O03_01

Construction of Flavivirus database and therapeutic antibody discovery
using machine learning algorithm

Piyatida NATSRITA *, Kenji MIZUGUCHI

Laboratory for Computational Biology, Department of Biological Sciences, Osaka University
( * E-mail: u131573g@ecs.osaka-u.ac.jp)

Flavivirus infection responsible for approximately 50 - 100 million apparent diseases and 300 million infections per year. Member of the Flavivirus group include Zika (ZIKV), Japanese encephalitis (JEV), West Nile (WNV), Yellow Fever (YFV), and dengue viruses, including 4 serotypes (DENV-1, DENV-2, DENV-3, DENV-4). Several studies hypothesized that cross-reactivity among distinct serotypes and other flaviviruses is a major problem of severe caused by antibody dependent enhancement (ADE) phenomenon. In this study, we aim to develop a novel dataset of CDR-H3-epitope sequences together with IC50 values and sequence-based ML approach to predict the potential neutralizing antibodies against Flavivirus towards the characterization and analysis of our obtained sequences in terms of neutralizing levels, cross-reactivities, and important features for broad neutralization. Firstly, we generated the dataset of CDR-H3 sequences together with epitope sequences and labeled with IC50 values toward the characterization and analysis of our obtained sequence in terms of neutralizing levels. In a total of 3,767 pairs including 1,366 high neutralizing activity (IC ≤ 10 ng/μL) and 2,400 low neutralizing activity (IC > 10 ng/μL). In the dataset, we found 541 cross-reactive antibodies, and 826 non-cross-reactive antibodies. From ML analysis, we found 20 important features including
chiral carbon and aromaticity. The larger pool of CDR-H3-epitope-IC50 data lead to empower a ML model high-throughput screening performance for sequence classification. Further effort will focus on different encoding method comparison, antibody repertoire-level Flavivirus screening and classification. The potential antibody candidates against Flavivirus will be evaluated by performing molecular docking and MD simulations of each candidate with DENV, ZIKV, YFV, WNV, and TBEV to determine the location of binding, binding affinity, and stability. This finding might be useful for further development of therapeutic antibodies in the future.