AdaptMol: domain adaptation for molecular image recognition with limited supervision

Area of Science:

• Computer Vision
• Cheminformatics
• Machine Learning

Background:

• Optical Chemical Structure Recognition (OCSR) converts 2D molecular images to machine-readable formats like SMILES.
• Deep learning enhances OCSR, but models struggle with real-world hand-drawn data due to variations in drawing styles.
• Existing methods often require synthetic data and fail to generalize without manual annotations on target domains.

Purpose of the Study:

• To develop AdaptMol, an image-to-graph model for robust OCSR.
• To enable effective transfer learning from synthetic to real-world molecular images without target domain annotations.
• To improve the accuracy and generalizability of OCSR models for diverse chemical structure representations.

Main Methods:

• AdaptMol employs unsupervised domain adaptation and self-training to refine models trained on synthetic data.
• Class-conditional Maximum Mean Discrepancy (MMD) aligns domain-invariant bond features across synthetic and real-world datasets.
• A comprehensive data augmentation strategy and dual position representation enhance model robustness and atom localization accuracy.

Main Results:

• AdaptMol achieves 82.6% accuracy on hand-drawn molecular images, outperforming prior methods by 10.7%.
• The model demonstrates competitive performance across benchmarks including scientific literature and patent documents.
• Unsupervised domain adaptation successfully bridges the gap between synthetic and real-world molecular image data.

Conclusions:

• AdaptMol offers a significant advancement in OCSR, particularly for challenging hand-drawn chemical structures.
• The proposed method effectively transfers knowledge from synthetic to real-world data without manual annotation.
• AdaptMol provides a robust and accurate solution for converting diverse molecular image inputs into machine-readable formats.