Fold Change and False Discovery Rate

Fold change (FC) is used to calculate the differential multiples of gene expression values between cancer samples and normal samples, which is the basic method for detecting differential genes, and represents the expression values of feature i and sample j in cancer samples and normal samples; FC is defined as:

When FC exceeds the initial set threshold, it can be considered that the feature is different, and it is generally considered that there is a significant difference when the difference multiple is more than 2. FC can directly obtain the differentially expressed values, but in the absence of false-positive control, the rate of false-positive results is relatively high.

According to statistical theory, in multiple-hypothesis testing, it is important to control the probability of making mistakes in multiple statistical inferences, called FDR. FDR can be used to analyze deferentially expressed genes to control the proportion of false positives (Reiner-Benaim, 2010). Table 2 shows the confusion matrix for the statistical test. FDR can be defined as follows:

The optimal parameters for each step in the proposed method.

The number of false positives in multiple-hypothesis tests can be controlled by controlling that FDR is below the threshold q. In general, keep FDR below 0.01, or ensure that there is at most one false positive for every 100 positive hypotheses. Feature genes with significant differences can be identified by FC and FDR, but these two methods do not evaluate the classification performance of these features.

Fold change and FDR are applied to integrative data to select the differentially expressed genes. By comparing the classification balanced accuracy under different FC and FDR thresholds shown in Supplementary Tables 3, 4, the optimal value of FC and FDR thresholds is obtained: log (FC) > 1.0, FDR < 0.05.