Overview
Druglikeness anlysis
ADME/T evaluation
Similarity searching based on ADME/T database
Systematic ADME/T assessment
Overview
| What's ADMETlab

ADMETlab is a platform for systematic ADME/T evaluation of drug molecules. It provides a rapid way to predict pharmacokinetic properties by using 24 well-optimized SAR models. 5 druglikeness rules and 1 specific model are provided for users to analyze druglikeness. Further, it includes a well-organized ADME/T database (208,967 entries) that can be used to search similar compounds and make a comparison conveniently. It also can be used to make a comprehensive evaluation to obtain meaningful and constructive suggestions of molecular optimization by performing the systematic evaluation. In addition, ADMETlab created a user-friendly interface to accomplish these jobs just by clicking mouse.


| Main features
  • Compared large datasets of most properties.
  • Better and robust SAR models.
  • Focus on drug molecules.
  • Systematic analysis and comparison
  • Provide constructive suggestions for molecular optimization
  • Batch computation
  • User-friendly interface
| Main functionalities
ADMETlab can help researcher, PK specialists to do as follows:

Druglikeness analysis.

ADME/T evaluation.

Similarity searching based on ADME/T database.

Systematic ADME/T assessment.

Druglikeness analysis
| Druglikeness rules

Druglikeness rules are several expert criterions that are used in drug design for how "druglike" a substance is with respect to factors like bioavailability. Here, we provide 5 selected rules:

  • Lipinski's rules: MW<=500; logP<=5; Hacc<=10; Hdon<=5
  • Ghose's rules: - 5.6< MclogP < -0.4 mean: 2.52; 160 < MW < 480 mean: 357;
    40 < MR < 130 mean: 97; 20 < natoms < 70 mean: 48
  • Oprea's rules: nrings≥3; nrigidbond>=18; nRotbond≥6
  • Veber's rules: nRotbond<=10; tPSA<= 140 or Hacc and Hdon<=12
  • Varma's rules: MW<= 500; tPSA<=125; -5< logD < – 2; Hacc+Hdon<=9; nRotbond<=12

| Druglikeness model

We collected 6731 drugs from Drugbank database as samples of druglikeness. Then 6769 molecules was picked as negative samples from those molecules with IC50 or Ki less that 10000nm from CHEMBL database by using (Self-organizing feature Map,SOM) method. The SOM method ensures that the samples are picked from the not similar clusters compared with positive samples. Finally, we get a classification model with a accuracy of 0.801 for training set by 5-fold cross validation. This model can help us further distinguish between molecules with activities and potential drugs.

ADMET prediction
| Data summary

For all the ADMET-related properties, we collected corresponding data mainly by two approaches: the previous literatures and the DrugBank database (http://www.drugbank.ca). After several pretreatments, we totally obtained 24 datasets. The global overview of these ADMET datasets can be seen in the table below and you can click the "" to get the data files. Table S1.

Table S1. The number of end-points of each property

CategoryPropertyTotalPositiveNegativeTrainTest
Basic physicochemical propertyLogS2055--1541514
LogD7.41031--773258
LogP     
AbsorptionCaco-21182--886296
Pgp-Inhibitor229713729251723574
Pgp-Substrate1252643609939313
HIA57850078434144
F (20%)1013759254760253
F (30%)1013672341760253
DistributionPPB1822-
VD544--408136
BBB223754016971678559
MetabolismCYP 1A2-Inhibitor121455713643291453000
CYP 3A4-Inhibitor118935047684688933000
CYP 2C19-Inhibitor122725670660292723000
CYP 2C9-Inhibitor117203960776087203000
CYP 2D6-Inhibitor1272623421038497263000
CYP 2C9-Substrate784278506626156
CYP 2D6-Substrate816352464611205
ExcretionClearance544--408136
T1/2544--408136
ToxicityhERG655451204392263
H-HT217114357361628543
Ames76194252336757141905

| Data sources

Note: you can contact us to get the datasets.

