Triple-negative breast cancer (TNBC) is an oncology indication with urgent need for drug discovery and limited targeted therapies. This work benchmarked several graph neural network (GNN) models for binary drug-sensitivity prediction in TNBC under conditions of scaffold-aware evaluation. A benchmark dataset of 3433 TNBC drug–cell line samples from 373 distinct compounds was compiled from GDSC and labeled using thresholds on IC\(_{50}\) activity measurements. Five GNN architectures (Graph Convolutional Networks, Graph Attention Network v2, Graph Isomorphism Network with edge features, Message Passing Neural Networks, and Transformer–GCN) were trained with Optuna-based hyperparameter optimization and assessed with Bemis–Murcko scaffold-disjoint splitting. Under the scaffold-aware TNBC benchmark, GATv2 achieved the strongest internal performance with an AUROC of 0.701, followed by Transformer–GCN (0.644) and GCN (0.642). External validation was further conducted using FDA-approved anticancer drugs and a phytochemical library to evaluate out-of-distribution generalization across chemically diverse molecular spaces. On the FDA benchmark, GINE achieved the highest AUROC (0.630), while Transformer–GCN and MPNN demonstrated comparatively stronger ranking capability on the phytochemical dataset, achieving AUROCs of 0.704 and 0.685, respectively. The external validation results highlighted the intrinsic difficulty of scaffold-aware molecular extrapolation in low-resource TNBC prediction. Visualization-based explainability analyses with attribution further confirmed the models were paying attention to chemically intuitive motifs, such as aromatic rings, heteroaromatic motifs, and electronegative active substituents. In summary, we show that scaffold-aware benchmarking, external validation, and interpretability analyses lead to more trustworthy and realistic application of AI methods for TNBC drug sensitivity prediction in the context of low-resource drug data.
