OVERVIEW
ADMET Predictor® is an integrated suite of powerful cheminformatics tools for examining how compounds' molecular structures relate to their absorption, distribution, metabolism, excretion and toxicology, i.e., to their ADMET properties. Users can apply the advanced model-building tools in its ADMET Modeler™ Module to find quantitative structure/activity relationships (QSARs) in their own data and to assess the predictive performance of the resulting models. The MedChem Studio™ module contains complementary tools for curating data sets as well as exploring qualitative SARs and the effects of incremental structural changes on activity and ADMET properties.
Models can only be as good as the data used to build them. Because the quantity and quality of the data used both matter, Simulations Plus, Inc., has invested a tremendous amount of time and energy over the last seventeen years in collecting and thoroughly curating data relating to a wide range of biologically relevant endpoints (Waldman 2014). Predictive ADMET models that are the fruits of that effort are available to users through the program's Physicochemical and Biopharmaceutical (PhysChem), Transporters, Metabolism, and Toxicity modules. The predicted properties currently provided include pKa, logP, logD at a specified pH, and aqueous solubility in pure water or at a specified pH; kinetic parameters for the major cytochrome P450s (CYPs); mutagenicity, hepatotoxicity, and several measures of ecotoxicity; and much more. The program also integrates substrate/non-substrate, site of metabolism and kinetic predictions to generate maps of likely metabolites via the MedChem Designer™ auxiliary program.
History
ADMET Predictor (or, more concisely "Predictor") began life in 1999 as QMPRPlus®. The state-of-the-art pKa model destined to evolve into one of its flagship features (Fraczkiewicz 2015) was added in 2003 and the program was renamed in 2005. Version 2.0 (2006) featured the integration of ADMET Modeler™, which had been a separate product up until then.
Predictor's current GUI is based on the ChemTK code base, a technology that Simulations Plus acquired from Sage Informatics LLC in 2005. That code base provided the framework for ClassPharmer™, the rights to which Simulations Plus acquired from Bioreason, Inc. at the same time. That program was subsequently modified to allow access to ADMET Predictor's property predictions in 2008, and was renamed MedChem Studio in 2010 to better reflect the structural elaboration and analysis tools added since its acquisition. The program's companion sketching tool - MedChem Designer™ - was linked into Predictor in 2012. Together with this version's migration of ADMET Predictor into the ChemTK GUI framework, the integration with MedChem Studio represents the next logical step in the evolutionary convergence of our cheminformatics offerings.
Using ADMET Predictor
During virtual library screening in early drug discovery, ADMET Predictor can be used to estimate crucial physicochemical and biological properties of large numbers of candidate drug compounds. Such predictions are important for synthesized compounds as well, because many available ADMET assays are prone to error or are themselves surrogates for the endpoint of real interest, i.e., the assays are themselves models. Using MDCK monolayer permeability to estimate passive jejunal permeability (Peff) is an example of such a surrogate endpoint. ADMET Predictor is especially useful as a way to provide "reality checks" and complementary second opinions for high throughput assays.
In silico predictions have the added virtue of not consuming material, which is often a limiting resource early in development. This makes it useful for prioritizing resynthesis of active compounds identified in screens. Of course, property estimates can also be made for compounds that have yet to be synthesized, including analogs and novel structures generated using the tools in the MedChem Studio module. Applying property and toxicity predictions in this way reduces the risk of making and testing compounds that are likely to fail later in the development process.
Many of the models in ADMET Predictor estimate properties (e.g., mutagenicity, low solubility, or high lipophilicity) that can be serious problems for drug candidates. Applying them individually is generally not a good idea, so the program provides default ADMET Risk™ scores that aggregate multiple predictions for easy sorting and examination. This makes it possible to rank-order candidate compounds by how many development hurdles each is likely to face. An accompanying Risk Code indicates what the likely risks are. Ease of use for this capability has been greatly enhanced by the addition of the MedChem Studio capability to ADMET Predictor.
Licensing
ADMET Predictor uses Reprise® licensing to provide access to its various models and functions. The installation package will install the requisite license server and associated software if it is not already in place.
All but a few ADMET property predictions are licensed to the Physicochemical and Biopharmaceutical, Transporters, Metabolism or Toxicity module, and calculating their values will check out one or more of the corresponding licenses, as will calculating descriptor values. Similarly, advanced chemistry operations such as generating classes, R-tables or molecular pairs checks out a MedChem Studio license, as does using our molecular design tools. An ADMET Modeler license is checked out when that program is launched.
Licenses are not checked out until the feature that it enables is used. Once checked out, however, a license will remain checked out for the duration of the ADMET Predictor session or you switch to Viewer Mode. All licensed features are disabled in Viewer Mode, but you will still be able to open files and import attributes as well as to carry out basic functions like scrolling through the spreadsheet and searching it. You will not, however, be able to do anything that would check out a license until you have toggled back to Full Mode.
The program will come up in Viewer Mode automatically when no feature licenses are available.
