Prediction of Protein-Protein Interaction Sites Based on Stratified Attentional Mechanisms

Abstract

Proteins are the basic substances that undertake human life activities, and they often perform their biological functions through interactions with other biological macromolecules, such as cell transmission and signal transduction. Predicting the interaction sites between proteins can deepen the understanding of the principle of protein interactions, but traditional experimental methods are time-consuming and labor-intensive. In this study, a new hierarchical attention network structure, named HANPPIS, by adding six effective features of protein sequence, position-specific scoring matrix (PSSM), secondary structure, pre-training vector, hydrophilic, and amino acid position, is proposed to predict protein-protein interaction (PPI) sites. The experiment proved that our model has obtained very effective results, which was better than the existing advanced calculation methods. More importantly, we used the double-layer attention mechanism to improve the interpretability of the model and to a certain extent solved the problem of the "black box" of deep neural networks, which can be used as a reference for location positioning on the biological level.

Keywords: deep learning; feature fusion; multilevel attention mechanism; protein features; protein–protein interaction.

FIGURE 1

Structure of HANPPIS. It consists of three steps, including embedded representation, amino acid–level attention and K-mers–level attention. We obtain vector representations of protein sequence fragments through multidimensional features. The vector representation of the protein fragment is the input to the first layer of the attention mechanism, and then the vector representation of the protein sequence is obtained through the second layer of attention and finally input to the prediction layer.

FIGURE 2

Amino acid feature generation and expression. This figure illustrates the specific details of the amino acid signature generation. Among them, because the pre-training vector feature dimension is too large, a layer of feedforward neural network is used to reduce the dimension to 50 dimensions. Then, the remaining five features are spliced and finally the 101-dimensional amino acid feature vector as the input of the entire model is obtained.

FIGURE 3

Attention distribution. This figure shows the attention visualization result of one of the samples (from the 135th amino acid of protein 1Z0J_A in Dset186, the sample sequence is “RDAKDYA” and the target amino acid is “K”). (A) shows the proportion of K-mers–level attention distribution and (B) shows the distribution of amino acid–level attention. As shown in the figure, the center position has the highest proportion of attention, which is also consistent with the task of protein interaction sites.