New Features/Changes Between
ADMET Predictor™ 4.0.0 and ADMET Predictor™ 5.0.0

First, please read and act on this important announcement:

CALCULATION OF SOME EXISTING DESCRIPTORS HAS BEEN CHANGED. IT IS STRONGLY RECOMMENDED THAT CUSTOM MODELS CREATED BY OUR USERS BE RETRAINED WITH THE NEW DESCRIPTORS! OTHERWISE, MODEL PARAMETERS WILL NOT BE COMPATIBLE WITH THE NEW DESCRIPTOR VALUES.

1. General Changes:

• Updated software licensing for better compatibility with recent versions of Microsoft Windows™.
• Improved software installation process.
• Usage logging is now turned on by default.
• Added support for the (extended) SMI input format.

2. Predictive Models:

• New toxicity models: TOX_SKIN - allergenic skin sensitization in mice, TOX_BCF - environmental bioconcentration factor, TOX_RAT - acute lethal toxicity in rat.
• New Physico-chemical and Biopharmaceutical models: S+FaSSGF - solubility in simulated gastric fluid, S+FaSSIF - solubility in simulated intestinal fluid in fasted state, S+FeSSIF - solubility in simulated intestinal fluid in fed state, S+Pcornea - permeability through rabbit cornea.
• Major overhaul of the pK_a prediction and presentation. Added ~1800 new compounds (~2400 new experimental pK_a values, bringing the modeling database size to 13881) resulting in the expanded applicability domain. Fixed a large number of wrong pK_a assignments. These enhancements resulted in our best pK_a model ever performing with the global RMSE below 0.6 log units on both training and external test sets, as well as on AstraZeneca dataset that was not included in modeling.
• Much improved Acid/Base labeling of the predicted pK_a by a new algorithm. The pK_a values are now reported in three columns S+Acidic_pKa, S+Mixed_pKa, S+Basic_pKa.
• Using our newly developed in-house tool for data curation, called Directory of Molecules, reprocessed some of the modeling data sets removing surviving duplicates, curating questionable structures, etc. Rebuilt all models for better predictive performance.
• Modified the HIA format: a) enabled column merging, b) added new RefAllPKa keyword for cases of multiple pK_a columns.

3. Descriptors:

• Upgraded the model of partial atomic charges and Fukui indices with hundreds of new training examples.
• Implemented different scaling of Fukui function descriptors for confirmed better performance in predictive models.
• Added boron to the list of "organic" elements, with proper parametrization, since boron-containing compounds gain wider recognition in pharmaceutical applications. New descriptors: N_Boron, SsssB, SaasB.
• Added a new amphoteric type of ionizable nitrogen: -NH- atoms that exist in a very special arrangement: Z(Y)A=B-NH- where A,B=C,N; Z,Y=-X(=O)- or -C#N; X=C,S,N, excluding O=X-OH
• Removed sulfonamide N from the list of potential basic centers.
• Certain rings containing a hypervalent S or P atom are no longer considered as "aromatic", in accordance with advanced ab initio results.
• Updated the default calculation of T_PSA descriptor by excluding S and P from the list of polar atoms
• Added recognition of carboniums and aromatic sulfoniums.
• Improved handling of unusual valence states of the nitrogen atom: completely changed the way tetravalent N and P are handled: hydrogens are never added to complete valence. Instead, some cases of N and P receive a +1 formal charge.
• The previous "Nitroso" descriptor is now reserved for the proper nitroso groups, while N-oxides are now under the new "N-oxide_>[N+][O-]" banner. Expanded the definition of ionizable N-oxides.
• Expanded the list of atomic descriptors.

4. Graphical User Interface:

• Much simplified File/Open menus. File types are now chosen in the common dialog box and automatically recognized from the filename extension.
• Added new "Plot pKa" button on the Prop./Desc. Correlations tab. It triggers the sophisticated automatic matching algorithm between observed pK_a and predicted pK_a.
• Added structure display activated by clicking points in the Prop./Desc. Correlations graphs.
• Added active update to pK_a microstate and structure displays triggered by row change in the Molecular Spreadsheet.
• Much improved display of ionization microstates. Added microconstants to the list of display options. Changed default minimal microstate contribution from 10% to 1%.
• Added support for a pH column (custom pH value per compound) in the Run Options.
• Added column title search to the Molecular Spreadsheet.

