Deep Learning in Biology and Medicine

The following sections are included:

• Preface
• About the Editors
• Acknowledgements
• Contents

The vast community of the life sciences encompasses research fields studying all aspects of the processes underlying and sustaining life, including biology, genetics, chemistry and medicine. In their modern interpretations, these large and diverse fields of study all share a common trait: their daily research and professional activities generate large amounts of digitized data comprising noisy, heterogeneous and often interrelated information. Such a data deluge has necessarily shifted attention towards data-driven or data-informed activities and processes, provoking a transformational effect on these fields, and as a result, producing novel developments, such as personalized medicine, intelligent molecule design, multi-omics, to name a few. The availability of data-driven computational models capable of processing, making sense of and exploiting such information is a key enabler to the progress of the life science field. This book is intended to provide an organic introduction to a groundbreaking data-driven methodology, deep learning, and to its manifold applications across the whole field. This chapter provides a high-level introduction to deep learning and reflects on the issues and challenges underlying the application of deep learning models to the life sciences. We conclude by providing a key to reading the technical essays in the book.

Deep learning has attracted the attention of researchers in the last few years due to its impressive performance on a plethora of computer vision tasks; and medical image analysis is no exception. In this chapter, we discuss the general deep learning strategies that have been considered in medical imaging; widespread preprocessing, processing, and postprocessing schemes; their targets and applications; and discuss key challenges to address in the future for easing their applicability in clinical practice.

Electronic health records (EHRs) contain a wealth of biomedical data and offer great promise to improving healthcare for individuals and providing essential data for enhancing population health. During each clinical visit, medical personnel capture a snapshot of a person’s health, and these records accumulate in massive storehouses of structured and unstructured data. As the size and scope of these repositories have expanded, opportunities abound to investigate biological and medical questions in new modes and on an unprecedented scale. One powerful tool to this end is deep learning (DL), a subfield of machine learning, which aims to algorithmically represent structure within data using layers of neural networks. Here, we review applications of DL with EHRs, in particular focusing on how the advent of DL changed how we mine EHRs for research.

In this chapter, we face the problems associated with natural language processing in the biomedical domain. We strive to provide enough technical background to allow the reader to develop real applications in this area while also providing the needed theoretical basis for supporting practical techniques. For the sake of self-containment, we organize the chapter as a general to specific tour following two orthogonal dimensions: the dimension of the techniques and the dimension of the domains. In the first dimension, we begin by presenting the generalities of natural language processing, i.e. linguistic units, tasks, applications, resources. After superficially discussing the symbolic approaches, we move into empirical methods, machine learning, shallow neural models and, finally, deep learning. In parallel, we start with open domain scenarios to further move into domain-restricted ones and, finally, into the biomedical domain.

Gene expression microarrays provide a characterisation of the transcriptional activity of a particular biological sample. Their high-dimensionality hampers the process of pattern recognition and extraction. Several approaches have been proposed for gleaning information about the hidden structure of the data. Among these approaches, deep generative models provide a powerful way for approximating the manifold on which the data reside. Here, we develop GEESE, a deep learning-based framework that provides novel insight into the manifold learning for gene expression data, employing a metabolic model to constrain the learned representation. We evaluated the proposed framework, showing its ability to capture biologically relevant features and encoding these features in a much simpler latent space. We showed how using a metabolic model to drive the autoencoder learning process helps in achieving better generalisation to unseen data. GEESE provides a novel perspective on the problem of unsupervised learning for biological data.

Deep Learning (DL) is among the most promising modeling techniques to tackle hard Cheminformatics problems, whose solutions continue to have profound humanitarian and societal implications. This chapter covers two relevant applications of deep learning in the Cheminformatics landscape. The first is QSAR/QSPR analysis, which concerns the prediction of chemical properties or biological activity of molecular structures. Along with a review of the main results obtained through the use of deep learning in the field, we provide an overview of neural network models capable of processing structured data, such as sequences, trees, and graphs, which allow one to represent the rich structure of chemical data for predictive purposes. The second is de novo drug design, which concerns the generation of novel molecular structures with desired chemical properties to speed up the drug discovery pipeline. Here, we present the family of deep generative models for molecule generation, by which it is possible to learn the distribution of molecules from data and to generate novel chemical structures through sampling.

Lives on earth are regulated by a complex system of interactions. Modelling these interactions through the network paradigms allows researchers to discover and understand the fundamental molecular mechanisms which drive the biological processes and lead to human diseases. The advancement made in the development of sequencing technologies has produced a growing amount of biological data. The aforementioned preconditions are at the base of a flourishing production of deep learning methods able to cope with the complexity and data abundance of this domain. For these reasons, this chapter provides a comprehensive overview of the recent advancements in the deep learning network-based approaches focusing on biology, medicine and pharmacological crucial research problems. At first, the needed biological and network science backgrounds are presented. Second, a comprehensive overview of the biological networks and resources is provided. Finally, we discuss the most recent methods in the field, organising them into three broad categories: the interactome, the network pharmacology, and the frontier biological problems.

From the widespread implementation and use of electronic health records to basic research in omics science, from the popularization of health wearables to the digitalization of procedures at the point of care, the domains of medicine and healthcare are bringing data to the fore of their practice. The abundance of data in turn calls for methods that allow transforming such raw information into novel knowledge that is usable in the domain. These methods for data analytics can be provided by statistics and also from machine learning and its currently most popular incarnation: deep learning. Effective as the latter may prove itself to be, its practical application is curtailed by a limitation that is particularly sensitive in the medical domain: its lack of interpretability, leading to limited explainability. In this chapter, we aim to map the many aspects of this problem and review research aimed at overcoming it, paying special attention to three questions, namely, reliability of data-driven methods in the medical field, relevance of deep learning as one of the most appealing approaches within machine learning, and need of interpretability and explainability for an accepted application of deep learning to clinical problems.

The recent availability of massive amounts of personal data that directly or indirectly describe individuals’ health status opened unprecedented opportunities to develop deep learning models able to leverage them to provide a wide range of benefits in healthcare. However, the development and the application of deep learning systems in such a critical domain raises many ethical and legal concerns. This chapter examines the practical implications of the ethical and legal guidelines for deep learning in healthcare. First, it provides an overview of the different approaches to AI ethics across the world. Then, focusing on the EU ethical and legal framework, it analyzes the pieces of legislation relevant for DL applications in healthcare. Finally, the key ethical values of privacy, fairness and explainability are mapped to the different stages of the machine learning lifecycle, and several state-of-the-art approaches to inscribe them in deep learning systems are overviewed.

The following section is included:

• Index