Introduction: Triple-negative breast cancer (TNBC) remains a major clinical challenge due to its aggressive behavior, molecular heterogeneity, and lack of targeted therapies. Accurate early prediction of response to neoadjuvant chemotherapy is critical for guiding personalized treatment and avoiding unnecessary toxicity. While prior approaches rely on tissue-level features from H&E WSIs, they often overlook the cellular and biomarker-specific information embedded within the tumor microenvironment.

Methods: We propose a model that performs virtual profiling by inferring spatial proteomic signals directly from routine H&E WSIs and leveraging them to predict pathologic complete response (pCR) in TNBC. Specifically, our model estimates cell-level biomarker expression patterns and captures their spatial organization within the tumor microenvironment, enabling a richer representation of tumor biology beyond conventional histomorphology. The resulting biomarker-informed cellular features are used to predict pCR from pre-treatment H&E WSIs. We evaluate our model on a cohort of 160 TNBC patients using 5-fold cross-validation and compare its performance against conventional: 1) clinical feature-based AI model and 2) Multimodal Histoclinical Model (MHM), that combines clinical features with imaging features using AI.

Results: Compared to both the clinical feature-based AI model and the MHM, our model shows improvements across all evaluation metrics, with relative gains of 19.71% and 10.10% in AUC, 25.91% and 4.07% in accuracy, 10.62% and 5.35% in sensitivity, and 43.24% and 8.63% in specificity, respectively.

Conclusion: This study presents a virtual profiling model that uses H&E images to estimate spatial proteomics information for predicting pCR in TNBC. Our results show the potential of AI-driven cellular inference to bridge the gap between histology and molecular characterization, offering a promising direction for assessment of tumor biology.

Learning Objectives:

1. Understand how deep learning can infer spatial proteomic signals from H&E WSIs for virtual tumor profiling.
2. Learn how biomarker-informed cellular features improve pCR prediction in TNBC over clinical and multimodal baselines.
3. Recognize the clinical value of AI-based virtual profiling for early, personalized chemotherapy response prediction.