Predictive Accuracy
No computer program will exactly reproduce experimental property values from 2D or 3D structures. Indeed, reproducing such values exactly is a hallmark of "look up" programs that perform well on compounds taken from the literature but tend to work much more poorly when applied to novel chemistry. Detailed statistics for performance of each model on "held out" test sets are provided elsewhere in this Manual, but typical predictive root mean square errors (RMSEs) for quantitative estimates are ±0.3-0.5 log units, i.e., within 2- to 3-fold. Classification accuracy for qualitative models typically is 85-90% or better.
The English Cardinal John Henry Newman (1801-1890) once said, "Nothing would be done at all, if a man waited till he could do it so well, that no one could find fault with it.” In most cases ADMET Predictor's estimates are more than sufficiently accurate to rank order large numbers of molecular structures by overall quality, with those compounds in the top 30% being consistently better than those in the bottom 30%. Identifying likely problems as early as possible will reduce the time and effort wasted on them, an approach sometimes referred to as “failing fast.”
This approach is especially powerful when users augment the "canned" models provided by adding their own predictive models, constructed from in-house data, using the ADMET Modeler module.
How This Manual Is Organized
The remainder of this Manual is organized into five parts. Part 1 describes the basics of running ADMET Predictor, e.g., how to load data into the program and control its displays. It also describes how to modify input structures, create new attributes from existing ones, and use the basic data curation tools. Blue sans serif fonts are used for dialogs, dialog items, menu items, etc. and boldface numbers (e.g., (1)) refer to figure annotations. Model and descriptor names are italicized. Other formatting conventions are laid out here.
Note that some of the images and discussions in this section may show results corresponding to earlier versions of our current models. These images and the discussions around them are meant to illustrate the overall usage and functionality of the program rather than to represent specific results obtained with the current version. However, every effort has been made to update performance plots and tables to represent current model performance as well as the results obtained running all the tutorials.
Part 2 describes how to generate ADMET property predictions and discusses the nature of the endpoints addressed by the various models, which are consolidated into chapters based on the attribute group to which they belong - PhysChem, Transporters, Metabolism or Toxicity. Specific details of how each model was built and how well it performs statistically are laid out.
Models for some endpoints (e.g., mutagenicity and CYP metabolism) group together naturally. Documentation for such groups has been consolidated to minimize repetition and facilitate access. Sections devoted to each group begin with a general overview and a summary of menu and toolbar options that can be used to access the corresponding predictions. This is followed by a discussion of how the models covered in the section relate to each other and tables that lay out the respective performance statistics. Topics close with detailed specifications such as where the data used to build the model came from. Theory and discussion of technical details are generally deferred to the end of each chapter or section.
Part 3 describes how ADMET Modeler is used. The mechanics of building your own models is described, as well as how to assess the likely reliability of the predictions those models generate. That information is important for putting the property predictions available through the Physicochemical and Biopharmaceutical, Transporters, Metabolism and Toxicity modules into an appropriate theoretical context for understanding the scope of their applicability. This makes the material relevant even if you do not have occasion to use the tools described therein to build your own models.
Part 4 describes how the analytical tools available in MedChem Studio can be accessed from the ADMET Predictor GUI. These include the particular functionalities that require a MedChem Studio license, including structural including structural class generation, R-group tables, and molecular pair analysis. Chemistry tools such as substructure searching that are available to all Predictor licensees are described in the general discussion of the new interface.
Part 5 describes the Design Module introduced in ADMET Predictor 10.4. The tools in this Module include virtual combinatorial chemistry, R-group explosion, and Artificial Intelligence driven Drug Design. AIDD generates a series of molecules by iterative application of chemical SMIRKS, keeping those that are the best by multiple Pareto optimization criteria that you specify. You specify a few seed structures, a target scaffold and selection criteria such as activity, ADMET Risk and synthetic accessibility. The program then returns a library of candidate molecules, each of which is superior to the others by at least one of your chosen Pareto criteria and many of which represent an appealing balance among them.
Part 6 describes the details of the HTPK Simulation module, which predicts the fraction absorbed (%Fa), oral bioavailability (%Fb) and other PK parameters (Cmax, Tmax and AUC) using the basic ACAT™ intestinal absorption simulation scheme used in GastroPlus® linked to a single central compartment augmented with liver and kidney clearances.
It also describes how to assess the sensitivity of %Fa, %Fb and other PK parameters to changes in compound solubility and permeability. Similar to parameter sensitivity analysis (PSA) in GastroPlus, this feature allows you to assess how changes in estimates of these physicochemical properties affect pharmacokinetic predictions.
Part 7 provides tutorials and illustrative applications showing how ADMET Predictor can be used. It is followed by appended notes, file format specifications, a glossary and a bibliography.
Many illustrative examples are based on the Demo2D data set, which is installed in your application directory:
C:\Users\\<user_name>\AppData\Local\Simulations Plus, Inc\ADMET_Predictor\<X.Y>
where \<user_name> has been assigned to you by your systems administrator and \<X.Y> is the current ADMET Predictor version number. The AppData folder may be hidden on your system. If it is, you can still access it by copying and pasting the path into a Windows Explorer navigation window.