Featurizers

Many of the descriptors implemented in mofdscribe have been discussed in our 2020 Chem. Rev. article.

ID	Considers Geometry	Considers Structure Graph	Encodes Chemistry	Scalar	Scope
AccessibleVolume	True	False	False	False	global
AcidGroupCounter	False	True	False	True	bu
AMD	True	False	optionally	False	local
APRDF	True	False	optionally	False	global
Asphericity	True	True	False	True	bu
AtomCenteredPH	True	False	optionally	False	global
BaseGroupCounter	False	True	False	True	bu
BUMatch	True	True	False	True	bu
DiskLikeness	True	True	False	True	bu
Eccentricity	True	True	False	True	bu
EnergyGridHistogram	True	False	True	False	global
GuestCenteredAPRDF	True	False	optionally	False	global
Henry	True	False	True	False	global
InertialShapeFactor	True	True	False	True	bu
LSOP	True	False	False	False	bu
MOFDescriber	True	True	True	False	global
NConf20	True	True	False	True	bu
NPR1	True	True	False	True	bu
NPR2	True	True	False	True	bu
PairwiseDistanceHist	True	False	False	False	bu
PairwiseDistanceStats	True	False	False	False	bu
PartialChargeHistogram	True	False	True	False	global
PartialChargeStats	True	False	True	False	global
PHHist	True	False	optionally	False	global
PHImage	True	False	optionally	False	global
PHStats	True	False	optionally	False	global
PHVect	True	False	optionally	False	global
PMI1	True	True	False	True	bu
PMI2	True	True	False	True	bu
PMI3	True	True	False	True	bu
PoreDiameters	True	False	False	False	global
PoreSizeDistribution	True	False	False	False	global
PriceLowerBound	False	False	True	True	global
RACS	False	True	optionally	False	local
RadiusOfGyration	True	True	False	True	bu
RayTracingHistogram	True	False	False	False	global
RodLikeness	True	True	False	True	bu
SmartsMatchCounter	False	True	False	True	bu
SphereLikeness	True	True	False	True	bu
SpherocityIndex	True	True	False	True	bu
SurfaceArea	True	False	False	False	global
VoxelGrid	True	False	optionally	False	global

Warning

Note that different featurizers have different ways in which they deal with solvent molecules. The RACs and SBU-centered features will ignore floating solvent molecules. Other featurizers, e.g. the APRDF, or pore geometry descriptors will do consider floating solvent molecules in the same way as framework molecules.

If you do not want the solvent molecules to impact the featurization, you’ll have to remove it from the structure.

Bound solvent is not suppressed by any of the featurizers.

To identify bound and unbound solvents, you might find the moffragmentor package useful.

Note

You might be surprised to see new folders appear and disappear in your working directory when you run certain featurizers. This is expected. For some featurizers we need to create temporary files, for which we will create the temporary folders. You can ignore this, they will be cleaned up automatically.

Atom-centered featurizers

A key approximation for machine learning in chemistry is the locality approximation. Effectively, this allows to train models on small fragments which then (hopefully) can be used to predict the properties of larger structures.

• Average minimum distance
• Atom-centered Persistent Homology
• Revised autocorrelation functions (RACs)

Global featurizers

In particular for porous materials, some properties are not local. For instance, the pore geometry (key for gas adsorption) cannot be captured by descriptor that only considers the local environment (of e.g., 3 atoms). For this reason it can make sense to also consider features that consider the full structure as a whole.

• Atomic-property labeled radial distribution function (APRDF)
• Partial charge statistics and histogram
• Topological descriptors
• Pore descriptors
• Price descriptors
• Based on elemental prices
• Molecular simulation derived/inspired
• Text-based
• Voxelization

BU-centered featurizers

Reticular chemistry describes materials built via a tinker-toy approach. Hence, a natural approach is to focus on the building blocks.

mofdscribe can compute descriptors that are BU-centred, for instance, using RDKit descriptors on the building blocks. However, you are not limited to descriptors operating on molecules – you can convert any featurizer into an BU-centered fearturizer:

from matminer.featurizers.site import SOAP
from matminer.featurizers.structure import SiteStatsFingerprint
from pymatgen.core import Structure
from mofdscribe.featurizers.bu import BUFeaturizer

base_feat = SiteStatsFingerprint(SOAP.from_preset("formation_energy"))
base_feat.fit([hkust_structure])
featurizer = BUFeaturizer(base_feat, aggregations=("mean",))
features = featurizer.featurize(structure=hkust_structure)

For this, you can either provide your building blocks that you extracted with any of the available tools, or use our integration with our moffragmentor package. In this case, we will fragment the MOF into its building blocks and then compute the features for each building block and let you choose how you want to aggregate them.

• Distance-based description of BU shape
• Flexibility of BUs
• Local structure order parameters
• RDKit-derived features
• BU Match

Host-Guest Featurization

If you have structures loaded with guest molecules as input, you might find the HostGuestFeaturizer convenient. This featurizer will automatically extract the host and guest structures from the input structures and then featurize them separately. If there are multiple guests in the input structures, the featurizer will featurize each guest separately and then aggregate the features.

from matminer.featurizers.structure.sites import SiteStatsFingerprint

from mofdscribe.featurizers.hostguest import HostGuestFeaturizer

featurizer = HostGuestFeaturizer(
    featurizer=SiteStatsFingerprint.from_preset("SOAP_formation_energy"),
    aggregations=("mean",),
)
featurizer.fit([structure])
features = featurizer.featurize(structure)

If you are interested in surface chemistry features, you might also find suitable featurizers in the matminer package.

mofdscribe also implements some featurizers that are specifically designed for host-guest systems.

• Guest-centered atomic-property labeled radial distribution function (APRDF)

Encoding of chemistry

Many featurizers that traditionally do not capture the chemistry of the structure are implemented in mofdscribe in a way that still allows to capture the chemistry. One example for this are the topology-based descriptors.

We encode the chemistry in those case by computing the descriptor for substructures of subsets of element types; for instance, for the metal- or organic-substructure.