ATTCry: Attention-based neural network model for protein crystallization prediction

Protein crystallization is the fundamental approach to solve the structure of protein. However, only a few (2%–10%) of these protein can be good crystallization. Recently, several computational methods have been proposed to predict protein crystallization. However, their model needs to select and extract thousands of physicochemical and structural handcrafted features, and the performances are modest. According to the properties of protein structure, we proposed a novel end-to-end attention-based deep neural network protein crystallization predictor called ATTCry. To capture the local k-mers feature of the raw protein sequence, We designed multi-scale convolutional neural networks (CNN) layer. Furthermore, to obtain more complex global spatial long-distance dependence of protein structure, we add multi-head self-attention layers to joint information from different representation subspaces at different positions parallelly. By integrating multi-scale and multi-head self-attention mechanisms, our method can capture both local and global features of protein sequences efficiently, thus enhance the robustness and generalization of protein crystallization prediction. Compared with other deep learning models for protein crystallization prediction, ATTCry reduces the amount of training parameters, and the model can be trained more efficiently. The experiments demonstrate that ATTCry outperforms significantly on three different test sets than all known crystallization predictors. It shows that ATTCry obtains relatively good predictive performance and outperforms existing methods. ATTCry is free available at https://github.com/zhanglabNKU/ATTCry