I want to know about PLS regression. QSAR model is built with PLS regression with 100 descriptors using LOO cross validation and 4 latent variables. I want to know the contributions of each descriptors in final model. If I know few most contributing descriptors, I can propose their role in model.
Yes, you can use it. Other nice commercial software: Pirouette (expensive) and Molegro Data Modeller (affordable). If you need a nice and simple(FREE)to use QSAR software for linear or quadratic models (PLS not included) see http://www.profanderson.net/files/buildqsar.php
Dear Bruno,
Can you help me in using PLS_toolbox in MATLAB? Though your suggestions are well taken, I am not sure I could use any other software for this purpose. I have tried to use BuildQSAR by Prof. Anderson but could not succeed with my dataset. My objective is to use PLS model with dataset of more than 16,000 descriptors with pIC50 as dependent variable. With eliminating non contributing descriptors having correlation less than 0.3 I have the dataset of around 600 descriptors. I have built a model with PLS analysis using training set and choosing LOO cross validation along with external validation of model with test set. The best model I am having now with r2 calculated 0.98..., r2 CV= 0.87..., r2 predicted= 0.85...
Now with this model, I am stuck in finding which descriptor is contributing most. Suppose I want to propose top 10 descriptors out of 600 descriptors used in model building, how do I do it? I have search all HELP in both MATLAB and in PLS_Toolbox but unable to find the answer.
I would appreciate if you give me few instructions.
How do I load the data in Build QSAR? I am having .exls data with me but could not load it in BuildQSAR module.
Regards
Rajesh Patil
(1)PLS_Toolbox: To variable selection I recommend that you read the Manual Chapter.11 related to the use of Genetic Algorithm or, better than that, the use of IPLS(Interval PLS) which can help you to find the best descriptors of your QSAR.
(2) BuildQSAR opens two type of data files: *.txt and *.cvs. Please ask Prof. Anderson about the details for saving the file format correctly, he is a very accessible person and he will help you. The correlation matrix of BuildQSAR is an excellent tool to check the relevance of your descriptors in relation to your pIC50 but you will be able only to use its Genetic Algorithm and MLR. I do not recommend to use more than 100 descriptors with this software.
I hope this helps.
Thank you very much. I will go through your recommendations. Regarding BuildQSAR already I have communicated Prof. Anderson and hope he will help me.
regards
Rajesh
Can I suggest you to try R or python?
R is similar to matlab, but it is free.
Python is also free and have a long list of libraries to use PLS, plus it does have libraries to read in mols anc alculate the descriptors on the fly.
- Badges
- Science topic