| Model summary

To obtain robust and reliable QSAR models for ADMET properties prediction, we constructed a series models and aimed to find a best one. Six methods (RF, SVM, RP, PLS, NB, DT) and seven types of descriptors (2D, Estate, MACCS, ECFP2, ECFP4, ECFP6, FP2) were applied in the modeling process. The best model for each property and its performance can be seen in tables below (Table S2, Table S3). You can click the “Detail” to get the concrete information about all obtained models if you need.


| Model Results

Table S2. The best regression models for the coressponding ADME/T related properties

propertyMethodTrain sizeTest sizemtryR2Q2R2TRMSEFRMSECVRMSET
LogSRF1541514100.9950.9670.9570.1380.3690.436
LogD7.4RF773258140.9830.8770.8740.2280.6140.605
LogP          
Caco-2RF886296140.9730.8450.8240.1210.2890.290
PPBRF1368454-0.9540.6910.6827.12418.44318.044
VDRF408136100.9500.6340.5560.2810.7620.948
CLRF408136100.91370.36940.43120.34670.93720.9047
T1/2RF408136120.90680.32540.25370.34170.91911.0055

Table S3. The best classification models for the coressponding ADME/T related properties

PropertyMethodfingerprint Five-fold cross validationExternal validation dataset
SensitivitySpecificityAccuracyAUCSensitivitySpecificityAccuracyAUC
HIASVMMACCS0.9920.9290.9840.9951.0001.0001.0001.000
F (20%)SVMMACCS0.9070.4500.7920.7490.9040.4310.7820.727
F (30%)SVMECFP60.9270.4020.7530.7900.8770.4000.7060.720
BBBSVMECFP20.9620.8130.9260.9480.9930.8540.9620.975
Pgp-inhibitorSVMECFP40.8870.7890.8480.9080.8630.8020.8380.913
Pgp-substrateSVMECFP40.8390.8070.8240.8990.8260.8540.8400.905
CYP1A2-IhibitorSVMECFP40.8330.8640.8490.9280.8530.8800.8670.939
CYP3A4-IhibitorSVMECFP40.7590.8580.8170.9010.7880.8600.8290.909
CYP2C19-IhibitorSVMECFP20.8260.8190.8220.8930.8120.8250.8190.899
CYP2C9-IhibitorSVMECFP40.7190.8980.8370.9000.7300.8820.8300.894
CYP2D6-SubstrateSVMECFP40.3110.9860.8600.8720.2870.9890.8650.867
CYP2C9-SubstrateSVMECFP40.9190.4230.7510.7740.9150.4270.7250.746
CYP2D6-SubstrateSVMECFP40.8460.5820.7300.8020.8840.7000.8110.847
hERGRF2D0.9080.7000.8440.8790.8880.7620.8480.873
H-HTRF2D0.7760.5200.6890.7100.7850.4870.6810.683
AmesRFMACCS0.8340.8000.8200.8900.8160.8480.8340.897
Search & Database
| Database contents

The database integrated all the ADMET entries from ChEMBL database, Drugbank database, EPA database and related records from several literatures along with all the data described above. We manually checked the correctness of the values and dropped redundant information which resulted in 208,967 entries. Each entrie including basic molecular properties( eg. , common name, SMILES, ALogp, PSA) and ADMET activities.


| Similarity search

QSAR models and similarity searches are both useful strategies to predict ADMET properties. Compared with QSAR modesl, The similarity searches in databases are fast and can easily be extended to include new information. Here, we provides 7 kinds of fingerprints to represent molecular information and 2 kinds of similarity metrics. Users can input molecules to estimate their properties by comparing with similar compounds.

Systemic evaluation
| Summary

Not just one property affects the behavior of drugs in body. Usually we are looking for molecules that possess relatively good performance through every stage of ADME/T. Here, we developed this module that allows users to evaluate most aspects of ADME/T process of one molcule. The results provides users an full impression and lead to constructive suggestions of molecular optimization.