Genome-Wide Repeat Element Landscapes Enhance Non-Invasive Detection and Characterization of Cancer

Introduction: Liquid biopsies that leverage cell-free DNA (cfDNA) fragmentomes have shown great promise as non-invasive, cost-effective blood tests for early detection and monitoring of cancer. We previously developed DELFI (DNA Evaluation of Fragments for Early Interception) which detects subtle changes in cancer patients even in low-coverage sequencing from < 1mL of blood by looking genome-wide at millions of fragments released into the circulation from dying cells. While our approach has shown great promise for detection of lung, liver and other cancers, a limitation to its sensitivity is that it does not comprehensively integrate information from the ~50% of the genome comprising repeat elements, which have long been hypothesized to contain important genetic hallmarks of cancer. For example, epigenetic changes in cancer are thought to drive movement of transposable elements that cause oncogene activation, and satellite instability may drive chromosomal mis-segregation. Historically, repeats have been under-studied due to incomplete genome references and challenges aligning repetitive sequences. Long-read sequencing can allow alignment and study of this “dark genome” in tissue, but in the blood, nearly all cfDNA fragments are 150-180bp in length necessitating study with short-read technology. Here, we develop ARTEMIS (Analysis of Repeat Elements in Disease) a novel, alignment-free approach for reconstruction of repeat landscapes comprising 1280 individual repeat types occurring genome-wide across 57 subfamilies and 6 families. We demonstrate ARTEMIS in cfDNA and further, ensemble it with DELFI to demonstrate improved utility of fragmentomic technologies in tumor detection, tissue of origin determination, and monitoring.

Materials and

Methods: We employed a de novo search of short sequences (kmers) in the telomere to telomere (T2T-CHM13) reference genome to identify 1.2 billion 24-mers uniquely defining 1280 repeat types from the satellite, DNA and RNA transposable element, short and long interspersed nuclear element (SINEs and LINEs) and Long Terminal Repeat (LTRs) families. We performed an alignment-free analysis of ARTEMIS kmers in low coverage (1-2x) whole genome sequencing (WGS) of cfDNA from 1425 individuals. To do this, we obtained all sequencing reads for each sample, counted each kmer and its reverse complement, aggregated kmer counts for each repeat type, and normalized the counts to sequencing coverage. We defined this vector as the kmer repeat landscape. We compared cfDNA kmer repeat landscapes in replicates from different sequencing platforms and to those computed from tumor tissue sequencing in The Cancer Genome Atlas. We then used cfDNA kmer repeat landscapes to train and cross-validate logistic regression models for non-invasive multi-cancer detection and tissue of origin determination (n=208 individuals with lung, bile duct, ovarian, colorectal, gastric, duodenal, pancreatic or breast cancer, and n=215 without cancer), lung cancer detection (diagnostic cohort individuals with (n=129) and without (n=158) lung cancer), and liver cancer detection (individuals with liver cancer (n=75) or with high-risk conditions (n=133)). We validated the locked lung cancer model in a separate cohort (individuals with (n=88) and without (n=400) lung cancer) and in patients undergoing lung cancer therapy (n=19, at 75 longitudinal timepoints). We ensembled ARTEMIS with models of cfDNA fragmentation (DELFI) to enhance performance.

Results, Conclusions, and Discussions: In a simulation of low-coverage short-read WGS incorporating typical sequence error rates, we found that 98% of kmers counted were found in reads originating from their true repeat type. For a subset of healthy cfDNA samples (n=155) sequenced on multiple sequencing platforms, kmer repeat landscapes were highly concordant (average spearman correlation coefficient between replicates, p=0.998), despite the use of alignment-free methodology at low coverage. As a multi-cancer early detection test, ARTEMIS detected 92% of cancers at a specificity of 90%, and 93% of cancers when ensembled with DELFI. The joint ARTEMIS-DELFI model determined tissue of origin with 68% and 83% accuracy for the highest or top two predictions, respectively. In single-cancer screening and surveillance settings, ARTEMIS continued to demonstrate high performance (AUC 0.82, 95% CI 0.78-0.87 and AUC 0.87, 95% CI 0.82-0.93 for lung and liver cancers, respectively). When ensembled with the DELFI approach, ARTEMIS-DELFI further enhanced sensitivity and specificity for detection (AUC 0.91, 95% CI 0.88-0.94 and AUC 0.9, 95% CI 0.86-0.94 for lung and liver cancers, respectively). We then externally validated the locked model for lung cancer detection. When the ARTEMIS-DELFI model score threshold to classify a sample as having cancer while maintaining 60% specificity was set in the cross-validated cohort, performance was maintained in the external validation cohort (sensitivity 70%, 69%, 82% for stage I, II and III/IV cancers with 60% specificity). This highlights the robust and generalizable nature of these cfDNA modeling approaches. Next we applied the locked model to patients under treatment for late-stage lung cancer. Here, ARTEMIS-DELFI scores at the first timepoint after initiation of treatment (median = 6 days) stratified progression free survival (median = 1.4 months for patients in the high score group vs 8.9 months for low score group, p< 0.001, log-rank test, two sided). Comparison of repeat elements quantified in cfDNA versus publicly available tumor tissue data suggested that repeat element representation in the circulation is impacted by both structural and epigenetic mechanisms. These results demonstrate the potential for ARTEMIS to evaluate genome-wide repeat landscapes in the circulation, and to enhance liquid biopsies for non-invasive detection and characterization of cancer.

Acknowledgements (Optional):