Discovering patterns from drug protein interactions based on their molecular fingerprints and applications in drug design

Abstract

Unveiling rules that govern drug-protein interactions is of paramount importance in drug discovery. However, up to now, not much work has been done in this field. Unlimited amount of knowledge in this field are acquired, with much left unknown. To fill this gap, we propose a novel method called Drug Protein Analysis (DPA) to discover patterns from the drug protein interactions. Given a set of drug-protein interactions, DPA performs its tasks in several steps: (i) for each drug involved, its substructures are converted into its fingerprints; (ii) for each protein involved, its protein domains are converted into its fingerprints; (iii) a dependency measure between each drug substructure and protein domain is then computed based on the known interactions between the drugs and proteins, (iv) the dependency measures are then used to predict previously unknown drug-protein interactions. DPA is superior to other extent methods due to the following advantages: first, it is able to perform its tasks effectively without any 3D information about drug and protein structures, which are required in other methods; second, it makes use of molecular fingerprints, which are information-rich and fast to compute. From a set of known drug-targets interaction data including enzymes, ion channels and protein-coupled receptors, DPA extracts a set of chemical substructures, which are shared by drugs and can be bind to a set of protein domains. Experiment results show that it is useful for predicting new drug-protein interaction as well as protein-ligand interactions. It can also be used to tackle problems such as ligand specificity thereby facilitating the drug discovery process. To show the capability of DPA, a novel Evolutionary Algorithm in Drug molecular Design (EvoDD) is proposed. EvoDD is an Evolutionary Algorithm (EA), which can determine a molecular structure with certain desirable properties. It possesses the following characteristics. (i) it encodes molecular designs in graphs and therefore does not require post-processing of a string- or tree-based genetic structures into molecular design, (ii) it uses a novel crossover operator that can facilitate the exchange of characteristics between two molecular graphs and does not require chemistry rules known in advanced; (iii) it uses a set of novel mutation operators to facilitate the introduction of minor variations in each molecular graph in either the labeling or the structure, (iv) it uses atomics-based and fragment-based approaches to handle different size of molecule, and (v) the value of the fitness function it uses is made to depend on the property descriptors of the design encoded in a molecular graph. The patterns discovered from DPA can be used as a descriptor for the fitness function for EvoDD to guide the EA to converge. EvoDD has been tested with different data sets and the results are very promising.