ADH-PPI

ADH-PPI predicts protein–protein interactions using an attention-based deep hybrid model that combines unsupervised FastText embeddings with LSTM, CNN, and self-attention to improve PPI prediction accuracy.

Key Features:

• Innovative Statistical Representation: Uses transfer learning in an unsupervised manner via FastText embedding generation to represent protein sequences for downstream prediction.
• Hybrid Neural Architecture: Integrates LSTM, convolutional neural network (CNN), and self-attention layers, where LSTM captures temporal dependencies, CNN extracts local sequence features, and self-attention focuses on relevant sequence regions.
• Superior Predictive Performance: Reports a 4% overall accuracy improvement and a 6% increase in Matthews correlation coefficient across two species benchmark datasets, and a 7% accuracy increase on four independent test sets.

Scientific Applications:

• Functional Genomics: Enables inference of protein functions and roles within cellular processes through predicted PPIs.
• Disease Research: Supports identification of disease-associated interaction networks and potential therapeutic targets.
• Drug Discovery: Facilitates identification of candidate molecules that may modulate protein–protein interactions.

Methodology:

Protein sequences are transformed into FastText embeddings using unsupervised transfer learning; the hybrid model combining LSTM, CNN, and self-attention layers is trained on benchmark datasets to learn discriminative PPI features; performance is evaluated using accuracy and Matthews correlation coefficient on benchmark and independent test sets.