4/30/2021 Theory - ADMET Descriptors

Small Molecules tools > Theory - Small Molecules > Theory - ADMET Descriptors

Theory - ADMET Descriptors

ADMET - Human Intes nal Absorp on

This model predicts human intes nal absorp on (HIA) a er oral administra on. Intes nal absorp on is defined as a
percentage absorbed rather than as a ra o of concentra ons (cf. blood-brain penetra on). A well-absorbed compound is
one that is absorbed at least 90% into the bloodstream in humans. The intes nal absorp on model includes 95% and
99% confidence ellipses in the ADMET_PSA_2D, ADMET_AlogP98 plane.
The ellipses define regions where well-absorbed compounds are expected to be found: 95% of well-absorbed
compounds are expected to fall within the 95% ellipse, while 99% of well-absorbed compounds should fall within the
99% ellipse. Note that the loca on of any par cular compound does not necessarily imply whether it will be well,
moderately, or poorly absorbed. In general, however, absorp on tends to drop off quite rapidly outside the 95% ellipse.
There are four predic on levels:
0 (Good)
1 (Moderate)
2 (Poor)
3 (Very Poor)
These levels are defined by the 95% (blue line) and 99% (magenta line) confidence ellipsoids in Figure 1. The upper limit
of PSA_2D value for the 95% confidence ellipsoid is at 131.62, while the upper limit of PSA_2D value for the 99%
confidence ellipsoid is at 148.12.

Development of model
The model was developed using 182 compounds in the training set, with descriptors that include AlogP98 and 2D polar
surface area (PSA_2D) [Egan et al., 2000] and [Egan and Lauri, 2002]. Well absorbed compounds have at least 90%
absorp on into human bloodstream, and by this model will generally reside within the ellipse regions of 95% and 99%
confidence levels.
Figure 1: The Development of Human Intes nal Absorp on Model

file:///D:/Program Files/BIOVIA/Discovery Studio 2016/share/doc/DS/DS.htm#DocDS/TSSmallMolecules/admet_the.htm?Highlight=admet 1/10

4/30/2021 Theory - ADMET Descriptors

The human absorp on model was validated against several datasets. Results showed that for drug-like compounds
within the 99% absorp on ellipsoid, the following ra os were predicted correctly:
Physicians Desk Reference (PDR): 87.4% of 438 orally delivered compounds
World Drug Index: 82.9% of 8,504 with USAN or INN
Comprehensive Medicinal Chemistry: 83.5% of 5,836 filtered by class
Pharmacopeia's Caco-2 data (446 compounds with low, moderate, and high permeability (i.e., Papp), used in many
PCOP Labs collabora ve projects):
Moderate/High Papp: 91.5% of compounds lie within 99% ellipse
Low Papp: 20.6% of compounds lie within 95% ellipse

Computed proper es
Property Descrip on
ADMET_PSA_2D Fast polar surface area
ADMET_AlogP98 ALogP
ADMET_Unknown_AlogP98 The count of unknown AlogP atoms found in the molecule
ADMET_Absorp on_Level Categorical absorp on level

Note: In calcula ng ADMET_AlogP98, atoms with unknown AlogP98 types do not contribute to the AlogP98
calcula on.

Key to human intes nal absorp on (ADMET_Absorp on_Level)

Level Value Descrip on
0 ADMET_Absorp on_T2_2D < 6.1261 (inside 95%) Good absorp on
1 6.1261 ≤ ADMET_Absorp on_T2_2D < 9.6026 (inside 99%) Moderate absorp on
2 9.6026 < ADMET_Absorp on_T2_2D (outside 99%) Low absorp on
file:///D:/Program Files/BIOVIA/Discovery Studio 2016/share/doc/DS/DS.htm#DocDS/TSSmallMolecules/admet_the.htm?Highlight=admet 2/10
4/30/2021 Theory - ADMET Descriptors

3 ADMET_PSA_2D ≥ 150.0 or ADMET_AlogP98 ≤ -2.0 or ADMET_AlogP98 ≥ 7.0 Very low absorp on

Note: ADMET_Absorp on_T2_2D is the Mahalanobis distance for the compound in the ADMET_PSA_2D,
ADMET_AlogP98 plane. It is referenced from the center of the region of chemical space defined by well absorbed
compounds

ADMET - Aqueous Solubility

This model uses linear regression to predict the solubility of each compound in water at 25 oC.

