Instructions for plugin teaching

General remarks about the pK_a teaching

What do you need to do in order to improve the accuracy of the pK_a calculation?
First, you need to see clearly which ionization center(s) was predicted inaccurately by the pK_a calculator. You need to collect experimental data for that ionization center(s). The learning algorithm is based on linear regression analysis, therefore you need to collect a certain amount of experimental pK_a data otherwise the regression analysis will fail. There is no rule of thumb for what a large pool of data is required to perform a reliable pK_a teaching. If your purpose is to create a local model with the scope only for a certain types of chemical environment of the ionization center then it may be enough to collect a few representative structures. A more robust model, however, requires as many as possible diverse structures and pK_a values of the ionization center in question.

The next step of the teaching process is the input of the collected data into an sdf file. The file can be easily created by using the graphical user interface of Instant JChem.

What kind of information should be included in the sdf file?
The structure of the molecules and their experimental pK_a value(s) and atomic ID's which are assigned to the appropiate pK_a value(s).

After preparing the sdf file you can run the teaching algorithm that creates a correction library from your data. This correction library will be used by the pK_a calculation of the ionization center in question.

Training of the pK_a plugin

1. Create a training set in sdfile (.sdf) format from your experimental data. The file must contain the following fields:
   • structure
   • pK_a value 1 (field name: pKa1)
   • ID of the atom which has the pKa1 value (field name: ID1)
   Additional pK_a values are optional (recommended for handling multiprotic compunds):
   • pK_a value 2 (pKa2)
   • ID2
   • etc.
   Definition of only one pK_a value is enough to apply the training data, but more values in case of multiprotic compounds will enhance the reliability of the pK_a teaching.

   In this example this file is mydata.sdf.
   The picture below is a detail from the training file. ID1 is the index of the atom with the experimental pK_a1 value (ID2 would be the index of the second measured pK_a value /pK_a2/, etc.). This atom index can be viewed by checking the Atom number option in the molecule editor (menu: View->Misc).
   

2. Generation of the knowledge base: Execute the following command from command line (in this example with the home directory path for Linux):

   cxcalc -T pKa -o /home/myaccount/ChemAxon/MarvinBeans mydata.sdf

   (option -o gives the location of the folder.) A 'pKaReg' folder will be created containing the training data. Create a folder called config in the Marvin installation directory.
   If Marvin's default installation was followed, Marvin installation directory is located in:
   • Windows: C:\Program Files\ChemAxon\MarvinBeans
   • Linux: USERHOME/ChemAxon/MarvinBeans (e.g. /home/myaccount/ChemAxon/MarvinBeans)
   • OS X: /Applications/ChemAxon/MarvinBeans
   
   Copy the created pKaReg folder to the config folder.
3. Use this knowledge base via cxcalc, Chemical Terms or Marvin. The training data helps to calculate more accurately the pK_a of the molecules and the results are closer to the experimental values.

Usage in Marvin

check the Use correction library box to activate the training option:

pKa calculation without training data	pKa calculation with training data

Usage with cxcalc

without correction library:

              cxcalc pKa "CC1=NC2=C(N1)C(C)=NC(C)=N2"
              id      apKa1   apKa2   bpKa1   bpKa2   atoms
              1       11.08           3.67    -2.38   6,9,3

with correction library:

              $ cxcalc pKa -c "CC1=NC2=C(N1)C(C)=NC(C)=N2"
              id      apKa1   apKa2   bpKa1   bpKa2   atoms
              1        9.90           3.67    -2.46   6,9,3

For more options see this page.

General remarks about the logP teaching

You can create your own logP calculator with the supervised learning method built into the logP calculator.

What you need to do is just simple collect experimental logP data and create a sdf file from them. Details about the expected file format given below in the technical help.

What do you need to see clearly in logP model building?

If you create a local logP model then the scope of the logP calculator will be limited. It means that the calculated logP will only provide reasonable prediction for a few types of structures. Practically only those types of structures will be predicted correctly which were lauched?? into the training set during the teaching process. For example, if the training set contains only certain types of carbohydrates (carbohydrogens?) and no other functional groups are present in the training set then it's not to expect that the predicted logP of any amine-like structure will be accurate.

In other words, you need to be aware that a more robust general logP model requires a large, diverse training set with thousands of structures.

Training of the logP plugin

1. Create a structure file of any molecule file format from your experimental data. The file must contain the following information:
   • structure
   • logP values in a property field named LOGP
   In this example this file is trainingset.sdf.
2. Execute the following command from command line:

   cxcalc -T logP -t LOGP -o logPparameters.txt trainingset.sdf

   With the -o option you can define a path for the file generated. Create a folder called config in the Marvin installation directory
   If Marvin's default installation was followed, Marvin installation directory is located in:
   • Windows: C:\Program Files\ChemAxon\MarvinBeans
   • Linux: USERHOME/ChemAxon/MarvinBeans (e.g. /home/myaccount/ChemAxon/MarvinBeans)
   • OS X: /Applications/ChemAxon/MarvinBeans
3. Save the file to the config folder with the name logPparameters.txt.
4. Use this data via via cxcalc, Chemical Terms or Marvin.