Title:Artificial neural networks for immunological recognition

Abstract:Adaptive immunity is analogous to machine learning as the immune system learns about dangerous pathogens and discriminates them from safe self-molecules. Each immune cell has a unique receptor for specific immunological recognition. The binding affinity between the receptors and antigenic peptides mainly determines immunological recognition followed by subsequent responses. Is the integrated binding affinity sufficient to probe the digital information of peptide sequences? To answer this fundamental question, we examined whether the affinity-based discrimination of peptide sequences is learnable and generalizable by artificial neural networks (ANNs) that process the digital information of receptors and peptides. We made use of large scale high-throughput data of T-cell receptors (TCRs) and open-source peptides amino acid sequences as the input layer data, and computationally examined the success and failure of immune recognition based on the pairwise binding energy of their amino acid sequences as output layer data. Machine learning successfully captured the relevant information of molecular interactions and discriminated between strong and weak affinity pairs. This suggests the potential applicability of artificial neural networks to predict the immune responsiveness of certain antigenic peptides once their sequence information is provided. In addition, we used two methods to describe amino acids: with and without considering the amino acid strength. The sequences with heterogeneous strengths of amino acids diminished TCR "reading" of the exact information of each amino acid site on the peptides for the immune response. This may be why natural TCRs have moderate amino acid compositions after thymic selection.