HER2-LADDER: A Spatially Interpretable AI Framework Predicts Neoadjuvant Dual HER2-Targeted Therapy Response in Breast Cancer
核心洞察
The HER2 (搜索)-LADDER model, integrating H&E and HER2 IHC digital pathology with clinical data, achieved an AUC of 0.944 in predicting pathological complete response to neoadjuvant TCbHP/PCbHP in HER2-positive breast cancer (搜索).
Each 0.1-unit increase in the HER2 (搜索)-LADDER score increased the odds of non-pCR nearly sevenfold (OR 6.628, 95% CI: 4.445–9.883), outperforming conventional predictors like HER2 IHC 3+ status and HR negativity.
The model stratified patients into Low, Medium, and High groups, suggesting potential for treatment de-escalation (THP, TCbH/PCbH) in Low-score patients and alternative regimens (ADCs, TKIs) in High-score patients.
A novel artificial intelligence framework, HER2 (搜索)-LADDER, has demonstrated robust predictive performance in identifying which patients with HER2-positive breast cancer (搜索) are likely to achieve a pathological complete response (pCR) following neoadjuvant dual HER2-targeted therapy. Developed and validated across multiple cohorts totaling 1,249 patients, the model integrates spatial features extracted from routine H&E and HER2 immunohistochemistry (IHC) whole-slide images with clinical-pathological variables, offering a clinically translatable tool for precision-guided treatment optimization.
Model Development and Validation
The HER2 (搜索)-LADDER model was developed using a core cohort of 358 patients with HER2-positive breast cancer (搜索) who completed standard-of-care taxane, carboplatin, trastuzumab, and pertuzumab (TCbHP/PCbHP) neoadjuvant regimens at Fudan University Shanghai Cancer Center (FUSCC) between 2020 and 2024. Of these, 276 cases treated from 2020 to 2022 formed the model construction set, while 82 cases treated after 2023 served as a temporal validation set.
Two specialized deep-learning algorithms were employed for single-cell segmentation: HoVer-Net distinguished tumor microenvironment components—including tumor cells, stromal cells, lymphocytes, neutrophils, and macrophages—on H&E slides, while D-PathAI characterized tumor cells on HER2 (搜索) IHC slides according to HER2 membrane staining intensity and completeness. A single-cell morphological and topological profiling (sc-MTOP) framework then extracted 69 features from H&E images and 70 from HER2 IHC images, encompassing cell proportions, spatial interactions, and cellular distributions.
The model demonstrated strong and consistent predictive performance. The area under the curve (AUC) was 0.944 (95% CI, 0.921–0.962) in the model construction set, 0.903 (95% CI, 0.829–0.955) in the temporal validation set, and 0.869 (95% CI, 0.814–0.929) in the FASCINATE-N PCbHP trial-based validation set. In an independent multicenter external validation cohort from Chongqing University Cancer Hospital (CQUCH), HER2 (搜索)-LADDER achieved an AUC of 0.798, demonstrating preserved discriminative performance when extrapolated to an external medical center.
Predictive Strength Outperforms Conventional Markers
Logistic regression analyses revealed that each 0.1-unit increase in the HER2 (搜索)-LADDER score increased the odds of a non-pCR after TCbHP/PCbHP treatment nearly sevenfold (OR 6.628, 95% CI: 4.445–9.883, P = 1.706 × 10⁻²⁰). This association remained robust after adjusting for age and clinical stage (adjusted OR 6.923, 95% CI: 4.537–10.563; P = 2.837 × 10⁻¹⁹). Notably, HER2-LADDER outperformed conventional clinicopathologic predictors, including HER2 IHC 3+ status (OR 0.112, 95% CI: 0.05–0.249), HR negativity (OR 0.217, 95% CI: 0.126–0.347), and TILs score (OR 0.839, 95% CI: 0.686–1.027, P = 0.089).
Layered Treatment Recommendations
Using a tertile-based division, patients were stratified into Low, Medium, and High HER2 (搜索)-LADDER score groups. In the Low group, pCR rates were similarly high for TCbHP/PCbHP (96.2%), THP (85.7%), and TCbH/PCbH (84.9%), suggesting potential appropriateness for regimen de-escalation. In the Medium group, TCbHP/PCbHP significantly outperformed THP (pCR 80.0% vs. 7.7%, P = 6.382 × 10⁻⁷) and TCbH/PCbH (47.1%, P = 6.856 × 10⁻⁵), supporting maintenance of the standard regimen. In the High group, patients who received the novel anti-HER2 ADC SHR-A1811 (54.2%) and TH plus pyrotinib (50.0%) had relatively better outcomes compared to TCbHP/PCbHP (19.6%).
The model also demonstrated prognostic value in the adjuvant setting. Patients in the High group had significantly worse overall survival (HR 7.17; 95% CI: 1.52–33.83; log-rank P = 0.013) and worse disease-free survival (HR 2.94; 95% CI: 1.14–7.61; P = 0.026) compared to the Low group.
Biological Interpretability Through Spatial Transcriptomics
Xenium in situ spatial transcriptomic profiling of 21 tumor samples provided mechanistic insights. HER2 (搜索)-LADDER-High tumors exhibited significantly greater heterogeneity in HER2 membrane staining, characterized by mosaic-like distribution patterns, with lower proportions of tumor cells with complete membrane HER2-strong positivity (P = 7 × 10⁻⁸) and higher proportions of HER2-weak complete membrane staining (P = 2.6 × 10⁻⁵).
In the tumor immune microenvironment, tumors with better therapeutic responses demonstrated an "immune-hot" phenotype, including increased aggregation scales between neutrophils and lymphocytes (Neutro_Lymph_Nsubgraph; P = 2.4 × 10⁻⁴) and reduced intercellular distances (Neutro_Lymph_minEdgeLength; P = 5.8 × 10⁻⁶). Ligand–receptor analyses revealed a spatial immunological cascade initiated by neutrophil–helper T-cell interactions that enhances effector immune responses, thereby potentiating antitumor immunity.
The study integrated both real-world treatment cohorts and clinical trial patients, with all slides processed through a standardized digital pathology workflow. The findings, published in Nature Signal Transduction and Targeted Therapy, position HER2 (搜索)-LADDER as a conceptually innovative framework for precision-guided optimization of anti-HER2 therapies, supporting potential de-escalation for low-risk patients and alternative regimens for high-risk patients.