Proteins are molecules of great importance for all living beings. Their role is to partake in biological processes and provide organisms with their shape, but at times they are also the cause of health disorders - in these cases we aim to be able to manipulate their activity. With modern computer based methods we can study proteins in detail and answer difficult questons without complex and expensive laboratory procedures. In this work we focus on determining protein locations to which small molecules - ligands - bind. Since ligand binding can alter protein function and consequently stop its negative effect on the organism, important motivation in this type of research is novel drug discovery. To predict protein binding sites we use a three-dimensional convolutional neural network which takes the spatial structure of a protein into account. For our dataset we choose proteins from the PDBbind and sc-PDB databases. In the cases of similar proteins we combine available ligand information using all the possible ligands and not only a special sub-selection, as a way to generalize the work of existing research. We analyze our model's performance through various metrics to notice that for 54% of larger and pharmacologically more relevant binding sites the distance between their real and predicted centers amounts to less than 4Å.
|
