Home Computer Science Computational systems pharmacology and toxicology
Drawing Conclusions from Multiple Systems
each in silico model that adheres to the oECD (Q)sar validation principles should include information concerning its performance, often based on a cross-validation exercise or using an external test set.66 it has been shown that using two or more complementary in silico models may improve the detection of positive compounds (resulting in a higher sensitivity). This conservative approach minimizes the chance of missing a positive outcome; however, this approach also increases the number of false positive (resulting in lower specificity). For example, it was recommended that combining the results from two complementary (Q)sAR methodologies (one expert rule based and one statistical based) should be used as part of the iCH M7 guideline to assess pharmaceutical impurities.105,1°6 the performance statistics resulting from combining the two methodologies are illustrated here using the Leadscope expert rule-based and statistical-based models. Table
9.2 shows the results of running the two methodologies over the Hansen set107 after removing any compounds contained in the training set of the statistical-based model. The results of the individual models are shown in Table 9.2, along with the overall conclusion column, by combining the results from the two models. The sensitivity of the overall conclusion increased by 8% (compare with the expert rule-based methodology); however, the specificity decreased.
Figure 9.4 Results of a database search returning mutagenicity data for 4-pyridylamine.
Table 9.2 Combining multiple in silico methodologies to increase overall sensitivity.0
aQSAR: quantitative structure-activity relationship model.
High sensitivity is usually a desired objective when combining the results from more than one methodology and this can be achieved using the general rule that a positive from any of the methodologies results in an overall positive conclusion. A clear negative conclusion may be generated when all methodologies generate a negative response. However, it is possible that any prediction methodology may be unable to generate a classification call with adequate confidence, i.e. the prediction is inconclusive. In addition, one or more of the models may be unable to generate a prediction when the test compound is outside the applicability domain of the model. In these situations it may be necessary to analyze the results further to construct an expert review to reach an overall conclusion.
|< Prev||CONTENTS||Next >|