基于卷积神经网络算法的X射线晶体衍射实验数据筛选

X-ray crystallography experimental data screening based on convolutional neural network algorithms

X射线串行晶体学作为一种解析蛋白质晶体结构的新方法,因为拥有室温采集、辐射损伤低、时间分辨等优势而得到迅速的发展。利用串行晶体学方法解析蛋白质结构需要在整合大量晶体衍射图的基础上筛选出有效的衍射数据,然而常规的数据筛选方法在处理衍射图时存在准确度不高且效率低的问题。基于卷积神经网络的数据筛选方法具有流程自动化的优势,并且已经被证明相对于传统的“找点法”具有更高的分类准确度。因此在比较5种不同卷积神经网络筛选晶体学衍射图的准确度和效率的基础上,选择并构建一个准确率高且运行速率快的卷积神经网络数据筛选工具,用于不同蛋白质晶体样品衍射图的筛选。结果显示:MobileNets不仅具有ResNet、GoogleNet-Inception等大型网络相似的准确度,而且运行速率更快,为串行晶体学实验提供了一个有效便捷的数据筛选工具。

[Background]Serial X-ray crystallography has developed rapidly due to its advantages of data collection at room temperature,low radiation damage and time resolution.To solve protein structures by using the serial X-ray crystallography,a large amount of produced diffraction data needs to be screened for finding the effective diffraction patterns.The use of convolutional neural networks(CNN)can not only automate the data screening process,but also improve the accuracy of data classification comparing with the traditional"point finding method".[Purpose]This study aims to explore five types of popular convolutional neural networks,i.e.,AlexNet,GoogleNet,MobileNets,Vgg16,ResNet,for screening crystallographic diffraction patterns,and compare the accuracy and efficiency of them to build up a fast and accurate convolutional neural network tool for screening the diffraction patterns of different protein crystal samples.[Methods]Firstly,the primitive data for model training extracted from the coherent X-ray image database,collected by Linac Coherent Light Source(LCLS)and Spring-8 Angstrom Compact free electron laser(SACLA),were pre-processed by gray level equalization and data enhancement.The deep learning models were trained by iteration of the preprocessed data.Then,the selected convolutional neural network through the comparison of accuracy and efficiency was used to process further the experimental data of protein crystals diffractions.[Results]The results show that MobileNets not only has the accuracy similar to large networks such as ResNet,GoogleNet-Inception,but also runs faster.[Conclusions]MobileNets provides an effective and convenient screening tool for serial X-ray crystallography experimental data.