LSTM-PHV

LSTM-PHV predicts protein-protein interactions (PPIs) between human and viral proteins using Long Short-Term Memory (LSTM) networks with word2vec embeddings applied to amino acid sequences.

Key Features:

• Model architecture: Uses Long Short-Term Memory (LSTM) networks combined with word2vec embeddings to learn from sequence data.
• Input representation: Employs word2vec to transform amino acid sequences into vector representations by treating sequences as tokens in a language model context.
• Structure-independence: Predicts PPIs without requiring 3D structural data.
• Imbalanced data handling: Trains effectively on highly imbalanced datasets where positive PPI samples are far outnumbered by negatives.
• Performance: Reported Area Under the Curve (AUC) scores of 0.976 and 0.973 and accuracies of 0.984 and 0.985 on training and independent test datasets.
• Generalization: Capable of predicting PPIs involving new or previously uncharacterized viruses by learning from sequence contexts alone.
• Visualization: Applies uniform manifold approximation and projection (UMAP) to illustrate separation between positive and negative PPI samples.
• Comparative performance: Reportedly surpasses existing state-of-the-art predictors in performance.

Scientific Applications:

• Interspecies PPI identification: Identifies human–viral PPIs to support understanding of viral infection mechanisms.
• Viral pathogenesis research: Facilitates exploration of human–virus interactions relevant to viral pathogenesis.
• Antiviral strategy development: Provides predicted host–virus PPIs to inform development of antiviral strategies.

Methodology:

Transforms amino acid sequences with word2vec into vector embeddings, trains LSTM networks on PPI datasets including highly imbalanced data, and employs UMAP for visualization.