A Foundation Model for the Cancer Genome

Abstract

AO_SCPLOWBSTRACTC_SCPLOWCancer is a disease of the genome, in which somatic mutations and copy-number alterations determine tumour identity, clinical behaviour, and response to therapy. Consortium-scale sequencing has profiled hundreds of thousands of tumours,1,2 yet clinical interpretation still proceeds one alteration at a time against hand-curated knowledgebases,3,4 often ignoring co-occurring alterations and the genome-wide copy-number pattern. Self-supervised foundation models pretrained on unlabelled corpora5 have produced transferable representations in adjacent biological domains6-8 by learning joint structure across many features, yet no comparable model exists for the cancer genome. Here we present TESSERA (Tumour Embeddings via Self-Supervised Encoding and Reconstruction of Alterations), a foundation model for the cancer genome; we pretrain it on somatic single-nucleotide variants and copy-number segments through masked-token reconstruction within each modality and a contrastive objective across modalities. A single representation, produced once and reused without retraining, supports variant pathogenicity prediction, pan-cancer tumour typing, unsupervised molecular subtyping, prognostic stratification, and counterfactual treatment-effect estimation that yields predictive chemotherapy-selection biomarkers in real-world cohorts. These biomarkers are interpretable: each surfaces the co-occurring alterations underlying the prediction, exposing biology that single-gene rules miss. In metastatic colorectal cancer, where the FOLFOX-vs-FOLFIRI choice is currently guided by toxicity rather than tumour biology, the model uncovers a candidate predictive biomarker: a three-feature rule (TP53+/KRAS+/17p-) selecting patients who derive substantially greater benefit from FOLFOX than FOLFIRI.