Additional File 2 — Use Case: Investigate Input Features for Carcinogenicity Models

Additional file for the article " CheS-Mapper 2.0 for Visual Validation of (Q)SAR models " We visually validate the effect of exchanging the descriptors used by a (Q)SAR algorithm. To this end, we select a subset of the Carcinogenic Potency Database (CPDB) [1] for various species. The database contains 86 compounds that have an activity value for hamster carcinogenicity assigned (active or inactive). We compute two different sets of features for these compounds with CheS-Mapper: 308 physico-chemical (PC) descriptors using CDK and Open Babel, and 287 structural fragments. The structural features have been calculated by matching the compounds with three predefined SMARTS lists included in Open Babel. We choose the random forest implementation from the WEKA workbench as classification algorithm and compare two different approaches: (Q)SAR-1 is built using only the physico-chemical descriptors, while (Q)SAR-2 exploits a combination of both feature sets. We apply a 5-times repeated 10-fold cross-validation to validate both variants. (Q)SAR-2 achieved a classification accuracy of 0.75, and significantly outperformed (Q)SAR-1 that had a classification accuracy of only 0.67. Apparently, using both feature types allows to build a more predictive model. As CheS-Mapper's 3D embedding is based on the features, we start the program twice to (simultaneously) compare the effect of using different feature sets. When highlighting the actual endpoint value, we note that the compounds are roughly separated according to their class value. The separation (and thus the decision boundary) is less distinctive when using only PC features (Figure 1) compared to adding structural fragments (Figure 2). This indicates that it is easier for (Q)SAR-2 to predict the endpoint, than for the (Q)SAR-1 approach. Comparing the misclas-sifications of both approaches, we detect two compounds that have always been correctly classified by (Q)SAR-2, but not by (Q)SAR-1. The inactive compound Isonicotinic acid (DSSTox-RID 20757) is selected in Figure 2 (marked with a label and drawn as 2D picture at the top right-hand side). In the embedding based on both feature types it is located in entirely inactive space. It is correctly classified by (Q)SAR-2 in 5 of 5 repetitions of the cross-validation. In contrast, this compound was misclassified as active 2 out of 5 times by (Q)SAR-1. As previously described, the feature list at the top right-hand side is sorted according to specificity. Hence, carboxylic acid is the structural feature that distinguishes this compound the most from the remaining dataset compounds. With the help of CheS-Mapper, we …