Development of model
Cheng and Merz reported a predic ve model for aqueous solubility that was generated using a dataset containing 775
compounds (molecular weight between 50 and 800) [Cheng and Merz, 2003]. The training set compounds cover
numerous classes, including alkanes, alkenes, alkynes, halogens, amines, alcohols, N-containing compounds, ketones,
aldehydes, organic acids, etc. A plot of the predicted versus experimental data gives a linear regression of LogSw (25 oC,
pH=7.0) R2 = 0.84, and standard devia on = 0.87. The test set consisted of 34 compounds, and the regression sta s cs
are R2 = 0.88 and standard devia on = 0.79 (see Figure 2). A valida on test consis ng of 1,615 compounds from the PDR,
Comprehensive Medicinal Chemistry database (CMC), and other sources was also performed. Results yield an overall
RMSE (SD) of 1.0.
Figure 2 shows the regression plots for the training set and test set applicable to the deriva on and tes ng of this model
Figure 2: Development of Aqueous Solubility Model

Computed proper es
Property Descrip on
ADMET_Solubility The base 10 logarithm of the molar solubility as predicted by the regression
ADMET_Solubility_Level Categorical solubility level

Key to aqueous solubility levels (ADMET_Solubility_Level)

Level Value Drug-likeness
0 log(Sw) < -8.0 Extremely low
1 -8.0 < log(Sw) < -6.0 No, very low, but possible
2 -6.0 < log(Sw) < -4.1 Yes, low

file:///D:/Program Files/BIOVIA/Discovery Studio 2016/share/doc/DS/DS.htm#DocDS/TSSmallMolecules/admet_the.htm?Highlight=admet 3/10

4/30/2021 Theory - ADMET Descriptors

3 -4.1 < log(Sw) < -2.0 Yes, good

4 -2.0 < log(Sw) 0.0="" Yes, op mal
5 0.0 < log(Sw) No, too soluble
6 -1000 Warning: molecules with one or more unknown AlogP98 types

Note: In calcula ng aqueous solubility, atoms with unknown AlogP98 types do not contribute to the AlogP98
calcula on.

ADMET - Blood Brain Barrier

This model predicts blood-brain penetra on (blood brain barrier, BBB) a er oral administra on. This model contains a
quan ta ve linear regression model for the predic on of blood-brain penetra on, as well as 95% and 99% confidence
ellipses in the ADMET_PSA_2D, ADMET_AlogP98 plane. These ellipses are not the same as those associated with the
ADMET - Human intes nal absorp on, although they have an analogous interpreta on. They were derived from over 800
compounds that are known to enter the CNS a er oral administra on [Egan and Lauri, 2002].
There are four predic on levels within the 95% and 99% confidence ellipsoids:
0 (Very high penetrant)
1 (High)
2 (Medium)
3 (Low)
4 (Undefined)
The plot depicts the five levels with logBB (ADMET_BBB) values as follows:
Very high penetrants (logBB ≥ 0.7);
High penetrants (0 ≤ logBB < 0.7),
Medium penetrants (-0.52 < logBB < 0),
Low penetrants (logBB ≤ -0.52).
No predic on is made for compounds outside the 95% and 99% confidence ellipsoids (undefined level = 4).

Development of model
Scien sts at BIOVIA recently developed two robust models for the predic on of BBB penetra on. A regression model to
predict logBB values was derived from a training set of 102 compounds (R2 = 0.7329, RMSE = 0.3638, N = 102), and a test
set of 86 compounds (R2 = 0.8892, RMSE = 0.3064, N = 86). The model was validated against 881 compounds designated
as CNS compounds in the CMC database (Figure 3). Further tes ng against a collec on of 124 compounds with known
logBB values yielded an R2 = 0.889 and SD = 0.306.
Figure 3: Development of Blood-Brain Barrier Model

file:///D:/Program Files/BIOVIA/Discovery Studio 2016/share/doc/DS/DS.htm#DocDS/TSSmallMolecules/admet_the.htm?Highlight=admet 4/10

4/30/2021 Theory - ADMET Descriptors

The plot highlights four BBB penetra on levels:

Very high penetrants (blue; logBB ≥ 0.7);
High penetrants (green; 0 ≤ logBB < 0.7),
Medium penetrants (cyan; -0.52 < logBB < 0),
Low penetrants (orange; logBB ≤ -0.52).
No predic on is made for compounds outside the 95% and 99% confidence ellipsoids (gray triangles).
The regression is based on ADMET_PSA_2D and ADMET_AlogP98, and the training set lies en rely within the 99%
confidence ellipse. This implies that predic ons of ADMET_BBB are less reliable for compounds that lie outside the 99%
ellipse.

