Exploring Multi-Modal Learning Approaches Towards Precision Medicine

Abstract

Recent years have testified unprecedented advances in the field of molecular systems biology. The increasing amount of data produced from disparate sources is giving us the possibility to study biological systems from a wide variety of angles at an incredibly fine scale. In this context keeping up with the pace of data production and fully exploiting the availability of information from multiple modalities is fundamental. In this work, methods to extract information from different data types are investigated and developed in order to improve complex diseases understanding and develop explainable personalized models for patient stratification. The different data modalities considered ranging from molecular data, such as genomics, transcriptomics, and proteomics to data from the literature, either in form of natural language from publications or structured data from databases. The common denominator to handle these different modalities is machine learning based on graph topologies summarizing molecular interactions that provide a high-level representation of the molecular processes governing cells behavior. The main focus is the study of these molecular interaction networks and their potential applications to personalized medicine. The thesis is structured around three main pillars: network reconstruction, integration of network information in interpretable machine learning algorithms and the development of personalized models. Network reconstruction is analyzed on two data modalities: on molecular data by implementing state-of-the-art inference models and investigating consensus strategies to ensure robust prediction and on natural language proposing a novel deep learning-based methodology. Integration of network information into machine learning models is tackled by making use of a multiple kernel learning algorithm that exploits pathway-induced kernels, a concept introduced in this work. Going towards patient personalized models, networks are also exploited in a dynamic perspective to perform large-scale logical modeling through the implementation of a framework for accelerated attractor analysis. Model personalization is also tackled at genomic level by considering two data modalities: copy number alterations and somatic mutations. These data are used to feed a novel inference algorithm, implemented during this work, to estimate patient-specific phylogenetic trees.