Recent breakthroughs in high-throughput genomic and biomedical data are transforming biological sciences into "big data" disciplines. In parallel, progress in deep neural networks are revolutionizing fields such as image recognition, natural language processing and, more broadly, AI. This course explores the exciting intersection between these two advances. The course will start with introduction to deep learning and overview the relevant background in genomics and high-throughput biotechnology, focusing on the available data and their relevance. It will then cover the ongoing developments in deep learning (supervised, unsupervised and generative models) with the focus on the applications of these methods to biomedical data, which are beginning to produced dramatic results. In addition to predictive modeling, the course emphasizes how to visualize and extract interpretable, biological insights from such models. Recent papers from the literature will be presented and discussed. Students will work in groups on a final class project using real world datasets.
College calculus, linear algebra, basic probability and statistics such as CS109, and basic machine learning such as CS229. No prior knowledge of genomics is necessary.
Lecture Venue and Times
09/26/2016 - 12/09/2016 Mon, Wed 3:00 PM - 4:20 PM at Hewlett Teaching Center 201
Recitation. Fridays 10:30am - 11:20am at Hewlett 102
James Zou: Wednesdays 5-7pm (Packard 253).
Anna Shcherbina and Nadine Hussami: Mondays 5-7pm (Lane L339)
Jayanth and Alon: Thursdays 10:30am-12:30pm (Huang Basement)
Course project (50%): the students will form teams of 4-6 and choose from one of the suggested projects or select their own project. Teams will be given Microsoft Azure credits to implement algorithms and perform analysis. Teams are expected to work on the research project throughout the second half of the quarter and produce conference-style papers. Each team will present the paper to the entire class at the end of the semester.
A significant portion of the class will be based on reading and discussing the latest literature. Every student should read the assigned papers before class and participate in discussions.
Paper presentation (20%): each team selects one of the suggested papers to present in detail to the class.
Paper review (20%): each team selects 2 other papers to review. The review concisely summarize the key findings of the paper, highlight interesting ideas, weaknesses and potential extensions.
Class participation and quizzes (10%): every student should actively engage in paper discussions in class and in the online forum. We will also have a few in class quizzes.
[10 weeks of instruction; 20 classes]
Module 1: introduction to deep learning and demos (7 classes).
Module 2: Applications of deep learning to regulatory genomics, variant scoring and population genetics (4 classes)
Module 3: Applications of deep learning to predicting protein structure and pharmacogenomics (3 classes)
Module 4: Applications of deep learning to electronic health records and medical imaging data (4 classes)
Project presentations (Exam period)
|Date||Topic||Primary instructor||papers||paper URLs||Other relevant links|
|9/26||Intro to neural networks, backprop||James||Stegle Review||http://msb.embopress.org/content/12/7/878 , http://neuralnetworksanddeeplearning.com/|
|9/28||Convolutional neural network + intro to func genomics||Anshul||DeepBind
|10/3||Conv nets for genomics and imaging (contd)||Anshul, Serafim||DeepCpG||http://biorxiv.org/content/early/2016/05/27/055715|
|10/5||Interpretation of deep learning models||Avanti||See Files Section|
|10/10||Recurrent neural network + autoencoders + EHR data||James||DeepNano: Deep Recurrent Neural Networks for Base Calling in MinION Nanopore Reads||http://arxiv.org/abs/1603.09195|
|10/12||Training deep neural networks + protein structures||James|
|10/17||Azure, Tensorflow and Keras demo||James|
The human splicing code reveals new insights into the genetic determinants of disease
Learning structure in gene expression data using deep
|10/31||PopGen, Small molecules||James, Serafim||
Deep Learning for Pop Gen Inference
Automatic chemical design using a data-driven continuous representation of molecules
Protein secondary structure prediction using deep convolutional neural fields
|11/7||Project proposal||James, Serafim||project proposal lightning talks||https://github.com/greenelab/deep-review/issues/45|
Convolutional LSTM Networks for Subcellular Localization of Proteins
Protein contact map prediction using ultra deep residual nets
|DeepTox: Toxicity Prediction using Deep Learning: http://journal.frontiersin.org/article/10.3389/fenvs.2015.00080/full|
Molecular Graph Convolutions: Moving Beyond Fingerprints
AtomNet: A Deep Convolutional Neural Network for Bioactivity Prediction in Structure-based Drug Discovery
Written presentation only:
|11/16||Med Records/Clinical data||Anshul, Serafim||
Deep Patient: An Unsupervised Representation to Predict the Future of Patients from the Electronic Health Records;
Deep Kalman Filters
|11/28||Med Records/Clinical data||Serafim, James||
DeepCare: A Deep Dynamic Memory Model for Predictive Medicine
Deep Survival analysis
|11/30||Medical Imaging||Anshul, James||
DeepCyTOF: Automated Cell Classification of Mass Cytometry Data by Deep Learning and Domain Adaptation
Deep Learning for Identifying Metastatic Breast Cancer
Efficient Multi-Scale 3D CNN with Fully Connected CRF for Accurate Brain Lesion Segmentation
|Exam period||Project presentations||Anshul, James|
The syllabus page shows a table-oriented view of the course schedule, and the basics of course grading. You can add any other comments, notes, or thoughts you have about the course structure, course policies or anything else.
To add some comments, click the "Edit" link at the top.