Transformer-based artificial neural networks for the conversion between chemical notations

Area of Science:

• Computational chemistry
• Artificial intelligence in chemistry

Background:

• Chemical notation systems like SMILES and IUPAC are crucial for representing molecular structures.
• Rule-based systems for translating between these notations can be complex and difficult to maintain.

Purpose of the Study:

• To develop and evaluate a Transformer-based artificial neural network for translating between SMILES and IUPAC chemical notations.
• To demonstrate the feasibility of using neural networks for chemical notation translation in production environments.

Main Methods:

• Development of a Transformer-based neural network architecture.
• Training and evaluation of models for both SMILES to IUPAC (Struct2IUPAC) and IUPAC to SMILES (IUPAC2Struct) translation.
• Comparison of performance against existing rule-based solutions.

Main Results:

• The developed neural network models achieve performance comparable to established rule-based solutions.
• The models demonstrate high accuracy, computational speed, and robustness, making them suitable for production.
• A showcase highlights the model's ability to facilitate rapid development while maintaining accuracy.

Conclusions:

• Neural-based solutions can effectively translate between chemical notations like SMILES and IUPAC.
• These findings suggest that artificial neural networks can replace complex rule-based systems, potentially reducing development costs and effort.
• The study encourages the adoption of neural approaches for chemical informatics tasks.