[DL Edition] T034: GNN based molecular property prediction

[PaulaKramer]

Details

• Talktorial ID: 034
• Title: [DL Edition] T034: GNN based molecular property prediction
• Original authors: Paula Kramer
• Reviewer(s): XXX
• Date of review: DD-MM-YYYY

Content

• One line summary: Introduction to Graph Neural Networks for Property Prediction
• Potential labels or categories (e.g. machine learning, small molecules, online APIs): Machine learning, small molecules, graph neural networks
• Time it took to execute (approx.): 7 min
• I have used the talktorial template and followed the content and formatting suggestions there
• Packages must be open-sourced and should be installable from conda-forge. If you are adding new packages to the TeachOpenCADD environment, please check if already installed packages can perform the same functionality and if not leave a sentence explaining why the new addition is needed. If the new package is not on conda-forge, please list them and their intended usage here.
  • numpy, matplotlib: Already in TeachOpenCADD
  • pytorch 1.12.1, pytorch-cluster 1.6.0, pytorch-scatter 2.1.0, pytorch-sparse 0.6.15, pyg 2.2.0 (conda-forge): I use it for implementing graph neural networks
• Data must be publicly available, preferably accessible via a webserver or downloadable via a URL. Please list the data resources that you use and how to access them:
  • QM9 dataset: Access via (torch-geometric)

Content style

• Talktorial includes cross-references to other talktorials if applicable
• The table of contents reflects the talktorial story-line; order of #, ##, ### headers is correct
• URLs are linked with meaningful words, instead of pasting the URL directly or linking words like here.
• I have spell-checked the notebook
• Images have enough resolution to be rendered with quality, without being too heavy.
• All figures have a description
• Markdown cell content is still in-line with code cell output (whenever results are discussed)
• I have checked that cell outputs are not incredibly long (this applies also to DataFrames)
• Formatting looks correctly on the Sphinx render (bold, italics, figure placing)

Code style

• Variable and function names follow snake case rules (e.g. a_variable_name vs aVariableName)
• Spacing follows PEP8 (run Black on the code cells if needed)
• Code line are under 99 characters each (run black-nb -l 99)
• Comments are useful and well placed
• There are no unpythonic idioms like for i in range(len(list)) (see slides)
• All 3rd party dependencies are listed at the top of the notebook
• I have marked all code cell with output referenced in markdown cells with the label # NBVAL_CHECK_OUTPUT
• I have identified potential candidates for a code refactor / useful functions
• All import ... lines are at the top (practice part) cell, ordered by standard library / 3rd party packages / our own (teachopencadd.*)
• I have used absolute paths instead of relative paths

  HERE = Path(_dh[-1])
  DATA = HERE / "data"

Website

We present our talktorials on our TeachOpenCADD website (https://projects.volkamerlab.org/teachopencadd/), so we have to check as well if the Jupyter notebook renders nicely there.

• If this PR adds a new talktorial, please follow these steps:
  • Add your talktorial to the complete list of talktorials here (at the end).
  • Add your talktorial to one or multiple of the collections here. Or propose a new collection section in your PR.
  • Add your talktorial's nblink file by running python generate_nblinks.py from within the directory teachopencadd/docs/talktorials.
  • Please complile the website following the instructions here.
• Check the rendering of the talktorial of this PR.
• Is your talktorial listed in the talktorial list?
• Is your talktorial listed in the talktorial collections?
  • Add a picture for your talktorial in the collection view by following these instructions.

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[gerritgr]

• GNNs should be defined first as differentiable and trainable, permutation equi(/in)variant functions. The architectures should be introduced as specific instances of such functions.  
• The relationship between massage passing as a general and powerful framework and GCN/GIN as (less powerful) instances could be clarified: Also the relationship between the aggregation/pooling function (input is a set) and the permutation invariance. 
• refer to T33 in the introduction.
• d is overloaded with the degree and the feature dimension.
• The advantages of a GNN library should be stated (sparse matrices, graph batching), also mention www.dgl.ai.
• bessere property?  -> ChatGPT suggests: Electronegativity, Ionization potential, Bond angles and distances, no idea if they make sense, though.
• Say explicitly that the pooling layer is invariant to the order of the input (the same as the aggregation function).
• True vs predicted value -> would say Ground truth vs prediction
• Can you give an intuition on what makes GIN more powerful?