To the Editor, We read with great interest the recent article by Lin et al titled “A Foundation Model for Predicting Outcomes of Neoadjuvant Chemotherapy in Breast Cancer”1. The study presents a compelling multimodal artificial intelligence (AI) model that integrates histopathological whole-slide images with clinicopathological data to predict pathological complete response (pCR) and disease-free survival (DFS) in breast cancer patients undergoing neoadjuvant chemotherapy (NAC). We commend the authors for their innovative and clinically relevant work, which represents a significant step toward precision oncology in the neoadjuvant setting. This letter was prepared in accordance with the TITAN 2025 Guidelines2. The hybrid transformer–CNN architecture, pretrained on The Cancer Genome Atlas and fine-tuned on institutional cohorts, demonstrates remarkable predictive performance, with AUCs reaching 0.994 for pCR and 0.885 for 4-year DFS in the blinded test cohort. These results underscore the potential of foundation models in medical imaging and multimodal data integration. Importantly, the model not only predicts outcomes but also provides interpretable attention heatmaps, linking regions of lymphocyte infiltration to treatment response – a finding consistent with established immunophenotypic correlates of chemosensitivity3,4. From a clinical perspective, the model’s ability to stratify non-pCR patients into risk categories and identify those who may benefit from intensive adjuvant chemotherapy is particularly noteworthy. This addresses a critical unmet need in post-NAC management, where decisions regarding adjuvant therapy intensity remain challenging. The integration of routinely available H&E slides and standard immunohistochemical markers (ER, PR, HER2 and Ki67) enhances the model’s feasibility and scalability compared to costly multigene assays or multiomics platforms. We do, however, offer several considerations for future research. First, while the retrospective design and rigorous validation across three cohorts are strengths, prospective multicenter validation will be essential to confirm generalizability across diverse populations and healthcare settings. Second, expanding the model to incorporate dynamic data – such as serial imaging or circulating tumor DNA or next-generation sequencing – could further refine predictive accuracy and monitoring of treatment response, in parallel with ongoing scientific advances. Finally, as the authors note, ethical and practical implementation challenges, including computational resource requirements and model interpretability in real-time clinical workflows, warrant continued attention. In summary, Lin et al have developed a robust and interpretable AI foundation model that advances the personalization of NAC and adjuvant therapy in breast cancer. Their work exemplifies the transformative potential of multimodal AI in oncology and sets a strong foundation for future research aimed at improving patient outcomes through data-driven decision support.
Chen et al. (Wed,) studied this question.