5. ADMET Modeler:

• Upgraded ANNE classification methodology: added advanced methods of handling unbalanced data sets and new statistics for evaluating predictive models (Specificity, Sensitivity, Youden index, Matthews Correlation Coefficient).
• Created a novel method of computing Descriptor Sensitivity Analysis for binary ANNE classification models.
• Added single threshold optimization method to ANNE classification models.
• Added new test set selection method in ADMET Modeler based on k-means algorithm.
• Extended the practical applicability of the Genetic Algorithm to large datasets by implementing k-means subsetting.
• Chemical structures are now depicted in the model graphing window.
• SALI data can now be sent to clipboard.
• Expanded model retraining options by enabling all test selection methods.
• Simplified options in the Advanced Settings tab.
• Exchanged blue/red colors of test and verification statistics on the Ensemble Statistics tab.
• Allowed the choice of Negative binary category in classification model building.

New Features/Changes Between
ADMET Predictor™ 3.0.0 and ADMET Predictor™ 4.0.0

First, please read and act on this important announcement:

CALCULATION OF SOME EXISTING DESCRIPTORS HAS BEEN CHANGED. IT IS STRONGLY RECOMMENDED THAT CUSTOM MODELS CREATED BY OUR USERS BE RETRAINED WITH THE NEW DESCRIPTORS! FORTUNATELY, THE NEW MODEL RETRAINING FEATURE MAKES IT MUCH EASIER THAN BEFORE :)

1. Added a new GI simulation model, SimDOSE, keyword "DOSE", for predicting the optimal drug dose in human matching the desired blood plasma concentration. The model works with experimental as well as in silico inputs.

2. Together with SimHIA models (simulated fraction absorbed), the SimDOSE models is a part of the new and optional Simulation Module. These models are no longer a component of the Physicochemical and Biopharmaceutical Module. This brings the total number of ADMET Predictor modules to five.

3. New, much more flexible format of the .hia files used by the Simulation Module.

4. Simulation Module now uses the new ODE integrator, CVODE, for compatibility with GastroPlus(TM).

5. All GI simulation models can now be run from the command line. The results are saved in a new output file .osm (Output of Simulation Module) in tab-delimited format. In Pipeline Pilot(TM) setting these results are passed in the data stream.

6. Tab-delimited files (.dat, .out, .txt) are now accepted as input in the command line mode. The new command line switch "-t DAT" indicates this input type.

7. Another useful command line switch, "-x", enables skipping predictions outside the individual models' applicability domains.

8. Added a unique model of predicting the pH-dependent distribution coefficient, log D, in the octanol-water solvent system. Previous versions of S+logD used constant correction factors for predicting partition of ionized species into n-octanol. Present version predicts these factors from molecular structure using Artificial Neural Network Ensembles greatly enhancing estimation accuracy.

9. New predictive models in the Toxicity Module: Human liver adverse effects of drugs: classification of molecules according to the likelihood of causing elevation in diagnostic liver enzymes: alkaline phosphatase, SGOT, SGPT, LDH, GGT - five models derived from the FDA database.

10. New predictive models in the Enslein Metabolism Module: classification of whether a molecule will be glucuronidated by one of the seven isoforms of the Uridine 5'-Diphosphate-Glucuronosyltransferase (UGT): 1A1, 1A3, 1A4, 1A6, 1A9, 1A10, and 2B7 (provided by Kurt Enslein)

11. New model of blood/plasma ratio belonging to the Physico-chemical and Biopharmaceutical Module; derived from published data.

12. New model of supersaturation ratio (ratio of kinetic to intrinsic solubility) in water belonging to the Physico-chemical and Biopharmaceutical Module; derived from published data.

13. Enabled multiple instances of ADMET Predictor running on the same network server if the number of network licenses allow it.

14. Upgraded the 4D Data Mining feature to the most recent version 7.2.6 of Miner3D™ software. Vastly improved handling of data columns by making full data matrix available to the Miner3D interface. Special characters in column headers are now supported.

15. The error log file, ADMET_Predictor_Errors.log, is now stored in a folder containing the user's input file instead of the ADMET Predictor installation folder.

16. Complete overhaul of subroutines handling the calculation of logD and water solubility profiles as a function of pH. Primarily, the definitions of the related base quantities, logP and intrinsic solubility, have been updated to cover all species in the zero charge state. Moreover, these routines consider the macroscopic, rather than microscopic, definition of acid and base dissociation. As a result, the calculations of logD and pH-dependent solubility variables has been simplified and improved. The effect is particularly visible in the values of intrinsic solubilities and solubility factors calculated for zwitterions.