Computed proper es
Property Descrip on
ADMET_PSA_2D Fast polar surface area
ADMET_AlogP98 Atom-based LogP
ADMET_Unknown_AlogP98 The count of unknown AlogP atoms found in the molecule
ADMET_BBB Base 10 logarithm of (brain concentra on)/(blood concentra on) as predicted by a robust
(least-median-of-squares) regression derived from literature in vivo brain penetra on
data
ADMET_BBB_Level Categorical level

Key to BBB penetra on levels (ADMET_BBB)

Level Value Descrip on
0 Very High Brain-Blood ra o greater than 5:1
1 High Brain-Blood ra o between 1:1 and 5:1

file:///D:/Program Files/BIOVIA/Discovery Studio 2016/share/doc/DS/DS.htm#DocDS/TSSmallMolecules/admet_the.htm?Highlight=admet 5/10

4/30/2021 Theory - ADMET Descriptors

2 Medium Brain-Blood ra o between 0.3:1 and 1:1

3 Low Brain-Blood ra o less than 0.3:1
4 Undefined Outside 99% confidence ellipse
5 AlogP98 Warning: molecules with one or more unknown AlogP98 types

Note: Atoms with unknown AlogP98 types do not contribute to the AlogP98 calcula on.

ADMET - Cytochrome P450 2D6 (CYP2D6)

The cytochrome P450 2D6 (CYP2D6) model predicts CYP2D6 enzyme inhibi on using 2D chemical structure as input.
CYP2D6 is involved in the metabolism of a wide range of substrates in the liver and its inhibi on by a drug cons tutes a
majority cases of drug-drug interac on. Therefore, CYP2D6 inhibi on experiment is required as part of the regulatory
procedures in the drug discovery and development process.

Development of model
A computa onal model for compounds inhibi ng the CYP2D6 enzyme was developed from a training set of 151
structurally diverse compounds with known CYP2D6 inhibi on constants [Susnow and Dixon, 2003]. However, the
modeling methodology described in that paper was not used; instead, modified Bayesian learning [Xia et al., 2004] was
performed. A leave-one-out cross valida on was conducted and the following table lists the performance of the model:
Model Name ROC ROC True False False True Sensi vity Specificity Concordance
Score Ra ng Posi ve Nega ve Posi ve Nega ve Rate
ADMET_EXT_CYP2D6 0.877 Good 56 13 12 70 0.812 0.854 0.834

Note: For defini on of these terms, see Receiver Opera ng Characteris c (ROC) Plot Analysis.

Figure 4: ROC Plot of model ADMET_EXT_CYP2D6 from Leave-One-Out Cross Valida on

file:///D:/Program Files/BIOVIA/Discovery Studio 2016/share/doc/DS/DS.htm#DocDS/TSSmallMolecules/admet_the.htm?Highlight=admet 6/10

4/30/2021 Theory - ADMET Descriptors

Computed proper es
Property Descrip on
ADMET_EXT_CYP2D6 Bayesian score from the model
ADMET_EXT_CYP2D6#Predic on The classifica on whether a compound is an CYP2D6 inhibitor using the cutoff
Bayesian score of 0.161 (obtained by minimizing the total number of false
posi ves and false nega ves)
ADMET_EXT_CYP2D6_Applicability The applicability of the model on the predicted compound
ADMET_EXT_CYP2D6#MD The Mahalanobis distance (MD) to the center of the training data. MD is a
generaliza on of the Euclidean distance that accounts for correla on among the
proper es. The larger the MD, the less reliable the predic on
ADMET_EXT_CYP2D6#MDpvalue The p-value associated with the MD. The smaller the p-value, the less reliable the
predic on

ADMET - Hepatotoxicity
The hepatotoxicity model predicts poten al organ toxicity for a wide range of structurally diverse compounds.

Development of model
The model was developed from available literature data of 436 compounds known to exhibit liver toxicity (i.e., posi ve
dose-dependent hepatocellular, cholesta c, neoplas c, etc.), or trigger dose-related elevated aminotransferase levels in
more than 10% of the human popula on [Cheng and Dixon, 2003]. From this, a model was generated from a
classifica on structure ac vity rela onship (SAR) technique that uses a modified Bayesian learning [Xia et al., 2004] The
model was validated using leave-one-out cross valida on. The following table lists the performance of the model:

file:///D:/Program Files/BIOVIA/Discovery Studio 2016/share/doc/DS/DS.htm#DocDS/TSSmallMolecules/admet_the.htm?Highlight=admet 7/10

4/30/2021 Theory - ADMET Descriptors

Model Name ROC ROC True False False True Sensi vity Specificity Concordance
Score Ra ng Posi ve Nega ve Posi ve Nega ve Rate
ADMET_EXT_Hepatotoxic 0.888 Good 157 15 66 198 0.913 0.750 0.814

Note: For defini on of these terms, see Receiver Opera ng Characteris c (ROC) Plot Analysis.

