AstraZeneca / chemicalx

A PyTorch and TorchDrug based deep learning library for drug pair scoring. (KDD 2022)

https://chemicalx.readthedocs.io

Apache License 2.0

700 stars 89 forks source link

Add GCN-BMP #71

Closed mughetto closed 2 years ago

mughetto commented 2 years ago

Summary

First draft of implementation for GCN-BMP. Closes Issue #21 .

[ ] Unit tests provided for these changes
[x] Documentation and docstrings added for these changes

Changes

Add GCN-BMP model. Some parts might be moved in other files/projects:
- Circular correlation function.
- Highway update
- Attention-based pooling/readout
Add example for the model.

Caveats

The number of relationships (for the Relational Conv layers) is configured in the model initialization. It should be done via some dataset-level constant.

Differences with the original paper

In the original GCN-BMP the feature vectors are initialized randomly based on the atomic number of the node. I couldn't find easily the implementation in their code so backed up on the features already provided in the PairedBatches.
Rel Conv layers are followed by a sigmoid activation before the Highway Update. This should only affect the stability.
Highway Update used is the one proposed in the original Highway network paper. Best for reusability imho and shouldn't affect the gating behavour.

Performance

Example's ROC AUC: 0.749128

benedekrozemberczki commented 2 years ago

@cthoyt I overtook this.

cthoyt commented 2 years ago

@benedekrozemberczki let me know when you're done and I'll take a final pass

benedekrozemberczki commented 2 years ago

@cthoyt Ready for review.

codecov-commenter commented 2 years ago

Codecov Report

Merging #71 (42eadbf) into main (2fb3421) will increase coverage by 0.35%. The diff coverage is 97.87%.

@@            Coverage Diff             @@
##             main      #71      +/-   ##
==========================================
+ Coverage   94.01%   94.36%   +0.35%     
==========================================
  Files          29       29              
  Lines         869      959      +90     
==========================================
+ Hits          817      905      +88     
- Misses         52       54       +2

Impacted Files	Coverage Δ
chemicalx/models/gcnbmp.py	`97.61% <97.59%> (-2.39%)`	:arrow_down:
tests/unit/test_models.py	`100.00% <100.00%> (ø)`

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update 2fb3421...42eadbf. Read the comment docs.