PAN-CANCER PREDICTION TRANSFORMER-BASED GENOMIC MODELING FOR IDENTIFYING CANCER-DRIVER MUTATIONS AND TREATMENT RESISTANCE ACROSS HETEROGENEOUS TUMOR TYPES

Complex genomic changes are common and can vary between different tumour types and different patients with cancer, and are a common cause of cancer progression and treatment failure. Here, we propose a Pan-Cancer Transformer-based genomic modelling system to identify cancer driving mutations and predict treatment resistance across different and complex cancer types. The framework leverages transformer architectures to capture long-range dependencies in genomic sequences, mutation profiles, copy-number alterations, patterns of gene-expression, and clinically-relevant molecular features. This proposed approach could be applied to model tumor specific and common pan-cancer genomic signatures, to find driver mutations with high impact that may contribute to tumour initiation, progression, metastasis and escape from therapy. The model is designed to include multi-omics data and clinical markers of resistance to targeted therapy, chemotherapy, immunotherapy and combination therapy, in order to more accurately predict resistance to these therapies. In the same way, attention-based interpretability can be used to extract important genes, pathways and mutation regions that impact model decisions. The results of this study show the promise of deep learning based on transformer models to advance precision oncology through precise detection of actionable mutations, patient stratification based on resistance risk, and personalized clinical decision-making. The proposed framework offers a scalable and interpretable trajectory to future pan-cancer genomic analyses, particularly when dealing with the statistical and machine learning difficulty of tumor heterogeneity.