Circulating tumour DNA (ctDNA) is found in the bloodstream of cancer patients and can be used to monitor genomic evolution to cancer therapies. However, tumours also adapt to therapeutic pressure by altering their transcriptome through ‘non-genomic evolution’. It is challenging to investigate these changes via invasive serial tumour biopsies, hence, it would be ideal to utilise ctDNA to dynamically monitor the transcriptome of cancer cells in vivo.
ctDNA fragmentation is non-random and closely linked with nucleosomal architecture. Notably, the region surrounding the transcriptional start sites of actively expressed genes are under-represented in ctDNA, as without protection of the nucleosome, these fragments are rapidly digested by plasma nucleases. By comprehensively mapping ctDNA fragments through low-coverage whole-genome sequencing (WGS), we are able to generate detailed nucleosome occupancy maps which can be used in a machine learning framework to infer the transcriptional state of all genes.
We showcase this novel strategy using longitudinal plasma samples from 12 patients undergoing azacitidine therapy for myelodysplasia (MDS), a cancer of the bone marrow. Comparing RNA-seq from matched bone-marrow biopsies, our approach is able to classify genes into high and low expression categories at >85% median accuracy in plasma, even at 2.5x WGS coverage. Through quantifying changes in ctDNA fragmentation, we can also monitor dynamic changes in the transcriptome of MDS patients on therapy, identifying distinct signatures at times of response and subsequent progression.
Furthermore, we have extended this analysis to advanced breast cancer patients where we have uncovered treatment specific resistance mechanisms from plasma upon progression to palbociclib therapy. This demonstrates the broad applicability of our method across different malignancies as a minimally invasive and accurate means to monitor non-genomic evolution.