A large-scale retrospective study in metastatic breast cancer patients using circulating tumour DNA and machine learning to predict treatment outcome and progression-free survival.

Abstract

Monitoring levels of circulating tumour-derived DNA (ctDNA) provides both a noninvasive snapshot of tumour burden and also potentially clonal evolution. Here, we describe how applying a novel statistical model to serial ctDNA measurements from shallow whole genome sequencing (sWGS) in metastatic breast cancer patients produces a rapid and inexpensive predictive assessment of treatment response and progression-free survival. A cohort of 149 patients had DNA extracted from serial plasma samples (total 1013, mean samples per patient = 6.80). Plasma DNA was assessed using sWGS and the tumour fraction in total cell-free DNA estimated using ichorCNA. This approach was compared with ctDNA targeted sequencing and serial CA15-3 measurements. We identified a transition point of 7% estimated tumour fraction to stratify patients into different categories of progression risk using ichorCNA estimates and a time-dependent Cox Proportional Hazards model and validated it across different breast cancer subtypes and treatments, outperforming the alternative methods. We used the longitudinal ichorCNA values to develop a Bayesian learning model to predict subsequent treatment response with a sensitivity of 0.75 and a specificity of 0.66. In patients with metastatic breast cancer, a strategy of sWGS of ctDNA with longitudinal tracking of tumour fraction provides real-time information on treatment response. These results encourage a prospective large-scale clinical trial to evaluate the clinical benefit of early treatment changes based on ctDNA levels.