LinkLlama: Enabling a Large Language Model for Chemically Reasonable Linker Design
The work presents LinkLlama, a fine-tuned Meta Llama 3 model supervised on a curated corpus of drug-like molecules from ChEMBL that accepts natural language prompts specifying geometric constraints such as distances and angles alongside physicochemical targets like Lipinski's rules and rotatable bond limits to generate tailored molecules for input fragments; benchmarking on the ZINC and HiQBind data sets shows competitive geometric fidelity relative to strictly 3D-aware models together with an approximately two-fold increase in the proportion of chemically reasonable designs, rising from 35% to over 80% under structural filters including PAINS, non-drug-like chemical patterns, and complex ring systems, with versatility illustrated through small-molecule scaffold hopping and PROTAC linker d