Pro-GAT: Reconnecting Fragmented PROTACs Using Graph Attention Transformer

Proposes Pro-GAT, 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. Prior diffusion-based generative models operate in continuous coordinate space and lack explicit mechanisms for enforcing discrete bond connectivity under fixed anchor constraints, routinely producing linker structures that are geometrically plausible yet chemically disconnected or valence invalid; Pro-GAT explicitly formulates repair as a geometry-preserving task on graphs. The paper describes the model architecture (geometry-aware graph attention layers, bounded coordinate corrections, constrained atom-type modifications) and training strategy (trained on PROTAC data sets with introduced disconnections), but the loaded text does not provide ablation studies or architectural comparison details.

On the DiffPROTAC test set, Pro-GAT recovers 31.58% of disconnected structures. This number directly quantifies repair capability, showing that systematic disconnection artifacts in diffusion generation can be partially recovered by graph attention repair rather than only through naive resampling. Quantitative result on the DiffPROTAC test set with a specific percentage, but the text does not state the test set size or the distribution of disconnection types.

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. This indicates that the repair step improves chemical validity without sacrificing diversity, converting otherwise unusable disconnected diffusion outputs into viable PROTAC candidates. Quantitative comparison across two independent generative models, reporting both validity and uniqueness metrics, but the text does not provide statistical significance tests or confidence intervals.

Across three experimentally characterized ternary complexes 7Z76 (SMARCA2-VHL), 7JTO (WDR5-VHL), and 6W8I (BTK-cIAP1), repaired candidates yielded docking scores comparable to reference crystal poses; in the 6W8I case, repair improved the docking score from -9.9 to -11.3 kcal/mol, with anchor coordinates fixed throughout repair by construction. Provides biological relevance evidence beyond purely computational metrics by comparing repair results against experimentally characterized crystal structures, and the anchor-fixing construction guarantees that repair does not disrupt the original anchoring geometry. Docking validation across three experimentally characterized ternary complexes, with a specific energy improvement value for 6W8I; however, the text does not specify the docking software, scoring function, or number of replicates.

This work targets the repair of PROTAC candidates generated by diffusion models under fixed anchor constraints, applicable to computational design pipelines that need to convert chemically disconnected diffusion outputs into viable candidates. Its validation scope includes recovery rate on the DiffPROTAC test set, chemical validity and uniqueness metrics when combined with DiffPROTAC and DiffLinker fine-tuned on PROTAC-specific data, and docking evaluation on three experimentally characterized ternary complexes 7Z76, 7JTO, and 6W8I. Applicability to other molecular generation tasks or other anchor constraint conditions requires further investigation.

The loaded text is abstract-level content and does not include figures, ablation studies, statistical significance tests, or test set size details, so the robustness of the repair method across different disconnection types or anchor constraint conditions cannot be assessed. The binding activity of repaired candidates in real experiments has not been reported, and the consistency between docking scores and experimental activity still requires follow-up work. Additionally, the 31.58% recovery rate means a substantial proportion of disconnected structures remain unrepaired, and their failure modes warrant attention.