Book Image

Deep Learning for Genomics

By : Upendra Kumar Devisetty
Book Image

Deep Learning for Genomics

By: Upendra Kumar Devisetty

Overview of this book

Deep learning has shown remarkable promise in the field of genomics; however, there is a lack of a skilled deep learning workforce in this discipline. This book will help researchers and data scientists to stand out from the rest of the crowd and solve real-world problems in genomics by developing the necessary skill set. Starting with an introduction to the essential concepts, this book highlights the power of deep learning in handling big data in genomics. First, you’ll learn about conventional genomics analysis, then transition to state-of-the-art machine learning-based genomics applications, and finally dive into deep learning approaches for genomics. The book covers all of the important deep learning algorithms commonly used by the research community and goes into the details of what they are, how they work, and their practical applications in genomics. The book dedicates an entire section to operationalizing deep learning models, which will provide the necessary hands-on tutorials for researchers and any deep learning practitioners to build, tune, interpret, deploy, evaluate, and monitor deep learning models from genomics big data sets. By the end of this book, you’ll have learned about the challenges, best practices, and pitfalls of deep learning for genomics.
Table of Contents (18 chapters)
Part 1 – Machine Learning in Genomics
Part 2 – Deep Learning for Genomic Applications
Part 3 – Operationalizing models

Use case – Model interpretability for genomics

In this hands-on exercise section, we will build a similar convolutional NN (CNN) model that we built in Chapter 9, Building and Tuning Deep Learning Models, but unlike in Chapter 9, here we will use a simulated dataset of DNA sequences of length 50 bases (whereas in Chapter 9, we have DNA sequence of length 101 bases). In addition, the binding sites in this example are not just for Transcription Factors (TFs) but any protein. The labels are designated as 0 and 1, corresponding to positive and negative binding sites (0 = no binding site and 1 = binding site).

The goal of this is to train a CNN model to predict the DNA binding site of the protein and visualize it in the predictions. Since these are artificial sequences, we have injected the AAAGAGGAAGTT motif into the positive sequence, but don’t worry—the CNN doesn’t know that.

Data collection

For this hands-on tutorial, we will use the simulated data...