CUPiD: a machine learning approach for determining tissue-of-origin in cancers of unknown primary from cell-free DNA methylation profiles

Patients with cancer of unknown primary (CUP) present with metastases, but without an identifiable primary tumour and typically have poor outcomes. Determining the primary site enables access to type-specific therapies, potentially leading to therapeutic benefit. We developed CUPiD, a machine learning classifier for non-invasively and accurately determining CUP tissue-of-origin (TOO) from cell-free DNA (cfDNA) methylation profiles obtained from a blood sample. We used a data augmentation strategy that enabled data arising from over 9,000 TCGA tumour tissue samples (rather than cfDNA samples) to be leveraged for classifier training. To generate the training dataset, we performed in silico mixing of reads from tumour tissue DNA with reads from non-cancer control (NCC) cfDNA, to mimic low tumour content of cfDNA, resulting in 276,108 in silico mixture samples, across 29 cancer types. These were used to train an ensemble of 100 XGBoost classifiers for TOO determination from cfDNA, with an AUROC of 0.98 on held-out mixture samples. We further tested CUPiD on 143 cfDNA samples from patients with metastatic cancer of known type, giving an overall multi-class sensitivity of 85% and TOO accuracy of 97%. In an additional cohort of 41 patients with CUP, CUPiD predictions were made in 32/41 (78%) cases, with 88% of the predictions clinically consistent with a subsequent or suspected primary tumour diagnosis. Our data demonstrate that CUPiD can accurately predict TOO from a single liquid biopsy and has potential to support treatment stratification and improve clinical outcomes for a significant proportion of patients with CUP.