Investigation of chemical structure recognition by encoder-decoder models in learning progress.

Area of Science:

• Computational Chemistry
• Machine Learning in Chemistry
• Cheminformatics

Background:

• Encoder-decoder (ED) models utilizing SMILES strings offer continuous chemical descriptors and structural restorability.
• The internal learning mechanisms of ED models regarding chemical structure recognition remain unclear.
• Existing evaluation methods may not fully capture the nuances of ED model performance in descriptor generation.

Purpose of the Study:

• To investigate the relationship between structural information and learning progress in ED models.
• To understand how ED models learn chemical structures from SMILES input during training.
• To assess the sensitivity of current evaluation metrics for ED-based descriptor generation.

Main Methods:

• Development of ED models with varying learning progress.
• Monitoring downstream task accuracy and input-output substructure similarity.
• Utilizing substructure-based descriptors to analyze learned chemical representations.

Main Results:

• Compound substructures are learned early in the ED model training process.
• Structure restoration is a more time-consuming aspect, with insufficient learning leading to inaccurate estimations.
• Determining the precise endpoint of a chemical structure presents a significant challenge for ED models.

Conclusions:

• The study provides the first comprehensive link between ED model learning progress for SMILES and chemical structure representation.
• Current downstream task accuracy metrics may be insufficient for evaluating ED models in descriptor generation.
• Further research is needed to improve the structure restoration capabilities of ED models.