Prediction of tumor patterns through the integration of clinical and transcriptomics data

Abstract

Cancer remains a leading cause of mortality worldwide, largely due to its remarkable heterogeneity and the lack of robust molecular classifiers that can capture the complexity of tumor biology beyond histopathological criteria. Despite major advances in molecular oncology, most transcriptomic studies have focused primarily on global gene expression, overlooking other regulatory layers such as promoter activity, alternative splicing, or tumor microenvironment composition. This perspective limits the discovery of biomarkers and constrains the development of predictive tools for precision oncology. Here, we present tumorProfiler, a modular analytical framework that integrates multiple transcriptomic dimensions, including promoter activity, alternative splicing (Percent Spliced-In, PSI) and gene expression, into predictive models for tumor classification. Using high-quality RNA-seq data from the Pan-Cancer Analysis of Whole Genomes (PCAWG) cohort (n = 305 donors), we systematically characterized differential promoter activity, exon inclusion patterns, gene deregulation, and immune–stromal profiles across ten tumor types and intra-organ progression subtypes. Six supervised learning models were constructed, combining interpretable machine learning algorithms such as Random Forest with automated frameworks for benchmarking. Our results reveal that promoter activity and gene expression consistently outperform splicing events and cell composition in multiclass tumor prediction, achieving high accuracy and generalization capacity (AUC > 0.95; OOB error < 7%). While splicing events-based models captured biologically meaningful variation, their predictive power was more limited. Importantly, variable importance analyses highlighted a reduced subset of promoters, splicing events, and genes as candidate biomarkers with potential translational relevance. Altogether, this work demonstrates that transcriptomic regulation in cancer operates through complementary layers of molecular information, each contributing differently to tumor identity and progression. By integrating these layers, tumorProfiler provides a flexible and interpretable platform for patient stratification, biomarker discovery, and the design of precision therapies. Although currently a computational proof of concept, its modular design and discovery potential open avenues for experimental validation and future clinical translation.