Pro-GAT: Reconnecting Fragmented PROTACs Using Graph Attention Transformer.

PROTACs recruit an E3 ligase to a protein of interest via a bifunctional linker, and their activity is critically dependent on the geometry and chemical integrity of that linker. Diffusion-based generative models, which operate in continuous coordinate space, lack explicit mechanisms for enforcing discrete bond connectivity under fixed anchor constraints; as a result, they routinely produce linker structures that are geometrically plausible yet chemically disconnected or valence invalid. Because these generation artifacts arise from a systematic gap between the continuous output space and the discrete requirements of chemical bonding, they recur across samples and are not remedied by naive resampling. Pro-GAT is a graph attention-based framework for geometry-preserving molecular graph repair that operates on chemically disconnected diffusion-generated PROTAC candidates by predicting bounded coordinate corrections and constrained atom-type modifications through geometry-aware graph attention layers. The model is trained on PROTAC data sets with introduced disconnections to address systematic connectivity challenges inherent to diffusion-based PROTAC generation under fixed anchor constraints and recovers 31.58% of disconnected structures in the DiffPROTAC test set. When combined with DiffPROTAC and DiffLinker fine-tuned on PROTAC-specific data, Pro-GAT improves the proportion of chemically valid candidates in the aggregated output from 76.70% to 83.92% and from 63.16% to 68.73%, while maintaining uniqueness levels of 80.18% and 63.80% among valid candidates, respectively, thereby converting otherwise unusable disconnected diffusion outputs into viable PROTAC candidates. Structural validation was performed across three experimentally characterized ternary complexes: 7Z76 (SMARCA2-VHL), 7JTO (WDR5-VHL), and 6W8I (BTK-cIAP1), where repaired candidates yielded docking scores comparable to reference crystal poses across all three systems; in the BTK-cIAP1 case (6W8I), repair produced an improvement from -9.9 to -11.3 kcal/mol, with anchor coordinates fixed throughout repair by construction.