Hall Vincent, Sklepari Meropi, Rodger Alison
MOAC, Department of Chemistry and School of Engineering, University of Warwick, Coventry, UK.
Chirality. 2014 Sep;26(9):471-82. doi: 10.1002/chir.22338. Epub 2014 Jun 2.
Collecting circular dichroism (CD) spectra for protein solutions is a simple experiment, yet reliable extraction of secondary structure content is dependent on knowledge of the concentration of the protein--which is not always available with accuracy. We previously developed a self-organizing map (SOM), called Secondary Structure Neural Network (SSNN), to cluster a database of CD spectra and use that map to assign the secondary structure content of new proteins from CD spectra. The performance of SSNN is at least as good as other available protein CD structure-fitting algorithms. In this work we apply SSNN to a collection of spectra of experimental samples where there was suspicion that the nominal protein concentration was incorrect. We show that by plotting the normalized root mean square deviation of the SSNN predicted spectrum from the experimental one versus a concentration scaling-factor it is possible to improve the estimate of the protein concentration while providing an estimate of the secondary structure. For our implementation (51 data points 240-190 nm in nm increments) good fits and structure estimates were obtained if the NRMSD (normalized root mean square displacement, RMSE/data range) is <0.03; reasonable for NRMSD <0.05; and variable above this. We also augmented the reference database with 100% helical spectra and truly random coil spectra.
收集蛋白质溶液的圆二色性（CD）光谱是一个简单的实验，但二级结构含量的可靠提取取决于蛋白质浓度的相关知识——而这并非总能准确获得。我们之前开发了一种自组织映射（SOM），称为二级结构神经网络（SSNN），用于对CD光谱数据库进行聚类，并使用该映射从CD光谱中确定新蛋白质的二级结构含量。SSNN的性能至少与其他可用的蛋白质CD结构拟合算法一样好。在这项工作中，我们将SSNN应用于一系列实验样品的光谱，这些样品怀疑其标称蛋白质浓度不正确。我们表明，通过绘制SSNN预测光谱与实验光谱的归一化均方根偏差相对于浓度缩放因子的关系图，可以在提供二级结构估计的同时改进蛋白质浓度的估计。对于我们的实现（51个数据点，波长范围为240 - 190 nm，以nm为增量），如果归一化均方根位移（NRMSD，均方根误差/数据范围）<0.03，则可获得良好的拟合和结构估计；NRMSD <0.05时结果合理；高于此值则结果可变。我们还用100%螺旋光谱和真正的无规卷曲光谱扩充了参考数据库。
