Abstract
Protein protein interactions (PPIs) are essential to most of the biological processes. The prediction of PPIs is beneficial to the understanding of protein functions and thus is helpful to pathological analysis, disease diagnosis and drug design etc. As the amount of protein data is growing fast in the post genomic era, high-throughput experimental methods are expensive and time-consuming for the prediction of PPIs. Thus, computational methods have attracted researcher’s attention in recent years. A large number of computational methods have been proposed based on different protein sequence encoders.
Introduction
Proteins rarely act alone; they interact in complex networks to facilitate vital biological functions from metabolism to DNA replication. Mapping these protein-protein interactions (PPIs) is essential for drug discovery and disease diagnosis. However, traditional laboratory methods like yeast two-hybrid screens are slow, expensive, and prone to noise.
While many computational models exist to predict these interactions from amino acid sequences, most treat the problem as a simple binary choice. This ignores the confidence score—a numerical value indicating the likelihood of an interaction. The authors of this study argue that there is "ordinal information" hidden in these scores that standard models fail to exploit. By treating PPI prediction as an ordinal regression task, the proposed OR-RCNN framework achieves significantly higher accuracy.
The OR-RCNN Framework
The OR-RCNN (Ordinal Regression and Recurrent Convolutional Neural Network) framework consists of two primary modules: an encoder for feature extraction and a predictor based on ordinal regression.
Encoder: Recurrent Convolutional Neural Network
To process protein sequences, the model uses two identical RCNNs with shared parameters. Convolutional layers extract robust local features, while Bidirectional Gated Recurrent Units (GRUs) capture long-range sequential information. Unlike models that use residual shortcuts, OR-RCNN uses a concatenate operator to preserve and reuse features more effectively. The two protein embeddings are then merged via element-wise multiplication.
Predictor: Ordinal Regression
Instead of predicting a single continuous number, the model divides the confidence score range (0 to 1) into 20 sub-intervals. It trains 20 binary sub-classifiers to determine if the interaction score is higher than a specific threshold. The results of these sub-classifiers are aggregated to produce a final, highly accurate confidence score.
Experimental Results
The study utilized high-throughput data from the STRING database, focusing on S. cerevisiae (Baker’s yeast) and Homo sapiens (Human).
Embedding Performance
The choice of sequence embedding is critical for model performance. The researchers compared different embedding strategies across various sub-classifiers to identify the most effective feature representation.
Key Insights and Accuracy
OR-RCNN significantly outperformed existing methods like Random Forest (RF), XGBoost, and Support Vector Machines (SVM). On the S. cerevisiae dataset, the model reduced the Mean Absolute Error (MAE) by 49.78% and the Mean Squared Error (MSE) by 57.33% compared to the best-performing baseline. Similar gains were observed in human data, demonstrating the model's robustness across species.
Conclusion
By reframing PPI prediction as an ordinal regression task and utilizing a specialized RCNN architecture for feature reuse, OR-RCNN provides a more accurate and stable method for predicting interaction confidence. While the model is computationally intensive, its ability to capture the "rank" of interaction certainty represents a significant leap forward in computational proteomics. Future work may involve integrating 3D structural data to further refine these predictions.
Summary updated