Multimodal Bond Reconstruction toward Generative Molecular Design

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Generative models accelerate de novo drug design by creating novel 2D and 3D molecular structures.
• Accurate chemical bond reconstruction, particularly from distorted geometries, is a significant challenge in generative chemistry.

Purpose of the Study:

• To develop a robust framework for chemical bond reconstruction from various molecular data formats.
• To address the limitations of existing methods in handling imperfect geometries generated by de novo drug design tools.

Main Methods:

• Introduced YuelBond, a multimodal graph neural network.
• Applied YuelBond to three scenarios: bond recovery from accurate 3D coordinates, reconstruction in crude de novo generated compounds (CDGs) with perturbed geometries, and bond order reassignment in 2D graphs.

Main Results:

• YuelBond achieved a 98.4% F1 score on standard 3D structures, outperforming traditional methods.
• Demonstrated strong performance (92.7% F1 score) on distorted CDGs where RDKit failed.
• Successfully enabled accurate bond reconstruction from imperfect molecular data.

Conclusions:

• YuelBond provides accurate and reliable bond reconstruction, crucial for generative drug discovery.
• The framework bridges a critical gap by handling imperfect molecular data, enhancing de novo drug design pipelines.