A COMPUTATIONAL BASED DEEP LEARNING APPROACH TO ASSESS POST-VACCINE MYOCARDITIS RISK VIA SPIKE PROTEIN CARDIAC ANTIGEN CROSS-REACTIVITY

Abstract

Post-vaccine myocarditis particularly following administration of mRNA-based COVID-19 vaccines has emerged as a clinically significant yet rare adverse event. The immunopathological mechanisms driving this condition remain largely elusive. One prevailing hypothesis centers on molecular mimicry, wherein structural similarities between the SARS-CoV-2 spike protein and endogenous cardiac antigens may trigger autoimmune-mediated myocardial inflammation. In this study, we present a robust deep learning-based framework to model and predict potential antigenic cross-reactivity between viral spike proteins and human cardiac peptides. Utilizing transformer-based protein embeddings (ProtBERT), we generate high-dimensional feature representations of 9-mer peptide sequences from both the SARS-CoV-2 spike protein and key cardiac antigens. These embeddings are input into a Siamese neural network architecture for pairwise similarity evaluation, followed by a residual attention-enhanced convolutional classifier to assess the likelihood of immunogenic mimicry. The proposed model demonstrated strong performance across five-fold cross-validation, achieving an accuracy of 93.1% and an AUC-ROC of 0.96. Several high-probability mimic peptides were identified, offering candidates for future in vitro and in vivo immunological validation. This framework provides a scalable, interpretable, and data-driven methodology to support vaccine safety surveillance and deepen our understanding of autoimmune responses associated with molecular mimicry.