LinkLlama: Enabling Large Language Model for Chemically Reasonable Linker Design

Area of Science:

• Computational Chemistry
• Drug Discovery
• Artificial Intelligence in Medicine

Background:

• Fragment-based drug discovery (FBDD) requires effective linker design to create potent drug leads.
• Current generative models often produce linkers with high strain and poor drug-likeness.

Purpose of the Study:

• To introduce LinkLlama, a novel large language model for generating chemically viable linkers in FBDD.
• To bridge text-based generation with 3D spatial awareness for improved linker design.

Main Methods:

• Fine-tuning Meta Llama 3 on drug-like molecules from ChEMBL.
• Utilizing natural language prompts for geometric and physicochemical constraints.
• Supervised fine-tuning to capture chemical grammar and ensure validity.

Main Results:

• LinkLlama achieves competitive geometric fidelity with 3D-aware models.
• Demonstrates a two-fold increase in chemically reasonable designs (35% to over 80%).
• Successfully applied to scaffold hopping and PROTAC linker design, validated by docking and simulations.

Conclusions:

• LinkLlama overcomes structural limitations of purely 3D generative methods.
• Offers a controllable and chemically robust framework for accelerating linker design.
• Highlights the potential of large language models in drug discovery.