DECIMER: towards deep learning for chemical image recognition.

Area of Science:

• Cheminformatics
• Artificial Intelligence
• Computational Chemistry

Background:

• Automatic recognition of chemical structures from literature is crucial for chemical data mining and open-access databases.
• Existing methods often rely on specific assumptions about chemical diagram structures.

Purpose of the Study:

• To develop and evaluate Deep lEarning for Chemical ImagE Recognition (DECIMER), a novel deep learning approach for chemical structure image recognition.
• To assess the performance of DECIMER using different chemical representation formats.

Main Methods:

• Utilized a deep learning model based on show-and-tell neural networks.
• Trained the model to translate bitmap chemical structure images into SMILES strings.
• Compared the efficacy of SMILES, DeepSMILES, and SELFIES as input data representations.

Main Results:

• DECIMER's current performance is preliminary but shows potential to rival traditional methods with sufficient training.
• DeepSMILES representations demonstrated superiority over standard SMILES.
• SELFIES emerged as a potentially superior representation compared to DeepSMILES.

Conclusions:

• DECIMER offers a flexible deep learning approach for chemical structure recognition with minimal problem-specific assumptions.
• Further training with extensive datasets (50-100 million structures) is projected to yield near-accurate predictions.
• The project emphasizes open-source software and open data accessibility.