Skills › Scientific Agent Skills › molfeat › available_featurizers.md
available_featurizers.md
77 lines · skills/molfeat/references/available_featurizers.md at 5f2bc03
Most marks come from matching words in the file, so treat them as pointers.
1# Available Molfeat 1.0 representations
3Use this as a checked starting catalog, not a claim that every historical model-store
4entry runs on the current release. Consult the [1.0 calculators](https://github.com/datamol-io/molfeat/tree/1.0.0/molfeat/calc)
5and [pretrained adapters](https://github.com/datamol-io/molfeat/tree/1.0.0/molfeat/trans/pretrained).
Available Molfeat 1.0 representations
Use this as a checked starting catalog, not a claim that every historical model-store entry runs on the current release. Consult the 1.0 calculators and pretrained adapters.
Raw lines 1 to 67## Fingerprints
9All names below are accepted by `FPCalculator`. The local CPU probe ran each family
10except external MAP4. Defaults refer to `FPCalculator`, not `FPVecTransformer`.
12| Name | Default width | Interpretation / settings |
13| --- | ---: | --- |
14| `ecfp`, `fcfp` | 2048 | Radius 2; ECFP atom invariants versus FCFP feature invariants; chirality off by default |
15| `ecfp-count`, `fcfp-count` | 2048 | Folded count vectors |
17| `atompair`, `atompair-count` | 2048 | Topological distances by default (`use2D=True`); not inherently 3D |
18| `topological`, `topological-count` | 2048 | Topological torsions, default four atoms |
19| `maccs` | 167 | 166 keys plus unused bit zero |
20| `avalon`, `avalon-count` | 512 | Avalon substructure fingerprint |
21| `pattern`, `layered` | 2048 | RDKit pattern / layered fingerprint |
22| `secfp` | 2048 | Folded molecular shingles using RDKit's MHFP encoder |
23| `erg` | 315 | Reduced-graph features; not simply a binary bit vector |
24| `estate` | 79 | E-State atom-type counts |
25| `map4` | 2048 | External `map4.MAP4Calculator` required; not installed by Molfeat extras |
27For current supported parameters call `FPCalculator.default_parameters(name)`.
28MAP4's standalone package has its own defaults; do not transfer them to Molfeat.
29Its external backend was source-reviewed but not installed or exercised.
Fingerprints
All names below are accepted by FPCalculator. The local CPU probe ran each family except external MAP4. Defaults refer to FPCalculator, not FPVecTransformer.
| Name | Default width | Interpretation / settings |
|---|---|---|
ecfp, fcfp | 2048 | Radius 2; ECFP atom invariants versus FCFP feature invariants; chirality off by default |
ecfp-count, fcfp-count | 2048 | Folded count vectors |
rdkit, rdkit-count | 2048 | Path fingerprints |
atompair, atompair-count | 2048 | Topological distances by default (use2D=True); not inherently 3D |
topological, topological-count | 2048 | Topological torsions, default four atoms |
maccs | 167 | 166 keys plus unused bit zero |
avalon, avalon-count | 512 | Avalon substructure fingerprint |
pattern, layered | 2048 | RDKit pattern / layered fingerprint |
secfp | 2048 | Folded molecular shingles using RDKit's MHFP encoder |
erg | 315 | Reduced-graph features; not simply a binary bit vector |
estate | 79 | E-State atom-type counts |
map4 | 2048 | External map4.MAP4Calculator required; not installed by Molfeat extras |
For current supported parameters call FPCalculator.default_parameters(name). MAP4's standalone package has its own defaults; do not transfer them to Molfeat. Its external backend was source-reviewed but not installed or exercised.
31## Descriptors and pharmacophores
33| Factory name | Concrete calculator | Requirements / contract |
34| --- | --- | --- |
35| `desc2D` | `RDKitDescriptors2D` | Named 2D features; 223 in the tested RDKit build, not a stable count |
36| `desc3D` | `RDKitDescriptors3D` | Conformer coordinates required; inspect `columns` and `len(calc)` |
37| `mordred` | `MordredDescriptors` | `molfeat[mordred]` installs mordredcommunity; default `ignore_3D=True`; descriptor count and finite values depend on configuration |
38| `cats2D` | `CATS(use_3d_distances=False)` | 21 pair types × 9 default bins = 189 values |
39| `cats3D` | `CATS(use_3d_distances=True)` | Conformer needed; 21 × 6 default bins = 126 values |
40| `pharm2D` | `Pharmacophore2D` | `factory="default"`, `"cats"`, `"gobbi"`, or `"pmapper"`; default folded length 2048 |
41| `pharm3D` | `Pharmacophore3D` | Conformers/pmapper; raw output in 1.0 is a sorted 1D integer array |
42| `usr`, `usrcat` | `USRDescriptors` | 12 / 60 shape values; conformer required |
43| `electroshape` | `ElectroShapeDescriptors` | 15 values; conformer and declared charge model |
44| `scaffoldkeys`, `skeys`, `scaffkeys` | `ScaffoldKeyCalculator` | 42 scaffold descriptors |
46Import `CATS` from `molfeat.calc`; `CATSCalculator(mode="2D")` is not the API.
