ProLinker–Generator: Design of a PROTAC Linker Base on a Generation Model Using Transfer and Reinforcement Learning

In PROTAC molecules, the design of the linker directly affects the formation efficiency and stability of the target protein–PROTAC–E3 ligase ternary complex, making it a critical factor in determining degradation activity. However, current linker data are limited, and the accessible chemical space remains narrow. The length, conformation, and chemical composition of linkers play a decisive role in drug performance, highlighting the urgent need for innovative linker design. In this study, we propose ProLinker-Generator, a GPT-based model aimed at generating novel and effective linkers. By integrating transfer learning and reinforcement learning, the model expands the chemical space of linkers and optimizes their design. During the transfer learning phase, the model achieved high scores in validity (0.989) and novelty (0.968) for the generated molecules. In the reinforcement learning phase, it further guided the generation of molecules with ideal properties within our predefined range. ProLinker-Generator demonstrates the significant potential of AI in linker design.