Computational methods for disease characterization and patient stratification

In biomedical research, technological advancements have fostered an increasing availability of heterogeneous data spanning diverse domains: from transcriptomic and genomic data, to high-resolution images, and clinical records. Each data type holds great potential in elucidating different but complementary aspects of human biology and disease. In order to leverage this potential, the development of tailored computational methods is required. This doctoral thesis relied on bioinformatics analyses of gene expression profiles, artificial intelligence (AI)-based approaches for digital pathology (DP), and healthcare data integration to characterize cardiovascular diseases and support patient stratification in oncology. Among omics technologies, transcriptome analysis allows measuring and comparing gene expression profiles across different conditions. It also lays the groundwork for the identification of functionally related genes, namely pathways, significantly enriched among a gene list of interest. In the first contribution to this thesis, a novel statistical framework for the meta-analysis of biological pathways was developed. The developed framework was used to analyze transcriptomic data of physiological and pathological cardiac hypertrophy with the aim of identifying gene signatures characterizing heart hypertrophy over time, and highlighting key differences between the two hypertrophy conditions.Transcriptomics, with its ability to identify disease-specific molecular features, has been adopted also in cancer diagnostics, for instance to stratify patients belonging to the same cancer entity into distinct molecular subtypes. Here, the main source of RNA material is represented by formalin-fixed paraffin-embedded (FFPE) tissues, whose degraded nature is widely recognized. In the second work, we proposed a bioinformatics analysis workflow optimized for the analysis of FFPE-derived RNA from archival tissue material. The workflow application to a muscle-invasive bladder cancer (MIBC) cohort demonstrated its effectiveness in accurately recapitulating established MIBC molecular subtypes. In parallel to the increasing use of omics data, cancer diagnostics and research has been revolutionized by advancements in the fields of AI and deep-learning (DL). For example, AI-based tools have been developed as alternative approaches for the prediction of molecular features directly from hematoxylin and eosin (H&E) slides. In the third study on upper-tract urothelial carcinoma (UTUC), we showed how a DL model can be trained to reliably predict UTUC subtypes on the basis of digitized H&E slides alone. Despite the increasing number of published AI-based tools, challenges in their deployment in real-world settings often make them underutilized. Thus, in an effort to fully leverage the potential of AI-based solutions in routine scenarios, in the last study we developed a proof-of-concept framework to integrate diagnostic and prognostic DL models in a fully digitized pathology department relying on the internationally recognized Health Level Seven standard. Collectively, the four contributions presented in this thesis show how computational methods can be harnessed for disease characterization and patient stratification, with the ultimate goal of leveraging their potential beyond academic research.