PROteolysis TArgeting Chimeras (PROTACs) represent a novel class of therapeutic agents that induce targeted protein degradation by hijacking the ubiquitin–proteasome system. Despite their potential, the rational design of PROTACs, particularly the design of flexible and diverse linkers connecting the ligands, still remains a major bottleneck due to the lack of comprehensive structure–activity relationships and limited experimental data. Fragment-based drug design has been proposed to guide PROTAC development by exploiting their modular structures. Meanwhile, recent breakthroughs in artificial intelligence, notably diffusion models and Transformer architectures, have opened new avenues for molecular generation, enabling the design of complex chemical structures through data-driven learning of both topology and geometry.
Recent models like EDM and DiffPROTACs have leveraged Transformers within diffusion frameworks to generate PROTAC linkers from given ligands. However, rational PROTAC design requires precise control over linker length and chemical properties, which are essential for effective ternary complex formation and favourable drug-like characteristics (e.g. aqueous solubility, cell permeability, and stability). Existing approaches lack mechanisms to enforce these constraints, leading to suboptimal results. To overcome this, we propose a novel diffusion-based model that incorporates linker length and chemical properties as controllable conditions. Our method first encodes PROTACs using a GNN-based Transformer, then introduces a multi-condition encoder based on feature-wise linear modulation to integrate conditional inputs, and finally performs constrained generation through a tailored diffusion process. Experiments on the PROTAC-DB demonstrate state-of-the-art performance in validity and recovery. Additionally, we generate 10 novel PROTACs targeting therapeutically relevant proteins such as TRIM24, TRIM33, BAZ1A, and PBRM1.