MMSSC-Net: multi-stage sequence cognitive networks for drug molecule recognition

Area of Science:

• Computational chemistry
• Cheminformatics
• Artificial intelligence in drug discovery

Background:

• Drug compound structures are often represented as 2D vector graphics, posing challenges for computational recognition and utilization.
• Current Optical Chemical Structure Recognition (OCSR) methods often process structures as isolated entities, limiting fine-grained analysis.

Purpose of the Study:

• To develop a multi-stage cognitive neural network model for fine-grained prediction and interpretation of molecular vector graphics.
• To enhance the computer-readability and analytical capabilities of drug structure representations.

Main Methods:

• A bottom-up, staged cognitive approach is employed, starting with atomic and bond representations as discrete label sequences.
• Subsequent stages build molecular graphs from label sequences and evolve them into machine-readable formats.
• The model, MMSSC-Net, utilizes a sequence cognitive method for improved interpretability and transferability.

Main Results:

• MMSSC-Net demonstrated superior performance compared to existing advanced methods on multiple public datasets.
• Achieved high accuracy rates (75-94%) in cognitive recognition across various resolutions.
• The model proved reliable in interpretability and transferability.

Conclusions:

• MMSSC-Net offers a more effective method for computer-aided recognition of molecular vector graphics.
• This approach provides new avenues for drug information discovery and exploration of chemical space.
• The staged, cognitive methodology enhances the reliability and applicability of molecular structure analysis.