Installation

NFP can be installed with pip into a suitable python environment with

pip install nfp

The only dependency that is not pip-installable is rdkit, which can be installed via conda-forge.

Tutorials

Tokenizer usage examples

The Tokenizer class allows transforming arbitrary inputs into integer classes

from nfp.preprocessing import Tokenizer

tokenizer = Tokenizer()
tokenizer.train = True

The 0 and 1 classes are reserved for the <MASK> and missing labels, respectively.

[tokenizer(item) for item in ['A', 'B', 'C', 'A']]

[2, 3, 4, 2]

When train is set to False, unknown items are assigned the missing label

tokenizer.train = False
[tokenizer(item) for item in ['A', 'D']]

[2, 1]

The total number of seen classes is available from the num_classes property, useful to initializing embedding layer weights.

tokenizer.num_classes

4

nfp

Exceptions

MissingDependencyException

nfp.frameworks
nfp.layers	Keras layers that handle the node, edge, and connectivity features returned by the preprocessor classes.
nfp.models	Tensorflow loss functions that accept masked true values (with np.nan)
nfp.preprocessing	Utilities for converting graphs into sets of tensors for machine learning models

Neural Fingerprint (nfp), tf.keras layers for molecular graphs

A collection of tensorflow.keras layers, losses, and preprocessors for building graph neural networks on molecules.