ADMET Predictor® User Manual
ADMET Predictor® User Manual
Breadcrumbs

ENDNOTES

ENDNOTES

  1. Throughout this document, a blue sans serif font is used to indicate widget labels and menu items. It is highlighted in bold when the interface text is in bold - e.g., for Dialog Titles. Hyperlinks are emphasized by underlining. Keyboard shortcuts and user-entered data are shown in a typewriter font. Model and Descriptor names are italicized. Most figures are only referenced by the text immediately above or below them and so are not numbered. Where figures are labeled, it is only for cross-reference within a section. Hence the indexing is local - i.e., Figure 1 rather than Figure 2.4.1.1.

  2. Many chemistry operations in ADMET Predictor will work on complex compounds like salts. The program will simply ignore ignore waters of hydration, counterions (e.g., Na+), and any other fragment containing only one heavy atom (e.g., HCl) when calculating ADMET properties, but other secondary components have to be removed. Salts dissociate in aqueous solution, so most of their ADMET properties are essentially additive. Moreover, the counterion in most pharmaceutical salts is selected from a group of innocuous "spectator" ions that are generally recognized as safe. Many of these have been preloaded in ADMET Predictor and can be removed by applying [CHEMISTRY >> Remove Common Fragments].

  3. If you suspect that both components in a salt are active, you can split its component molecules by selecting it and loading it into MedChem Designer (Shift+D). The compound will come up selected, and Structure >> Ungroup will break it into its constituent molecular fragments. Make sure all fragments of interest have a unique name, select them and return them to ADMET Predictor by clicking on one of the "back arrows" in the MedChem Designer toolbar. Each will now have its own row in the ADMET Predictor spreadsheet.

  4. ADMET Predictor does not calculate molecular descriptors or properties for charged forms of a molecule. Our ADMET property models are built on neutralized structures for consistency's sake, so all ionized forms get converted to their neutral forms before being sent to the model evaluation engine.

  5. Ordinary multiple linear regression (MLR) cannot be applied when the number of descriptors exceed the number of observations. Instead, TLA uses a pseudoinverse MLR method that can handle nearly singular matrices.

  6. Effective with the release of ADMET Predictor 8.5, the model building process is reseeded for each model architecture using a seed that is a function of the initial random seed, the number of inputs, and the number of neurons (or 1 for methods other than ANN). This makes the resulting models independent of the grid in which they are embedded, which was not the case for earlier versions of the ADMET Modeler module. In addition, model building in the earlier versions was perturbed by launching performance plots during the model building process; this is no longer the case.

  7. "Raw" descriptors are autoscaled before model building is begun. For artificial neural net (ANN) models, this entails subtracting off the minimum and dividing by the range, so values map to the [0,1] interval. For building support vector machine (SVM) models, descriptors are autoscaled to the interval [-1,1]. For kernel partial least squares (KPLS) and multiple linear regression (MLR), they are normalized by subtracting off the mean and dividing by the standard deviation.

  8. ADMET Risk scores might be not be integers if their assigned weights are non-integer in value or if soft thresholds have been called for, as they are in the Risk models that come with ADMET Predictor.

  9. Serum AST is also known as serum glutamic oxaloacetic transaminase (SGOT) and serum ALT is also known as serum glutamate-pyruvate transaminase (SGPT). They were referred to by the older names in previous versions of ADMET Predictor.

  10. Within the .ro5 file, “hard” thresholds are represented by a single number. Soft thresholds are represented by a pair of numbers, the second of which is set off by in curly brackets. The final criterion above, for example, is denoted by “N_Bonds > 35{41}”.

  11. One way to measure the balance between sensitivity and specificity is Youden’s index, which is obtained by adding the other two statistics and subtracting 1 (Part II, Section 3.4). Here, reducing the maximum allowed scores yields Youden index values of 0.689 and 0.425 for Absn_Risk and RuleOf5, respectively.

  12. Unless otherwise explicitly indicated, all models included in ADMET Predictor were developed by Simulations Plus, Inc.

  13. The positions of the menu bar and toolbar have been switched from their usual locations so that the pulldown menu doesn't cover up the toolbar icons. This is accomplished by clicking on the dappled bar at the right end of the toolbar and dragging it above the menu bar or vice versa. In fact, the toolbar and menu bar can be detached completely from the main window.

  14. Note that formaldehyde exists mostly as its acetal dihydroxymethane (HOCH2OH) in aqueous solution.

  15. To qualify as a descriptor for model building purposes, an attribute must be numeric; it must have no missing values within the row selection passed to Modeler; and it must take on more than one value within that row selection.

  16. In most lists in Predictor, Shift+click will extend your selection and Ctrl+click will add an item or deselect it if it is already highlighted.

  17. There are several situations in which verification set performance metrics are not split out from training set statistics, i.e., where a single consolidated statistic is reported for the training pool as a whole. This is done when the verification set's influence on the model is more direct than that involved in early stopping or - in the case of MLR - where no verification set is held out. Another case is SVM models, where the verification set is included when calculating model hyper-parameters for the model; there the performance grid has "-" as a placeholder for the verification set metric. A third case is the averaging method in ANNE classification models, where the verification set helps set the ensemble decision threshold. Finally, the training pool statistics are not split out when CFA is applied to the entire training pool to refine the voting threshold. In these cases, the two performance grid entries are set to the value for the training pool as a whole.

  18. Adding a descriptor to a model cannot expand an applicability domain (AD). When descriptors are selected by TLA, ITLA or GA per Grid, each row in the grid draws descriptors sequentially from the same list of descriptors, so all cells within a column share the same AD and each additional descriptor reduces it somewhat. When descriptor selection is by input gradient or GA per Row, the AD shrinks as one moves to the right within rows but not necessarily between rows - i.e., it is possible for a particular 6 input x 3 neuron model to have a larger AD than a 5 x 2 model. Finally, descriptors are not chosen sequentially by GA per Cell even within rows, so more complex models may have larger applicability domains. That said, more inputs do usually imply a more restricted AD.

  19. The discussion here is cast in terms of ANNEs because that is the model type where we have the most experience in the method's application. You can substitute "submodel" for "network" for other ensemble types of models (e.g., SVM and KPLS).

  20. LIBSVM was downloaded from https://www.csie.ntu.edu.tw/~cjlin/libsvm/ (© 2000-2014 C.-C. Chang and C.-J. Lin, all rights reserved). The terms of the license they provide requires inclusion of the following disclaimer in the program documentation:

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OLIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.