Acta Biochimica et Biophysica Sinica
Identification and experimental validation of core genes associated with breast cancer brain metastasis via machine learning
Breast cancer (BC) is the most common malignancy among women, with approximately 2.3 million new cases diagnosed annually, accounting for approximately 11.6% of all cancer cases worldwide. Distant metastasis is the primary cause of mortality in BC patients, with nearly 50% of patients ultimately developing metastatic disease. The predominant metastatic sites of BC include the lung, liver, brain, and bone, each exhibiting distinct biological characteristics that drive the organ-specific tropism of cancer cells. Among these, brain metastasis represents a significant cause of mortality in BC patients and is particularly prevalent in those with human epidermal growth factor receptor 2 (HER2)-positive or triple-negative breast cancer (TNBC) subtypes. Breast cancer brain metastasis (BCBM) can manifest in three forms: choroid plexus metastasis (rare), leptomeningeal metastasis (approximately 8%), and parenchymal metastasis, the most common presentation, with multiple lesions in 78% of cases and solitary lesions in 14%. Distinct anatomical regions of the brain provide different micro-environments, which in turn shape epidemiological patterns, biological behaviors, and therapeutic vulnerabilities of metastatic cancer. With the continuous advancement of systemic therapies and imaging surveillance, brain metastases from BC have become increasingly prevalent, accounting for approximately 10%â30% of all metastatic breast cancer (MBC) cases. The continuous progression of BCBM often compromises patientsâ cognitive and sensory functions, leading to neurological impairment and severely limiting quality of life (QOL). Notably, the mortality rate within one year after diagnosis remains at 80%. Current therapeutic strategies for BCBM primarily include surgery, whole-brain radiotherapy (WBRT), stereotactic radiosurgery (SRS), chemotherapy, or combinations thereof. Although these approaches provide some clinical benefit, the efficacy remains limited due to the blood-brain barrier (BBB), which restricts drug penetration and contributes to chemoresistance. Therefore, elucidating the molecular mechanisms underlying BCBM is imperative to identify novel diagnostic biomarkers and therapeutic targets, with the ultimate goal of improving treatment efficacy and patient prognosis. Bioinformatics provides a powerful platform and data foundation for exploring the mechanisms of tumor initiation and progression. High-throughput platforms for gene expression analysis have gained significant popularity, with next-generation sequencing (NGS) and microarray analysis now widely applied as essential tools in medical oncology. These techniques have diverse clinical applications, including molecular cancer classification, prediction of therapeutic response, prognostic assessment, molecular diagnostics, and the discovery of novel drugs and therapeutic targets. Weighted gene coexpression network analysis (WGCNA) has been widely applied in studies of gene regulatory networks, biomarker discovery, and elucidation of the molecular mechanisms underlying complex phenotypes. In this study, we utilized the BCBM microarray dataset GSE43837. We performed differential expression analysis and WGCNA clustering using the R packages limma and WGCNA to identify potential gene modules and candidate targets. GSE43837 consists of 19 nonmetastatic primary breast tumor samples and 19 breast cancer brain metastasis samples. Differential expression analysis, with thresholds set at |logFC| > 1 and P < 0.05, identified 245 upregulated and 188 downregulated genes (Supplementary Table S1 and Supplementary Figure S1A). WGCNA further confirmed that the constructed network satisfied the scale-free topology criterion, with the optimal soft-threshold power determined to be 14 based on model fit and mean connectivity (Supplementary Figure S1B). Using the dynamic tree cut method, we clustered genes into multiple modules, each representing a group of coexpressed genes with varying degrees of correlation among modules (Supplementary Figure S1C,D). Notably, the midnightblue and black modules showed stronger correlations, and a significant positive relationship was observed between gene significance (GS) and module membership (MM) within these modules (Supplementary Figure S1E). This finding suggests that the core genes in these modules are highly representative and stable within the coexpression network. A total of 89 BCBM-related candidate genes were extracted from these key modules (Supplementary Table S2). To further identify key feature genes associated with BCBM, we applied two machine learning methods, LASSO regression and random forest (RF), to the 29 overlapping genes obtained from the intersection of DEGs and hub module genes (Figure 1A and Supplementary Table S3). In the LASSO regression analysis, the optimal penalty parameter λ was determined by cross-validation, yielding a set of candidate genes with nonzero regression coefficients (Figure 1B). Concurrently, in the RF model, 500 decision trees were constructed, and the classification ...
Graphical AbstractVerified
