Association of a peptide with Major Histocompatibility Complex (MHC) is crucial for adaptive immune system in vertebrates. Upon binding to an MHC molecule, the peptide is presented to T-cells, which triggers an immune response, if the peptide is recognized as foreign. In recent years personalized treatment approaches such as cancer immunotherapy based on the knowledge of MHC selectivity of a particular patient started to emerge. Detection of peptides bound to MHC on the tumor cell surface, containing cancer driver mutations (neoantigens) is essential for efficiency of the treatment, and therefore, it is important to understand mechanisms, which drive MHC-peptide complex formation. Over the years many complexes have been crystallized and several approaches were developed for bound peptide structure prediction, being, however, not suitable for a large scale structural analysis of multiple peptides (~103) due to large execution times, hence faster and more efficient modeling techniques are required. Machine Learning approaches have been demonstrating promising results in protein structure and ligand binding prediction. Here we present an ultra-fast peptide-MHC docking method based on 3D Convolutional Neural Network (CNN) scoring and constrained inverse kinematics peptide sampling. Our algorithm is suitable for docking of multiple peptide-MHC complexes and can provide insights for selectivity and preferential binding of different MHC alleles and facilitate structure based binding analysis.