Explainable machine learning-guided integrated multiomics analysis reveals macrophage-driven immune suppression in breast cancer.

Despite thorough characterizations of cellular compositions within the breast tumor microenvironment (TME), their implications for disease progression and patient prognosis remain poorly understood. Unraveling these effects is vital for identifying potential targets to improve treatment outcomes. In this study, we devise an explainable machine learning (XML) pipeline to scrutinize the associations between TME cellular constituents and relapse-free survival (RFS). By applying this pipeline to estimated cell fractions in the METABRIC and TCGA datasets and comparing these results with associations to pathological complete response (pCR) after neoadjuvant chemotherapy (NAC), we create a comprehensive catalog of the TME's role based on 5000 patient samples. Our findings reveal an unexpected dichotomy in which macrophages correlate positively with pCR but negatively with RFS, particularly within estrogen receptor-positive (ER+) and Luminal A and B (LumA/B) cancer subtypes. We show that this pattern is driven by heterogeneity in breast tumors characterized by increasing levels of macrophage infiltration. Through imaging mass cytometry (IMC) data analysis, we find that macrophages tend to accumulate in the vicinity of HLA-ABC