A method for constructing interpretable hidden Markov models for the task of identifying binding cores in sequences

Solving the problem of predicting the immune response against foreign protein sequence fragments processed by cells is one of the major milestones on the road to the personalized cancer vaccine development. The selection of peptides participating in the immune response is a complex multi-stage process of filtering initial sequences to present their fragments on the cell surface. The most studied task regarding this filtering nowadays is the prediction of the binding probability of peptides to major histocompatibility complex molecules. Modern methods for predicting this stage are usually based on algorithms using artificial neural networks, which make it impossible to interpret the result predictions of such models. One of the methods to overcome this limitation is the use of interpretable hidden Markov models. In this work, an analysis of the binding prediction task is performed. As a result, a method for constructing interpretable models that consider domain-specific constraints and requirements is proposed. A method for the constriction, training and interpretation of hidden Markov models was proposed for each class of molecules. The construction and training are based on maintaining the model architecture capable of extracting and visualizing the binding core of the peptide. Interpretation is possible through the analysis of the model graph. The proposed method is tested in the task of training a model that not only enables prediction but also facilitates determining the position of the peptide binding core and the distribution of amino acids within the core. Prediction models were trained for two types of molecules using binding data. The distributions of amino acids in the binding core match the state distributions of the model. Sequence patterns of such regions extracted using the trained models for two sets of peptide data correspond to patterns from public databases, confirming the successful validation of the method. Interpretable models provide a better description of the problem domain and help to draw a conclusion about peptide characteristics based on information extracted from the model. This information will allow researchers to better understand other steps of peptide processing involved in the immune response. For example, one can study relationships between these steps or perform a transfer of knowledge from models trained for one step to others. Using this knowledge will allow the training of the models under conditions of limited training data.