Canonical recipes

Directory that contains all the prime examples of Anvil recipe.yamls for each model in the OpenADMET suite.

Current available models are:

canonical_recipes/
├── multitask/
  ├── chemeleon
  └── chemprop
├── single_task/
  ├── chemeleon
  ├── chemprop
  ├── lgbm
  ├── tabpfn # still in progress
  └── xgboost
  └── catboost
  └── random forest
  └── dummy

Models that should really only be run on GPU are:

• Chemeleon (DL)
• Chemprop (DL)
• TabPFN (huge featurizer)

Data splitting (random vs. scaffold)

Every recipe in this repo uses a random split by default:

split:
  type: ShuffleSplitter
  params:
    random_seed: 42
    train_size: 0.7
    val_size: 0.1
    test_size: 0.2

ShuffleSplitter wraps sklearn's train_test_split, so train and test are drawn from the same random shuffle. Molecules sharing a scaffold can land on both sides of the split, which lets the model "recognize" near-duplicates of training compounds at test time. The reported metrics are therefore an optimistic estimate of how the model generalizes to genuinely novel chemistry. This is a fine default for runnable examples and quick iteration, but it is not the right benchmark for prospective generalization.

Making a scaffold-split variant

To estimate generalization to unseen chemical series, copy a recipe and change only the splitter type — the params (fractions, seed) stay the same:

split:
  type: ScaffoldSplitter   # was: ShuffleSplitter
  params:
    random_seed: 42
    train_size: 0.7
    val_size: 0.1
    test_size: 0.2

The structure-aware splitters registered in openadmet.models.split are:

Splitter	Behavior
ShuffleSplitter	Random (default; optimistic)
ScaffoldSplitter	Holds out whole Bemis–Murcko scaffolds — no scaffold appears in both train and test
PerimeterSplitter	Pushes the most dissimilar pairs into the test set
MaxDissimilaritySplitter	Test set maximally dissimilar from train
ClusterSplitter	Splits by chemical cluster

These operate on the SMILES input_col, so they work for any recipe whose data.input_col is a SMILES column.

Caveat — cross-validation stays random. The structure-aware splitters apply only to the train/val/test holdout split. The *RepeatedKFoldCrossValidation evaluators in the report.eval block always run a random k-fold over the full dataset (they are not scaffold-aware, and the splitters log a "... is not available for cross-validation" warning). So in a scaffold-split variant, treat the holdout test metrics as the out-of-distribution estimate and read the CV metrics as random-split (optimistic) numbers.