Fig. 2

CNN architecture of DEGnext. The input to the model is a 1D input vector (\(x_{1}\), \(x_{2}\), \(\dots\), \(x_{n}\)), which represents each gene row of a cancer dataset. This 1D vector is converted to a 2D matrix of channel 1 using np.reshape(). We used a sequence of eight 2D convolutional neural network (CNN) layers (\(C_{1}\), \(C_{2}\), \(\dots\), \(C_{8}\)) with ReLU() as activation function. Each CNN layer uses kernel-size (3, 3), stride of 1, and padding equal to 1. We used a 2D Maxpool layer of kernel-size 2. In order to make the model inclusive for any input size, we used a 2D AdaptiveMaxPool layer with target output size of 1 \(\times\) 1. The output of the CNN layers is fed to a sequence of 5 linear layers (\(L_{1}\), \(L_{2}\), \(\dots\), \(L_{5}\)) with ReLU() as activation function. We used Softmax() to the output of linear layers, to find the probabilities of each class in the range of [0, 1]