Deep Learning Training Using PyTorch: SOT Supplemental Training for Education Program

By David Umbaugh posted 02-17-2022 15:46


This blog was crafted by the Secretary of the Graduate Student Leadership Committee Communications Subcommittee to share the experience and encourage graduate students to take advantage of this opportunity.

I received funding through the SOT Supplemental Training for Education Program (STEP) to attend a weeklong workshop on deep learning and its biomedical applications. This workshop was hosted virtually through the National Institute of Health’s Foundation for Advanced Education in the Sciences (NIH/FAES). The workshop consisted of morning didactic lessons followed by afternoons with hands-on coding sessions in Python, using powerful deep learning packages such as Pytorch. The workshop culminated in a final project where we applied deep learning algorithms to our own datasets to answer a relevant question of interest. My goal in taking this course was to be introduced to this exciting field and to begin brainstorming ways to apply this to my research.

Broadly speaking, I am interested in mixing wet lab and computational approaches to derive novel biological insights into liver injury and regeneration with the aim of identifying novel therapeutic targets. As part of my dissertation work, I have collected large amounts of single cell RNA-sequencing (scRNAseq) data, which is an exciting area to utilize deep learning strategies to analyze these massive datasets. “Deep learning” is a general term often used to refer to a neural network (with more than one hidden layer!), essentially an algorithm designed to make predictions. For example, a neural network could be used to identify cell populations in scRNAseq data. Throughout the course, we were introduced to various types of neural networks, such as convolutional neural networks (often used in image processing and microscopy applications), recurrent neural networks (often used text-mining/time series), generative adversarial networks (two neural networks competing against one another to improve each other), and graph neural networks (GNN, increasingly used to analyze scRNAseq data).

This workshop served as an excellent introduction to deep learning and offered practical strategies to implement deep learning into my own data analysis. For example, one immediate use is using an autoencoder to identify critical features in the scRNAseq data for subsequent clustering (hopefully improving the ability to easily discriminate cell populations). In the future, I hope to implement a GNN to study cell-cell interactions in my own datasets.

I am grateful for the opportunity to have had this training experience, which was generously supported by the SOT Education and Career Development Committee. I look forward to expanding my knowledge in this area and integrating what I learned throughout the course to my own research projects. I encourage any interested students to apply for the Supplemental Training for Education Program; more information can be found on the SOT website.