Prediction of neoadjuvant therapy response in breast cancer based on interpretable artificial intelligence

Abstract

Background: To develop an AI-based predictive model for neoadjuvant therapy (NAT) efficacy in breast cancer, we integrated multimodal data and analyzed tumor microenvironment (TME) features to provide interpretability.

Methods: We retrospectively analyzed H&E-stained whole-slide images (WSIs) from a multicenter cohort of breast cancer patients receiving NAT to develop an AI predictive model. The cohort was stratified into training, test, internal validation, and external validation sets. Feature extraction used UNI and classification employed a multiple instance learning (MIL) framework. Model performance was evaluated via ROC curve analysis (AUC, precision, specificity, recall). Molecular mechanisms underlying model predictions were explored using TCGA multimodal data, integrating differential gene expression profiling with pathway enrichment analysis (GO, KEGG). TME component correlations with model scores were also investigated.

Results: The AI model demonstrated robust discriminative capacity across three residual cancer burden (RCB)-based classification tasks in 826 patients from two centers, achieving peak performance in subtask 2 (NAT-sensitive: RCB 0-1 vs. NAT-resistant: RCB 2-3). For subtask 2, AUCs were 0.901 (training), 0.858 (test), 0.808 (internal validation), and 0.819 (external validation). Molecular analysis linked the model's predictive efficacy to tumor cell cycle processes. TME analysis revealed positive correlations between model scores and activated immune cells (M0/M1 macrophages, dendritic cells), and negative correlations with inhibitory cells (M2 macrophages, resting mast cells). Crucially, the model's predictive scores were closely related to tumor-infiltrating lymphocytes (TILs), with spatial colocalization observed between classification weights and TILs distribution. Significant differences in TILs levels occurred across model score strata, validating the model's biological plausibility in predicting NAT response mechanisms.

Conclusion: We developed an interpretable AI model that predicts response to neoadjuvant therapy in breast cancer using H&E slides. The model's predictions are biologically interpretable, correlating with TME dynamics and spatial TIL patterns, offering a novel strategy for personalizing NAT treatment strategies.

Keywords: breast cancer; evaluation of therapeutic effect; explainable artificial intelligence; neoadjuvant therapy; tumor microenvironment.

Figure 1.

The data collection process of a pathomics-based AI model for predicting the prognosis of neoadjuvant therapy in breast cancer using WSI.

Figure 2.

Receiver operating characteristic (ROC) curves for the predictive performance of three RCB classifications and the confusion matrix of the external validation dataset. (A) The ROC cures of subtask 1 (RCB 0 vs. RCB 1, 2, 3); (B) the ROC cures of subtask 2 (RCB 0 and 1 vs. RCB 2 and 3); (C) the ROC cures of subtask 3 (RCB 0, 1, 2 vs. RCB 3); (D) the confusion matrices of subtask 1; (E) the confusion matrices of subtask 2; (F) the confusion matrices of subtask 3.

Figure 3.

Differential gene expression analysis and KEGG/GO enrichment analysis related to NAT efficacy in breast cancer. (A) Volcano plot of differentially expressed genes; (B) KEGG pathway enrichment plot of differentially expressed genes; (C) GO enrichment analysis plot of differentially expressed genes.

Figure 4.

Utilizing a predictive model to analyze changes in molecular mechanism mutations in breast cancer during NAT.

Figure 5.

Screening of tumor microenvironment components associated with neoadjuvant therapy in breast cancer. (A) Expression profiles of stromal and leukocyte fractions in the tumor microenvironment of NAT-sensitive and NAT-resistant breast cancer samples; (B) expression levels of six immune cell types within the leukocyte compartment in NAT-sensitive and NAT-resistant breast cancer samples; (C) expression patterns of immune cell subsets in NAT-sensitive and NAT-resistant breast cancer samples.

Figure 6.

TILs score in breast cancer patients receiving NAT. (A) Comparison of TILs scores between NAT-sensitive and NAT-resistant groups; (B) classification weight map for NAT-sensitive group; (C) lymphocyte aggregation areas in NAT-sensitive group; (D) classification weight map for NAT-resistant group; (E) lymphocyte aggregation areas in NAT-resistant group; (F) low magnification H&E staining of NAT-sensitive group (magnification × 1); (G) high magnification H&E staining of NAT-sensitive group (magnification × 10); (H) low magnification H&E staining of NAT-resistant group (magnification × 1); (I) high magnification H&E staining of NAT-resistant group (magnification × 10).

Figure 7.

Results from the external validation cohort on the correlation between sensitivity to NAT and the distribution characteristics of TILs in breast cancer. (A) Analysis of breast cancer patient grouping based on TILs percentage and NAT sensitivity. TILs are categorized into three levels: low TILs percentage (0–10%), intermediate TILs percentage (11–59%), and high TILs percentage (60–100%); (B) lymphocyte distribution weight map of the tumor microenvironment in the NAT-sensitive group; (C) lymphocyte distribution weight map of the tumor microenvironment in the NAT-resistant group; (D) low-magnification H&E staining section of tumor tissue in the NAT-sensitive group (magnification × 1); (E) high-magnification H&E staining section of tumor tissue in the NAT-sensitive group, showing the distribution of lymphocyte aggregation (magnification × 10); (F) low-magnification H&E staining section of tumor tissue in the NAT-resistant group (magnification × 1); (G) high-magnification H&E staining section of tumor tissue in the NAT-resistant group, showing sparse distribution of lymphocytes (magnification × 10). (H) Exemplar high attention patches from NAT-sensitive group and NAT-resistant group cases with corresponding cell labels. (I) Quantitative cellular profiling of the NAT-sensitive and NAT-resistant groups within the highly attended regions revealed comparative changes in tumor, lymphocytes, stromal cells, and necrosis and epithelial cells.