17. The "GastroPlus(TM) Acid/Base Table" option in the "Save" menu now adds several extra columns (NumAcidGroups, FormalCharge, FractionZwitterionic, PAPN, PCPN, PZPN - the last three are structural parameters of the new logD model) to the saved pK_a table for GastroPlus(TM) and DDDPlus(TM) import. These parameters are used by the new logD and solubility routines.

18. New descriptors:

• Herndon = Herndon resonance energy (created to replace N_Kekule in modeling)
• SssssP = Phosphonium E-type
• T_RDmtr = Relative topological diameter of the molecule: maximal element of the distance matrix divided by the number of atoms
• T_PSA = Topological polar surface area, as defined by P. Ertl, et al.
• FUnion, FZwitter = fraction un-ionized and fraction zwitterionic. In principle, these are old descriptors but now calculated according to new definitions. FUnion now represents the pH-dependent fraction of _all_ species with the total electric charge of zero. FZwitter now represents the pH-independent fraction of FUnion species that exist in the zwitterionic state.
• Oliferenko's Hydrogen Bonding Acidity/Basicity atomic descriptors (structure display only)

19. Improvements in molecular structure processing and predictive modeling:

• Better detection of pi systems
• D and T in SMILES inputs are now treated as implicit hydrogens
• Recognition of aromatic sulfonium [S+] atoms
• Nonzero default values for EEM electronegativity and hardness for unknown E-types
• Added detection of aliphatic diazo groups
• Updated E-type assignment and charge state for coordinates oxygen [O-] in sulfates and phosphates
• Improved perception of charge state shifts by the presence of formal charge
• Updated parameters for EEM+Huckel model of partial atomic charges
• Retrained all predictive ADMET models

20. Absent pK_a values are now indicated by the "None" keyword instead of an empty spredsheet cell.

21. Added facilites for retraining previously built models in ADMET Modeler. This feature saves a lot of effort and time when good predictive models are obsoleted by new data or minor changes in molecular descriptors. Instead of doing full-fledged model building each time, dealing with descriptor selection and optimal model architecture, the existing models can be simply reoptimized with minimal effort.

22. Added an option for SALI (Structure-Activity Landscape Index) curves calculation and display after model building. The SALI and the related concept of SCI (SALI curve integral) have been created by R. Guha and J. H. Van Drie to address issues of capturing Maggiora's "activity cliffs" by predictive in silico models. Option available from the ADMET Modeler(TM) menu.

23. Improved ANNE classification algorithm for handling unbalanced binary data.

24. Extended Descriptor Sensitivity Analysis to SVM and KPLS models.

25. Added best fit line in the Graphic View of Model Performance.

26. Increased the maximum number of predictive models to 500.

27. Known bugs:

• Simulation models are not working (yet) with user pK_a values supplied in multiple columns

New Features/Changes Between
ADMET Predictor™ 2.4.0 and ADMET Predictor™ 3.0.0

First, please read and act on this important announcement:

CALCULATION OF SOME EXISTING DESCRIPTORS HAS BEEN CHANGED. IT IS STRONGLY RECOMMENDED THAT CUSTOM MODELS CREATED BY OUR USERS BE RETRAINED WITH THE NEW DESCRIPTORS! OTHERWISE, MODEL PARAMETERS WILL NOT BE COMPATIBLE WITH THE NEW DESCRIPTOR VALUES.

1. A completely new model of partial atomic charges has been implemented based on Natural Population Analysis of ab initio wavefunctions calculated at a high level of quantum theory. It differs from our previous S+CHARGE method in several respects. First, the database of 472 training molecules has been extended by 200 ionized (protonated/deprotonated) forms to better parameterize ionized atoms. Second, each partial atomic charge has been split into its sigma and pi components. Each of these components is calculated by a different method: sigma part by the Electronegativity Equalization Method while the pi part is computed by a modified Hückel treatment. This separation resulted in much better reproduction of pi electron densities as well as providing a multitude of useful descriptors.

2. Significant changes have been made in pi system detection: triple bonds are now counted as two perpendicular pi systems: pz and py. Both pz and py can conjugate with other systems and lone pairs in the respective z and y planes. These changes allow for the correct estimation of pi atomic charges and proper treatment of cumulenes.

