In the initial phase of drug discovery, computer aided methods are routinely applied. They are used to tackle many interesting problems that are challenging even for a physicist. For instance, algorithms for effectively searching in a database of millions of chemical compounds do exist (e.g. similarity search, substructure search, etc.), however their speed and/or accuracy is still not satisfactory in various cases. Problems such as retrieving a set of chemical compounds from a database based on some criteria (i.e. non-exact optimization problems) or discovering and predicting new molecules and their properties by machine learning and AI are under research and yet to be explored. Such models can be potentially useful for the prediction of physicochemical properties of organic molecules. For example, a tool for accurately predicting solubility of diverse, organic compounds is currently not available. The application of neural networks can be advantageous if sufficient amount of experimental data is supplied. Another interesting topic is the prediction of synthetic feasibility of compounds. The so-called building block compounds would be the core target for such a prediction algorithm, i.e. whether or not a reaction can take place between two or more of these instances leading to a drug-like molecule. It is very tempting to create a neural network based on some kind of topological fingerprint which would capture exciting (abstract) features in a molecule. Eventually it would lead us to areas in the chemical space which are yet to be discovered and thus numerous potential drug candidates would become available. The applicant will dive into the details of these problems. He will study the existing solutions and try to come to a better and more useful one.