Dear researchers,
In a ligand-based drug discovery, in which ligands are aligned to obtain a pharmacophore a pharmacophore-fit score is obtained. When I tried to do a QSAR whit that ligands, I was not able to use Pka or IC50s as independent variable. It was due to low fitness and inadequate adjusted R-squared. I tried to transform data. However, predictability did not increased significantly and Std.Errs were still high. When I tried to use Pharmacophore-fit score as Y variabel, 5 different models were retrieved with >98% predictivities and Std.Err range from 0.5 to 0. Is it possible to chose other quantitative variables instead of bio-activity measures? Nevertheless, I should mention that the Pka of selected inhibitors were from different cell culture and different methodologies. I think that is why the linearity didn't observed.
Regards
As a drug design practitioner I would just recommend to proceed with the screening and experimental testing. However, if you are doing QSAR for the sake of QSAR, then I would suggest to spend some time on the model validation. A model with an R^2 of 0.92 looks too good to be true. Usual ways to validate the model are cross-validation, splitting the data set into training and test sets, and y-randomization.
Hello
you can try to fit the model with values obtained from one single assay could be in different labs but following the same protocol. If you are able to see a good fit then the problem could be the diffrent sources of your experimental values.
Best Regards
There are no reasons to believe that fitness to a pharmacophore may anyhow be related to some quantitative expression of biological activity (especially, the activity in cell culture). Any pharmacophore obtained from available structure-activity data would be neither a necessary nor a sufficient set of molecular features of an active. It should only be considered as a useful filter to preselect compounds for a more relevant test (either computational or experimental).
Yes I used pharmacophore model for screening. By taking IC50 as dependent variable R-squared became 0.926 (by transforming data i mean log(1/x) and 1/ln(y)). Mean of residuals was 0.0213+/-0.288. But i think this is not enough. Unfortunately my training set for modeling pharmacophore (those compounds with structural homology) were only 4.
Dear @Dmittri_Kireev, as you recommended I reviewed more inhibitors in literature and performed ligand-based pharmacophore creation once again. By MLR analysis on more than 2400 descriptors. 7 models were obtained, by which the last one had good fitness. Here I attached residual plot. I hope to hear your recommendations.
As a drug design practitioner I would just recommend to proceed with the screening and experimental testing. However, if you are doing QSAR for the sake of QSAR, then I would suggest to spend some time on the model validation. A model with an R^2 of 0.92 looks too good to be true. Usual ways to validate the model are cross-validation, splitting the data set into training and test sets, and y-randomization.
Thank you for the replay. I found 34 compounds in the literature. The they were clustered structurally. Of those, 10 compounds were flagged as training set and other used for validation of the pharmacophore model (test set). I'd like to know that if i can use those test set to validate QSAR model. Do I need to know the biological activity of test set compounds?
Regards
Of course you need to know the activity. Otherwise, how would you know whether your predictions for them are correct?
I am very thankful for your recommendations. I would like to know your opinions about test-set. Here I attached the prediction of biological activity for test-set with the same formula used for training set. The residuals are varied and high furthermore predicted values are far different from the actual ones.
The x axes is one of the predictors and the y axes is predicted/actual values.
Then, what's the meaning of the two sets of points (blue and orange)?
Yes you are right, I tried to show the difference between Actual and Predicted values with different colors.
- Badges
- Science topic
- Similar topics
- Cognitive Science
- Thinking