We have long been accustomed to the fact that neural networks help in almost all areas of our lives. It seems that soon they will be even more popular, because new approach in their work will allow to predict the behavior of the substances in the interaction, using the minimum set of input data.
According to the editorial Board of journal of Physics: Condensed Matter, neural network, developed by a group of scientists from Russia, Estonia and the UK, allows you to “predict one of the most unpredictable factors” biological concentration of a substance. It is extremely important to assess the effectiveness of any biological compounds from medicine to poison, and indicates the degree of accumulation of the test substance in a living organism. Earlier it was used of laboratory animals. Now you can apply a more modern approach.
How to achieve this? There are 2 variants: in the first case, the neural network need to generate all possible behaviors of molecules and their interactions with living organisms. This approach will be quite accurate, but extremely expensive because of the huge data array. In the second case, the neural network must be trained to work with the molecular theory of fluids. In this case, the interpretation of the results will be quick (as it is based on a pre-defined and fairly simple formulas), but not as accurate as we would like.
A group of experts from the University of Tartu (Estonia), University of Strathclyde (UK) and the SKOLKOVO Institute of science and technology (Russia) developed a hybrid method. First, the calculations are based on molecular theories, and it already applied the algorithms behavior and interaction of molecules. According to one of the authors, graduate students of the SKOLKOVO Institute of science and technology Sergey Sosin,
“Our method allows to simplify the prediction of environmental load from a particular substance. But the most important thing is that we have developed a universal method of the description of molecules for the transmission of its “image” in precise three-dimensional neural network. In the future, our development will allow us to predict the properties of various exotic molecules and radically new compounds, for which existing methods of finding correlations “structure — property” not applicable.”