DeephageTP

DeephageTP identifies phage-specific proteins in metagenomic datasets using a convolutional neural network to classify Portal, TerL (large terminase subunit), and TerS (small terminase subunit) proteins.

Key Features:

• Alignment-free CNN: Employs a convolutional neural network that learns amino acid sequence patterns without relying on sequence alignment.
• High prediction precision: Reports precision of 98.8% for Portal, 98.6% for TerL, and 97.8% for TerS on evaluated data.
• Cutoff loss value optimization: Determines per-protein cutoff loss values (TerL: -5.2; Portal: -4.2; TerS: -2.9) to maintain precision with larger datasets.
• Novel sequence identification: Detects phage sequences with remote homology that are undetectable by alignment-based methods such as DIAMOND and HMMER.
• Generalization of amino acid-based models: Acts as a natural generalization of amino acid-based models, enabling identification of similarities among viral proteins at high latitudes.

Scientific Applications:

• Phage genome identification and annotation: Large-scale identification and annotation of phage proteins within metagenomic assemblies.
• Novel protein discovery: Detection of novel phage proteins with low sequence conservation or remote homology to known proteins.
• Viromics and protein landscape analysis: Exploration of complex protein landscapes in viromic datasets to expand viral protein catalogs.
• Ecology, evolution, and therapeutic research: Application in metagenomic studies addressing viral ecology, viral evolution, and potential therapeutic phage protein discovery.

Methodology:

Processes one-hot encoded protein sequences as input to a CNN that extracts predictive features during training, uses optimized cutoff loss values per protein category, and was validated on three real metagenomic datasets.