THEMIS: a cfDNA Methylome & Fragmentome Tool

The multimodal epigenetic characterization of cell-free DNA (cfDNA) in blood plasma (from "liquid biopsies") could improve minimally invasive approaches to early cancer detection and reduce treatment morbidity/mortality. One way forward involves the integrative profiling of cfDNA methylome and fragmentome data, which remains technologically challenging; however, a recent study from Vaisvila et al. that reported the development of enzymatic methyl-seq (EM-seq) (Vaisvila et al.) supported a sensitive and non-destructive means of detecting DNA methylation (as opposed to conventional bisulfite-based methylation sequencing; Tanaka and Okamoto) while simultaneously evaluating fragmentation.

News of this newly developed protocol prompted researchers guided by Fengwei Tan and Shugeng Gao (Chinese Academy of Medical Sciences and Peking Union Medical College) to adapt EM-seq to analyze the small amounts of cfDNA generally obtained from the blood plasma of cancer patients. In their new study in Nature Communications (Bie and Wange et al.), the Tan and Gao groups now report how the integration of crucial tumor genetic and epigenetic characteristics defined through the analysis of cfDNA supported the development and validation of "THEMIS" or "THorough Epigenetic Marker Integration Solution" as a multicancer detection test that facilitates sensitive and accurate early detection.

To develop THEMIS, Bie and Wange et al. applied shallow whole-methylome sequencing to plasma samples isolated during a multicenter, case-control, observational MONITOR (Multi-Omics Noninvasive Inspection of TumOr Risk) study that involved seven common cancer types (780 patients) and healthy controls (497 patients). As the mild enzymatic reactions minimized DNA damage during the analysis of cfDNA methylation, the team could integrate the simultaneous profiling of multiple types of cancer genomic features - fragmentation and copy number alteration - to enhance cancer detection. They also developed computational methods that extracted DNA methylation (Allen Chan et al.), fragment size (Cristiano et al.), copy number alterations (Leary et al.), and fragment end motif (Jiang et al.) cancer-associated features from the plasma cfDNA genome to then construct an ensemble machine learning classifier (THEMIS) that distinguished cancer patients from healthy controls with high sensitivity.

The team first analyzed associations among fragmentation, methylation, and copy number alteration profiles in the plasma cfDNA genome of healthy controls and colorectal cancer samples to investigate whether multimodal genomic features generate complementary signals in cancer detection. Overall, the modest-to-weak associations between multimodal genomic features suggest that they provide concordant and complementary information, implying that integrative analyses may enhance the detection power of cancer-associated signals.

The team next applied whole-methylome sequencing via EM-seq to plasma cfDNA samples from all cancer patients and healthy controls of the MONITOR study to develop and validate computational approaches for multicancer detection. Encouragingly, data on each detected feature separated cancer patients from healthy controls, supporting their utility in cancer detection. While late-stage samples displayed more apparent cancer-type clustering than early-stage samples (which suffer from lower levels of cfDNA), the data still separated early-stage cancers from healthy controls.

The authors subsequently applied machine learning methods to the cfDNA-derived data to increase the sensitivity of cancer detection, capture weak tumor signals more sensitively, and avoid problems related to technical and biological noise, and then constructed the ensemble THEMIS classifier to boost the performance of their multicancer detection model. Encouragingly, the application of THEMIS reduced the number of healthy individuals misclassified as cancer patients and increased the number of correctly identified cancer patients, which confirmed the author's hypothesis that integrative analysis of multimodal cancer genomic features would improve the detection specificity and sensitivity over profiling individual features alone. THEMIS consistently and accurately detected seven cancer types in early-stage cancer patients, achieving high sensitivity across disease types and stages; however, prediction score and detection sensitivity increased with increasing clinical stage (as occurred for individual modalities).

Finally, and perhaps most interestingly, the team also profiled plasma cfDNA methylation and fragmentation patterns at cancer tissue-specific accessible chromatin regions, which allowed the accurate localization of the tissue origin of cancer signals, which represents a crucial component of a multicancer test to guide follow-up diagnostic workups.