison to the non-cancer CBS controls. 52 out of 61 cancer plasma samples were correctly identified as such, which corresponds to 85.2% sensitivity. This is higher than the sensitivity of any of the three data modalities. B Multimodal scores against cancer stage and type (Bi) and monotonic increase of median multimodal scores with cancer stage (Bii). C In silico validation of multimodal analysis at increasing ctDNA fractions. At each ctDNA fraction, we simulated 1000 controls and 1000 cancer plasma samples using actual non-cancer and cancer plasma samples as templates (see Methods). The area under the receiver operating characteristic (ROC) curve (AUC) is 86% at ctDNA fractions 0.7%. CTRL CBS controls (n = 9 subjects), CRC colorectal (n = 36 subjects), OES oesophageal (n = 8 subjects), PNCR pancreatic (n = 6 subjects), RNL renal (n = 5 subjects), OVR ovarian (n = 4 subjects), BST breast (n = 2 subjects). For each boxplot in (Bi, C), the box bounds, and centre correspond to the 25th, 50th (median) and 75th percentiles of the data in each corresponding group, and the whiskers extend to 1.5 times the interquartile range (IQR) above and below the box bounds. Source data is provided as a source data file.

Fig. 6. Multimodal ctDNA detection for post-operative MRD and adjuvant therapy response tracking in colorectal cancer without matched tumour.

A Tracking post-operative MRD in case GEL193. ctDNA was detectable in the plasma 1 year after surgery and this correlated with inoperable metastatic rectal cancer and a possible lung adenocarcinoma suggested through radiological examination, both of which were recorded ~3 years after the post-operative plasma sample was collected. B Tracking response to adjuvant therapy following surgery. Case GEL282 (Bi) did not have detectable ctDNA immediately after the last cycle of treatment. However, low ctDNA burden was detected 5 months later, which correlated with the presence of tubular adenomas with low grade dysplasia in the sigmoid colon at around the same time. Case GEL432 (Bii) did not have detectable ctDNA shortly after the last cycle of treatment and was still alive ~6 years after the last plasma sample was collected. C Confusion matrix (Ci) and event-free (i.e., no recurrence, metastasis, or precancerous adenomas present) survival (Cii) in 9 patients with colorectal cancer. In 8 out 9 patients, ctDNA burden after the end of surgery and/or adjuvant therapy correlated with the presence/absence of clinical events, such as recurrence or pre-cancerous adenomas (Ci). Absence of ctDNA detection after the end of surgery/adjuvant treatment correlated (hazard ratio: 8.2; 95% CI: 1.3–53.1; two-sided log-rank test p value = 0.02) with longer survival times (Cii). Dx diagnosis, Sx surgery, Tx0 first cycle of adjuvant therapy, Tx1 last cycle of adjuvant therapy, RR recurrence, LGD low-grade dysplasia. Source data is provided as a source data file.