AI-Driven stratification of cancer patients using The Cancer Genome Atlas whole-genome sequencing data

Recent genomic foundation models have advanced DNA sequence interpretation, yet most remain constrained to local sequence patterns and fail to produce the patient-level insights required for clinical decision-making. To address this limitation, we developed a framework that extends beyond sequence-level inference, enabling robust patient stratification through a Cancer Foundation Model. Our approach begins with “DNAChunker”, which employs a dynamic H-net-based tokenization strategy that divides the genome into variable-length segments, preserving high-resolution detail in regulatory and coding regions while efficiently compressing repetitive sequences. When evaluated on the Nucleotide Transformer and Genomic Benchmarks, DNAChunker achieved performance comparable to the state-of-the-art GENERator (1.2 billion parameters) while using only 156 million parameters. To translate these genomic embeddings into patient-level insights, we implemented a transformer-based Cancer Aggregation Model that integrates mutation embeddings with somatic copy-number alteration (SCNA) features. The framework was evaluated on large whole-genome sequencing (WGS) cohorts, including PCAWG (n=2,040) and CUBRICS breast cancer samples (n=1,053), with TCGA-BRCA (breast cancer; n=920) serving as an external validation cohort. The model effectively stratified patients by cancer type (accuracy, 96.89%), homologous recombination deficiency (HRD; accuracy, 92.83%), and PAM50 subtype (accuracy, 84.05%). Notably, it classified PAM50 intrinsic subtypes using only DNA-level information, eliminating the conventional reliance on RNA-based expression profiling. The Cancer Foundation Model demonstrates that patient-level representation learning from whole-genome data can achieve clinically meaningful stratification across diverse tumor types. By bridging the gap between genomic sequence interpretation and actionable phenotypic classification, this framework establishes a foundation for AI-based precision oncology. With further validation, it will facilitate biomarker discovery and patient stratification in clinical trials directly from WGS data.