This study provides a clinically translatable AI prediction framework for individualized treatment of HER2-positive breast cancer. By integrating spatial topological features from routine H&E and IHC slides, it offers a direct paradigm for guiding anti-HER2 therapeutic strategies using digital pathology in experimental design.
Literature Overview
The article titled 'Spatially interpretable artificial intelligence framework to tailored neoadjuvant dual HER2 blockade in HER2-positive breast cancer,' published in Signal Transduction and Targeted Therapy, systematically explores how artificial intelligence can integrate spatial information from routine histopathological slides to accurately predict response to neoadjuvant dual-targeted therapy in patients with HER2-positive breast cancer. The research team developed the HER2-LADDER model, which not only demonstrates high predictive performance but also offers biological interpretability, providing a new tool for personalized treatment decisions.Background Knowledge
HER2-positive breast cancer accounts for approximately 15–20% of all breast cancer subtypes, characterized by amplification of the HER2 gene or overexpression of the HER2 protein, leading to activation of downstream PI3K/AKT and MAPK pathways that promote tumor proliferation and invasion. Although anti-HER2 monoclonal antibodies such as trastuzumab have significantly improved patient outcomes, neoadjuvant dual-targeted therapy (e.g., trastuzumab combined with pertuzumab) has become the standard regimen, achieving pathologic complete response (pCR) rates exceeding 45%. However, treatment response is highly heterogeneous—some patients are overtreated, while others exhibit primary resistance. Currently, there is a lack of widely applicable clinical prediction tools. Traditional biomarkers such as HER2 IHC 3+, HR status, and TILs have limited predictive accuracy. Genomic sequencing methods are costly and less accessible, limiting their broad application. Therefore, there is an urgent need for a prediction model based on routine tests that is reproducible and biologically interpretable. This study addresses this gap by leveraging widely available H&E and HER2 IHC whole-slide images, combining spatially resolved single-cell morphology and topological features to build an interpretable AI model, overcoming limitations of existing approaches.
Research Methods and Experiments
The research team constructed a multi-cohort dataset comprising 1,249 patients with HER2-positive breast cancer, encompassing both real-world and prospective clinical trial cohorts. The model was developed using 358 patients from FUSCC who received TCbHP/PCbHP therapy, with paired H&E and HER2 IHC slides digitized. Deep learning algorithms—HoVer-Net and D-PathAI—were employed for single-cell segmentation of tumor microenvironment components (e.g., tumor cells, lymphocytes, neutrophils) and HER2 membrane expression intensity, respectively. Subsequently, the sc-MTOP framework extracted 69 spatial features from H&E and 70 from IHC, including cell proportions, aggregation (Nsubgraph), intercellular distances (MinEdgeLength), and connectivity (Degree). These features were integrated with clinical variables (age, clinical stage, HR status) to build the HER2-LADDER scoring model using a multi-model ensemble voting approach. The model was independently validated in a temporal validation cohort (N=82) and the FASCINATE-N trial cohort (N=85), achieving AUCs of 0.903 and 0.869, respectively. Further validation of its generalizability was conducted in external cohorts ddAC-THP and CQUCH.Key Conclusions and Perspectives
Research Significance and Prospects
This study introduces a novel biomarker strategy for drug development—predicting anti-HER2 therapy response using spatial topological features rather than single molecular markers. Its interpretable AI framework enhances understanding of resistance mechanisms and guides the design of novel ADCs or combination immunotherapies.
In terms of clinical monitoring, HER2-LADDER relies solely on routine slides without requiring additional testing, making it easily integrable into existing pathology workflows and highly accessible in clinical practice. Future validation through multicenter RCTs could assess its value in guiding treatment escalation or de-escalation, potentially updating standards for personalized therapy.
For disease modeling, this study illustrates how digital pathology can be combined with spatial omics to build 'virtual patient' models. This approach could be extended to other cancer types, establishing a universal spatial AI prediction platform to advance precision oncology.
Conclusion
The HER2-LADDER model developed in this study represents a significant advancement in precision therapy for HER2-positive breast cancer. By integrating spatial information from routine H&E and HER2 IHC slides, this AI framework not only achieves high-accuracy prediction of neoadjuvant treatment response but, more importantly, provides biologically interpretable mechanistic insights. The model stratifies patients into distinct risk groups, directly guiding individualized adjustments in treatment intensity—from de-escalated chemotherapy to switching to next-generation ADCs—significantly enhancing the scientific rigor of clinical decisions. Its prognostic value in the adjuvant cohort further supports its potential as a long-term risk stratification tool. The success of HER2-LADDER demonstrates the immense value of combining digital pathology with spatial omics, setting a paradigm for translational medicine research moving from 'histology' to 'spatial digital phenotypes.' In the future, this model has the potential to become a routine clinical tool, optimizing treatment pathways for HER2-positive breast cancer, reducing overtreatment, improving patient quality of life, and ultimately reshaping the standard of care for this disease.
