Table 4.

Individual and knowledge-based group classification accuracies of 3DCNN and the FEATURE Softmax classifier

Method – Dataset	Single Class accuracy	Knowledge-Based Group Accuracy
3DCNN- train	0.55	0.6732
3DCNN- test	0.425	0.573
FEATURE vectors -train	0.245	0.416
FEATUR vectors -test	0.237	0.405

The deep 3DCNN achieves superior prediction performance compared to models that employ conventional structure-based hand-engineered biochemical features. A two-fold increase in prediction accuracies is achieved by the 3DCNN compared to the FEATURE Softmax Classifier. 3DCNN correctly predict amino acid types for structures in the test dataset, which are in proteins families different from the ones in the training dataset