Ligand graph fix

[t-kimber]

Fix issue #9 .

Error:

RuntimeError: stack expects each tensor to be equal size, but got [2, 56] at entry 0 and [2, 50] at entry 1

These shapes correspond to the COO format of the adjacency matrix in pytorch geometric (see graph data documentation).

How to reproduce the bug

First, featurize the compounds using

bash tests/features/test_featurization_ligand_only.sh

This command line will create parquet files which contain the featurized compounds.

Second, train the Graph Neural Network (which fails):

bash tests/experiments/test_model_ligand_only.sh

Notes on Jaime's implementation:

• Data for pytorch geometric specification is here:

  experiments-binding-affinity/experiments/001_example-ligand-only-graph-EGFR-subset.py

  Line 30 in 4034009

  DATALOADER_CLS = "torch_geometric.data.DataLoader"

• In the notebook template, there is a specification for different input shape:

    if COLLATE_FN:
        # IMPORTANT: This will be needed if your X tensors have different shapes across systems!
        # COLLATE_FN can be an import string, or a eval-able lambda
        # Read https://pytorch.org/docs/stable/data.html#dataloader-collate-fn
        try:
            collate_fn = import_object(COLLATE_FN)
        except ImportError:
            collate_fn = eval(COLLATE_FN)

General Notes:

• The ML model we'd like to apply from kinoml can be found here: https://github.com/openkinome/kinoml/blob/afe169e6eefc6fddbaaf81cf720a4e224c82696d/kinoml/ml/torch_geometric_models.py#L11
• Mini-batches in pytorch geometric doc: https://pytorch-geometric.readthedocs.io/en/latest/notes/introduction.html#mini-batches
• A gist that I created as a simple example: https://gist.github.com/t-kimber/eb12d4638bbe3e2e5ebb9f1e6acef69b