Adapt-cMolGPT: A Conditional Generative Pre-Trained Transformer with Adapter-Based Fine-Tuning for Target-Specific...

Area of Science:

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

Background:

• Small-molecule drug design is vital for identifying compounds targeting specific proteins in early drug discovery.
• Generative models like GPT show promise for molecular compound generation, but face performance degradation with small pharmaceutical datasets.
• Existing methods struggle with generating target-specific compounds due to data limitations.

Purpose of the Study:

• To propose an enhanced target-specific drug generation model, Adapt-cMolGPT.
• To address the performance degradation of generative models in drug discovery caused by small datasets.
• To improve the generation of novel, valid, and target-specific molecular compounds.

Main Methods:

• Developed Adapt-cMolGPT, modifying molecular representation and optimizing the fine-tuning process.
• Introduced a novel fine-tuning method using an adapter module integrated into a pre-trained base model.
• Implemented alternating weight updates by sections during the fine-tuning process.

Main Results:

• Adapt-cMolGPT demonstrated performance improvements over previous models in target-specific compound generation.
• The model generated a higher number of novel and valid compounds compared to existing methods.
• Generated compounds exhibited properties similar to real molecular data, indicating high efficacy.

Conclusions:

• Adapt-cMolGPT is highly effective for designing drugs targeting specific proteins.
• The proposed fine-tuning strategy enhances generative model performance in data-scarce pharmaceutical settings.
• This approach advances the field of computational drug discovery and AI-driven molecular design.