ConfRetro: A 3D-aware Template-free Method for Enhancing Retrosynthesis via Molecular Conformer Information

Area of Science:

• Computational chemistry
• Organic synthesis
• Drug discovery

Background:

• Retrosynthesis is vital for organic synthesis and drug discovery, identifying reactants for target molecules.
• Current methods often neglect 3D conformer and spatial structure, limiting stereochemical and chemical plausibility.
• Exploiting 3D information is key to advancing retrosynthetic prediction accuracy.

Purpose of the Study:

• To introduce ConfRetro, a template-free retrosynthesis method leveraging 3D molecular conformer and spatial structure information.
• To improve the accuracy and chemical plausibility of predicted reactants in retrosynthetic analysis.
• To establish a new state-of-the-art in template-free retrosynthesis.

Main Methods:

• ConfRetro utilizes a Transformer architecture incorporating an Atom-align Fusion module for 3D positional information.
• A Distance-weighted Attention mechanism guides self-attention, focusing on chemically relevant atom pairs in 3D space.
• The method is template-free, directly predicting reactants from the target molecule's structure.

Main Results:

• ConfRetro significantly outperforms existing template-free methods on USPTO-50K and USPTO-FULL datasets.
• Achieved new state-of-the-art performance in retrosynthetic prediction.
• Demonstrated capability in predicting accurate, chemically plausible reactants for complex molecules and recovering synthetic routes for drugs like Camptothecin.

Conclusions:

• ConfRetro effectively integrates 3D molecular information for superior retrosynthetic predictions.
• The method offers advancements in accuracy, chemical plausibility, and applicability to complex drug molecules.
• ConfRetro represents a significant step forward in computational approaches to organic synthesis and drug discovery.