In silico screening in modern drug discovery research

1
In silico screening in modern drug discovery
research

• Presented by Olga Komina
• Department of Computer Science Engineering
• University of Nebraska Lincoln
• July 2004

2
Modern Drug Discovery

• Multidisciplinary area of research
• Combinarotial chemistry
• Chemoinformatics
• Molecular biology
• Biochemistry
• Medicine
• Macromolecular modeling
• Pharmacology
• Drug Discovery is a goal of research. Methods and
  approaches from different science areas can be
  applied to achieve the goal.

3
Drug Discovery Pipeline

• Target identification and validation
• Assay development
• Virtual screening (VS)
• High throughput screening (HTS)
• Quantitative structure activity relationship
  (QSAR) and refinement of compounds
• Characterization of prospective drugs
• Testing on animals for activity and side effects
• Clinical trials
• FDA approval

4
Computer-aided Drug Design Strategy
5
Mechanism of Drug Action
6
Virtual Screening (VS)

• In silico screening of large compound databases
  in order to reduce the scale of high-throughput
  screening.
• Conceptual diversity
• Small molecule screening
• Protein structure based screening
• Algorithmic diversity
• Similarity searching
• Clustering and partitioning
• Simple filters
• Artificial intelligence
• Integration of different computational approaches
• Similarity paradox

7
Similarity Paradox
8
Descriptors of Molecular Structure Properties

• 1D-descriptors encode chemical composition
  physicochemical properties
• MW, CmOnHk ,hydrophobicity
• 2D-descriptors encode chemical topology
• Connectivity indices, degree of branching, degree
  of flexibility, of aromatic bonds
• 3D-descriptors encode 3D shape, volume,
  functionality, surface area
• Pharmacophore the spatial arrangement of
  chemical groups that determines its activity

9
Connectivity Indices

• Connectivity of an atom
• of atoms connected to it
• Connectivity of a bond -
• the reciprocal of the square root of the product
  of the connectivities of the atoms
• Connectivity index of a molecule summation of
  all bond connectivities

Isobytul alcohol
10
Classification of Atoms to Atom Types

• Developed for prediction of log P values
• A molecule characterized by the count of 120 atom
  types
• Atom type commonly occurring atomic states of
  C, H, O, N, S, P, Se and
• halogens (F, Cl, Br, I)

11
Atom Types (Carbon)
12
Example Description of a Molecule by the Count
of Atom Types
13
Molecular Fingerprints

• Molecule A 00011101010
• Molecule B 00101111000
• Tanimoto coefficient Tc

Nc
3
Tc

NA NB - Nc
5 5 - 3
Nc the number of common bits set on NA the
number of bits set in A NB the number of bits
set in B.
14
Drugs vs. Non-drugs

• Enriching screening libraries with drug-like
  compounds
• fail fast, fail cheap strategy
• Manual classification is time-consuming and bias
• Computational approaches speeds up the screening,
  reduce the size and improves the quality of
  combinatorial libraries
• Assumption typical drugs have something in
  common that other compounds lack

15
Lipinski Rule of Five (1997)

• Poor absorption and permeation are more likely to
  occur when there are more than 5 hydrogen-bond
  donors, more than 10 hydrogen-bond acceptors, the
  molecular mass is greater than 500, or the log P
  value is greater than 5.
• Further research studied a broader range of
  physicochemical and structural properties
• Related problems
• Compound toxicity
• Compound mutagenicity
• Blood-brain barrier penetration
• Central nervous system activity

16
Data Sets

• Drug Databases
• World Drug Index (WDI)
• Comprehensive Medical Chemistry (CMC)
• MACCS-II Drug Data Report (MDDR)
• Non-drug Databases
• Available Chemical Directory (ACD)
• Quality of training sets

17
Artificial Neural Networks

• ANNs are self learning systems which learn from
  experience
• Biologically inspired
• Neuron is a processing element
• Artificial neuron simulates four basic functions
  of a natural neuron
• Receives input from other sources
• Combines those inputs in some way
• Performs nonlinear operations on the result
• Outputs the final result

18
Artificial Neuron
19
Network Topology
20
ANN Training

• Supervised both inputs and outputs are provided
• Initial weights chosen randomly
• Errors propagated back through the system to
  adjust weights
• Most common algorithm backward-error propagation
  (back-propagation)

21
ANNs for Drug Classification (1998)

• Input Counts of atom types
• Topology 92 x 5 x 1
• Feedforward with backpropagation
• Training 5000 ACD and 5000 WDI
• Accuracy 83 - ACD, 77 - WDI

22
ANNs for Drug Classification (1998)

• Input seven 1D descriptors (MW, log P, aromatic
  density) and ISIS fingerprints
• Topology 173 x 0/5/10 x 1
• Bayesian learning procedure
• Training 3500 ACD and 3500 CMC
• Accuracy 90 - CMC, 80 - MDDR, 90 - ACM

23
Misclassification Examples
Misclassified non-drug
Misclassified drug
24
ANNs to Predict Biological Activity

• Applications
• CNS-active compounds
• Protein kinase inhibitors
• G protein-coupled receptor ligands
• Best prediction accuracy 80
• Advantage capable of predicting structurally
  diverse compounds
• Disadvantage no definite rules

25
Recursive Partitioning

• Statistical method for analyzing and mining large
  data sets that consists of active and inactive
  molecules
• HTS data analyzed to discover SAR
• Easy to visualize and interpret
• Applicable to a variety of classification problem
• A problem of assigning chemical compounds to
  property classes based on their structural and
  physicochemical features

26
Partitioning Problem Definition

• Given a training set of D descriptor values and P
  property values for each molecule in the set, the
  question is to create a set of yes/no questions
  which are organized into hierarchical tree from
  with one question per node and class predictions
  at leaf nodes with minimum classification error.

