Jacobsson Micael, Karlén Anders
Department of Medicinal Chemistry, Faculty of Pharmacy, University of Uppsala, Box 574, SE-751 23 Uppsala, Sweden.
J Chem Inf Model. 2006 May-Jun;46(3):1334-43. doi: 10.1021/ci050407t.
A total of 945 known actives and roughly 10 000 decoy compounds were docked to eight different targets, and the resulting poses were scored using 10 different scoring functions. Three different score postprocessing methods were evaluated with respect to improvement of the enrichment in virtual screening. The three procedures were (i) multiple active site correction (MASC) as has been proposed by Vigers and Rizzi, (ii) a variation of MASC where corrections terms are predicted from simple molecular descriptors through PLS, PLS MASC, and (iii) size normalization. It was found that MASC did not generally improve the enrichment factors when compared to uncorrected scoring functions. For some combinations of scoring functions and targets, the enrichment was improved, for others not. However, by excluding the standard deviation from the MASC equation and transforming the scores for each target to a mean of 0 and a standard deviation of 1 (unit variance normalization), the performance was improved as compared to the original MASC method for most combinations of targets and scoring functions. Furthermore, when the molecular descriptors were fit to the mean scores over all targets and the resulting PLS models were used to predict mean scores, the enrichment as compared to the raw score was improved more often than by straightforward MASC. A high to intermediate linear correlation between the score and the number of heavy atoms was found for all scoring functions except FlexX. There seems to be a correlation between the size dependence of a scoring function and the effectiveness of PLS MASC in increasing the enrichment for that scoring function. Finally, normalization by molecular weight or heavy atom count was sometimes successful in increasing the enrichment. Dividing by the square or cubic root of the molecular weight or heavy atom count instead was often more successful. These results taken together suggest that ligand bias in scoring functions is a source of false positives in structure-based virtual screening. The number of false positives caused by ligand bias may be decreased using, for example, the PLS MASC procedure proposed in this study.
总共945种已知活性化合物和大约10000种诱饵化合物与八个不同靶点进行对接，对接得到的构象使用10种不同的评分函数进行打分。针对虚拟筛选中富集度的提高，评估了三种不同的分数后处理方法。这三种方法分别是：(i) Vigers和Rizzi提出的多活性位点校正（MASC）；(ii) MASC的一种变体，其中校正项通过偏最小二乘法（PLS）从简单分子描述符预测得到，即PLS MASC；(iii) 尺寸归一化。结果发现，与未校正的评分函数相比，MASC一般不会提高富集因子。对于某些评分函数和靶点的组合，富集度有所提高，而对于其他组合则没有。然而，通过从MASC方程中排除标准差，并将每个靶点的分数转换为均值为0、标准差为1（单位方差归一化），与原始MASC方法相比，对于大多数靶点和评分函数的组合，性能得到了改善。此外，当将分子描述符拟合到所有靶点的平均分数，并使用所得的PLS模型预测平均分数时，与原始分数相比，富集度提高的情况比直接使用MASC更为常见。除了FlexX之外，所有评分函数的分数与重原子数量之间都存在高度到中度的线性相关性。评分函数的尺寸依赖性与PLS MASC在提高该评分函数富集度方面的有效性之间似乎存在相关性。最后，通过分子量或重原子数进行归一化有时能够成功提高富集度。相反，除以分子量或重原子数的平方根或立方根往往更有效。综合这些结果表明，评分函数中的配体偏差是基于结构的虚拟筛选中假阳性的一个来源。使用例如本研究中提出的PLS MASC程序，可以减少由配体偏差引起的假阳性数量。
