Skip to content

Robust representation of semantically constrained graphs, in particular for molecules in chemistry


Notifications You must be signed in to change notification settings


Repository files navigation


GitHub release versions License Maintenance GitHub issues Documentation Status GitHub contributors

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 1, 045024 (2020), extensive blog post January 2021.
Talk on youtube about SELFIES.
A community paper with 31 authors on SELFIES and the future of molecular string representations.
Blog explaining SELFIES in Japanese language
Code-Paper in February 2023
SELFIES in Wolfram Mathematica (since Dec 2023)
Major contributors of v1.0.n: Alston Lo and Seyone Chithrananda
Main developer of v2.0.0: Alston Lo
Chemistry Advisor: Robert Pollice

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.

SELFIES validity in a VAE latent space


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, before upgrading:

pip install selfies --upgrade



Please refer to the documentation, which contains a thorough tutorial for getting started with selfies and detailed descriptions of the functions that selfies provides. We summarize some key functions below.

Function Description
selfies.encoder Translates a SMILES string into its corresponding SELFIES string.
selfies.decoder Translates a SELFIES string into its corresponding SMILES string.
selfies.set_semantic_constraints Configures the semantic constraints that selfies operates on.
selfies.len_selfies Returns the number of symbols in a SELFIES string.
selfies.split_selfies Tokenizes a SELFIES string into its individual symbols.
selfies.get_alphabet_from_selfies Constructs an alphabet from an iterable of SELFIES strings.
selfies.selfies_to_encoding Converts a SELFIES string into its label and/or one-hot encoding.
selfies.encoding_to_selfies Converts a label or one-hot encoding into a SELFIES string.


Translation between SELFIES and SMILES representations:

import selfies as sf

benzene = "c1ccccc1"

# SMILES -> SELFIES -> SMILES translation
    benzene_sf = sf.encoder(benzene)  # [C][=C][C][=C][C][=C][Ring1][=Branch1]
    benzene_smi = sf.decoder(benzene_sf)  # C1=CC=CC=C1
except sf.EncoderError:
    pass  # sf.encoder error!
except sf.DecoderError:
    pass  # sf.decoder error!

len_benzene = sf.len_selfies(benzene_sf)  # 8

symbols_benzene = list(sf.split_selfies(benzene_sf))
# ['[C]', '[=C]', '[C]', '[=C]', '[C]', '[=C]', '[Ring1]', '[=Branch1]']

Very simple creation of random valid molecules:

A key property of SELFIES is the possibility to create valid random molecules in a very simple way -- inspired by a tweet by Rajarshi Guha:

import selfies as sf
import random

alphabet=sf.get_semantic_robust_alphabet() # Gets the alphabet of robust symbols
rnd_selfies=''.join(random.sample(list(alphabet), 9))

These simple lines gives crazy molecules, but all are valid. Can be used as a start for more advanced filtering techniques or for machine learning models.

Integer and one-hot encoding SELFIES:

In this example, we first build an alphabet from a dataset of SELFIES strings, and then convert a SELFIES string into its padded encoding. 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))  # ['[=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)}

dimethyl_ether = dataset[0]  # [C][O][C]

label, one_hot = sf.selfies_to_encoding(
# label = [1, 3, 1, 4, 4]
# one_hot = [[0, 1, 0, 0, 0], [0, 0, 0, 1, 0], [0, 1, 0, 0, 0], [0, 0, 0, 0, 1], [0, 0, 0, 0, 1]]

Customizing SELFIES:

In this example, we relax the semantic constraints of selfies to allow for hypervalences (caution: hypervalence rules are much less understood than octet rules. Some molecules containing hypervalences are important, but generally, it is not known which molecules are stable and reasonable).

import selfies as sf

hypervalent_sf = sf.encoder('O=I(O)(O)(O)(O)O', strict=False)  # orthoperiodic acid
standard_derived_smi = sf.decoder(hypervalent_sf)
# OI (the default constraints for I allows for only 1 bond)

relaxed_derived_smi = sf.decoder(hypervalent_sf)
# O=I(O)(O)(O)(O)O (the hypervalent constraints for I allows for 7 bonds)

Explaining Translation:

You can get an "attribution" list that traces the connection between input and output tokens. For example let's see which tokens in the SELFIES string [C][N][C][Branch1][C][P][C][C][Ring1][=Branch1] are responsible for the output SMILES tokens.

selfies = "[C][N][C][Branch1][C][P][C][C][Ring1][=Branch1]"
smiles, attr = sf.decoder(
    selfies, attribute=True)
print('SELFIES', selfies)
print('SMILES', smiles)
for smiles_token in attr:

# output
SELFIES [C][N][C][Branch1][C][P][C][C][Ring1][=Branch1]
AttributionMap(index=0, token='C', attribution=[Attribution(index=0, token='[C]')])
AttributionMap(index=2, token='N', attribution=[Attribution(index=1, token='[N]')])
AttributionMap(index=3, token='C', attribution=[Attribution(index=2, token='[C]')])
AttributionMap(index=5, token='P', attribution=[Attribution(index=3, token='[Branch1]'), Attribution(index=5, token='[P]')])
AttributionMap(index=7, token='C', attribution=[Attribution(index=6, token='[C]')])
AttributionMap(index=8, token='C', attribution=[Attribution(index=7, token='[C]')])

attr is a list of AttributionMaps containing the output token, its index, and input tokens that led to it. For example, the P appearing in the output SMILES at that location is a result of both the [Branch1] token at position 3 and the [P] token at index 5. This works for both encoding and decoding. For finer control of tracking the translation (like tracking rings), you can access attributions in the underlying molecular graph with get_attribution.

More Usages and Examples


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 -- --trials=10000 --dataset_samples=10000

By default, selfies is tested against a random subset (of size dataset_samples=10000) on various datasets:

  • 130K molecules from QM9
  • 250K molecules from ZINC
  • 50K molecules from a dataset of non-fullerene acceptors for organic solar cells
  • 160K+ molecules from various MoleculeNet datasets
  • 36M+ molecules from the eMolecules Database. Due to its large size, this dataset is not included on the repository. To run tests on it, please download the dataset into the tests/test_sets directory and run the tests/ script.

Version History



We thank Jacques Boitreaud, Andrew Brereton, Nessa Carson (supersciencegrl), Matthew Carbone (x94carbone), Vladimir Chupakhin (chupvl), Nathan Frey (ncfrey), Theophile Gaudin, HelloJocelynLu, Hyunmin Kim (hmkim), Minjie Li, Vincent Mallet, Alexander Minidis (DocMinus), Kohulan Rajan (Kohulan), Kevin Ryan (LeanAndMean), Benjamin Sanchez-Lengeling, Andrew White, Zhenpeng Yao and Adamo Young for their suggestions and bug reports, and Robert Pollice for chemistry advices.


Apache License 2.0