Proteolysis-targeting chimeras (PROTACs) represent an emerging modality for targeted protein degradation with broad therapeutic potential. However, the risk of off-target protein degradation remains a major concern in the development of PROTAC-based therapeutics. Here, we present SENTINEL, a graph-based deep learning framework that predicts the off-target propensity of PROTAC warheads based on their involvement levels in drug-target interactions as determined from established databases and the literature. By encoding warheads as molecular graphs using path-augmented graph transformer networks (PAGTNs), we show that graph attention-based neural networks (GATs) achieve accurate modelling of binding count-based off-target effects with an area under the ROC curve (AUC) of 0.9600 and an F1-score of 0.6983, outperforming classical machine learning algorithms such as random forests (AUC=0.840, F1-score=0.2778). SENTINEL provides a scalable strategy to prioritise lower-risk warheads in a low-data setting, supporting early-stage evaluation of PROTAC off-target risk. Results should be interpreted with the dataset size in mind and will benefit from larger external validation.