[mahdi-shafiei]

Summary

Molecular graphs carry the AttentiveFP featurization (Xiong et al. 2019), as DGL-LifeSci's
AttentiveFPAtomFeaturizer and AttentiveFPBondFeaturizer (Apache-2.0) define it.

 smiles_to_graph / batch_smiles_to_graphs
-  atoms: element(12) degree 0-6(7) charge -2..2(5) hybridization(4) aromatic H 0-4(5)   = 34
-  bonds: type(4)                                                                         = 4
+  atoms: element(15 + unknown) degree 0-5(6) charge radicals hybridization(5 + unknown)
+         aromatic H 0-4(5) chiral-centre CIP R/S(2)                                      = 39
+  bonds: type(4) conjugated in-ring stereo none/any/Z/E(4)                               = 10

• The value sets are tuples matched against RDKit enum names, so DEFAULT_ATOM_FEATURES and
  DEFAULT_BOND_FEATURES are derived without importing RDKit. Every operator default follows
  them (since fix(drug_discovery): operator input widths default to the featurizer's #48).
• AtomFeatureConfig, DEFAULT_ATOM_CONFIG and ATOM_TYPES are removed; get_atom_features
  takes no configuration. This is a clean break with no alias.
• Docs: the API and user-guide feature tables, the shapes, the config row naming the removed
  class, and the fingerprint example's shape output.

Evidence

Before: on main, test_attentivefp_features.py and the updated test_molecular_graphs.py
fail 19 and pass 6 (the padding, x64 and refusal tests).
After: 25 passed.

ethanol atoms == DGL-LifeSci's documented 'CCO' rows      hot [1,17,26,34] [1,18,26,33] [3,17,26,32]
ethanol bonds == [single, -, -, no stereo] both directions
N[C@@H](C)C(=O)O -> S, N[C@H](C)C(=O)O -> R, both chiral-centre
C/C=C/C -> E, C/C=C\C -> Z
the six BBBP stereoisomer pairs -> distinct graphs
[Na+].[Cl-] -> unknown element and hybridization, charge +1 / -1;  [CH3] -> 1 radical
SF6 sulfur (degree 6) -> degree block all zero, as in the reference

BBBP identical-graph groups (every split, the shipped builder, all groups stereo-only):

	main	this branch
SMILES as MolNet supplies it	6 groups, 12 molecules, 2 with conflicting labels	0
canonical SMILES	20 groups, 41 molecules, 4 with conflicting labels	0

Speed, RTX 4090, jax 0.11.1. Main and the branch alternate, 2 rounds each:

	main	branch
Adam step, nnx.value_and_grad through nnx.vmap, 64 padded BBBP molecules: MolecularPropertyPredictor	41.2, 42.9 ms	46.8, 42.6 ms
same step, CircularFingerprint	15.5, 16.7 ms	15.7, 15.7 ms
same step, AttentiveFP	11.0, 10.8 ms	9.5, 10.4 ms
host batch_smiles_to_graphs, 256 BBBP molecules, CPU backend, best of 15	87.3, 85.3 ms	126.8, 136.6 ms

The training steps are within run-to-run noise at these shapes. The host cost per molecule is
about the same (_host_graph over 256 molecules: 59.6 and 61.3 ms against 66.7 and 71.0 ms). The
batch is slower because its dense (B, N, N, F_bond) edge array grows from 41 MB to 102 MB with
the bond width: the NumPy fill goes from 4.4 to about 12 ms, the transfer to the GPU from about 5
to 13 ms. The cost belongs to the dense layout; the padding change that follows works on that
layout.

Gate: tests/operators/drug_discovery and tests/operators/test_drug_discovery.py 146 passed;
the MolNet and DTI benchmark tests and the experimental-boundary test 21 passed; pre-commit
over all files, ruff, ruff format, lint-imports (6 kept), pyright on the changed files, and
mkdocs build --clean --strict all pass.

Merge Danger

Door: two-way

A revert restores the 34 and 4 widths. Nothing persists graphs or trained weights at these
widths in the repository.

Blast Radius: featurization

Every graph, and every operator built with default widths, changes width; models trained on the
old graphs need retraining. Code that passed a config to get_atom_features, or imported
AtomFeatureConfig, DEFAULT_ATOM_CONFIG or ATOM_TYPES, breaks. Within the repository, the
benchmarks and scripts refer to the widths by symbol only. The random-init numbers printed on
the example pages were already stale on main (the fingerprint page shows 0.4735 where main
computes 0.4362), because scripts/generate_example_outputs.py fails at import on the installed
matplotlib (plt.BarContainer). That defect is left for the docs change.