Application of machine learning techniques to biological big data

Abstract

To date, has been the primary driver of global innovation, compet- itiveness and cultural development. It is also a powerful engine

for creating new job opportunities, expanding market segments and inspiring new horizons where new skills and specialties can compete. From this perspective, we are constantly pushed to investigate the ICT industry and its interconnections with other areas, such as new biomedical technology. In the past twenty years, the development and increase of new diagnostic methods in the medical field has made available a huge amount of data capable of being stored and analyzed in order to extract important new knowledge. In the biological field, the data produced by the sequencing

techniques and the available databases provide a lot of informa- tion on multiple levels that can be integrated with each other.

The ability to integrate and analyze data from multiple sources is vital in order to collect real benefits and speed up outputs thanks to the high computational possibilities of some tools.

Technology has the potential to dramatically change the con- ception of medicine and, at the same time, it plays a critical role

in advanced diagnostics systems in making decisions intrinsic to

patient care. Developing high-quality, accurate Artificial Intelli- gence (AI) resources improves work of clinicians by intervening

on prevention, diagnosis and treatment of many pathologies. Some modern AI and computer science technologies, in general, encompass the power of clinical laboratory devices, allowing diagnostic activity to be carried out even outside of laboratories.

Medical aids and new advanced diagnostic equipment are in- creasingly relying on qualified experts in the field to supplement

medical evaluations and assist in diagnosis. In this context, the focus of this work was on two main topics. First one, we explored additional Machine Learning and Deep Learning techniques that can guarantee a better classification

of melanoma images even on clinical datasets with lower im- age quality. The goal is to improve melanoma early detection,

vii

which is now a limiting factor for first-line therapies in this tumor pathology. Many of the research in the literature utilize similar

strategies but use various approaches: some try to extract infor- mation directly from the image (such as color, plot and pixel

density), while others try to extract functions based on guided lines of dermatologists (such as ABCDE and the Seven-Point

Checklist). The majority of these researches are conducted us- ing higher-resolution dermoscopic pictures. The purpose of the

research is to identify novel features for melanoma classifica- tion that may be applied to less detailed images using advanced

learning techniques.

The second contribution of this thesis is addressed to the clas- sification of proteins. Researches focused on the possibility of

exploring further molecular descriptors in addition to those al- ready present in the literature to classify these proteins and to

build new tools able to explore the complex interaction between proteins in a visual and intuitive way. In this line of research,

the visualization of biological data was also taken into considera- tion. The work has mostly concentrated on the presentation tools

of biological ontologies in order to develop user-friendly sys- tems that allow end users to interact and extrapolate information

more easily. This is useful for the complexity of the biological

system that can be explored by the integration of omics disci- plines. These sciences attempt to analyze the biological system

holistically using biological Big Data, mainly proteomic, genomic, transcriptomic and metabolomic data. The latter are the most important groupings of organic compounds for the study of the functioning of living organisms. [edited by Author]