Complete genomic characterization in patients with cancer of unknown primary origin in routine diagnostics

Abstract

Background: In ∼3%-5% of patients with metastatic disease, tumor origin remains unknown despite modern imaging techniques and extensive pathology work-up. With long diagnostic delays and limited and ineffective therapy options, the clinical outcome of patients with cancer of unknown primary (CUP) remains poor. Large-scale genome sequencing studies have revealed that tumor types can be predicted based on distinct patterns of somatic variants and other genomic characteristics. Moreover, actionable genomic events are present in almost half of CUP patients. This study investigated the clinical value of whole genome sequencing (WGS) in terms of primary tumor identification and detection of actionable events, in the routine diagnostic work-up of CUP patients.

Patients and methods: A WGS-based tumor type 'cancer of unknown primary prediction algorithm' (CUPPA) was developed based on previously described principles and validated on a large pan-cancer WGS database of metastatic cancer patients (>4000 samples) and 254 independent patients, respectively. We assessed the clinical value of this prediction algorithm as part of routine WGS-based diagnostic work-up for 72 CUP patients.

Results: CUPPA correctly predicted the primary tumor type in 78% of samples in the independent validation cohort (194/254 patients). High-confidence predictions (>95% precision) were obtained for 162/254 patients (64%). When integrated in the diagnostic work-up of CUP patients, CUPPA could identify a primary tumor type for 49/72 patients (68%). Most common diagnoses included non-small-cell lung (n = 7), gastroesophageal (n = 4), pancreatic (n = 4), and colorectal cancer (n = 3). Actionable events with matched therapy options in clinical trials were identified in 47% of patients.

Conclusions: Genome-based tumor type prediction can predict cancer diagnoses with high accuracy when integrated in the routine diagnostic work-up of patients with metastatic cancer. With identification of the primary tumor type in the majority of patients and detection of actionable events, WGS is a valuable diagnostic tool for patients with CUP.

Keywords: cancer of unknown primary; diagnostic tool; tumor type prediction; whole genome sequencing.

Figure 1

The cancer of unknown primary prediction algorithm (CUPPA). Created with BioRender.com. GIST, gastrointestinal stromal tumor; NET, neuroendocrine tumor; NSCLC, non-small-cell lung cancer; WGS, whole genome sequencing

Figure 2

Predictive performance of CUPPA. (A) Confusion matrix showing model performance across 29 tumor types in the external validation cohort. The confusion matrices for the reference set and internal validation set can be found in the supplementary figures (Supplementary Figures S1 and S2). (B) Receiver operator curves and precision-recall curves (C) showing the overall model performance. (D) By using the probability score generated by the model as a cut-off, a high predictive precision (95%) was reached at a score of 0.8. The corresponding metrics at this cut-off are plotted in panel B and C. (E) In low-confidence predictions (<0.8), predictive precision was lower. The high top-3 and top-5 model accuracy demonstrated, however, that low-confidence predictions can be used as a mean to derive a differential diagnosis to correlate with clinicopathological differential diagnoses. (F) In total, 64% of samples reached a high-confidence prediction. Distribution of high- and low-confidence predictions varied across tumor types. AP, average precision; AUC, area under the curve; CUPPA, cancer of unknown primary prediction algorithm; GIST, gastrointestinal stromal tumor; NET, neuroendocrine tumor.

Figure 3

Application of CUPPA in diagnostic work-up of 72 CUP patients. For 37 (51%) patients, a high-confidence prediction was reached (panel A). All high-confidence predictions were consistent with the differential diagnosis (green borders) before WGS. For 35/37 cases, a final diagnosis could be reached (dark green boxes under ‘Review by expert pathologists’). For two patients, the prediction did not provide additional information for the diagnosis (red boxes). For the remaining 35 cases, a low-confidence prediction was reached (panel B). When integrated with prior clinicopathological differential diagnoses, this prediction proved to be informative to reach a diagnosis in 12 patients (patients 38-49). In two additional patients (50 and 51), a disease-defining genomic event was detected with WGS. A description of the clinical value of CUPPA in low-confidence predictions can be found in Supplementary Table S6. the Supplementary material, available at https://doi.org/10.1016/j.esmoop.2022.100611. CUPPA, cancer of unknown primary prediction algorithm; GIST, gastrointestinal stromal tumor; NET, neuroendocrine tumor; NSCLC, non-small-cell lung cancer; WGS, whole genome sequencing.

Figure 4

Biomarker-based therapy options detected with WGS in CUP patients. In 47% of patients, an actionable event was identified (panel A). In patients with a definitive CUP (n = 21), an actionable event was identified in 33% (7 patients). For 10 patients, multiple therapy options were identified. In panel B, each line represents 1 of these 10 patients, showing the multiple therapy options identified in each patient. CUP, cancer of unknown primary; WGS, whole genome sequencing.