A cfDNA methylation-based tissue-of-origin classifier for cancers of unknown primary

Abstract

Cancers of Unknown Primary (CUP) remains a diagnostic and therapeutic challenge due to biological heterogeneity and poor responses to standard chemotherapy. Predicting tissue-of-origin (TOO) molecularly could help refine this diagnosis, with tissue acquisition barriers mitigated via liquid biopsies. However, TOO liquid biopsies are unexplored in CUP cohorts. Here we describe CUPiD, a machine learning classifier for accurate TOO predictions across 29 tumour classes using circulating cell-free DNA (cfDNA) methylation patterns. We tested CUPiD on 143 cfDNA samples from patients with 13 cancer types alongside 27 non-cancer controls, with overall sensitivity of 84.6% and TOO accuracy of 96.8%. In an additional cohort of 41 patients with CUP CUPiD predictions were made in 32/41 (78.0%) cases, with 88.5% of the predictions clinically consistent with a subsequent or suspected primary tumour diagnosis, when available (23/26 patients). Combining CUPiD with cfDNA mutation data demonstrated potential diagnosis re-classification and/or treatment change in this hard-to-treat cancer group.

Fig. 1. CUPiD, an accurate tissue-of-origin classifier applicable to cfDNA.

a Schematic of CUPiD development (DMR=Differentially Methylated Region). b Two-dimensional Uniform Manifold Approximation and Projection for Dimension Reduction (UMAP) using 22,179 DMRs selected by tumour class by class comparison, applied to 9,017 converted methylation arrays. Class labels superimposed over centroid of members of that class. c Per-class Receiver Operator Characteristic (ROC) curves for CUPiD, each using 276,108 mixture sets. Tissue-of-origin predictions for each mixture set are averaged over predictions made by those sub-classifiers not using that mixture set for training. Colours represent 30 individual classes as in Fig. 1b. Class abbreviations are defined in Table 1. Source Data are provided as a Source Data file.

Fig. 2. CUPiD performance testing in cfDNA from cohort of patients with known tumour types.

a Alluvial plot showing CUPiD performance in test cohort (170 cfDNA samples; 143 known primary tumour types, 27 NCCs). b Estimated Tumour Fraction (TF) of 143 cfDNA samples from known primary tumour types, grouped by concordance of CUPiD predictions, coloured by correct class. Dotted line denotes limit of detection of ichorCNA (3%). Boxes mark the 25th percentile (bottom), median (central bar) and 75th percentile (top); whiskers extend to 1.5 times the interquartile range. Class abbreviations are defined in Table 1. Source Data are provided as a Source Data file.

Fig. 3. Application of CUPiD to cfDNA from 41 patients with CUP.

a Alluvial plot showing how the tumour type-enriched alterations (TTEAs) (left) and CUPiD predictions (right) correspond to clinical classifications (centre) (QC=Quality control). b Distribution of tumour types predicted by CUPiD (n = 32), excluding unclassified predictions. c Alluvial plot showing CUPiD predictions and correlation with subsequent primary tumour diagnosis or clinical suspicion (iCCA intrahepatic cholangiocarcinoma, Gyn gynaecological cancer, CRC colorectal cancer, cCUP confirmed Cancer of Unknown Primary). Class abbreviations are defined in Table 1. Source Data are provided as a Source Data file.

Fig. 4. Swimmers plot of the 15 ‘clinically resolved’ patients.

Timeline of diagnostic investigations from point of CUP diagnosis to death or data lock. Time of final primary tumour diagnosis colour coded. Annotated by final diagnosis, CUPiD prediction and concordance. MDT multidisciplinary team meeting, AFP alpha feto-protein, TARC thymus and activation-regulated chemokine. Class abbreviations are defined in Table 1. Source Data are provided as a Source Data file.