Grain Protein Function Prediction Based on CNN and Residual Attention Mechanism with AlphaFold2 Structure Data

The prediction of grain protein function is essential for the advancement of food science. Traditional experimental methods are associated with high costs and significant time requirements. Computational methods are recognized for their efficiency and reduced time demands. A new multimodal deep learning method, MMSNet, is proposed in this study, and protein data of four types of grains (japonica, indica, maize, and wheat) are analyzed. This method fuses the protein structure information predicted by AlphaFold2 and combines a multiscale one-dimensional convolutional neural network (1DCNN) with a two-dimensional convolutional neural network (2DCNN) to enable the model to capture sequence and structural information effectively. We used a residual attention mechanism to replace the traditional pooling layer, thereby improving the feature extraction capability of the network layers in 2DCNN. The experimental results indicate that secondary structure and spatial structure information contribute to improving model performance. Compared with two classical methods, MMSNet demonstrates optimal performance, which validates the effectiveness of our approach in integrating complex grain protein data and highlights its potential to open new avenues for grain protein function prediction.