A unique feature of ChemBrain is its inbuilt capability to learn from the data stored in the molecule database. For this purpose, it provides a number of architectures of artificial neural networks (ANN), designed for specific tasks and applying the well-known Kohonen mapping and back-propagation strategies.
ChemBrain's scope of tasks encompasses classification, modeling, selection and prediction, Classification in this context means the association of molecules to groups according to their 3-dimensional structure or other properties; the modeling task associates experimental values of a selected property to the molecules' 3-dimensional structure, also called "quantitative structure-activity relationship" (QSAR); the selection option enables you to select a representative subset of molecules from a larger set, based on their 3-dimensional structure. This latter option has been proven in many ways to be far superior over ordinary random selection methods, because it evenly covers the whole space of structural varieties.
Since each of these tasks requires for its solution various prerequisites, ChemBrain guides you through these input requirements in the neural-network strategy form in a self-explanatory way. The results of the corresponding ANN calculations are then represented either as a Kohonen map or as a list in the case of the back-propagation technique (see examples in the screenshots).
Finally, the prediction capabilities enable ChemBrain to predict the value of a certain property of a given molecule, provided there is a series of structurally similar molecules in the database for which the property in question is known. In order to use this application the query molecule has to be called from the database (if present or otherwise entered as a new input), which opens its datasheet. There, you find the Predict single-value property button which opens the property and ANN strategy selection form, where you will be asked to select the requested property, the ANN strategy and optionally further limitations. If you select a new ANN calculation or a pre-trained one which ChemBrain does not find for the selected property, the program immediately opens the neural-network strategy form mentioned above.
The result of the prediction process is shown in the prediction sheet (see last screenshot). On the left side it shows the query molecule and the predicted value; in the right list, which formed the basis of the prediction value, the structurally most closely related molecule is framed in red. (The experimental value of the isoelectric point for the example is 5.96. The example was not part of the training set.)