In silico ADME modelling 2: Computational models to predict human serum albumin binding affinity usi

1
In silico ADME modelling 2 Computational models
to predict human serum albumin binding affinity
using ant colony systems

• Sitarama B. Gunturi, Ramamurthi Narayanan and
  Akash Khandelwal
• Bioorganic Medicinal Chemistry, June 2006
• 14(12)4118-4129
• presented by Martin Kircher

2
Outline

• ADMETox and Human serum albumin (HSA)
• QSPR/QSAR
• Ant Colony Systems and the model derivation for
  HSA
• Model validation and conclusions
• Discussion

3
IntroductionADMETox properties

• Pharmacokinetic processes a drug undergoes
• A Absorption
• D Distribution
• M Metabolism
• E Elimination
• ADME properties of a compound define its
  bioavailability
• ADMETox fail early, fail cheap

D Distribution
4
IntroductionHuman serum albumin (HSA)

• Most frequent protein in blood plasma
• Transport of non-water soluble compounds like
  hormones, fatty acids, drugs,
• Maintenance of oncotic pressure
• Binds calcium ions (Ca2) and buffers pH of blood

II Indole-benzodiazepine site
I Warfarin site
PDB1E7H
HSA with six palmitic acid molecules
5
IntroductionQSPR/QSAR

• Quantitative Structure-Property/Activity-Relations
  hip
• Aim Prediction of properties from molecular
  structure without need of performing experiments
• Assumption Similar structures cause similar
  properties
• Hansch equation
• Intercept c
• Descriptors D any molecular/numerical
  property
• Coefficients k Contribution of the
  Descriptor

6
IntroductionSet of investigated compounds

• 94 drugs and drug-like compounds (Colmenarejo et
  al. 2001) with HSA affinity values
  (high-performance affinity chromatography)
• 84 training set, 10 test set
• Diverse set with molecular weight from 129.0935
  to 764.9488 g/moland log(Khsa) values from-1.39
  to 1.34

Acetylsalicylic acid
Clotrimazole
5-Fluorocytosine
Digitoxin
7
IntroductionAvailable descriptors

• Set of 396 descriptors
• 392 molecular descriptors created with in-house
  software (BioSuite) belonging to classes
  structural, physico-chemical, topological
• 4 physico-chemical descriptors created with
  'PreADME' (web-based)
• 327 kept after pre-filtering removed descriptors
  with constant value in more than 95 of all
  compounds

8
Aim of this study

• Build optimal QSPR models for HSA binding
  affinity prediction
• Train models with 1 to 8 out of lt 300 available,
  computable descriptors
• Use a computational fast algorithm for descriptor
  selection (ACS)
• Present a general applicable strategy for
  QSPR/QSAR model creation
• Evaluate the quality of the obtained models
• Interpret the physical meaning of selected
  descriptors

9
Derivation of HSA binding models Ant Colony
Systems (ACS)

• Ants are social insects behavior directed to
  survival of colony
• Deposit pheromones to mark trails (nest/food)
• Pheromones evaporate over time
• Individual ants select paths by random, but favor
  paths with high pheromone concentrations
• Colony finds (almost) optimal paths

10
Derivation of HSA binding models Ant Colony
Systems (ACS)
Shortest path ? fastest path ? path with the
highest concentration of pheromones Evaporation
and random choice restrict a rapid drift towards
same (suboptimal) part of search space ?
Stochastic optimization process
11
Derivation of HSA binding modelsAnt Colony
Systems

• Ant colony algorithm first used for QSAR
  descriptor selection by Izrailev and Agrafiotis
  2001
• Build models with 1-8 descriptors, judge fitness
  of model by R2 value (coefficient of
  determination)
• Inter-correlated descriptors are rejected if r2
  0.75

12
Derivation of HSA binding modelsAnt Colony
Systems
Set weights Wj 0.01 IT 0
Calc. probabilities Pj
Select descriptors by weighted random Reject
inter-correlated descriptors
Select k descriptors
Train model
Calc. new weights
Update IT 1
IT lt 20000
Return best model
13
Derivation of HSA binding modelsAnt Colony
Systems

• Roulette wheel selection
• Generate uniform distributed random number in
  0,1
• Rotate wheel by 2p times random number
• Return segment on stop position
• Allows weighted random selection

14
Derivation of HSA binding modelsAnt Colony
Systems

• Transform wheel to array
• Generate uniform distributed random number in
  0,1
• Sum probabilities of descriptors until it exceeds
  the selected random number
• Return last descriptor

15
Derivation of HSA binding modelsAnt Colony
Systems
Train linear regression models on training
set Calculate R2 value
Update weights
16
Derivation of HSA binding modelsAnt Colony
Systems
Set weights Wj 0.01 IT 0
Calc. probabilities Pj
Select k descriptors
Train model
Calc. new weights
Update IT 1
IT lt 20000
Return best model
17
Derivation of HSA binding models