3. The molecular descriptor pool has been increased by 55 new descriptors. The total number of descriptors is now 321 (2D) and 352 (2D+3D)

4. Coordination bonds in the following groups are now stored in charged, as opposed to double-bonded, representation (Nitro, NNitro, N-oxide, Nitrosamine, Azoxy, Azide, Nitrate, Oxadiazooxide, Nitrosohydroxyamine, aromatic Sulfoxide)

5. As a direct consequence of descriptor changes, ALL predictive models, including ADMET Risk, have been retrained resulting in similar or better performance compared to version 2.4.0. We have also observed the new models use in general less descriptors than the old ones. This is a very positive trend since it less descriptors in a model correlates with wider coverage of the chemical space by this model.

6. Simulated Fraction Absorbed now uses updated intestinal volumes in accord with the most recent release of GastroPlus(TM)

7. Parameters of the S+logD model have been updated resulting in better predictive performance.

8. The S+Vd model has been trained with a much larger database of experimental steady-state volumes of distribution in human.

9. Added 14 brand new models. This brings the total number of predictive models up to 77.

10. Added two new algorithms for descriptor selection in model building:

• ITLA = iterative truncated linear analysis based on Spearman ranks.
• GA-KPLS = a novel form of genetic algorithm with KPLS objective function.

11. In previous versions of ADMET Modeler, the train/verify split was 1:1 for non-Kohonen methods of test set selection. This has been changed to 2:1 in accord with the Kohonen method.

12. Multicorrelated descriptors can now be removed either randomly (default), or by the lowest TLA rank.

13. Removed file paths from Reload.txt for greater flexibility. The program assumes current path to contain all the files for model building reload.

14. Significant enhancements in 4D Data Mining tool:

• Added Miner3D(TM) graphics module for better, more advanced data visualization.
• Enabled multiple selections of the coloring "C" variable for use with Miner3D.
• Added new Chemically Relevant Principal Component "Atomic Reactivity".

15. Added the MRU (most recently used) files list to the "File" menu

16. Added new options to the pK_a microstate display: atom numbers and fully deprotonated structure.

17. Significantly enhanced Structure Visualization: 21 types of atomic properties (charges, polarizabilities, etc.) can now be displayed.

18. Enhanced pH-speciation plots: mouse-sensitive cursor enables reading the dominant fractions ionized as a function of pH. A full speciation list can be displayed on demand.

19. New right-click menu added to Molecular Record Spreadsheet with Copy, Paste, and Find options.

20. Outlier indicator color has been changed from red to magenta.

21. Enhancements in Ensemble Statistics display:

• Grid of model performances can now be saved as a data table or a BMP image, interactively or automatically.
• Single click on a model cell automatically opens performance graph for that model
• CTRL + click on a model cell automatically sets the model for export into ADMET Predictor

New Features/Changes Between
ADMET Predictor™ 2.3.0 and ADMET Predictor™ 2.4.0

1. Brand new set of 10 predictive models covering metabolism in human.

• Quantitative prediction of Km and Vmax (kinetic Michaelis-Menten constants) for hydroxylation reaction catalyzed by human CYP P450 enzymes: 1A2, 2C19, 2C9, 2D6, and 3A4.
• A result of collaboration with Enslein Research, Inc.

2. Program modularization. ADMET Predictor is now split into four independent modules:

• Physico-Chemical and Biopharmaceutical Module
• Toxicity Module
• Metabolism Module
• ADMET Modeler and Descriptors Module

New Features/Changes Between
ADMET Predictor™ 2.0.0 and ADMET Predictor™ 2.3.0

1. We have added a number of new descriptors and we have implemented S+CHARGE: an empirical method of estimating partial atomic charges trained against the results of Natural Population Analysis (NPA) of high-level ab initio wavefunctions. Consequently, many of our existing descriptors, based on partial charges, have changed. ADMET Predictor version 2.3.0 will not generate the same results for existing custom models as version 2.0.1.

2. We have added many new models for pK_a and toxicological properties prediction.

3. We have greatly enhanced the ability of the user to run ADMET Predictor from a command line and have now included options that mimic all of the run-time options found in the interactive version. We have now introduced a drag-and-drop component for Pipeline Pilot™. This module comes with a simple example for reading a molecule file, calculating descriptors and ADMET properties, and sending the result to an .html formatted browser page.

