Root-aligned SMILES: a tight representation for chemical reaction prediction

Area of Science:

• Computational Chemistry
• Organic Synthesis
• Machine Learning in Chemistry

Background:

• Chemical reaction prediction is crucial for organic synthesis, encompassing forward and retrosynthesis prediction.
• Current methods often use Simplified Molecular Input Line Entry System (SMILES) for molecule representation in sequence-to-sequence models.
• Standard SMILES overlooks the conserved molecular graph topology in chemical reactions, leading to suboptimal prediction performance.

Purpose of the Study:

• To propose a new molecule representation, root-aligned SMILES (R-SMILES), for enhanced chemical synthesis prediction.
• To improve the efficiency and accuracy of computational models for predicting chemical reactions.

Main Methods:

• Developed root-aligned SMILES (R-SMILES) to establish a strict one-to-one mapping between reactant and product SMILES.
• Utilized R-SMILES in a sequence-to-sequence framework, reducing the model's reliance on learning complex syntax.
• Compared R-SMILES performance against state-of-the-art baselines in chemical reaction prediction tasks.

Main Results:

• R-SMILES demonstrates a significantly reduced edit distance between reactant and product representations due to its strict alignment.
• The computational model, using R-SMILES, can focus more on learning chemical reaction knowledge rather than molecular representation syntax.
• R-SMILES achieved superior performance compared to all evaluated state-of-the-art methods in chemical synthesis prediction.

Conclusions:

• Root-aligned SMILES (R-SMILES) offers a more effective molecular representation for chemical reaction prediction.
• The proposed R-SMILES method significantly advances the state-of-the-art in computational organic synthesis prediction.
• This approach facilitates more efficient and accurate prediction of chemical transformations.