Assigning functional annotations to genes with previously unknown function is an important problem to solve in bioinformatics. A high fraction of both bacterial and especially bacteriophage still do not have functional annotations (e.g., unknown protein, hypothetical protein, etc.). However, phage structural protein functional assignment is difficult, with 50-90% of phage genes currently lacking functional assignments. Protein sequence homologybased methods often fail due to the high degree of divergence between phages, and experimental methods such as EM and crystallography are expensive and time consuming. Machine learning classifiers have been utilized to solve various forms of this problem with mixed results. The most sophisticated of these is PhANNs, or Phage Artificial Neural Networks, an ANN that classifies putative phage structural protein sequences as one of 10 common structural protein classes, or as an “others” class indicating none of the 10 classes apply, with an accuracy of 86.2% and an F-1 score of 0.875. This work describes three methods/approaches to improve the accuracy and utility of PhANNs: improving PhANNs input features, adding a post-network secondary classifier model, and changing the definitions of the predicted classes. Expanding the input features and changing the class definitions improved the accuracy of the model to 89% with an F-1 score of 0.90. The postnetwork secondary classifier was implemented as a proof of concept on a pair of classes that are often confused by PhANNs. This additional step improved the recall of the targeted class from 0.75 to 0.83.
