SELFIES

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.

<p align="center"> <img src="https://github.com/aspuru-guzik-group/selfies/blob/master/examples/VAE_LS_Validity.png" alt="SELFIES validity in a VAE latent space" width="666px"> </p>

Installation

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

Usage

Overview

Please refer to the documentation in our code-paper, 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. |

Examples

Translation between SELFIES and SMILES representations:

import selfies as sf

benzene = "c1ccccc1"

# SMILES -> SELFIES -> SMILES translation
try:
    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))
rnd_smiles=sf.decoder(rnd_selfies)
print(rnd_smiles)

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(
   selfies=dimethyl_ether,
   vocab_stoi=symbol_to_idx,
   pad_to_len=pad_to_len,
   enc_type="both"
)
# 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)

sf.set_semantic_constraints("hypervalent")
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)
print('Attribution:')
for smiles_token in attr:
    print(smiles_token)

# output
SELFIES [C][N][C][Branch1][C][P][C][C][Ring1][=Branch1]
SMILES C1NC(P)CC1
Attribution:
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

• 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 allows for [highly efficient exploration and interpolation of the chemical space](https://chemrxiv.org/articles/preprint/Beyond_Generative_Models_Superfast_Traversal_Optimization_Novel