Search

To screen differentially expressed genes quickly and efficiently in breast cancer, two gene microarray datasets of breast cancer, GSE15852 and GSE45255, were downloaded from GEO. By combining the Logistic Regression and Random Forest algorithm, this paper proposed a novel method named LR-RF to select differentially expressed genes of breast cancer on microarray data by the Bonferroni test of FWER error measure. Comparing with Logistic Regression and Random Forest, our study shows that LR-FR has a great facility in selecting differentially expressed genes. The average prediction accuracy of the proposed LR-RF from replicating random test 10 times surprisingly reaches 93.11 percent with variance as low as 0.00045. The prediction accuracy rate reaches a maximum 95.57 percent when threshold value α = 0.2 in the random forest algorithm process of ranking genes' importance score, and the differentially expressed genes are relatively few in number. In addition, through analyzing the gene interaction networks, most of the top 20 genes we selected were found to involve in the development of breast cancer. All of these results demonstrate the reliability and efficiency of LR-RF. It is anticipated that LR-RF would provide new knowledge and method for biologists, medical scientists, and cognitive computing researchers to identify disease-related genes of breast cancer.

Exploring the intrinsic differences among breast cancer subtypes is of crucial importance for precise diagnosis and therapeutic decision-making in diseases of high heterogeneity. The subtypes defined with several layers of information are related but not consistent, especially using immunohistochemistry markers and gene expression profiling. Here, we explored the intrinsic differences among the subtypes defined by the estrogen receptor, progesterone receptor and human epidermal growth factor receptor 2 based on the decision tree. We identified 30 mRNAs and 7 miRNAs differentially expressed along the tree's branches. The final signature panel contained 30 mRNAs, whose performance was validated using two public datasets based on 3 well-known classifiers. The network and pathway analysis were explored for feature genes, from which key molecules including FOXQ1 and SFRP1 were revealed to be densely connected with other molecules and participate in the validated metabolic pathways. Our study uncovered the differences among the four IHC-defined breast tumor subtypes at the mRNA and miRNA levels, presented a novel signature for breast tumor subtyping, and identified several key molecules potentially driving the heterogeneity of such tumors. The results help us further understand breast tumor heterogeneity, which could be availed in clinics.

Breast cancer is highly heterogeneous. The subtypes defined using immunohistochemistry markers and gene expression profilings (GEP) are related but not equivalent, with inter-connections under investigated. Our previous study revealed a set of differentially expressed genes (diff-genes), containing 1015 mRNAs and 69 miRNAs, which characterize the immunohistochemistry-defined breast tumor subtypes at the GEP level. However, they may convey redundant information due to the large amount of genes included. By reducing the dimension of the diff-genes, we identified 119 mRNAs and 20 miRNAs best explaining breast tumor heterogeneity with the most succinct number of genes found using hierarchical clustering and nearest-to-center principle. The final signature panel contains 119 mRNAs, whose superiority over diff-genes was replicated in two independent public datasets. The comparison of our signature with two pioneering signatures, the Sorlie's signature and PAM50, suggests a novel marker, FOXA1, in breast cancer classification. Subtype-specific feature genes are reported to characterize each immunohistochemistry-defined subgroup. Pathway and network analysis reveal the critical roles of Notch signalings in [ER+|PR+]HER2- and cell cycle in [ER+|PR+]HER2+ tumors. Our study reveals the primary differences among the four immunohistochemistry-defined breast tumors at the mRNA and miRNA levels, and proposes a novel signature for breast tumor subtyping given GEP data.