Selection of key descriptors is very important in QSPR analysis. Presence of noise in the subset of descriptors reduces the quality of predictions. A complete set is considered as perfect when it does not include irrelevant or redundant elements. This paper reports complete sets of descriptors used to develop QSPR models for 1786 ¹³C NMR chemical shifts (δ_C parameters) of carbon atoms in 125 diverse chemical compounds. PBE1PBE/6-311G(2d,2p) and B3LYP/6-31G(d) basis sets were used for quantum chemistry calculations after the molecular structures were optimized with semi-empirical AM1 and B3LYP/6-31G(d). The two complete sets consisting of magnetic shielding elements (σ_XX, σ_YY, σ_ZZ) and the chemical shift principal values (σ₁₁, σ₂₂, σ₃₃) were used as the inputs for support vector machine (SVM) models of δ_C parameters. The four SVM models obtained have the mean root mean square (rms) errors of about 4.5–4.6 ppm. The results suggest that SVM models are accurate and acceptable compared with previous models, although our models are based on a relatively large set of compounds. Our approach is valuable in the selection of important descriptors for QSPR studies of δ_C parameters.

在QSPR分析中，关键描述符的选择非常重要。描述符子集中存在噪声会降低预测的质量。当一个完整的集合不包含无关或多余的元素时，它被认为是完美的。本文报道用于开发QSPR模型1786点的描述符的成套¹³ C NMR化学位移（δ _Ç在125种的不同化合物的碳原子的参数）。在用半经验AM1和B3LYP / 6-31G（d）优化分子结构后，将PBE1PBE / 6-311G（2d，2p）和B3LYP / 6-31G（d）基础集用于量子化学计算。由磁屏蔽元件的两个成套（σ _XX，σ _YY，σ_ZZ），化学位移主值（σ ₁₁，σ ₂₂，σ ₃₃）用作用于支持向量机（SVM）的模型输入δ _Ç参数。获得的四个SVM模型的均方根（rms）均方根误差约为4.5-4.6 ppm。结果表明，尽管我们的模型基于相对较大的一组化合物，但与以前的模型相比，SVM模型是准确且可以接受的。我们的做法是对的定量构效关系研究的重要描述选择有价值δ _Ç参数。