Drug properties (ADMET) prediction using AI

ADMET Property Prediction
Platform using AI
1

Agenda
•What is ADMET?
•What are the different ADMET Properties?
•Data Collection?
•Atom Features/Bond Features
•Models /Algorithms
•What is our Solution?
•Explanability of model?
2

ADMET Property Prediction Platform using AI
ADMET
(Absorption, Distribution,
Metabolism, Excretion
and Toxicity). The
prediction of the ADMET
properties plays an
important role in the drug
design process because
these properties account
for the failure of about
60% of all drugs in the
clinical phases.
3

Dataset #Task Task Type #Compound Metric
Clearance 1 Regression 4000 RMSE
Lipophilicity 1 Regression 4200 RMSE
PPB 1 Regression 1700 RMSE
LogP 1 Regression 3900 RMSE
LogD 1 Regression 7000 RMSE
Solubility 1 Regression 11000 RMSE
t1/2 1 Regression 3293 RMSE
hERG 1 Regression 1102 RMSE
Cyp450 5 Regression 7-8000 RMSE
Toxicity 32 Regression/
Classification
8000 RMSE/ROC/F
Score
What are the ADMET Properties?
4

Clearance
Clearance is a pharmacokinetic measurement of the volume of plasma(blood) from
which a substance is completely removed per unit time.
Units: clearance is measured in L/h or mL/min.
renal clearance(kidney) + hepatic clearance(liver) + lung clearance = total body
clearance.
Why clearance important in drug discovery?
It defines the rate of administration and dosages formulation of drug required to
maintain a plateau drug concentration in the body.
DrugBlood
Blood
Blood
Blood
Liver
lungs
Kidney
5

Cyp450 Inhibition:
 CYPs are the major enzymes
involved in drug metabolism,
accounting for about 75% of the
total metabolism.
 Most drugs undergo deactivation
by CYPs.
Why cyp450 important in drug
discovery ?
Effects on CYP isozyme activity are a
major source of adverse drug
interactions, since changes in CYP
enzyme activity may affect
the metabolism and clearance of
various drugs. For example, if one
drug inhibits the CYP-mediated
metabolism of another drug, the
second drug may accumulate within
the body to toxic levels.
Proportion of antifungal
drugs metabolized by
different families of CYPsCyp450 Inhibition
6

What is the use of AI Predictor:
Increasing the success rate of compounds reaching development.
In silico approaches accelerates the properties prediction process. Unfavorable
absorption, distribution, metabolism and elimination (ADME) properties have been
identified as a major cause of failure for candidate molecules in drug development.
Log P
Log D
Clearance
PPB
…………….
Known properties
In silico approach for ADMET
properties prediction
Molecules with
unknown properties
Unknown properties
7

Data collection
Data are collected from Open Sources publically available databases
Data contains:- Chemical molecules in form of smiles, binary values
and target labels.
Road-Blocks in data collection:-
For ADMET properties prediction, labels having different units
because bioassay performs are in different Exp. for same properties.
CSV/JSON/
xml file
Database Sources
8

9
SMILES (Log P Data) LogP
COc1ccc(CC(=O)O[C@H](CC=C(C)C)C2=CC(=O)c3c(OC)ccc(OC)c3C2=O)cc1 3.39
COc1ccc(OC)c2C(=O)C(=CC(=O)c12)[C@@H](CC=C(C)C)OC3CCCO3 1.8
CC(=O)N[C@@H](Cc1cc(F)cc(F)c1)[C@H](O)CNC2(CCCCC2)c3cccc(c3)C4CCCO4 1.371
SMILES (Aqueous Solubility) Value Unit
C[C@H](O)CNc1cc(ccn1)c2[nH]c3cccnc3c2c4ccc(F)cc4 25 ug.mL-1
CNc1nccc(n1)c2cccnc2Oc3ccc(NC(=O)c4cc(F)ccc4Nc5ccccc5)cc3 16 ug.mL-1
COc1ccc(cc1C#Cc2cccc(C)n2)C(=O)N3CCN(CC3)c4ccccn4 18 ug.mL-1
SMILES (t_half) Value Unit
Fc1cc(cc(F)c1c2ccnc3c(c4cccc5[nH]ncc45)c(nn23)c6ccncc6)N7C[C@@H
]8C[C@H]7CN8C9CCC9
0.0833 hr
Cc1ccc(c(Cl)c1)n2nnnc2SCC(=O)Nc3ccccc3Cl 0.0333 hr
Data Sample

Data Pre-processing
CSOCNP Features
Atom type
Chirality
Formal charge
Partial charge
Ring sizes
Hybridization
Aromaticity
Hydrogen bonding
One hot encoded vector
10

0 1 0 0 1 0 - - - -
- - - -
- - - -
- - - -
- - - -
- - - -
C1
C2
O
C4
C3
N
Hybridization
C1C2C3C4ON
Atom type
One-hot encoded vector
Feature matrix (H0)
0 1 1 0 0 0
1 0 0 1 1 0
1 0 0 0 0 1
0 1 0 0 0 0
0 1 0 0 0 0
0 0 1 0 0 0
Adjacency matrix(A)
C1C2C3C4ON
C1 C2 C3 C4 O N

