Abstract
BACKGROUND: Breast cancer, a heterogeneous malignancy, comprises multiple subtypes and poses a substantial threat to women's health globally. Neoadjuvant therapy (NAT), administered prior to surgery, is integral to breast cancer treatment strategies. It aims to downsize tumors, optimize surgical outcomes, and evaluate tumor responsiveness to treatment. However, accurately predicting NAT efficacy remains challenging due to the disease's complexity and the diverse responses across different molecular subtypes. METHODS: In this study, we harnessed multimodal data, including proteomic, genomic, MRI imaging, and clinical information, sourced from multiple cohorts such as I-SPY2, TCGA-BRCA, GSE161529, and METABRIC. Post data preprocessing, Lasso regression was utilized for feature extraction and selection. Five machine learning algorithms were employed to construct diagnostic models, with pathological complete response (pCR) as the predictive endpoint. RESULTS: Our results revealed that the multi-omics Ridge regression model achieved the optimal performance in predicting pCR, with an AUC of 0.917. Through unsupervised clustering using the R package MOVICS and nine clustering algorithms, we identified four distinct multimodal breast cancer subtypes associated with NAT. These subtypes exhibited significant differences in proteomic profiles, hallmark cancer gene sets, pathway activities, tumor immune microenvironments, transcription factor activities, and clinical characteristics. For instance, CS1 subtype, predominantly ER-positive, had a low pCR rate and poor response to chemotherapy drugs, while CS4 subtype, characterized by high immune infiltration, showed a better response to immunotherapy. At the single-cell level, we detected significant heterogeneity in the tumor microenvironment among the four subtypes. Malignant cells in different subtypes displayed distinct copy number variations, differentiation levels, and evolutionary trajectories. Cell-cell communication analysis further highlighted differential interaction patterns among the subtypes, with implications for tumor progression and treatment response. CONCLUSION: Our multimodal diagnostic model and subtype analysis provide novel insights into predicting NAT efficacy in breast cancer. These findings hold promise for guiding personalized treatment strategies. Future research should focus on experimental validation, in-depth exploration of the underlying mechanisms, and extension of these methods to other cancers and treatment modalities.