Machine learning-based pathomics signature as a novel digital marker for C-X-C motif chemokine ligand 9 expression and prognostic assessment in breast cancer

Objectives: C-X-C motif chemokine ligand 9 (CXCL9) is a critical mediator of tumor-infiltrating lymphocyte recruitment and a potential biomarker in breast cancer (BC). This study aimed to develop a machine learning-based handcrafted pathomics model to forecast survival outcomes by characterizing CXCL9-related morphological signatures from routine hematoxylin and eosin (H&E) slides. The biological relevance and clinical utility of the derived pathomics score (PS) were subsequently evaluated to provide an objective framework for risk stratification. Material and Methods: We developed a pathomics model integrating quantitative H&E image features with machine learning in 791 Cancer Genome Atlas-BC patients. Feature selection using minimum redundancy maximum relevance and recursive feature elimination was performed exclusively within the training set, and a gradient boosting machine model was constructed to output a PS representing the predicted probability of CXCL9 expression. Model performance was assessed in the training, testing, and external validation cohorts. Associations of PS with survival, immune infiltration, cancer-immunity cycle (CI cycle) activity, tumor mutational burden, and molecular subtypes were investigated through integrated bioinformatics. Results: CXCL9 expression was confirmed as a significant prognostic factor for overall survival ( P = 0.037). The pathomics model demonstrated stable performance for predicting CXCL9 status in training, testing, and external cohorts (area under the curves: 0.741, 0.729, and 0.704, respectively). The PS significantly correlated with CXCL9 expression, stratified patients into distinct prognostic groups ( P = 0.012), and remained a significant protective factor in multivariate analysis (hazard ratio = 0.66, P = 0.048). Mechanistically, high PS was associated with immune-inflamed features, characterized by increased natural killer cells and CD8+ T-cell infiltration, enhanced CI cycle activity, and elevated TMB. Furthermore, the PS effectively reflected clinical heterogeneity, with triple-negative BC exhibiting the highest scores compared to other molecular subtypes ( P <0.001). Conclusion: This pathomics framework provides an imaging-derived indicator of CXCL9 expression and may support prognostic stratification in BC. Its association with immune-inflamed characteristics suggests potential relevance for immune phenotype assessment, but its value for predicting immunotherapy response requires validation in immunotherapy-treated cohorts.