Vitalii Aleksandr Stebliankin
Vitalii Aleksandr Stebliankin is a Ph.D. candidate in the Bioinformatics Research Group (BioRG) at the Knight Foundation School of Computing and Information Sciences, working on deep learning and bioinformatics under the supervision of Professor Giri Narasimhan. He holds a B.Sc. in Physics from Ural Federal University in Russia. He has worked as a data engineer intern at Assurant, a data analyst intern at Deloitte, and an applied scientist intern at Amazon. During his time as a Ph.D. student, he has published 11 peer-reviewed papers in high quality venues such as RECOMB, Microbial Genomics, AIDS, Scientific Reports, Viruses, Access Microbiology, Biophysical Journal, and Oxford Bioinformatics. After graduation, he has agreed to join the Buyer Risk Prevention team at Amazon as an Applied Scientist.
The binding of proteins plays an essential role in most critical biological processes. Investigating methods to study protein binding is of great practical importance in the development of modern vaccines, drugs, and therapeutics. Existing tools have been unable to study protein binding in a cost-effective manner. Laboratory experiments are time-consuming, labor-intensive, and expensive.
This dissertation presents state-of-the-art, scalable, and interpretable deep learning solutions to investigate protein binding for the study of molecular mimicry, protein docking, and binding affinity estimation. In the first part, we developed a computational pipeline called EMoMiS that predicts cross-reactivity events induced by molecular mimicry with particular emphasis on antibody-antigen binding. The EMoMiS pipeline used sequence similarity search and structural alignment to identify similar proteins, followed by the application of a deep learning model to evaluate the cross-reactive binding between an antibody (or antigen) and a mimicking protein. The resulting molecular mimicry search pipeline, EMoMiS, can be used as a tool for pandemic preparedness. When applied to the SARS-CoV-2 Spike protein and its antibodies, the pipeline identified many examples of molecular mimicry that can explain COVID-19-related side effects. In the second part, a deep learning approach called PIsToN was developed to classify and rank protein interfaces. The PIsToN can identify viable protein complexes from a large set of candidate complexes. It introduces a novel way to extract features to represent protein interfaces as 2D multi-channel images. The PIsToN was designed as a hybrid multi-attention transformer network endowed with explainability and significantly outperformed current state-of-the-art methods to classify and rank protein docking models. In the last part, we present a deep learning model called PIKD that incorporates protein molecular dynamics to predict the binding strength of protein complexes. PIKD achieved state-of-the-art performance in predicting the binding affinity of two macromolecules. Overall, this dissertation is a significant step toward the use of deep learning to investigate the complex world of protein binding in an efficient and accurate manner.