Table 8.
Zero-shot and trained prediction performance on specific classes with more than 100 annotations
| Ontology | Term | Name | AUC (test) | AUC (all) | AUC (trained) | AUC (trained mlp) |
|---|---|---|---|---|---|---|
| mf | GO: 0001227 | DNA-binding transcription repressor activity, RNA polymerase II-specific | 0.257 | 0.405 | 0.932 | 0.926 |
| mf | GO: 0001228 | DNA-binding transcription activator activity, RNA polymerase II-specific | 0.574 | 0.699 | 0.948 | 0.944 |
| mf | GO: 0003735 | Structural constituent of ribosome | 0.400 | 0.194 | 0.940 | 0.942 |
| mf | GO: 0004867 | Serine-type endopeptidase inhibitor activity | 0.972 | 0.967 | 0.985 | 0.984 |
| mf | GO: 0005096 | GTPase activator activity | 0.847 | 0.870 | 0.938 | 0.960 |
| bp | GO: 0000381 | Regulation of alternative mRNA splicing, via spliceosome | 0.855 | 0.865 | 0.906 | 0.886 |
| bp | GO: 0032729 | Positive regulation of interferon-gamma production | 0.870 | 0.919 | 0.932 | 0.906 |
| bp | GO: 0032755 | Positive regulation of interleukin-6 production | 0.719 | 0.819 | 0.884 | 0.873 |
| bp | GO: 0032760 | Positive regulation of tumor necrosis factor production | 0.861 | 0.906 | 0.925 | 0.867 |
| bp | GO: 0046330 | Positive regulation of JNK cascade | 0.855 | 0.894 | 0.904 | 0.916 |
| bp | GO: 0051897 | Positive regulation of protein kinase B signaling | 0.772 | 0.864 | 0.888 | 0.915 |
| bp | GO: 0120162 | Positive regulation of cold-induced thermogenesis | 0.637 | 0.789 | 0.738 | 0.835 |
| cc | GO: 0005762 | Mitochondrial large ribosomal subunit | 0.889 | 0.975 | 0.874 | 0.916 |
| cc | GO: 0022625 | Cytosolic large ribosomal subunit | 0.898 | 0.969 | 0.893 | 0.849 |
| cc | GO: 0042788 | Polysomal ribosome | 0.858 | 0.950 | 0.889 | 0.780 |
| cc | GO: 1904813 | Ficolin-1-rich granule lumen | 0.653 | 0.782 | 0.792 | 0.900 |
| Average | 0.745 | 0.804 | 0.898 | 0.900 |
Note: Evaluation measures are class-centric. AUC(test) is the zero-shot performance on the test set, i.e., neither the class nor the protein were included during model training; AUC(all) is the zero-shot performance on all proteins, i.e., the class was never seen during training but the model has seen the proteins (annotated with other classes) during training; AUC(trained) and AUC(trained mlp) is the performance of the DeepGOZero and MLP models on the testing set when trained with the class (i.e. the protein is not seen but other proteins with the class were used during training).