Rational design of protein inhibitors using molecular modelling and multivariate analysis
MetadataVis full innførsel
- Institutt for kjemi 
Cancer cells utilise signalling cascades involving protein kinases for their growth and survival. Hence, design of protein inhibitors that block the function of these signalling proteins is interesting for the development of new cancer therapies. This can be achieved through rational drug design. The purpose of this work was to develop drug design methods that can aid the discovery of selective drugs and to utilise these methods to design drugs that block the function of proteins involved in cancer development. This work has focused on development of drug design methods that can be used with protein structure models made by homology modelling, since this will significantly increase the number of protein targets for which the methods can be used. In this context, a review about the use of homology-based modelling in rational drug design was published. The design part has focused on design of selective inhibitors of protein kinases, in particular Tyrosine kinase 2 (Tyk2), a member of the Janus kinase (Jak) family of protein kinases. The interactions between the receptor kinase fibroblast growth factor receptor 1 (FGFR1) and a known inhibitor have also been studied, and several improvements of the inhibitor have been suggested, based on results from computational sensitivity analysis. Both in the case of Tyk2 and FGFR1, focus has been on inhibiting the binding of adenosine triphosphate (ATP) to the tyrosine kinase domain of the proteins. The interactions between E-selectin and a set of carbohydrates and peptide ligands were also studied with computational docking. The results from this study provide insight into some of the limitations of docking methods. In order to analyse the relationship between the target-template similarity and the accuracy of the obtained homology model, a large number of homology models for protein kinase structures were generated, and the accuracy of the homology models was evaluated by comparison to available experimental structures of the targets. Based on the obtained data, a new method for prediction of homology model accuracy with multivariate regression was developed, that predicts the model accuracy directly from the amino acid sequence alignment. This method can be used to assure that the optimal templates are chosen, and for identification of regions of the protein structure that are difficult to model, as well as errors in the alignment of the proteins. Here, this method has been applied to the protein kinase family, but the same approach can be used for other protein families. A new method for analysis of protein binding site properties, called Protein Alpha Shape Similarity Analysis (PASSA), and a new gaussian-based docking method suitable for use with homology modelled protein structures have been developed. Both methods use gaussian functions to represent atomic properties. This smooth representation makes them relatively robust against small structural errors. PASSA has been shown to be a useful method for identification of regions in a protein binding site that can be utilised to achieve selective binding of ligands to the protein. Interaction sites identified by PASSA to be important for selectivity have been shown to correspond to functional groups of known, selective inhibitors. The gaussian-based docking method developed here is relatively fast, and well suited for virtual screening, where the purpose is to seek out a set of promising drug candidates from a large amount of ligand structures. However, the accuracy of our docking method cannot be compared to that of other methods that use fewer approximations. In contrast to many other docking methods, our docking method predicts hydrophobic interactions better than hydrophilic interactions. PASSA was used to suggest functional groups for a selective inhibitor of Tyk2. The results from this study were used further in a screening of the database of the National Cancer Institute (NCI) for possible Tyk2 inhibitors. The proposed functional groups were also combined into drug candidates by de novo ligand design. The gaussian-based docking method developed here was applied to rank the drug candidate molecules resulting from the database screening and de novo ligand design according to binding to Tyk2. The selectivity of the compounds was tested by computational docking in seven other protein kinase structures. The results from the docking of the compounds from the NCI database were compared to the results obtained using another docking method, MOE-Dock. The two docking methods ranked the structures differently, but produced the same conclusion, namely that none of the compounds in the NCI database can inhibit Tyk2 selectively. One compound was found to inhibit Tyk2 and insulin receptor tyrosine kinase selectively, and five of the drug candidates from the de novo ligand design seem promising as selective Tyk2 inhibitors. These results have to be verified experimentally, of course. PASSA has also been used to model selectivity within the protein kinase family. In this way, the PASSA method may be used quantitatively to predict activities for a number of ligands within a set of closely related protein targets. This makes PASSA a promising method in screening for side effects. This method also allows for effective visualisation of the molecular basis for selectivity. The results presented here indicate that methods utilising gaussian functions to describe molecular properties have many applications in structure-based drug design, and will be useful supplements to other methods. These methods seem especially useful in the initial stages of a drug design process, when computational efficiency and robustness are most important.