Large language models enable tumor-type classification and localization of cancers of unknown primary from genomic data.

Abstract

Tumor-type classification is critical for effective cancer treatment, yet current methods based on genomic alterations lack flexibility and have limited performance. Here, we introduce OncoChat, an artificial intelligence (AI) model designed to classify 69 tumor types by integrating diverse genomic alterations. Developed on genomic data from 158,836 tumors sequenced with targeted cancer gene panels, OncoChat demonstrates superior performance, achieving a micro-averaged precision-recall area under the curve (PRAUC) of 0.810 (95% confidence interval [CI], 0.803-0.816), accuracy of 0.774, and an F1 score of 0.756, outperforming baseline methods. In a cancer of unknown primary (CUP) dataset of 26 cases whose types were subsequently confirmed, OncoChat correctly identified 22 cases. In two larger CUP datasets (n = 719 and 158), tumor types predicted by OncoChat were associated with survival outcomes and mutation profiles consistent with those of known tumor types. OncoChat offers promising potential for clinical decision support, particularly in managing patients with CUP.