ZeroGEN: leveraging language models for zero-shot ligand design from protein sequences

Area of Science:

• Computational chemistry
• Drug discovery
• Artificial intelligence in medicine

Background:

• Deep generative models show promise in ligand design but struggle with novel targets (zero-shot learning).
• Effective ligand generation for unseen targets is critical for accelerating drug discovery.
• Existing methods require significant adaptation for new protein targets.

Purpose of the Study:

• To introduce ZeroGEN, a novel zero-shot deep generative framework for ligand design.
• To enable the generation of effective ligands for previously unseen protein targets.
• To enhance the efficiency and applicability of ligand design in drug discovery.

Main Methods:

• Utilizes protein sequences and protein-ligand interaction data.
• Incorporates contrastive learning to align protein-ligand features.
• Employs self-distillation for data filtering and data augmentation for unseen targets.

Main Results:

• ZeroGEN successfully generates ligands with strong affinity and drug-like properties for unseen targets.
• Model visualizations confirm focus on key protein residues, indicating understanding of interactions.
• Demonstrates robust performance in zero-shot ligand generation scenarios.

Conclusions:

• ZeroGEN offers a powerful new approach for zero-shot ligand design.
• The framework accelerates the identification of potential drug candidates for novel targets.
• ZeroGEN enhances the capabilities of deep generative models in drug discovery.