4. We have added many new user convenience items.

5. Finally, we have improved and enhanced the built-in ADMET Modeler™ module.

New Features/Changes Between
ADMET Predictor™ 1.3.2 and ADMET Predictor™ 2.0.0

1. ADMET Predictor and ADMET Modeler programs are now seamlessly integrated under a common graphical user interface into one application. Previously, predictive model building first required launching ADMET Predictor to process molecular structures to generate and save descriptors into a tab-delimited text file. Next, ADMET Modeler had to be launched in succession to read the descriptor file and perform model training. The final step required semi-manual model export back to ADMET Predictor. The new integrated package eliminates the file transfer steps allowing for model construction immediately after reading the structure file and includes automatic model export.

2. The File->Load command now accepts generic tab-delimited files to be loaded into the molecular spreadsheet.

3. Internal policies of some of our customers require keeping usage logs of purchased software. Since manual recordkeeping can be tedious, ADMET Predictor now offers an option for automatic and transparent usage logging to be recorded in a disk file.

4. Name change: the somewhat obscure "J-Alert" and "J-Code" terms have been renamed to "ADMET_Risk" and "ADMET_Code", respectively. Moreover, individual ADMET_Risk rules can now be weighted by fractional factors, if desired.

5. Automatic descriptor selection is now provided in both the "Column Selection" window (invoked from the "Save Selected Columns as ASCII Text" command under the File/Save menu) and the list of "Descriptors selected for model training" in the Basic Modeler Settings tab.

6. Molecular spreadsheet has a new capability for performing simple mathematical operations on numerical columns (arithmetic, log and antilog) through a new menu option "Arithmetic operations on columns" under the Recalculate pulldown.

7. ADMET Predictor recognizes two new model types: classification support vector machines (CSVM) and kernel partial least squares (KPLS).

8. All main tabs are scalable with the size of main window.

9. File editor now features Refresh function.

10. Correlation graph images and data can now be saved and copied to the Windows clipboard.

11. Program now issues a warning when the 3D Open command is attempted on a 2D MDL (SDF, RDF, or MOL) file. WARNING: this works only for _properly_ formatted MDL files in which the 2D/3D dimensionality indicator is present.

New Features/Changes Between
ADMET Predictor™ 1.2.3 and 1.3.2

1. The scope of all predictive models in ADMET Predictor is now extendable to user's chemical space without retraining, thus improving their predictability. This is achieved through creation and processing of Associative Neural Networks (ASNN) first proposed by Igor Tetko. ASNN come in two flavors: (a) "Fixed k", where the optimal number of nearest neighbors is constant, as originally formulated by Tetko, and (b) "Fixed r", where the number of nearest neighbors is variable and determined by an optimal radius in the chemical space. For testing purposes, the leave-one-out (LOO) mode of ASNN can be activated via the RunOptions form.

2. A new Model Editor available from the Options menu enables user-friendly handling of predictive ADMET models. Models can be modified, added, deleted, turned on and off, all through an easy graphical interface without burdensome editing of the ModelProperties.inp and TooltipsAndFormats.inp databases. In addition, Model Editor launches ASNN extension runs. Program upgrades gracefully handle custom models generated by the user: existing ModelProperties.inp and TooltipsAndFormats.inp are renamed to ModelProperties.bak and TooltipsAndFormats.bak. In the very first run of ADMET Predictor, Model Editor will merge user's models into the current version of ModelProperties.inp and TooltipsAndFormats.inp.

3. Simulated Fraction Absorbed now uses Human Fasted Physiological ACAT model, in tune with the latest release of GastroPlus™. Absorption scale factors (ASF) are calculated with the "Optimal logD Model". ASF calibration parameters: C1-C4 are now accessible to the user in the model's *.hia files.

4. Added support for V3000-formatted molecular structures in MDL-type inputs (SDF, RDF, MOL, and MAC).

5. Spreadsheet improvements: compound table is now initialized with two splits: the first holds the structure column, the second holds the remaining columns. Column settings and positions are now remembered in each split and restored with each input load and data sorting.

6. Warning messages are useful when a user gets familiar with the program, but after a while they may become an annoyance. The new option "Display Warning Messages" allows for turning these off.

7. New descriptor Isocynd_[N+]#[C-] = number of isocyanide groups.

8. Improved prediction of the topological state of nitrogen atoms in heteroaromatic rings.

New Features/Changes Between
ADMET Predictor™ 1.2.0 and 1.2.3

1. New predictive model, TOX_hERG, of cardiac toxicity stemming from the hERG potassium channel blockage.

2. Corrected handling of batch runs in the command line mode.

3. Corrected handling of depiction and descriptor calculations for certain heterocyclic ring systems (eg. benzimidazole).

---

---