BIDpred: unraveling B cell Immunodominance hierarchical pattern using statistical feature discovery and deep learning prediction

Sungjin ChoiDongsup Kim^*

• Korea Advanced Institute of Science and Technology, Daejeon, Republic of Korea

Knowledge of B cell immunodominance is important for designing vaccines that may elicit effective immune responses. However, the prevalence and characteristics of B cell immunodominance remain poorly understood. In this study, we introduced an immunodominance score through novel data processing methods and identified statistically significant characteristics of B cell immunodominance at the residue and patch levels. Based on these findings, we developed BIDpred, a B cell ImmunoDominance predictor, that learns newly discovered features by leveraging protein language model embeddings and graph attention network to predict the immunodominance scores. BIDpred demonstrates superior performance in predicting immunodominance scores compared to existing methods while maintaining competitive accuracy with state-of-the-art methods for conventional B cell epitope prediction. To the best of our knowledge, this is the first study to systematically analyze and predict B cell immunodominance patterns, marking a significant advancement in vaccine -2 -design research.Immunoinformatics, B cell immunodominance, vaccine design, deep learning, protein language model X-ray crystallography data were collected from SAbDab (22) as of Mar.19, 2024. The data curation process is illustrated in Figure 1. For quality filtering, we used a filtering cutoff of resolution 3.0Å, R-factor 0.25, antigen size with at least 50 amino acids, and maximum antibody sequence identity of 99% (23). After quality filtering, we extracted the antigen sequences and performed sequence clustering using the mmseq2 easy-cluster command with a minimum sequence identity threshold of 0.70; all other options were set to default. (24) (Figure 1A). Clusters with at least 4 elements were used. One cluster can be viewed as a group of the