用岭回归法改善定量构效关系中量子化学参数的多重相关

Reduction of Multiple Correlativity of Quantum-Chemical Descriptors in QSPR by Using Ridge Regression

在多氯二苯-p-二氧杂芑(PCDD)类化合物极性气相色谱相对保留因子的QSPR分析中采用岭回归处理,取岭参数k=0.02时可有效抑制量子化学参数间多重相关性的不良影响。与传统的逐步回归法的结果相比,得到的QSPR函数表达更加符合微观过程的物理、化学原理。在外推区域对未知化合物性质的预报精度有明显的提高。证实岭回归方法得到的QSPR模型更为合理并具有更高的稳定性。

The ridge regression approach, with a ridge parameter k =0.02, was applied to the QSPR analysis of the relative retention indexes in polar gas chromatogram of the polychlorinated dibenzo-p-dioxin (PCDD) compounds. It was found that the approach could availably eliminate the bad effect of the multiple correlativity among the quantum-chemical descriptors. In comparison with the results from the conventional step regression, the QSPR function fitted by the ridge regression was in better agreement with the physical and chemical principles of the corresponding microcosmic process, and can thus give a prediction of higher accuracy to the unknown homologues in the extrapolative region. This shows that the QSPR model with ridge regression is of more reasonableness and higher stability.