DECIMER

DECIMER performs automated recognition and segmentation of chemical structure diagrams in scientific literature to enable Optical Chemical Structure Recognition (OCSR) and conversion of images into machine-readable chemical representations.

Key Features:

• Deep learning-based recognition: Uses show-and-tell neural networks to identify chemical structures within images for OCSR.
• Two-stage workflow: Detection generates masks that delineate structure positions and a post-processing step refines incomplete masks to ensure comprehensive coverage.
• Bitmap document support: Processes bitmap images from scanned or older publications that lack vector graphics.
• Performance and scalability: Reported performance is comparable to traditional methods with projected accuracy exceeding 90% given 60–100 million training structures and feasible training time on a single GPU.
• Chemical data representation: Uses SMILES, DeepSMILES, and SELFIES encodings, with SELFIES reported to outperform DeepSMILES and SMILES in training experiments.

Scientific Applications:

• Data mining: Automates extraction of chemical structure images from literature to enable large-scale retrieval and analysis of chemical knowledge.
• Database enrichment: Converts graphical chemical depictions into machine-readable formats suitable for populating chemical databases.
• Historical literature analysis: Enables recovery of chemical information from older scanned publications that predate vector graphics.

Methodology:

Applies show-and-tell neural networks in a two-stage pipeline where detection produces masks of structure depictions and post-processing refines incomplete masks; models are trained on chemical string representations (SMILES, DeepSMILES, SELFIES) with estimated datasets of 60–100 million structures and reported training on a single GPU.