A Multi-Omics Transformer Foundation Model For Ai-Driven Early Cancer Detection Using cfDNA And cfRNA: Implications For Precision Oncology And Early Intervention In U.S. Healthcare Systems

A non-invasive and scalable substitute for tissue biopsies is liquid biopsy-based cancer diagnosis using circulating tumor DNA (cfDNA) and RNA (cfRNA); however, appropriate integration of heterogeneous molecular signals is still a major obstacle. Using TCGA gene expression and copy number variation data, we present a Transformer-based multi-omics fusion framework for pan-cancer identification in this paper. The suggested method dynamically weights informative molecular properties and learns intricate cross-omics interactions using self-attention. A single representation was produced by preprocessing, normalizing, and combining the gene expression and copy number variation profiles of 5,408 tumor samples from 33 different cancer types. The Autoencoder, Multilayer Perceptron, Support Vector Machine, and Logistic Regression baselines were used to compare the model. The Transformer achieved an accuracy of 82.6%, an F1-score of 82.0%, and an AUC of 0.899, surpassing all other methods. The results of this study demonstrate that attention-based multi-omics integration is a powerful and reliable method for liquid biopsy-based cancer diagnosis, which encourages the creation of scalable, AI-powered precision oncology systems.