Modeling of Immune Sequences from COVID-19 Infected Human Blood Samples with Variational Autoencoders

Immune sequences - and more specifically T cell receptors (TCR) - repertoires analysis is one of the key tasks in adaptive immunology problem space. Probabilistic models, based on explicit V(D)J recombination rules are widely used for that purpose, with applications being: determining vaccination targets, environmental and genetic factors influence on repertoires evolution, TCR binding to specific peptides and many others. Here we show that neural network based models, namely variational autoencoders (here and further on referred to as VAE), can be efficiently used to model the sequences, generate new ones and quite successfully capturing the distribution of real sequences on an example of repertoires of infected with and vaccinated against COVID-19 disease persons. We proposed different network architectures, including fully connected and convolutional layers based and made a comparing analysis of their performance, assessing latent vectors clustering and generated sequences distributions. Results of the research can be further used for analysis of latent sequences representations, e.g. using them for prediction and/or classification tasks in adaptive immunology. The proposed modeling approach can be also used to generate synthetic sequences for research cases when access to real data, or the volume, is limited.