Figure 5: ROC Plot of model ADMET_EXT_Hepatotoxic from Leave-One-Out Cross Valida on

Computed proper es
Property Descrip on
ADMET_EXT_Hepatotoxic Bayesian score from the model
ADMET_EXT_Hepatotoxic#Predic on The classifica on whether a compound is hepatotoxic using the cutoff
Bayesian score of -4.154 (obtained by minimizing the total number of false
posi ves and false nega ves)
ADMET_EXT_Hepatotoxic_Applicability The applicability of the model on the predicted compound
ADMET_EXT_Hepatotoxic#MD The Mahalanobis distance (MD) to the center of the training data. MD is a
generaliza on of the Euclidean distance that accounts for correla on among
the proper es. The larger the MD, the less reliable the predic on
ADMET_EXT_Hepatotoxic#MDpvalue The p-value associated with the MD. The smaller the p-value, the less reliable
the predic on

ADMET - Plasma Protein Binding

The plasma protein binding model predicts whether a compound is likely to be highly bound (>= 90% bound) to carrier
proteins in the blood. Plasma protein binding of drug molecules can affect the efficiency of a drug, because the bound
file:///D:/Program Files/BIOVIA/Discovery Studio 2016/share/doc/DS/DS.htm#DocDS/TSSmallMolecules/admet_the.htm?Highlight=admet 8/10
4/30/2021 Theory - ADMET Descriptors

frac on is temporarily shielded from metabolism. On the other hand, only the unbound frac on exhibits
pharmacological effects.

Development of model
Two data sets with plasma protein binding level data [Dixon and Merz, 2001][Votano et al., 2006]were combined and
duplicates were removed to generate a training set of 854 compounds. Using a cutoff binding level of 90%, the training
set was divided into 329 binders and 525 non-binders. A modified Bayesian learning [Xia et al., 2004] was used to create
a binary classifica on model. The model was validated using leave-one-out cross valida on. The following table lists the
performance of the model:

Model Name ROC ROC True False False True Sensi vity Specificity Concordance
Score Ra ng Posi ve Nega ve Posi ve Nega ve Rate
ADMET_EXT_PPB 0.873 Good 285 44 121 404 0.866 0.770 0.807

Note: For defini on of these terms, see Receiver Opera ng Characteris c (ROC) Plot Analysis.

Figure 6: ROC Plot of model ADMET_EXT_PPB from Leave-One-Out Cross Valida on

Computed proper es
Property Descrip on
ADMET_EXT_PPB Bayesian score from the model
ADMET_EXT_PPB#Predic on The classifica on whether a compound is highly bounded (>= 90% bound) to plasma
proteins using the cutoff Bayesian score of -2.209 (obtained by minimizing the total
number of false posi ves and false nega ves)
ADMET_EXT_PPB_Applicability The applicability of the model on the predicted compound

file:///D:/Program Files/BIOVIA/Discovery Studio 2016/share/doc/DS/DS.htm#DocDS/TSSmallMolecules/admet_the.htm?Highlight=admet 9/10

4/30/2021 Theory - ADMET Descriptors

ADMET_EXT_PPB#MD The Mahalanobis distance (MD) to the center of the training data. MD is a
generaliza on of the Euclidean distance that accounts for correla on among the
proper es. The larger the MD, the less reliable the predic on
ADMET_EXT_PPB#MDpvalue The p-value associated with the MD. The smaller the p-value, the less reliable the
predic on

Further informa on
Receiver Opera ng Characteris c (ROC) Plot Analysis
Toxicity Predic on (Extensible) - Theory
Toxicity Predic on (TOPKAT) - Theory

Legal No ces BIOVIA Discovery Studio 2016 Help: Tuesday, December 01, 2015

file:///D:/Program Files/BIOVIA/Discovery Studio 2016/share/doc/DS/DS.htm#DocDS/TSSmallMolecules/admet_the.htm?Highlight=admet 10/10