Fig. 3

The network architecture of DeepGOA. The upper yellow subnetwork is the convolutional network part. The amino acids are extracted by convolution kernels of different sizes, and the fully connected layer is used to learn the mapping from sequence features to semantic representations of GO terms. The lower blue subnetwork is the graph convolution part, it uses the GO hierarchy \({H^{0}} \in {\mathbb {R}^{{\left | {\mathcal {T}} \right |} \times {\left | {\mathcal {T}} \right |}}}\) and empirical correlations between GO terms stored in \(A \in {\mathbb {R}^{\left |\mathcal {T} \right | \times {\left | {\mathcal {T}} \right |}}}\) to learn the semantic representation of each GO term. The dot product is finally used to guide the mapping between proteins and GO terms and to reversely adjust the representations of proteins and GO terms. In this way, the associations between GO terms and proteins are also predicted