RAPPPID

RAPPPID: Deep learning-based protein–protein interaction prediction

RAPPPID predicts protein–protein interactions (PPI) using a twin AWD-LSTM (Averaged Stochastic Gradient Descent with Adaptive Word Representations Long Short-Term Memory) network to model sequential protein features and improve generalization to unseen proteins.

Key Features:

• Twin AWD-LSTM Architecture: Implements a paired AWD-LSTM network to capture complex sequential patterns relevant to protein–protein interaction prediction.
• Regularisation Strategies: Applies multiple regularisation techniques during training to reduce information leakage and sampling bias in PPI datasets, enabling generalized weight learning.
• Generalization Capability: Maintains predictive performance for proteins not included in the training dataset.

Scientific Applications:

• Unseen Protein Interaction Prediction: Predicts interactions involving previously unobserved proteins to expand functional interaction networks.
• Benchmark Evaluation: Demonstrates high predictive accuracy on stringent interaction datasets compared to existing PPI prediction methods.
• Biologically Supported Edge Detection: Achieves higher accuracy for biologically supported interactions, supporting identification of functionally relevant protein interactions.

Methodology:

RAPPPID trains a twin AWD-LSTM network using protein sequence data and multiple regularisation techniques to mitigate dataset bias and information leakage, optimizing model parameters to enhance generalization and predictive accuracy across diverse PPI data splits.