Learning the interaction rules of antibody-antigen binding

Antibody-antigen binding is the basis of two fundamental biotherapeutic pillars: monoclonal antibodies (1) and vaccines (2). To accelerate therapeutics discovery, we need to perform antibody (Ab) and antigen (Ag) design in silico.

Specifically, we need to address a fundamental immuno-biotechnological challenge: understanding the interaction rules that predict Ab-Ag binding.

Solving this challenge demands the convergence of biotechnology, computational structural biology, and machine learning (ML).

My lab is one of the few worldwide to have this transdisciplinary expertise.Research problem: Currently, the predictive performance of Ab-Ag binding is poor, and an understanding of the underlying rules of Ab-Ag binding is mostly absent.

We previously showed that both unprecedentedly large datasets (>10^5 Ab-Ag sequence pairs) and extensive structural information on the Ab-Ag binding interface (paratope, epitope) are needed to increase prediction accuracy and recover binding rules.

Targeted breakthrough: To address the lack of large-scale Ab-Ag sequence and structural data, we will develop a method for high-throughput screening of >10^3 Ab paratope-mutated variants binding to >10^3 of Ag epitope-mutated variants, generating sequence data of Ab-Ag binding pairs at an unprecedented scale (>10^6 sequence Ab-Ag pairs).

Structural information of the entirety of the sequence-based Ab-Ag binding data will be generated by building and innovating on recent breakthroughs in computational structural biology.

To derive Ab-Ag interaction rules from the generated data, we will develop ML techniques for Ab-Ag binding prediction and rule recovery. We will demonstrate experimentally that we have begun to understand Ab-Ag interaction rules.

Impact: The proposed research generates the exact data necessary to recover the rules of Ab-Ag binding and provides a first groundbreaking insight into those rules, moving us closer to in silico on-demand antibody and vaccine design.