Fig. 2.

The figure provides a high-level overview and example of the DeepGOZero model. On the left, a protein P is embedded in a vector space using an MLP whereas the right side shows how GO axioms are embedded using the EL Embedding method; the MLP embeds the protein in the same space as the GO axioms. The example above shows a protein P which is annotated to positive regulation of protein kinase B signaling (GO: 0051897). This class is defined as biological regulation (GO: 0065007) and positively regulates (RO: 0002213) some protein kinase B signaling (GO: 0043491). This knowledge allows us to annotate proteins with GO: 0051897 even if we do not have any training proteins (zero-shot). Both the protein and the GO class embeddings are optimized jointly during training of DeepGOZero