Transformer-based molecular optimization beyond matched molecular pairs.

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Drug Discovery

Background:

• Molecular optimization is crucial for drug discovery, aiming to enhance a molecule's profile.
• Challenges include optimizing multiple properties and navigating vast chemical spaces.
• Current deep learning methods often mimic matched molecular pairs (MMPs), limiting exploration of structural modifications.

Purpose of the Study:

• To develop a general methodology for molecular optimization offering broader structural modifications beyond MMPs.
• To investigate how Transformer model behavior changes when trained on datasets reflecting different transformation types.
• To explore the potential of tailored datasets for diverse molecular optimization strategies.

Main Methods:

• Utilized a Transformer architecture trained on various datasets representing different molecular transformations.
• Constructed datasets beyond MMPs from ChEMBL using Tanimoto similarity and scaffold matching criteria.
• Analyzed how dataset characteristics influence the model's molecular transformation outputs.

Main Results:

• Models trained on different datasets exhibited distinct molecular transformation patterns.
• The nature of the training dataset directly influenced the model's output for molecular optimization.
• The Transformer architecture proved adaptable to diverse data representations for molecular modification.

Conclusions:

• The proposed methodology provides a more general approach to molecular optimization than MMPs.
• Tailoring training datasets allows for control over the types of structural modifications generated.
• These models can complement each other, offering chemists diverse options for improving drug candidates.