Pgp modelling

1
P-gp modelling

• Is a straight line always the best way..?

2
Short answer

• No...

3
But why not..?

• Linear methods, best applied when
• single binding site
• consistent binding site
• exposed binding site
• P-gp on the other hand has
• broad specificity
• multiple binding sites
• high mobility
• membrane-bound protein

4
So whats our view?

• For a general model then some form of
  non-linearity should be involved
• Use in vivo data where possible
• CF1 mouse vs wild-type, Brain-Blood ratios were
  compared
• 33 Compounds represented by our in-house 2D
  topologically based descriptors
• quick to calculate
• covers property as well as atom-type features
• Relate descriptors to activity

5
Traditional properties..
Polar Surface Area
6
Log P
ACD Labs Log D ACD Labs Log P KW Log P
7
Molecular Size
AP Total MW
8
So whats our view?

• For a general model then some form of
  non-linearity should be involved
• Use in vivo data where possible
• CF1 mouse vs wild-type, Brain-Blood ratios were
  compared
• 33 Compounds represented by our in-house 2D
  topologically based descriptors
• covers property as well as atom-type features
• PCA on this data matrix
• classification using volumes within the PC space

9
And the results..

• First 3 PCs plotted
• Explain 81 of data (1300 descriptors)
• Conical clustering
• non-substrates at the apex and down the sides
• PC contributors
• PC1 - Hydrophobicity, path length lt7 ve
• PC2 - Charge descriptors prominent, Fs -ve
• PC3 - AP-type short path ve
• Too few compounds to fully explore the PC space

10
Does it work..?

• A qualified Yes
• The test points are not well distributed

• Mis-predictions for a couple
• point A prediction too low,
• point B Rhodamine 123 vs B, 5 fold vs 12 fold,
• Training set compound in red ellipse now seems
  out of place

11
Conclusions

• So far, we have a predictive model for P-gp,
  in-vivo, in mouse
• Attempts to model this data in other
  traditional ways have failed to give a model
• Need to expand the data set
• project in compound collection
• In-vitro P-gp data does not behave the same as
  in-vivo
• This model works with in-vivo data, not in-vitro