MolMiner

MolMiner transforms 2D printed chemical structure depictions in scientific documents such as journal papers and patents into machine-readable molecular graphs using an object-detection approach to Optical Chemical Structure Recognition (OCSR).

Key Features:

• Object-detection framing: Reformulates the traditional vectorization problem of OCSR as an object detection task to extract chemical elements from 2D depictions.
• Deep neural networks: Uses deep neural networks originally developed for semantic segmentation and object detection to perform recognition.
• Atom and bond recognition: Detects atoms and bonds from printed depictions as discrete objects for downstream assembly.
• Annotated training data: Trains on well-labelled datasets annotated with atoms and bonds to learn element detection.
• Distance-based graph construction: Connects detected atomic and bonding elements into a coherent molecular graph using a distance-based construction algorithm.
• Benchmark validation: Demonstrates state-of-the-art performance across four benchmark datasets and an external dataset collected from scientific papers.

Scientific Applications:

• Optical Chemical Structure Recognition (OCSR): Converts non-machine-readable 2D chemical depictions into structured representations for computational use.
• Extraction from literature and patents: Extracts atom and bond information from journal papers and patents for downstream analysis.
• Machine-readable format generation: Produces machine-readable molecular graphs suitable for cheminformatics workflows.
• OCSR benchmarking: Serves as a method for evaluating OCSR performance on benchmark and external datasets.

Methodology:

MolMiner frames OCSR as object detection, employs deep neural networks developed for semantic segmentation and object detection, trains on datasets annotated with atoms and bonds to detect those elements, and uses a distance-based construction algorithm to assemble detected elements into a molecular graph.