• Steep increase of R2 up to six descriptors (?R2 gt
  0.015)
• Danger of over parameterization, further evaluate
  5/6-descriptor models

18
Derivation of HSA binding modelsQuality measures

• Leave One Out Cross Validation (LOOCV) approx.
  train error
• Standard Error standard deviation of the
  residual error
• F statistic testing utility of model

19
Derivation of HSA binding models

• Retrieve best three models with 5 and 6
  parameters based on the training set

best 5 descriptor model
best 6 descriptor model
20
Derivation of HSA binding models
r2 0.7778
r2 0.7322
21
Model validation

• Test set results and comparison to other models

22
Conclusions

• Impact of descriptors on HSA binding affinity
  analyzed by frequencies in models with R² gt 0.88
• Frequently occurring descriptors probably the
  ones with most importance
• Evaluation to better understand HSA binding

23
Conclusions

• - 311 SklogS predicted solubility
• Correlation of 0.5471 with log(KHSA)
• 307 AlogP98 predicted octanol water
  partition coefficient
• Highest correlation (0.7867) with log(KHSA)
• 263 E-state S-hydrophobic
• Atomic Type Electro-topological state index
  describing hydrophobicity
• - 159 Order 4 auto-correlation (Broto-Moreau)
  weighted by masses
• Topological descriptor encoding molecular
  structure and atomic mass
• 166 Order 5 auto-correlation (Broto-Moreau) w.
  by polarizability
• Topological descriptor encoding structure and
  atomic polarizability

24
Conclusions

• Hydrophobic interactions have high impact
• 307/267 in more than 90 of all models
• Colmenarejo et al. and Xue et al. identified
  hydrophobic descriptors to contribute to higher
  log(KHSA)
• Hall et al. observed positive contributionof
  electron accessibility of aromatic and aliphatic
  groups
• Site I and II consist of mainly hydrophobic amino
  acids

II Indole-benzodiazepine site
I Warfarin site
25
Conclusions

• High solubility (311) of a drug reduces the
  binding affinity
• log(KHSA) rendered by branching, flexibility and
  shape descriptors (55,79,88,92,108) as well as
  descriptors of polarizibility (163,166) or
  electron accessibility (262)
• Meaning of descriptors 155, 157, 159 and 175
  remains unclear
• Overall Similar observations for different
  descriptors by Colmenarejo et al. 2001, Hall et
  al. 2003, and Xue et al. 2004

26
Summary

• Application of ant colony algorithm on available
  descriptors resulted in best linear model with 6
  descriptors known so far
• Obtained descriptor types consistent with
  previous results
• Interpretation of selected descriptors can help
  to understand HSA binding and to design optimized
  drugs
• Linear model quality seems to be limited
• Selectivity of the two sites probably difficult
  to capture
• Support Vector Machine with halve of prediction
  error (Xue et al. 2004)
• Set of 94 compounds rather small

27
Thank you for your attention!

• References
• Colmenarejo G, Alvarez-Pedraglio A, Lavandera JL.
  Cheminformatic models to predict binding
  affinities to human serum albumin. J Med Chem.
  2001 Dec 644(25)4370-8
• Dorigo M, Di Caro G, Gambardella LM. Ant
  algorithms for discrete optimization. Artif Life.
  1999 Spring5(2)137-72
• Gunturi SB, Narayanan R, Khandelwal A. In silico
  ADME modelling 2 computational models to predict
  human serum albumin binding affinity using ant
  colony systems. Bioorg Med Chem. 2006 Jun
  1514(12)4118-29. Epub 2006 Feb 28
• Hall LM, Hall LH, Kier LB. Modeling drug albumin
  binding affinity with e-state topological
  structure representation. J Chem Inf Comput Sci.
  2003 Nov-Dec43(6)2120-8.
• Izrailev S, Agrafiotis D. A novel method for
  building regression tree models for QSAR based on
  artificial ant colony systems. J Chem Inf Comput
  Sci. 2001 Jan-Feb41(1)176-80
• Izrailev S, Agrafiotis DK. A method for
  quantifying and visualizing the diversity of QSAR
  models. J Mol Graph Model. 2004 Mar22(4)275-84
• Xue CX, Zhang RS, Liu HX, Yao XJ, Liu MC, Hu ZD,
  Fan BT. QSAR models for the prediction of binding
  affinities to human serum albumin using the
  heuristic method and a support vector machine. J
  Chem Inf Comput Sci. 2004 Sep-Oct44(5)1693-700

28
Discussion

• Questions?

29
(Multiple) Linear regression

• Formulate Hansch equation as regression problem
• Find ß that minimizes a Loss-function
• Quadratic function with a global minimum

Element of training set
Complete training set
30
Comparison Confidence scores