Colacino, Andrea (2025) Deep Fusion and Precision Outcome: An AI Framework for Integrating Multi-Omics Data in Personalized Medicine. [Tesi di dottorato]

Documento PDF Tesi_CQB_Review1_0.pdf Visibile a [TBR] Amministratori dell'archivio Download (3MB) | Richiedi una copia

Abstract

Cancer remains a disease of immense complexity, with prognosis and thera- peutic response dictated by multi-layered molecular interactions. While multi- omics assays provide rich molecular profiles, the vast size and heterogeneity of these data present a significant translational gap for clinical application. Current computational methods often fail to robustly integrate multiple molecular modal- ities and phenotypic data, thereby hindering the realization of truly personalized oncology. This doctoral thesis addresses this challenge by proposing and validating a novel, advanced Deep Learning framework for multi-modal data fusion. The core innovation lies in the synergistic combination of Contrastive Learning (CL) and the Transformer architecture. The CL component is utilized to train ro- bust, biologically coherent latent representations for each individual omics layer (e.g., genomics, transcriptomics). These representations are then input into a custom Transformer module, which leverages the Attention Mechanism to model the complex, non-linear dependencies between the different molecular layers and derive a unified prognostic signature. The methodology is tested on Breast Cancer (BC) multi-omics cohorts for the critical task of survival prediction and recurrence risk stratification. The findings establish a scalable, efficient, and generalizable framework that can be readily extended to other cancer types and data modalities (e.g., Pro- teomics). This work represents a significant step towards unlocking the full po- tential of multi-omics data, providing clinicians with powerful tools for making more informed and precise treatment decisions.

Downloads

Actions (login required)

Modifica documento