Structure based prediction of MHC II binding peptides

Responses of the immune system to combat pathogens are complex processes. A central part of the defense is initiated by antigen presenting cells, in which proteins derived from the pathogens are digested to short peptides. A few of those peptides bind to MHC-II proteins. Consequently, those MHC-II peptide complexes are presented on the cell surface in order to interact with T-cells and trigger canonical immune responses. These interactions and responses are also known in allergies and several severe autoimmune diseases. In order to develop therapies it is of increasing interest to specify which pathogen derived peptides will actually bind to MHC-II molecules. The experimental assessment which fragments bind is time consuming and thus costly. As an alternative, computational approaches have been developed over the last two decades enabling the prediction of MHC-II binding peptides. Approaches are either protein sequence or structure-based. Modern sequence-based methods utilize machine learning methods to extract sequence pattern from experimentally proven binders for a certain MHC-II variant (allotype). As experimental data for binders are limited in particular in terms of the investigated MHC-II allotypes, the power of such methods is also limited. Structure based methods work with known 3D structures of MHC-II molecules and employ physical principles and simulation methods to predict binders. As these methods do not require experimental binding data, they are considered universal. However, their power is still restricted due to the number of experimentally determined MHC II structures and in respect to prediction accuracy. It is therefore our aim is to design and test a new structure-based approach firstly to overcome these limitations. For this purpose we will employ optimized statistical scoring functions in combination with intrinsic peptide features in order to maximize the prediction accuracy. The final scoring scheme will be applied to antigen sequences by a fast threading approach. In parallel, we will create high quality structural models for all MHC-II allotypes thus obtaining a complete array of MHC-II structures. Working along this line the new method will be also comprehensively applicable. The emerging methods will be made publicly available in two versions: (i) as standalone program for high throughput predictions and (ii) as easy-to-use web-server for small scale experiments. In close collaboration with molecular research groups at the University of Salzburg specialized on allergy, we intend to employ the method to design hypo-allergens. Applications of hypo-allergens are considered greatly promising approaches to specifically treat allergy.

The immune system recognizes and processes foreign antigen proteins in different ways. We focus on MHC II immune molecules, which bind fragments of antigen proteins. Once loaded with an antigen peptide, the MHC II molecules are presented on the surface of certain cells which triggers further natural immune responses. This process is also observed in the context of several diseases such as cancer or allergy and is thus of high relevance in medicine. Essential for triggering an immune response is, that a distinct MHC II molecule can only recognize specific antigen peptides. Different individuals have varying MHC II genes and thus MHC II proteins. The range of distinct MHC II variants is large. Therefore, a computational method, that is able to accurately predict which antigen peptide is bound by which MHC II variant can be quite supportive in the detection of possible (misrouted) immune responses to certain antigens. In case of a disease this may guide specific treatments in the context of personalized medicine. Our computational method, named MHCII3D, is based on the known three-dimensional structure of a few MHC II molecules. Using those as templates we can calculate 3D models also for other MHC II variants. The 3D structural model of a certain variant and the application of statistical scoring function allows us to predict, if a certain antigen peptide can bind or not. In contrast to other predictions methods, the knowledge of experimentally determined MHC II binding data is not required to train our method. Therefore, MHCII3D can be applied to virtually any MHC II variant. This advantage had to be bought with a somewhat lower prediction accuracy compared to the leading method, but we are confident to be able to close this gap soon. In addition, we could show that MHCII3D complements other prediction methods and therefore improves the joint prediction success. MHCII3D is provided to the public as a command line program and as a web service. We are gradually adding new models to MHCII3D, which constantly increases the scope of application.