Figure 4

Overview of the deep neural network used to classify mosquito and non-mosquito sequences. (A) A simple overview of how information from normalized FCGRs passes through each branch of the network. Each branch begins with FCGRs being split into patches. The information from each patch then passes through attention layers and a small fully connected feed-forward network. The layers predicting target information (real/synthetic, genera) are the last layers of the network. During training, the loss between predictions and actual targets is minimized by gradually adjusting the weights and biases of each layer. When predicting unknown labels, the layer which classifies each FCGR as either real or synthetic is discarded and only the taxonomic classification layer is used. Colors have been added only to aid in visualization. (B) The meta-classifier creates random training sets using the training data. The weights of each model are initialized randomly. This helps train a diverse set of models which can be used to classify unseen data.