Membrane-specific HER2 expression by artificial intelligence-based quantitative scoring for prediction of efficacy of trastuzumab deruxtecan in biliary tract cancer (HERB trial): Exploratory analysis of a multicenter, single arm, phase II trial.

Abstract

4097 Background: Trastuzumab deruxtecan (T-DXd) showed promising results in patients with HER2-positive biliary tract cancer (BTC). With T-DXd's expanding indication into HER2-low cancers, quantitative Artificial Intelligence (AI)-based scoring of HER2 expression at the cellular level becomes increasingly important. This study investigated whether the intensity and subcellular pattern of HER2 staining would correlate with response to T-DXd. Methods: HER2 immunohistochemistry (IHC) whole slide images from the phase II HERB trial were analyzed. These participants had unresectable or recurrent BTC refractory or intolerant to gemcitabine-containing regimen and received T-DXd based on confirmed HER2-positive or low status. Lunit SCOPE universal IHC, a deep learning based IHC analyzer, was used to provide cell level classes (AI-H0, H1+, H2+, H3+) and continuous scoring of HER2 staining intensities of subcellular compartments (membrane, cytoplasm and nucleus) for each tumor cell. Membrane specificity was calculated for each cell as the ratio of membrane intensity to the sum of all three subcellular compartments. Results: The 29 patients analyzed showed continuous improvement in response rates with an increasing proportion of AI-H3+ cells. The ORR was 37.5%, 42.9% and 50.0% for patients with more than 10%, 25%, and 50% of tumor cells classified as AI-H3+, respectively. The HER2 intense cohort (n=4), defined by tumors with over 50% of tumor cells classified as AI-H3+, had a significantly better PFS (HR 0.15, p<0.05) and OS (HR 0.10, p<0.05) compared to the rest of the treatment group. The high membrane specificity group defined by ≥80% of tumor cells with membrane specificity ≥0.4 (N=6) had a confirmed ORR of 50%. These patients also demonstrated significantly longer PFS (HR 0.30, p<0.05) and OS (HR 0.27, p<0.05). The six cases identified by membrane specificity included all four cases of the HER2 intense cohort and two more cases, showing improved sensitivity in identifying likely responders. Conclusions: AI based quantification of HER2 intensity and membrane specificity was predictive of therapeutic response to T-DXd in HER2 expressing BTC. Membrane specificity analysis was more sensitive in identifying exceptional responders compared to intensity alone. Confirmed response and survival based on AI defined biomarkers. By AI-H3+ proportion By AI-MB specific cell proportion < 50% ≥ 50% < 80% ≥ 80% Sample size 25 4 23 6 ORR, % 28.0 50.0 26.1 50.0 mPFS 4.21 (2.83-4.40) 11.04 (5.68-12.91) 4.21 (2.83-4.40) 11.04 (1.45-12.91) HR (95% CI, p-value) REF 0.15 (0.03-0.67, <0.05) REF 0.30 (0.10-0.92, <0.05) mOS 7.00 (4.37-8.94) NR (5.68-NR) 7.00 (4.27-8.94) NR (9.63-NR) HR (95% CI, p-value) REF 0.10 (0.01-0.79, <0.05) REF 0.27 (0.08-0.93, <0.05)