MICER: a pre-trained encoder-decoder architecture for molecular image captioning.

Area of Science:

• Computational Chemistry
• Cheminformatics
• Artificial Intelligence

Background:

• Traditional chemical structure recognition methods struggle with stylistic variations in molecular images.
• Deep learning offers flexibility but is under-researched for molecular image analysis.

Purpose of the Study:

• To develop and evaluate MICER, an advanced deep learning model for automated molecular image captioning and structure recognition.
• To assess the impact of various factors on model performance, including stereochemical information and data volume.

Main Methods:

• Developed MICER, an encoder-decoder architecture incorporating transfer learning and attention mechanisms.
• Evaluated model performance across four datasets, analyzing the influence of image features and model architecture.

Main Results:

• MICER demonstrated superior performance compared to state-of-the-art methods on multiple datasets.
• Identified significant impacts of molecular image features and sub-model matching on recognition accuracy.

Conclusions:

• MICER provides a reliable, scalable, and interpretable framework for automated molecular structure identification.
• The findings guide the design of training datasets and encoders for enhanced chemical space exploration.