Abstract
Protein interactions play an essential role in studying living systems and life phenomena. A considerable amount of literature has been published on analyzing and predicting protein interactions, such as support vector machine method, homology-based method and similarity-based method, each has its pros and cons. Most existing methods for predicting protein interactions require prior domain knowledge, making it difficult to effectively extract protein features. Single method is dissatisfactory in predicting protein interactions, declaring the need for a comprehensive method that combines the advantages of various methods. On this basis, a deep ensemble learning method called EnAmDNN (Ensemble Deep Neural Networks with Attention Mechanism) is proposed to predict protein interactions which is an appropriate candidate for comprehensive learning, combining multiple models, and considering the advantages of various methods. Particularly, it encode protein sequences by the local descriptor, auto covariance, conjoint triad, pseudo amino acid composition and combine the vector representation of each protein in the protein interaction network. Then it takes advantage of the multi-layer convolutional neural networks to automatically extract protein features and construct an attention mechanism to analyze deep-seated relationships between proteins. We set up four different structures of deep learning models. In the ensemble learning model, second layer data sets are generated with five-fold cross validation from basic learners, then predict the protein interaction network by combining 16 models. Results on five independent PPI data sets demonstrate that EnAmDNN achieves superior prediction performance than other comparing methods.
Introduction
Proteins rarely act in isolation; they interact in complex networks to drive vital cellular processes like signal transduction and metabolism. Mapping these protein-protein interactions (PPIs) is crucial for understanding disease mechanisms and designing therapeutic interventions. Historically, researchers relied on experimental methods that are slow and expensive, or computational methods requiring extensive prior domain knowledge. The challenge addressed here is predicting whether two proteins will interact using only their raw amino acid sequences, bypassing the need for evolutionary or localization data.
Methodology: The EnAmDNN Framework
The proposed EnAmDNN (Ensemble Deep Neural Networks with Attention Mechanism) framework utilizes a multi-stage pipeline to automate feature extraction and improve prediction robustness.
Feature Encoding
To capture the diverse properties of protein sequences, the model employs four distinct encoding techniques:
- Auto Covariance (AC): Captures physicochemical properties over various sequence distances.
- Conjoint Triad (CT): Groups the 20 amino acids into seven categories based on their dipole and side-chain volume to represent the sequence as a vector of triad frequencies.
- Local Descriptor (LD): Analyzes the composition, transition, and distribution of amino acids across 10 specific sequence regions to capture local information.
- Pseudo Amino Acid Composition (PseAAC): Incorporates sequence-order information alongside traditional composition data.
Deep Learning and Attention
Following encoding, multi-layer Convolutional Neural Networks (CNNs) automatically learn abstract representations from these features. A key innovation is the inclusion of a multi-head attention mechanism. Unlike standard models that simply concatenate protein vectors, the attention mechanism analyzes the "internal links" between protein pairs, selectively focusing on the sequence segments most relevant to the interaction.
Ensemble Stacking
To ensure stability, the researchers trained 16 base learners using different combinations of the four encodings and four distinct DNN architectures. The outputs of these learners were then integrated into a second-level classifier to produce the final prediction.
Experimental Results
The model was evaluated on five independent PPI datasets related to major human diseases: Parkinson’s Disease (PD), Alzheimer’s Disease (AD), Cancer, Cardiac Disease, and Diabetes.
EnAmDNN consistently outperformed traditional methods like SVM and recent deep learning models like EnsDNN. On the Alzheimer’s dataset, the model achieved a peak accuracy of 94.67% and an AUC of 94.63%. Across all five disease datasets, EnAmDNN improved the AUC index by 0.41% to 3.90% compared to previous state-of-the-art methods. This high precision is critical for biological research, as it minimizes the risk of expensive, failed laboratory experiments resulting from false positives.
Conclusion
EnAmDNN demonstrates that combining diverse sequence-encoding techniques with attention mechanisms and ensemble learning significantly enhances the prediction of protein-protein interactions. While the computational cost of training 16 models is high, the resulting accuracy and stability offer a powerful tool for reconstructing protein interaction networks from raw sequence data alone. Future work will focus on optimizing network configurations and exploring more complex attention variants.
Summary updated