This study proposes a neural-multitask learning (N-MTL) framework, OncoMark, to predict the activity of all core cancer hallmarks for tumor profiling, using transcriptomic data from tumor biopsies. Single-cell RNA sequencing (ScRNA-seq) data were used to create synthetic bulk transcriptomic profiles and were then fed into a multi-task learning (MTL) framework capable of recognizing heterogeneous biopsy transcriptomes, which are those representing more than one hallmark. Transcriptomic profiles were taken from 941 tumors spanning 14 tissue types, with validation of the model on five external datasets demonstrating high generalizability and accuracy. Additionally, the ability of OncoMark to detect hallmark activity associated with metastatic progression was evaluated using primary tumor samples. Results of the evaluation include higher reports of the Activating Invasion and Metastasis hallmark across the samples, suggesting OncoMark’s ability to capture biologically relevant signatures indicating early metastatic competence. Thus, OncoMark is proposed as a novel computational approach to predict all hallmarks of cancer and as a clinical tool to diagnose tumor aggressiveness, design personalized care, and predict patient outcomes.