“Trained” Computer Will Help
Software, simplifying the process of drug development, was created at the Research Center of molecular mechanisms of aging and geriatric diseases. Knodle Software kit for determining links and the extent of links in molecules has been developed by an employee of Laboratory of structural biology of receptors linked to G-protein, Moscow Institute of Physics and Technology, Sergey Grudinin, and postgraduate student of MIPT, Maria Kadukova.
They “helped” the computer to assess the structure of links in molecules, using technology of macihine learning. Selecting a non-linear support vector machine (SVM), well-established in automatic recognition of handwriting and images, they sourced it with positions of neighbouring atoms, and at output the structure of links was obtained. Source materials for the program were selected by researchers and included 7605 compounds with characterized structure and atom positions. “The decisive benefit of our software is that as learning progresses, more accurate the results of recognition are available. Currently Knodle is one step ahead of similar programs: the error rate is just 3.9%, while the error rate of the closest competitor is 4.7%, according to Maria Kadukova.
The software is easily adjustable for a specific task. Currently, Knodle does not work with metal-containing substances, but if required, the database may just be supplemented with metal compounds, and it will be able to operate with metals.