27
Single Property RP

• Single property classification such as molecules
  classified active or inactive
• Drugs vs. Non-drugs
• C4.5, C5.0

28
Single Property RP (cont.)

• All possible questions are asked based on single
  descriptor values, scores of corresponding
  partitions are computed
• Descriptor resulting in the best score is used
  to grow the tree
• Loop to question asking until terminating
  condition is met

29
Gini Impurity Metric

• Impurity, I, of a node
• I ? pipj
• where pi and pj are the fractions of the members
  of a node that belong to class value i and j
  respectively
• Gini metric maximizes the decrease in Impurity,
  ?I, from a potencial node question
• ?I I pLIL pRIR
• where pL and pR are the fractions of the node
  members that partition to nodes L and R
  respectively for a given question, and IL and IR
  are the impurities of new nodes

30
Tree Growth
Entire Training set
Root
Descriptor 3
yes
no
Node L
Node R
Pruning phase metric R? R? ?Nleaf R? the
number of misclassifications in the training set
31
Application of Single Property RP for
Drug/Non-drug Classification

• Input 120 atom types
• C5.0
• Training 5000 WDI, 5000 ACD
• Prediction error 21
• The presence of alcohols, tertiary and secondary
  amines, phenols, enols, and carboxylic groups
  accounts for 75 of correct classifications for
  drugs.

32
Decision Tree for Drug/Non-drug Classification
33
Multiple Property RP

• SP is not sufficient in many biological systems
• ADMET properties
• Absorbtion
• Distribution
• Metabolism
• Excretion
• Toxicity
• Nonspecific binding to multiple targets causes
  side effects
• Dependent properties

34
Partially Unified Multiple Property RP

• Developed for prediction of multiple dependent
  properties
• Discover features that distinguishes the classes
  of different properties and make them similar
• Some node apply to all properties while others
  apply to only single properties
• Classes are NOT mutually exclusive
• Nodes are labeled with one class of a single
  property type

35
Mapping to SP Representation

• D descriptor values x1, x2, x3, , xD
• P property values y1, y2, y3, , yP
• New descriptor K is a property descriptor
• x1, x2, x3, 1, y1
• x1, x2, x3, y1, y2, y3 x1, x2, x3, 2,
  y2
• x1, x2, x3, 3, y3
• Every path from the root to a leaf has a split on
  the descriptor K

36
1. Pure Specific Tree
2. Generic node growth Max ( Min ?I k ) gt 0
k
37
PUMP-RP (cont.)

• A split with an improvement for each property is
  chosen
• The metric maximizes the minimum decrease in
  impurity from each potential node question
• A compound may appear in more than one leaf node
• Each K node is regrown recursively
• The resulting tree is overgeneralized

38
Finding the Best Tree

• R?? Ro ?(Nleaf - ?Ngeneric)
• Where ? is a generality parameter,
• Ngeneric is the number of generic nodes

39
Application of PUMP-RP for Drug Specificity

• Cyclooxygenase (COX) inhibitors
• COX-2 inhibitors are antiinflammatory agents
• COX-1 inhibitors damage gastrointestinal tract
• Good drug should be highly specific to COX-2
• Celebrex, Vioxx are widely prescribed
• Goal to obtain a model of activity and
  selectivity of COX-2 inhibitors as a function of
  their physicochemical properties

40
Data and Results

• 100 2D and 3D descriptors
• Each property has two classes active and
  inactive
• Gini Impurity score
• Accuracy
• on the training set
• 60-80 COX-2, 78-91 COX-1
• On the test set
• 50-89 COX-2, 60-100 COX-1
• Disadvantage not capable of predicting compounds
  with molecular scaffolds not yet discovered

41
Extension to PUMP-RP

• To model systems with more than two properties
• semi-generic node applies to more than one
  property but not all
• To model multiple properties with opportunity to
  observe what properties are more closely related
  than others
• Problems to apply
• ADMET properties
• Activity/ADMET properties
• COX-2/COX-1/Drugs
• Drug-drug interactions based on target
  specificity
• Modified Gini Impurity score

42
Gini Impurity for the Extended PUMP-RP

• Modified scoring function
• Max (Max (Min ?Ik)), where k P

k
43
Tree Built by the Extended PUMP-RP
44
Targeting RNA

• Emerging field in drug discovery
• RNA plays an essential role in many biological
  processes
• Natural antibiotics are RNA-targeting drugs
  (streptomycin, tetracycline, etc)
• Potential drug targets viral RNAs
• Antisense strategy

45
Targeting RNA

• HTS against RNA targets less successful that for
  protein targets
• Identification of new classes of RNA ligands are
  extremely rare
• Limited knowledge of the chemistry and structure
  of RNA recognition
• Consists of 4 nucleotides less diverse than
  proteins, RNA flexibility

46
What Can Be Done?

• Assumption compounds binding RNA have something
  in common that other compounds lack
• Dataset a comprehensive database containing
  examples of bindings between small molecules and
  RNAs
• Computational approaches to extract common
  features of such compounds and to train models
  for prediction (AI methods)

47
Concluding Remarks

• Drug Discovery is a goal of multidisciplinary
  research
• No algorithm to discover a drug
• Old problem given a compound structure, what are
  its properties?
• Computational approaches can assist drug
  discovery process
• Limitation lack of systematic biological data
• Market pressure and prospective profit bring more
  and more resources into drug discovery

48
Multilevel Neighborhoods of Atoms
phenol