Log In Sign Up
Models:
Marco Lenoir
MarcoTheBlack/
AIGenomics
0
Downloads: 1
[376c87]: / 2020 / README.md

Download this file

225 lines (111 with data), 8.6 kB

AI for Genomics 2020

Course Description:

Sponsored by the city of Montréal, with support from Mila – Quebec Artificial Intelligence Institute and IVADO (Institut de valorisation des données), the AI in Genomics program is a 12-week training that will allow participants to get hands-on experience in working with machine learning. The program will help to prepare participants with expertise in genomics to develop a foundational knowledge of advanced machine learning methodologies so that they can develop a better understanding of where and how these techniques could be used with genomics data.

Dates: 1/20/2020-4/13/2020

System to Q&A: https://piazza.com/class/k4rmacqhp136ae

Instructors:

  • Joseph Paul Cohen (Program Scientific Advisor)

  • Tariq Daouda

  • Paul Bertin

  • Julie Hussin

  • Ahmad Pesaranghader

  • Sydney Swaine-Simon

Lecture 1: Onboarding and Introduction to neural networks

(Tariq Daouda, January 24th @ 15 h -18 h)

Participants should get a basic understanding of neural networks and deep learning as well as enough practical knowledge to start building neural networks.

  • Datasets

  • Classification

  • KNN

  • Regression

  • Evaluation: Accuracy (train, test, validation)

  • Basics of Backprop (momentum?)

  • Fully connected layers

  • Non-linearities (Relu, tanh, sigmoid)

  • Conv (1D, 2D)

  • pyTorch introduction (Colab)

  • Practical: pyTorch feed forward: Fully connected & Conv

Slides (pdf): link

Slides: link

Lecture 2: Representation learning and backprop

(Joseph Paul Cohen, January 31st @ 15 h -18 h)

Location- John Molson School of Business - S2.445 - Classroom

Deep learning overview, representation learning methods in detail (sammons map, t-sne), the backprop algorithm in detail, and regularization and its impact on optimization.

  • (30min) What is deep learning overview (Slides, Slides (pdf))

    • Define supervised and self-supervised prob perspective

    • How to approach problems (use sklearn)

    • Examples of go-to methods: logistic regression, decision tree etc (use sklearn)

  • (45min) Backprop in more detail (Slides, Slides (pdf))

    • Work through an example of manually performing the algorithm

    • Backpropagation (visualizing the chain rule)

    • Intuition for applying gradient updates for arbitrary functions

  • (1hr) Representation learning (Slides, Slides (pdf))

    • Non-linear dim reduction

    • word2vec

    • Sammons map (tutorial code)

    • t-SNE

    • Regularization

Lecture 3: Challenges of Machine Learning for Transcriptomics

(Paul Bertin, February 7th @ 15 h -18 h)

Challenges facing machine learning and deep learning techniques when applied to transcriptomics: biases, high dimensionality, and interpretability of models. We will dive in the limitations of a machine learning assisted drug effect prediction pipeline and analyse each step to identify the challenges of ML for transcriptomics. (Slides)

  • From real world to input data (45min)

    • Dataset biases

    • Acquisition biases

    • Preprocessing

  • The supervised learning pipeline (45min)

    • The curse of dimensionality

    • Making the right assumptions: inspiration from Computer Vision

    • Which assumptions for transcriptomics?

    • Gene interaction graphs?

    • Parameter sharing among genes?

    • Similar response to perturbation in latent space?

  • Model interpretability (30min)

    • Feature importance for deep models

    • Simpson’s paradox

  • Practical: (30min)

    • Pytorch deep learning pipeline

    • Saliency Maps

    • Deep Dream

Lecture 4: Deep Learning Models in Genomics

(Ahmad Pesaranghader and Julie Hussin, February 14th @ 15 h -18 h)

In the first part of this lecture, we introduce the different DL architectures used in population and functional genomics. In the second part of this lecture, we then introduce generative models and explore how they can be beneficial in the context of genomics, mainly for the augmentation of the training data.

1. Deep learning in population genetics and multi-omics (1h)

  • Introduction to population and functional genomics

  • Simulations in population genetics.

  • Convolutional Neural Networks (CNNs) for population genetics inference

  • Motif-based approaches in functional genomics

  • DeepSEA (link) and state-of-the art models in functional genomics.

2. Advanced deep learning models for genomics (1h30)

  • Variational AutoEncoders (VAEs)

  • Generative Adversarial Networks (GANs)

  • Limitations of vanilla GANs and vanilla VAEs

  • GANs and VAEs in Genomics

  • Discussion of interesting applications in the field mainly with respect to different omics data-types (current state-of-the-art and guideline for future work)

3. Tutorial (30 mins)

  • Quick Implementation of vanilla VAE/GAN in PyTorch (Google Colab)

  • GANs from the Paper: Generating and designing DNA with deep generative models (https://arxiv.org/abs/1712.06148)

Lecture 5: Ethics

(Sydney Swaine-Simon, February 21st @ 15 h -18 h)

In this lecture we will discuss the ethics associated with Genomics data and developing machine learning algorithms.

Some papers and additional resources:

Functional genomics papers:

Population genetics papers:

Torrente, Aurora, et al. "Identification of Cancer Related Genes Using a Comprehensive Map of Human Gene Expression." PLOS ONE, edited by Paolo Provero, vol. 11, no. 6, Public Library of Science, June 2016, p. e0157484, doi:10.1371/journal.pone.0157484.

Ching, Travers, et al. "Opportunities And Obstacles For Deep Learning In Biology And Medicine." Journal of The Royal Society Interface, Cold Spring Harbor Laboratory, Jan. 2018, doi:10.1101/142760.

https://canvas.stanford.edu/courses/51037