MolScribe

MolScribe translates molecular images into molecular graph structures by explicitly predicting atoms, bonds, and their geometric layouts to enable accurate molecular structure recognition from varied chemical diagrams.

Key Features:

• Image-to-graph generation model: Generates molecular graph structures directly from molecular images.
• Explicit atom and bond prediction: Predicts atoms and bonds as graph elements, including their connectivity.
• Geometric layout prediction: Predicts geometric layouts for atoms and bonds to reconstruct molecular geometry.
• Symbolic chemistry constraints: Incorporates symbolic chemistry constraints directly into the prediction process.
• Chirality recognition: Accurately recognizes chirality in chemical structures.
• Abbreviation expansion: Expands abbreviated forms commonly used in chemical diagrams.
• Data augmentation strategies: Employs advanced data augmentation to improve robustness against domain shifts.
• Confidence estimation and atom-level alignment: Provides confidence estimations and atom-level alignment between predictions and input images.

Scientific Applications:

• Molecular structure recognition: Extracts complete molecular graphs from chemical literature and diagrams with varied drawing styles and conventions.
• Stereochemistry extraction: Enables extraction of stereochemical information, including chirality, from 2D depictions.
• Benchmarking and evaluation: Evaluated on synthetic and realistic molecular images and reported 76–93% accuracy on public benchmarks, outperforming previous models.
• Prediction verification: Supports atom-level verification of predicted structures via confidence scores and alignment with input images.

Methodology:

Uses an image-to-graph generation model that explicitly predicts atoms, bonds, and their geometric layouts, incorporates symbolic chemistry constraints into prediction, employs advanced data augmentation strategies to enhance robustness, and outputs confidence estimations with atom-level alignment.