[Paper Review] SELFIES: a robust representation of semantically constrained graphs with an example application in chemistry.

Mario Krenn, Florian Häse+3 more|arXiv (Cornell University)|May 31, 2019

Computational Drug Discovery MethodsComputer Science43 references54 citations

TL;DR

This paper introduces SELFIES, a 100%-valid string-based molecular representation that guarantees every generated string corresponds to a chemically valid molecule. By encoding molecular structures through a hierarchical, self-referential grammar, SELFIES enables robust generative machine learning in chemistry, increasing model memory diversity by two orders of magnitude and enabling interpretable, valid molecule generation without post-processing.

ABSTRACT

The discovery of novel materials and functional molecules can help to solve some of society's most urgent challenges, ranging from efficient energy harvesting and storage to uncovering novel pharmaceutical drug candidates. Traditionally matter engineering -- generally denoted as inverse design -- was based massively on human intuition and high-throughput virtual screening. The last few years have seen the emergence of significant interest in computer-inspired designs based on evolutionary or deep learning methods. The major challenge here is that the standard strings molecular representation SMILES shows substantial weaknesses in that task because large fractions of strings do not correspond to valid molecules. Here, we solve this problem at a fundamental level and introduce SELFIES (SELF-referencIng Embedded Strings), a string-based representation of molecules which is 100\% robust. Every SELFIES string corresponds to a valid molecule, and SELFIES can represent every molecule. SELFIES can be directly applied in arbitrary machine learning models without the adaptation of the models; each of the generated molecule candidates is valid. In our experiments, the model's internal memory stores two orders of magnitude more diverse molecules than a similar test with SMILES. Furthermore, as all molecules are valid, it allows for explanation and interpretation of the internal working of the generative models.

Motivation & Objective

• Address the fundamental limitation of SMILES, where a large fraction of generated strings are invalid molecules, by creating a representation that guarantees validity.
• Enable reliable and efficient inverse molecular design using machine learning by ensuring all generated candidates are chemically valid from the start.
• Support diverse and memory-efficient exploration of chemical space in generative models without validation filtering.
• Facilitate interpretation of generative model behavior by eliminating invalid molecule candidates from the search space.
• Provide a universal, grammar-based string representation that can be directly used in any machine learning model without architectural modifications.

Proposed method

• Design a hierarchical, self-referential grammar that encodes molecular structures using a recursive, context-free approach to ensure syntactic validity.
• Represent molecules as strings using a fixed set of production rules that enforce valency and connectivity constraints at the grammar level.
• Encode molecular substructures using self-referencing tokens that allow compact representation of complex fragments.
• Construct a string representation where every possible string corresponds to a unique, valid molecule by design.
• Integrate SELFIES directly into existing machine learning models without requiring model retraining or architectural changes.
• Use the grammar-based structure to enable efficient search and generation of molecules with high diversity in latent space.

Experimental results

Research questions

• RQ1 Can a string-based molecular representation be constructed such that every possible string corresponds to a valid molecule?
• RQ2 How does SELFIES compare to SMILES in terms of memory efficiency and diversity of generated molecules in generative models?
• RQ3 To what extent does using a 100%-valid representation improve the interpretability and reliability of machine learning models in molecular generation?
• RQ4 Can SELFIES be seamlessly integrated into existing deep learning frameworks without modifying model architecture?
• RQ5 Does the use of SELFIES lead to a significant increase in the number of unique, valid molecules explored during model training?

Key findings

• Every SELFIES string corresponds to a valid molecule, ensuring 100% validity without post-processing or filtering.
• The model's internal memory stored two orders of magnitude more diverse molecules compared to a similar SMILES-based model.
• SELFIES enables direct application in any machine learning model without architectural adaptation, simplifying integration.
• The representation supports the generation of complex, valid molecular structures through a self-referential grammar that enforces chemical valency and connectivity.
• The use of SELFIES allows for clearer interpretation of generative model behavior due to the absence of invalid candidates in the search space.
• SELFIES can represent every possible molecule, making it a universal and complete representation for molecular space.