Deep scaffold hopping with multimodal transformer neural networks.

Area of Science:

• Medicinal Chemistry
• Computational Drug Design
• Artificial Intelligence in Drug Discovery

Background:

• Scaffold hopping is crucial for rational drug design, aiming to discover novel molecules with preserved biological activity.
• Traditional methods rely on database searches, limiting exploration of vast chemical spaces.
• Generating molecules with similar 3D conformations but novel 2D scaffolds is key for drug discovery.

Purpose of the Study:

• To re-formulate scaffold hopping as a supervised molecule-to-molecule translation task.
• To develop a deep learning model capable of generating novel 2D scaffolds with similar 3D structures and improved bioactivity.
• To address the limitations of traditional scaffold hopping methods in exploring chemical diversity.

Main Methods:

• Curated a dataset of over 50,000 molecule pairs with enhanced bioactivity, similar 3D structure, and different 2D structure from public databases.
• Designed a multimodal molecular transformer architecture integrating 3D conformers (via spatial graph neural networks) and protein sequences (via Transformer).
• Trained the DeepHop model on kinase-targeted molecule data.

Main Results:

• The DeepHop model generated molecules with improved bioactivity, high 3D similarity, and low 2D scaffold similarity in approximately 70% of cases.
• This success rate was 1.9 times higher than state-of-the-art deep learning, rule-based, and virtual screening methods.
• Demonstrated model generalization to new target proteins via fine-tuning and showcased practical utility in case studies.

Conclusions:

• DeepHop effectively addresses the scaffold hopping challenge by generating novel molecules with desired properties.
• The model significantly outperforms existing methods in generating successful hopped molecules.
• DeepHop offers a powerful and practical tool for accelerating drug discovery and exploring new chemical entities.