Abstract
Quantifying the biological processes that drive cancer progression remains a key challenge in oncology. Although the hallmarks of cancer provide a foundational framework for understanding tumor behavior, existing diagnostic tools rarely measure these hallmarks directly. Here we present a neural multi-task learning-based framework that estimates hallmark activity using gene expression data from tumor biopsies. The model was trained on transcriptomic profiles from 941 tumors spanning 14 tissue types and tested on five independent datasets. It predicts the activity of ten cancer hallmarks simultaneously and with high accuracy. Additional validation on large-scale datasets including normal and cancer samples confirmed its sensitivity and specificity. Predicted hallmark activity was associated with clinical staging, suggesting biological relevance. A web-based tool was developed to facilitate integration into research and clinical workflows. This approach enables efficient analysis of transcriptomic data to inform understanding of tumor biology and support individualized treatment strategies.