Comparison of learning strategies based on percent accuracy.

The table lists, for each prediction task, the per-residue percent accuracy achieved via single-task training of the neural network with just the PSI-BLAST features (“Single”), single-task training that includes the amino acid embedding (“Embed”), multitask training just using the PSI-BLAST features (“Multi”), multitask training including the amino acid embedding (“Multi-Emb”), multitask training of one task along with the natural protein task (“NP”), multitask training without the PSI-BLAST embedding module but initializing the amino acid embedding by using the natural protein task (“NP only”), multitask training including the natural protein task (“All3”), “All3” with Viterbi post-processing (“All3+Vit”) and a previously reported method (“Previous”). Each row corresponds to a single task. The -value column indicates whether the difference between “Single” and “All3+Vit” is significant, according to a Z-test. The “CV” column is computed based on the accuracies separately for each cross-validation fold. It counts the percentage of CV folds in which the “All3+Vit” method outperforms the “Single” method. Rows labeled “(prot)” or “(seg)” report the protein- or segment-level accuracy, rather than residue-level accuracy. For the “NP” setting, the “*” in the “Embedding?” row indicates that this network uses the pre-trained embedding layer from the natural protein task.