Oops, you're using an old version of your browser so some of the features on this page may not be displaying properly.

MINIMAL Requirements: Google Chrome 24+Mozilla Firefox 20+Internet Explorer 11Opera 15–18Apple Safari 7SeaMonkey 2.15-2.23

OncoNPC Classifier Identifies CUP Subgroups Sharing Characteristics with Predicted Primaries and Significant Differences in Outcomes and Elevated Germline Risk

Through a machine-learning model, many carcinomas of unknown primary can be classified into meaningful subgroups with the potential to aid clinical decision making
06 Sep 2023
Genetic and Genomic Testing;  Cancer Intelligence (eHealth, Telehealth Technology, BIG Data)
Carcinoma of Unknown Primary Site (CUP)

Oncology NGS-based primary cancer-type classifier (OncoNPC) is a newly developed machine-learning model for the molecular classification of tumour samples using multicentre next-generation sequencing (NGS) panel data. Applied to 971 carcinoma of unknown primary (CUP) tumour samples, OncoNPC CUP subgroups showed significantly higher germline polygenic risk score for their predicted cancers, the first evidence of germline genetic correlation between CUP tumours and corresponding primary. Furthermore, OncoNPC CUP subgroups showed significant survival differences, consistent with those observed in the corresponding primary cancer types. OncoNPC predictions enabled a 2.2-fold increase in patients with CUP who could have received genomically guided therapies.

The findings suggest that CUPs share a genetic and prognostic architecture with known cancer types and may benefit from molecular classification. Findings are published by Alexander Gusev of the Dana-Farber Cancer Institute and Harvard Medical School in Boston, MA, US, and colleagues on 7 August 2023 in the Nature Medicine.

CUP represents about 3–5% of all cancers worldwide and is characterised by aggressive progression and poor prognosis. The hidden nature of the primary sites limits treatment options. Accurately identifying the latent primary site for CUP and demonstrating clinical benefit from site-specific therapies may open many existing treatment options for patients with CUP. Pathological diagnosis can be challenging for highly metastatic or poorly differentiated tumours.

For patients with CUP, immunohistochemistry results suggestive of a single primary diagnosis account for only 25% of tumours. Molecular tumour profiling has been proposed as an alternative for primary site classification, potentially for CUP. Clinical utility of NGS in diagnosing and aiding treatment for patients with CUP was not systematically investigated. Several recent studies have investigated the potential clinical benefit of molecular CUP classification, in non-randomised prospective studies and randomised clinical trials. These trials have often struggled to recruit enough representative patients and explore the full range of available therapies.

Retrospective electronic health record data, despite potential biases, can capture a larger and more heterogeneous patient population compared to prospective trials. When paired with tumour sequencing, these data can offer insights into the molecular workings of CUP tumours and how they relate to patient outcomes. As panel sequencing is often part of the standard of care, such insights also have the potential to assist diagnostic efforts and clinical management within existing molecular workflows.

The study team used multicentre, NGS-targeted panel sequencing data from 36,445 tumour samples with known primary cancers across 22 cancer types to train and evaluate a machine-learning classifier predicting a primary cancer type of a given tumour sample. They applied this classifier to 971 patients with CUP with clinical follow-up.

OncoNPC achieved a weighted F1 score of 0.942 for high confidence predictions (≥ 0.9) on held-out tumour samples, which made up 65.2% of all the held-out samples. When applied to 971 CUP tumours collected at the Dana-Farber Cancer Institute, OncoNPC predicted primary cancer types with high confidence in 41.2% of the tumours.

OncoNPC also identified CUP subgroups with significantly higher polygenic germline risk for the predicted cancer types and with significantly different survival outcomes. Notably, patients with CUP who received first palliative intent treatments concordant with their OncoNPC-predicted cancers had significantly better outcomes (hazard ratio 0.348; 95% confidence interval 0.210–0.570; p = 2.32 × 105). Furthermore, OncoNPC enabled a 2.2-fold increase in patients with CUP who could have received genomically guided therapies.

The authors commented that their findings suggest that routinely collected targeted tumour panel sequencing data have clinical utility in assisting diagnostic work-up and prognosis and may additionally inform treatment decisions. Through pathology-based evaluation, they discovered that 51.9% of CUP cases in the cohort had agreement between OncoNPC predictions and at least one pathology-based suspected primary. Despite being substantially higher than expected by chance, this relatively low agreement underscores the challenge that highly metastatic or poorly differentiated tumours pose to pathological diagnosis.

In several cases, they found that OncoNPC predictions could have been helpful where multiple primaries were pathologically suspected. Upon retrospective chart review, they found that only 12.7% of patients with CUP (20 of 158) received genomically guided targeted treatments, which could have potentially increase to 44 patients (27.8%) based on OncoNPC predictions.

In future work, the investigators envision a multimodal foundational framework that incorporates molecular sequencing together with patient pathology images, longitudinal physiological data, and clinical notes to directly predict optimal treatment regimens rather than just cancer types. This work paves a way for incorporating routine panel sequencing data into clinical decision support tools for clinically challenging cancers. OncoNPC thus provides evidence of distinct CUP subgroups and offers the potential for clinical decision support for managing patients with CUP.


Moon I, LoPiccolo J, Baca SC, et al. Machine learning for genetics-based classification and treatment response prediction in cancer of unknown primary. Nature Medicine 2023;29:2057-2067.


This site uses cookies. Some of these cookies are essential, while others help us improve your experience by providing insights into how the site is being used.

For more detailed information on the cookies we use, please check our Privacy Policy.

Customise settings
  • Necessary cookies enable core functionality. The website cannot function properly without these cookies, and you can only disable them by changing your browser preferences.