Model creation wizard
You can create a single model using web interface by accessing the "Models > Create a model" menu as shown below, which will open a wizard-like dialogue with a series of the necessary configuration steps.
First, you have to select a training and (optionally) a validation set. It is possible to have multiple validation sets. We assume that you have already prepared the sets.
On the same screen, it is required to select:
- the unit of measurement for the model
- the desired machine learning method
- the desired validation protocol (cross validation, bagging or no validation)
The next screen allows to configure the preprocessing of molecular structures. The options include:
- Standardization (e.g., of nitro-groups)
- Neutralisation of ions
- Removal of salts (by keeping the counter-ion)
- Cleaning of structures (by converting it to SMILES and back to SD-file)
Molecular descriptors
The next important step is the choice of molecular descriptors. OCHEM supports more than 20 descriptor packages provided by different vendors.
OCHEM policy is to integrate state-of-the art descriptors rather than to develop own solutions.
On the next step, we have to configure filtering of descriptors. The main filtering options include:
- elimination of constant or semi-constant descriptors
- unsupervised forward selection
- extraction of PCAs
- manual selection of the desired descriptors from a list
Machine learning method
Next, we have to configure the machine learning method. This step is method-specific. The screenshot below shows the configuration options for a neural network model.
Naturally, the other machine learning methods (e.g., linear regression, PLS, random forest, etc) will have different options.
Initiating the calculations
Now, we are ready to start calculations. In the dialog below, we have to give our model a name, define the priority of this task and, finally, initiate the calculations.
We recommend using high and extra-high priorities only for fast tasks (e.g., models with less than 300-500 molecules in the training set).
The next screen shows the progress of calculations. It is possible to fetch results later (in one hour, one day or more) from the registry of pending tasks.
The registry of pending tasks displays the status of all currently running calculation tasks. Once the task is ready, it will become "green", and it will be possible to fetch it by clicking a green checkbox icon on the right.
After you fetch the calculation task you will be presented with a model profile.
The model profile displays basic model statistics (RMSE, R2, etc), scatter plot, applicability domain plots, etc. On this step, you can either save your model or discard it in case if the model performance is unsatisfactory.
Accessing your model
After you save your model, it will become available from the browser of models:
The browser of models displays all the publicly available models developed by other OCHEM users as well as your own private models. Here, it is possible to search for models by various criteria.
The displayed information includes:
- model name
- training and validation sets
- machine learning method
Each available model has a model profile described above, which contains the detailed statistics, the scatter plots and confusion matrices for the classification models.
