Ligand-based progress tracking

[jaimergp]

This issue will track all the progress related to getting the core functionality for ligand-based models merged to master.

Core functionality

• [x] DatasetProvider exports to PyTorch (see #8)
  • [ ] We might need specific exports to PyTorch Geometric (very similar)
• [x] MeasurementType implements accurate observation models (see #8)
• [x] System is composed of typed MolecularComponents: Protein, Ligand, Complexes (see #8)
• [x] Featurizers can operate in batch mode or upon dataset access via DataLoader (see #8)
• [x] Filter invalid records (see note in #8 description)
• [ ] Cross validation

Standard practices

• [ ] Ensure documentation is accurate and up-to-date
  • [ ] Drop mkdocs for Sphinx + material theme?
• [ ] Unit tests have good coverage and are passing

Consolidate previous PRs

• [ ] Bring scripts from #6
• [x] Bring observation model notebook from #14

Pending scientific questions

• [ ] Unknowns. These could be nuisance parameters in the future, but for now:
  • [ ] [substrate]: we could estimate by cross-validation, use relative pIC50s, or add that as a nuisance parameter in the future
  • [ ] uncertainties: estimate per dataset (like Kramer's paper) or measurement class
• [ ] Verify % displacement equations. See discussion

[t-kimber]

Cross-validation in Core functionality could be solved using skorch package. Check PR #27 for some toy example using skorch with observation model :)

[t-kimber]

A few aspects which still need to be addressed:

• [x] When dealing with the XGBoost package, there should be a distinction between the observation model and the custom loss. See this notebook for more details.

• For the full detailed explanation of the observation models, losses, custom losses, and associated differentiation (gradient and hessian), please refer to the overleaf document (maybe useful @jaimergp ).

• [x] If we want to have a ligand-based only model, we can pass None as a kinase featurizer. However, this will not reflect a real ligand-based model, since all the systems would be used in this case. We should instead create one single model for each of the kinases, meaning that if there are N unique kinases in a data set, N models have to trained. -> This will probably be addressed by looping over all kinases.

• [x] Create toy example for dealing with multiple measurement types @t-kimber .

[schallerdavid]

Outdated, will create new issues when needed.