RAPPPID: towards generalizable protein interaction prediction with AWD-LSTM twin networks

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning

Background:

• Protein-protein interactions (PPIs) are crucial for cellular functions.
• Predicting PPIs computationally is vital but challenged by generalization to unseen proteins.
• Existing methods suffer from information leakage and sampling biases in prediction datasets.

Purpose of the Study:

• Introduce RAPPPID, a novel deep learning method for regularized PPI prediction.
• Address the challenge of model generalization to proteins not encountered during training.
• Provide standardized datasets to facilitate future assessment of PPI prediction methods.

Main Methods:

• Developed RAPPPID, a twin Averaged Weight-Dropped Long Short-Term memory network.
• Employed multiple regularization techniques during training to learn generalized weights.
• Utilized stringent datasets comprising proteins absent from training data for evaluation.

Main Results:

• RAPPPID significantly outperforms state-of-the-art methods on unseen proteins.
• Performance remains robust across different testing sets and protein compositions.
• Prediction accuracy is higher for experimentally validated PPIs (edges).

Conclusions:

• Appropriate regularization is key to developing generalizable PPI prediction models.
• RAPPPID offers a robust solution for predicting protein-protein interactions.
• The study provides valuable resources for the research community to assess PPI prediction methods.