Features Description Size
Atom type Type of atom (ex. C, N, O), by atomic number. 12
# bonds Number of bonds the atom is involved in. 5
Formal charge Integer electronic charge assigned to atom. 5
Chirality Unspecified, tetrahedral CW/CCW, or other. 4
# H bonds Number of bonded Hydrogen atom. 5
Hybridization sp, sp2 , sp3 , sp3d, or sp3d 2 5
Aeromaticity Whether this atom is part of an aromatic system. 1
Atomic Mass Mass of the atom, divided by 100. 1
Total 38
Atom Features
12

Bond Features
Features Description Size
Bond Type Single, double, triple, or
aromatic
4
Conjugated Whether the bond is
conjugated.
1
In-ring Whether the bond is part of
a ring.
1
Stereo None, any, E/Z or cis/trans 6
13

Models: Algorithm
Implementation of Graph Based Models-
 Try to build different types of models.
 ex-GCN, Graph SAGE,MPNN, Edge-Convolution……etc.
14

MPNN
MPNNs operate in two phases: a
message passing phase, which
transmits information across the
molecule to build a neural
representation of the molecule, and a
readout phase, which uses the final
representation of the molecule to
make predictions about the properties
of interest.
15

What is our Solution
ADME and Toxicity Property Prediction Platform using AI
. . . . .
.
Collect data
from various
Sources:ChEMBL
Clean data
Model Building:
Create different graph
based models for different
properties
Explanability:
Explaining of the
features & predictions
made
Collect more data:
Collect data on the go with
collaboration & improve
model
Predical model update:
Update the model with new
data & better representation
17

Three waves of AI:
Factors driving rapid advancement ofAI
Symbolic AI
Logic rules represent
knowledge
No learning capability and
poor handling of
uncertainty
StatisticalAI
Statistical models for specific
domains training on big data
No contextual capability and
minimal explainability
ExplainableAI
Systems construct
explanatory models
Systems learn and reason with
new tasks and situations
GPUs , On-chip
Neural Network
New
Algorithms
Cloud
Infrastructure
Data
Availability
Model Explanability
18

Black box AI creates confusion and doubt:
Why I am getting this
decision?
Can ItrustourAI
decisions?
Business Owner
Data Scientists
Are these AIsystem
decisions fair?
Internal Audit, Regulators
Customer Support
How do Ianswer this
customercomplaint?
Is this the bestmodel
thatcan be built?
Black-
box
AI
How can I get a better
decision?
Poor
Decision
19

What is explainable AI:
Black BoxAI
Data
Black-Box
AI
AI
product
Confusion with Today’s AI Black
Box
● Why did you do that?
● Why did you not do that?
● When do you succeed or fail?
● How do I correct an error?
Decision,
Recommendation
Clear & Transparent Predictions
● We got,why
● We got, why not
● We know why you succeed or
fail
● Got Explanable model, so we
trust on model
ExplainableAI
Data
Explainable
AI
Explainable
AI Product
Decision
Explanation
Feedback
20

Why need Explainability of model in healthcare domain:
Add potential effect
YesNot
Healthcare Domain
Accountability & Transparency
Q
Shame
 Improving interpretability and Explainability are important in more prosaic
business scenarios.
 Understanding how an algorithm is actually working can help to better align the
activities of data scientists and analysts
21

Why Explainability: How Improve ML model
PredictionPrediction
interpretability
Humaninspection
Verified prediction
Model/ data
improvement
Standard ML
ML model
data
Generalization error
ML model
data
Generalization error &
Human experience
Interpretable ML
Integrated
gradient, LRP,
GradCam
22

Achieving Explainable AI
Architecture:
I. Build GCN model for properties prediction.
II. Apply model interpretability method to get what model learned.
III.Make modification in data & model based on inference got from
Explainability of what model learned.
1)General Attribution: Evaluates contribution of each input feature to
the output of a model.
2)Layer Attribution: Evaluates contribution of each neuron in a given
layer to the output of the model.
3)Neuron Attribution: Evaluates contribution of each input feature on
the activation of a particular hidden neuron.
Building interpretable model:
Interpretability of model categories in three groups:
23

1) GCN Model:
Log p, log D, PPB, cyp450,
BBBP, herg
~ 40-50 properties
Molecular
explanation
For aqueous property
prediction
Forward propagation:
24
Sucrose(210gm/100ml)

LRP(Layer wise relevance propagation)
I. General attribution:
25
Interpretability
Y=1.80
Y_hat=1.959

EAGCN:
edge attention-based
multi-relational GCN
(EAGCN), jointly
learns attention
weights and node
features in graph
convolution. For each
bond attribute, a real-
valued attention matrix
is used to replace the
binary adjacency
matrix.
27

Physiological properties:
 1) Log P:
ratio of the concentrations of a solute between the two solvents (a bi-phase of liquid phases) for un-ionized solutes.
When one of the solvents is water and the other is a non-polar solvent, then the log P value is a measure
of lipophilicity or hydrophobicity.
2) Log D:
The distribution coefficient, log D, is the ratio of the sum of the concentrations of all forms of the compound
(ionized plus un-ionized) in each of the two phases, one essentially always aqueous; as such, it depends on the pH of
the aqueous phase, and log D = log P for non-ionizable compounds at any Ph.
28

Drug properties (ADMET) prediction using AI

More Related Content

What's hot (20)

Similar to Drug properties (ADMET) prediction using AI (20)

Recently uploaded (20)

Drug properties (ADMET) prediction using AI