PhANNs

PhANNs classifies bacteriophage open reading frames (ORFs) into ten structural protein classes or "others" to enable structural annotation of phage proteomes and to address limitations of homology-based methods.

Key Features:

• Classification targets: Predicts one of ten structural protein classes or assigns sequences to an "others" category for bacteriophage proteins.
• Model architecture: Uses an ensemble of Artificial Neural Networks (ANNs) for prediction.
• Training dataset: Trained on 538,213 manually curated phage protein sequences.
• Dataset partitioning: Employs an eleven-subset split created by a clustering method that ensures no homologous proteins between sets while maintaining sequence diversity.
• Validation strategy: Uses ten subsets for cross-validation and one subset reserved for independent testing.
• Performance: Achieved a test F1-score of 0.875 and a test accuracy of 86.2% on the reserved test set.
• Input data format: Operates on protein sequence data in multi-FASTA format.
• Homology-independent detection: Leverages machine learning to identify structural proteins when traditional homology-based methods are inadequate.

Scientific Applications:

• Functional annotation: Enables structural annotation of phage genes to reduce the fraction of uncharacterized ORFs in phage genomes.
• Phage genomics: Supports discovery of structural protein composition across bacteriophage genomes for comparative analyses.
• Microbial ecology: Facilitates investigation of phage contributions to microbial community structure and function.
• Phage therapy research: Assists evaluation of phage-encoded structural components relevant to therapeutic development.

Methodology:

An ensemble of ANNs was trained on 538,213 manually curated phage protein sequences; the dataset was split into eleven subsets by a clustering method that prevents homologous proteins between sets while maximizing sequence diversity, with ten subsets used for cross-validation and one subset held out for testing.