Abstract
The DESTINY-Breast06 trial demonstrated survival benefits of trastuzumab deruxtecan (T-DXd) in human epidermal growth factor receptor 2 (HER2)-low/ultralow metastatic breast cancer. However, the associated clinicopathological features, HER2 testing consistency, and utility of artificial intelligence (AI)-assisted interpretation remain unclear. We analyzed 1455 HER2-non-amplified cases and conducted a multicenter ring study across 45 laboratories, with 209 tissue sections evaluated by manual and AI-assisted scoring. Compared to HER2-ultralow cases, HER2-null tumors exhibited higher histological grade (p < 0.001), lower estrogen receptor (ER) expression, and elevated Ki-67 (all P < 0.05). Triple-negative breast cancer (TNBC) patients in the HER2-null subgroup were significantly younger (P = 0.040). Significant staining variability was observed between antibodies: the 4B5 clone was associated with more HER2 0 and 1+ scores, while MXR001 yielded higher rates of HER2-low results (P < 0.05). Inter-laboratory concordance was moderate (68.89%), being poorest for HER2 1+ (57.58%; Kappa = 0.566). AI-assisted interpretation improved observer agreement (Kappa: 0.703 vs. 0.610 in non-amplified cases) and reduced ambiguous immunohistochemistry (IHC) 2+/0 assignments. These findings delineate distinct clinicopathological characteristics of HER2-ultralow/null tumors, highlight HER2 IHC reproducibility challenges, and validate AI as an effective tool for standardizing scoring to optimize patient selection for T-DXd therapy.