Self-Referencing Embedded Strings (SELFIES): A 100% robust molecular string representation
Mario Krenn, Florian Haese, AkshatKumar Nigam, Pascal Friederich, Alan Aspuru-Guzik
Machine Learning: Science and Technology, (2020), arXiv.
Major contributors since v1.0.0: Alston Lo and Seyone Chithrananda

A main objective is to use SELFIES as direct input into machine learning models,
in particular in generative models, for the generation of molecular graphs
which are syntactically and semantically valid.

Use pip to install selfies.

pip install selfies

To check if the correct version of selfies is installed, use the following pip command.

pip show selfies

To upgrade to the latest release of selfies if you are using an older version, use the following pip command. Please see the CHANGELOG to review the changes between versions of selfies:

pip install selfies --upgrade 

The documentation can be found on ReadTheDocs. Alternatively, it can be built from the docs/ directory.

The selfies library has eight standard functions:

Please read the documentation for more detailed descriptions of these functions, and to view the advanced functions, which allow users to customize the SELFIES language.

import selfies as sf

benzene = "c1ccccc1"

# SMILES --> SELFIES translation
encoded_selfies = sf.encoder(benzene)  # '[C][=C][C][=C][C][=C][Ring1][Branch1_2]'

# SELFIES --> SMILES translation
decoded_smiles = sf.decoder(encoded_selfies)  # 'C1=CC=CC=C1'

len_benzene = sf.len_selfies(encoded_selfies)  # 8

symbols_benzene = list(sf.split_selfies(encoded_selfies))
# ['[C]', '[=C]', '[C]', '[=C]', '[C]', '[=C]', '[Ring1]', '[Branch1_2]']

In this example we first build an alphabet from a dataset of SELFIES, and then convert a SELFIES into a padded, label-encoded representation. Note that we use the '[nop]' (no operation) symbol to pad our SELFIES, which is a special SELFIES symbol that is always ignored and skipped over by selfies.decoder, making it a useful padding character.

import selfies as sf

dataset = ['[C][O][C]', '[F][C][F]', '[O][=O]', '[C][C][O][C][C]']
alphabet = sf.get_alphabet_from_selfies(dataset)
alphabet.add('[nop]')  # '[nop]' is a special padding symbol
alphabet = list(sorted(alphabet))
print(alphabet)  # ['[=O]', '[C]', '[F]', '[O]', '[nop]']

pad_to_len = max(sf.len_selfies(s) for s in dataset)  # 5
symbol_to_idx = {s: i for i, s in enumerate(alphabet)}

# SELFIES to label encode
dimethyl_ether = dataset[0]  # '[C][O][C]'

# [1, 3, 1, 4, 4]
print(sf.selfies_to_encoding(dimethyl_ether,
                             vocab_stoi=symbol_to_idx,
                             pad_to_len=pad_to_len,
                             enc_type='label'))

• More examples can be found in the examples/ directory, including a variational autoencoder that runs on the SELFIES language.
• This ICLR2020 paper used SELFIES in a genetic algorithm to achieve state-of-the-art performance for inverse design, with the code here.

SELFIES uses pytest with tox as its testing framework. All tests can be found in the tests/ directory. To run the test suite for SELFIES, install tox and run:

tox

By default, SELFIES is tested against a random subset (of size dataset_samples=100000) on various datasets:

• 130K molecules from QM9
• 250K molecules from ZINC
• 50K molecules from non-fullerene acceptors for organic solar cells
• 8K molecules from Tox21 in MoleculeNet
• 93K molecules from PubChem MUV in MoleculeNet
• 27M molecules from the eMolecules Plus Database. Due to its large size, this dataset is not included on the repository. To run tests on it, please download the dataset in the tests/test_sets directory and enable its pytest at tests/test_on_emolecules.py.

Other tests are random and repeated trials number of times. These can be specified as arguments

tox -- --trials 100 --dataset_samples 100

where --trials=100000 and --dataset_samples=100000 by default. Note that if dataset_samples is negative or exceeds the length of the dataset, the whole dataset is used.

See CHANGELOG.

We thank Jacques Boitreaud, Andrew Brereton, Matthew Carbone (x94carbone), Nathan Frey (ncfrey), Theophile Gaudin, Hyunmin Kim (hmkim), Minjie Li, Vincent Mallet, Kevin Ryan (LeanAndMean), Benjamin Sanchez-Lengeling, and Zhenpeng Yao for their suggestions and bug reports, and Robert Pollice for chemistry advices.

Apache License 2.0