Unintrusive multi-cancer detection by circulating cell-free DNA methylation sequencing (THUNDER): development and independent validation studies

Abstract

Background: Early detection of cancer offers the opportunity to identify candidates when curative treatments are achievable. The THUNDER study (THe UNintrusive Detection of EaRly-stage cancers, NCT04820868) aimed to evaluate the performance of enhanced linear-splinter amplification sequencing, a previously described cell-free DNA (cfDNA) methylation-based technology, in the early detection and localization of six types of cancers in the colorectum, esophagus, liver, lung, ovary, and pancreas.

Patients and methods: A customized panel of 161 984 CpG sites was constructed and validated by public and in-house (cancer: n = 249; non-cancer: n = 288) methylome data, respectively. The cfDNA samples from 1693 participants (cancer: n = 735; non-cancer: n = 958) were retrospectively collected to train and validate two multi-cancer detection blood test (MCDBT-1/2) models for different clinical scenarios. The models were validated on a prospective and independent cohort of age-matched 1010 participants (cancer: n = 505; non-cancer: n = 505). Simulation using the cancer incidence in China was applied to infer stage shift and survival benefits to demonstrate the potential utility of the models in the real world.

Results: MCDBT-1 yielded a sensitivity of 69.1% (64.8%-73.3%), a specificity of 98.9% (97.6%-99.7%), and tissue origin accuracy of 83.2% (78.7%-87.1%) in the independent validation set. For early-stage (I-III) patients, the sensitivity of MCDBT-1 was 59.8% (54.4%-65.0%). In the real-world simulation, MCDBT-1 achieved a sensitivity of 70.6% in detecting the six cancers, thus decreasing late-stage incidence by 38.7%-46.4%, and increasing 5-year survival rate by 33.1%-40.4%, respectively. In parallel, MCDBT-2 was generated at a slightly low specificity of 95.1% (92.8%-96.9%) but a higher sensitivity of 75.1% (71.9%-79.8%) than MCDBT-1 for populations at relatively high risk of cancers, and also had ideal performance.

Conclusion: In this large-scale clinical validation study, MCDBT-1/2 models showed high sensitivity, specificity, and accuracy of predicted origin in detecting six types of cancers.

Keywords: cell-free DNA (cfDNA); machine learning; methylation; multi-cancer early detection.