Assessing conformation validity and rationality of deep learning-generated 3D molecules.

Area of Science:

• Computational Chemistry
• Artificial Intelligence
• Drug Discovery

Background:

• Advancements in AI have spurred 3D molecule generation, but evaluation methods are lacking.
• Current techniques for evaluating 3D conformations are often too slow or lack accuracy.
• Existing methods fail to capture subtle conformational anomalies or detailed atomic/torsional information.

Purpose of the Study:

• To introduce a novel, two-stage evaluation system for AI-generated 3D molecules.
• To achieve both high-speed processing and quantum mechanical accuracy in conformational analysis.
• To address the limitations of current empirical and molecular mechanics-based evaluation metrics.

Main Methods:

• A two-stage approach: a validity test and a rationality test.
• The validity test uses an AI-derived force field to detect atoms in high-energy, implausible environments.
• The rationality test employs a deep learning network (trained with DFT accuracy) to identify high-energy rotatable bonds.

Main Results:

• The proposed system demonstrates high speed and accuracy comparable to quantum mechanical methods.
• The approach was successfully applied to five AI 3D molecule generation models.
• Evaluation was performed on 102 targets from the Directory of Useful Decoys-Enhanced dataset.

Conclusions:

• The developed two-stage system offers a significant improvement for evaluating 3D molecule conformations.
• This method enhances the reliability and accuracy of AI-driven molecular generation.
• The open-source package facilitates broader adoption and advancement within the research community.