PickyBinders/geometric-learning-protein-structures-course

Geometric Deep Learning for Protein Structure Data with PyTorch Lightning

• 10:00 - 10:15: Introduction
• 10:15 - 11:00: Notebook 1 - Proteins as Graphs
• 11:00 - 11:30: Break
• 11:30 - 12:30: Notebook 2 - Graph Datasets and DataLoaders
• 12:30 - 13:30: Lunch
• 13:30 - 13:45: Introduction to geometric deep learning
• 13:45 - 15:00: Notebook 3 - Geometric Deep Learning
• 15:00 - 15:30: Break
• 15:30 - 16:30: Notebook 4 - Training and Tracking
• 16:30 - 17:00: Wrap-up

Link to the google colab search for github repos: https://colab.research.google.com/github/ with the repository: https://github.com/PickyBinders/geometric-learning-protein-structures-course

Objectives

Develop a code-base for exploring, training and evaluating graph deep learning models using protein structures as input for a residue-level prediction task.

• Learn how to featurize protein structures as graphs using Graphein
• Understand the data loading and processing pipeline for graph datasets using PyTorch Geometric
• Learn how to implement graph neural networks using PyTorch Geometric
• Understand the typical deep learning training and evaluation loops using PyTorch Lightning

Task and Dataset

• Given an input protein chain, predict for each residue whether or not it belongs to a protein-protein interface.
• The dataset (in dataset.txt) is a subset of the MaSIF-site dataset.
• Each line is a PDB ID and a chain. We'll use these to extract residues at the interface with other chains and label them as positive examples. All other residues are negative examples.