47Use `max_dist`, `bins`, `scale`, and `use_3d_distances`. Extra keywords to CATS can
48be silently ignored, making a mistaken `mode="3D"` scientifically dangerous.
49`gobbi2D`, `pmapper2D`, `rdkit2D`, and `cats2D_pharm` are not factory aliases.
50For Gobbi use `Pharmacophore2D(factory="gobbi")`.
52Atom/bond feature calculators live in `molfeat.calc.atom` and `molfeat.calc.bond`.
53They return graph-related feature structures, not fixed-length molecular fingerprints.
54Use `molfeat[pyg]` and the maintained graph transformers for PyTorch Geometric;
55`atom-onehot` and `bond-default` are not `get_calculator` names.
Descriptors and pharmacophores
| Factory name | Concrete calculator | Requirements / contract |
|---|---|---|
desc2D | RDKitDescriptors2D | Named 2D features; 223 in the tested RDKit build, not a stable count |
desc3D | RDKitDescriptors3D | Conformer coordinates required; inspect columns and len(calc) |
mordred | MordredDescriptors | molfeat[mordred] installs mordredcommunity; default ignore_3D=True; descriptor count and finite values depend on configuration |
cats2D | CATS(use_3d_distances=False) | 21 pair types × 9 default bins = 189 values |
cats3D | CATS(use_3d_distances=True) | Conformer needed; 21 × 6 default bins = 126 values |
pharm2D | Pharmacophore2D | factory="default", "cats", "gobbi", or "pmapper"; default folded length 2048 |
pharm3D | Pharmacophore3D | Conformers/pmapper; raw output in 1.0 is a sorted 1D integer array |
usr, usrcat | USRDescriptors | 12 / 60 shape values; conformer required |
electroshape | ElectroShapeDescriptors | 15 values; conformer and declared charge model |
scaffoldkeys, skeys, scaffkeys | ScaffoldKeyCalculator | 42 scaffold descriptors |
Import CATS from molfeat.calc; CATSCalculator(mode="2D") is not the API. Use max_dist, bins, scale, and use_3d_distances. Extra keywords to CATS can be silently ignored, making a mistaken mode="3D" scientifically dangerous. gobbi2D, pmapper2D, rdkit2D, and cats2D_pharm are not factory aliases. For Gobbi use Pharmacophore2D(factory="gobbi").
Atom/bond feature calculators live in molfeat.calc.atom and molfeat.calc.bond. They return graph-related feature structures, not fixed-length molecular fingerprints. Use molfeat[pyg] and the maintained graph transformers for PyTorch Geometric; atom-onehot and bond-default are not get_calculator names.
57## Pretrained models and licensing
59- `PretrainedHFTransformer`: ChemBERTa, ChemGPT, MolT5 and other registered Hugging
60 Face models. Query exact current card names instead of assuming availability.
61 The publisher's ChemBERTa-77M-MLM config has hidden size **384**, not 768;
62 pooling and concatenated layers determine the final width.
63- `FCDTransformer`: ChemNet embeddings via `molfeat[fcd]`. An embedding alone is
64 not a Fréchet distance; distribution comparisons need a separate statistical step.
65- `CheMeleonTransformer`: maintained small-molecule graph adapter, checkpoint fetched
66 from the authors; documented width 2048. No additional Chemprop runtime is needed.
67- `MolJEPATransformer`: requires transformer/PyG extras, pinned remote code, and explicit
68 noncommercial-license acceptance. Check upstream terms before loading.
70DGL/DGLLife GIN and legacy Graphormer adapters were **removed in 1.0**. Their old cards
71are historical metadata. Do not try `molfeat[dgl]`, `molfeat[graphormer]`, or instantiate
72the pretrained base class with one of their names. Protein adapters are also removed.
74Pretrained and optional-backend inference was not executed during this review.
75Sources: [foundation-model guide](https://github.com/datamol-io/molfeat/blob/1.0.0/docs/foundation_models.md),
76[ChemBERTa publisher config](https://huggingface.co/DeepChem/ChemBERTa-77M-MLM/blob/main/config.json),
77[MAP4 upstream source](https://github.com/reymond-group/map4/blob/master/map4/map4.py).
Pretrained models and licensing
PretrainedHFTransformer: ChemBERTa, ChemGPT, MolT5 and other registered Hugging Face models. Query exact current card names instead of assuming availability. The publisher's ChemBERTa-77M-MLM config has hidden size 384, not 768; pooling and concatenated layers determine the final width.FCDTransformer: ChemNet embeddings viamolfeat[fcd]. An embedding alone is not a Fréchet distance; distribution comparisons need a separate statistical step.CheMeleonTransformer: maintained small-molecule graph adapter, checkpoint fetched from the authors; documented width 2048. No additional Chemprop runtime is needed.MolJEPATransformer: requires transformer/PyG extras, pinned remote code, and explicit noncommercial-license acceptance. Check upstream terms before loading.
DGL/DGLLife GIN and legacy Graphormer adapters were removed in 1.0. Their old cards are historical metadata. Do not try molfeat[dgl], molfeat[graphormer], or instantiate the pretrained base class with one of their names. Protein adapters are also removed.
Pretrained and optional-backend inference was not executed during this review. Sources: foundation-model guide, ChemBERTa publisher config, MAP4 upstream source.
Raw lines 57 to 77