Archives

  • 2019-07
  • 2019-08
  • 2019-09
  • 2019-10
  • 2019-11
  • 2020-03
  • 2020-07
  • 2020-08
  • br linear unit as an activation function In the

    2019-09-16


    linear unit) as an activation function. In the output layer, we created two nodes (a1 and a2, for alive and deceased, respectively). We applied a softmax function to each node, and designated y2 (probability of death; that is, the a2 node) as Y. We utilized cross Omadacycline hydrochloride error as a loss function (E) and optimized the value of each weight with Adam method (learning rate, 0·001; epochs, 1000). After the training, we used the weights of the nodes (“Gene_Weight”) to calculate mPS (sum-mation of Gene_Score x Gene_Weight for all 23 genes). We used the Python-based Keras library for this neural network training.
    For validation, we used the other half of the METABRIC cohort (METABRIC test set) and the independent cohorts GSE86166 and GSE96058. Within each cohort, we converted expression level to binary status (above or below the median), which was then converted to Gene_Score by the above-mentioned step function. The cutoff criterion (median value) was study specific and calculated for each cohort independently.
    Kaplan-Meier plots were constructed with the use of R (survival package). The median value was used as the cutoff between low
    and high expression levels of each gene. For mPS validation, we truncated the survival data at 10 years unless indicated otherwise. We computed OS from the date of diagnosis to the date of death from any cause. For most of the data (Figs. 3–5 and Supplementary Figs. S5–S11), survival outcomes were compared with the log-rank test. For the survival analysis shown in Fig. 2 and Table 2, the HR and its 95% CI were calculated by Cox regression analysis after proper evaluation of the assumptions of the Cox regression models with the use of the survival package. Statistical significance was de-termined at a two-sided P value of 0·05, with the exception of the TCGA discovery cohort, for which we adopted 0·01 as the cutoff criterion.
    2.6. Data availability
    All the data analyzed in this study are open to the public and can be downloaded from cBioPortal (http://www.cbioportal.org) and GEO (https://www.ncbi.nlm.nih.gov/geo). A Web-based tool we cre-ated in this study is freely available at our github page (https:// hideyukishimizu.github.io/mPS_breast).
    Please cite this article as: H. Shimizu and K.I. Nakayama, A 23 gene–based molecular prognostic score precisely predicts overall survival of breast cancer pati..., EBioMedicine, https://doi.org/10.1016/j.ebiom.2019.07.046
    Please cite this article as: H. Shimizu and K.I. Nakayama, A 23 gene–based molecular prognostic score precisely predicts overall survival of breast cancer pati..., EBioMedicine, https://doi.org/10.1016/j.ebiom.2019.07.046
    3. Results
    Table 1
    The 23 genes necessary and sufficient for calculation of mPS. For genes in red, patients
    3.1. Limitation of hypothesis-driven approaches
    with a high level of expression (above the median) are assigned a score of 1. Conversely,
    for genes in blue, patients with a low level of expression (below the median) are assigned
    Since the initial discovery of the cancer-causing Src gene, tremen-
    a score of 1. None of the 23 genes are included in existing indicators of relapse-free sur-
    vival such as Oncotype and MammaPrint.
    dous advances have been made in the field of oncogenes. Given that
    Symbol Gene ID Full name Score Score Weight
    MYC plays many important roles related to cancer development [27],
    we hypothesized that the expression of MYC might be associated with
    OS in cancer patients. Breast cancer patients in the TCGA cohort [28]
    were divided into two groups (low and high expression level) on the
    basis of the median mRNA abundance for MYC, and the difference in sur-
    vival outcome between the two groups was assessed. Unexpectedly, OS
    did not differ between the two groups (Fig. 2a). Similarly, the mRNA