Creating Feature Representations of Antibody–Antigen Complexes for Fast Binding Prediction with Machine Learning

Daniel Faissol | 19-FS-059

Overview

antibody—antigen interactions underlie immune system function. Numerous computational tools exist that aim to predict important quantities associated with antibody—antigen interactions, such as binding energy. Efficient and accurate prediction of such quantities would greatly accelerate vaccine (antigen) and therapeutic antibody design. Current methods and tools, however, rely on varying degrees of computationally expensive steps that limit their use on problems requiring many predictions. Moreover, no single method produces consistently accurate estimates.

In this project, we determined the feasibility of creating feature representations that concisely describe interacting antibody—antigen pairs as rapidly computable numerical vectors, given the antibody—antigen co-complex. These representations enable a statistical machine learning model that jointly models the results of disparate, existing computational tools available, enhancing their combined predictive performance. We developed a feature representation consisting of 86 descriptors of the antigen-antibody interface, which was reduced into a lower-dimensional representation using a neural network dimensionality reduction step. We demonstrated that this feature representation serves the intended purpose of rapidly describing antigen-antibody pairs for the task of binding strength prediction. While this study has proven that such a feature representation is feasible, additional research will likely improve its ability to support other tasks, including binding strength prediction.

Impact on Mission

The capability demonstrated in this feasibility study enhances current and future Lawrence Livermore National Laboratory efforts that depend on modeling these specific interactions. Determinations made will also contribute significantly to the broader antigen-antibody design community. Moreover, our feature representations might be applicable to protein-protein interactions, of which antibody—antigen interactions are a special case. This study supports the Laboratory Director's initiative in cognitive simulation and Livermore's core capabilities in high-performance computing, simulation, and data science. The work also supports Livermore's mission research challenge in chemical and biological countermeasures by providing a feature representation that can serve as the basis for the rapid development of vaccines and antibody-based therapeutics against biothreat agents. Moreover, Livermore has enhanced its existing molecular dynamics and bioinformatics portfolio, building an enabling capability for future antibody or antigen design efforts and positioning the Laboratory as a thought leader in the use of computing in response to biothreats.