Transformer-Based Generative Model Accelerating the Development of Novel BRAF Inhibitors

Area of Science:

• Medicinal Chemistry
• Computational Chemistry
• Artificial Intelligence

Background:

• De novo drug design using SMILES strings is a sequence-processing challenge.
• Recurrent neural networks (RNNs) struggle with long-range dependencies, leading to invalid molecular structures.
• Transformer architecture shows promise for enhanced sequence data handling.

Purpose of the Study:

• To develop a novel generative model for de novo drug design using a Transformer encoder.
• To improve the generation of chemically valid and structurally diverse molecules with desired activities.
• To optimize drug design through transfer and reinforcement learning.

Main Methods:

• Utilized a Transformer-encoder architecture for a generative model.
• Trained a predictive model for molecular activity assessment.
• Employed transfer learning and reinforcement learning for fine-tuning and optimization.
• Validated the pipeline by designing BRAF protein inhibitors.

Main Results:

• Achieved a higher percentage of chemically valid molecules (98.2% vs. 95.6% for RNNs).
• Enhanced structural diversity and feasibility of molecular synthesis.
• Generated small molecules with higher predicted activity and similar interaction sites compared to crystal ligands.
• Demonstrated successful inhibitor design against the human BRAF protein.

Conclusions:

• The Transformer-encoder model significantly outperforms RNN-based methods in de novo drug design.
• This approach offers a powerful tool for generating novel, active, and synthesizable drug candidates.
• The findings provide valuable insights for pharmaceutical chemists in drug discovery.