Decoding cancers of unknown primary through genomics-driven clustering: A pan-cancer framework for prognostic classification using AACR GENIE data.

e15006 Background: Cancer of unknown primary (CUP) is a heterogeneous metastatic entity lacking an identifiable tissue of origin, limiting evidence-based therapy. Leveraging large-scale genomic data, machine-learning approaches can stratify CUP into biologically informative subgroups. Methods: Using AACR GENIE data, we implemented a machine-learning framework to stratify tumors based solely on genomics. Unsupervised k-modes clustering was applied to single-nucleotide variant (SNV) and copy-number alteration (CNA) profiles from tumors with known primary sites (reference, REF; n = 52,289) to identify molecular subgroups, supported by t-SNE for visualization. A supervised Random Forest classifier trained on REF-derived labels assigned CUP samples (n = 1,637) to these clusters which were evaluated for survival associations. Results: Four distinct molecular clusters were identified in the REF dataset. Cluster 1 (n = 2,376; favorable prognosis) was enriched for APC, PIK3CA, and KRAS mutations, consistent with colorectal- and endometrial-associated biology. Cluster 2 (n = 6,587; poor prognosis) exhibited KRAS- and TP53-driven alterations with frequent SMAD4 mutations, reflective of aggressive pancreatic- and NSCLC-like profiles. Cluster 3 (n = 18,473; intermediate prognosis) showed pervasive TP53 mutations with recurrent EGFR and ERBB2 amplifications. Cluster 4 (n = 24,853; favorable prognosis) displayed wild-type and copy-number–neutral states, with selective enrichment of targetable alterations (VHL, GATA3, SPOP, BRAF, NRAS, CTNNB1). CUP samples exhibited similar overall survival trends to matched REF clusters, albeit with uniformly shorter median survival. Conclusions: This hybrid unsupervised–supervised approach provides a genomics-driven framework for CUP stratification, identifying biologically informative subgroups that may improve prognostication and support treatment decisions independent of tissue-of-origin prediction. Keywords: machine learning, genomics, cancer of unknown primary, precision oncology.