Energy-based generative models for monoclonal antibodies

TL;DR

This paper introduces energy-based generative models to optimize monoclonal antibodies by balancing properties like solubility, humanness, and affinity, aiming to streamline drug discovery.

Contribution

It presents a novel energy-based generative modeling approach for antibody optimization, addressing multi-property tradeoffs and generating candidates on an optimal Pareto front.

Findings

Successfully generated antibody candidates balancing multiple properties.

Identified tradeoffs between affinity, solubility, and humanness.

Demonstrated the model's ability to produce Pareto-efficient solutions.

Abstract

Since the approval of the first antibody drug in 1986, a total of 162 antibodies have been approved for a wide range of therapeutic areas, including cancer, autoimmune, infectious, or cardiovascular diseases. Despite advances in biotechnology that accelerated the development of antibody drugs, the drug discovery process for this modality remains lengthy and costly, requiring multiple rounds of optimizations before a drug candidate can progress to preclinical and clinical trials. This multi-optimization problem involves increasing the affinity of the antibody to the target antigen while refining additional biophysical properties that are essential to drug development such as solubility, thermostability or aggregation propensity. Additionally, antibodies that resemble natural human antibodies are particularly desirable, as they are likely to offer improved profiles in terms of safety, efficacy, and reduced immunogenicity, further supporting their therapeutic potential. In this article, we explore the use of energy-based generative models to optimize a candidate monoclonal antibody. We identify tradeoffs when optimizing for multiple properties, concentrating on solubility, humanness and affinity and use the generative model we develop to generate candidate antibodies that lie on an optimal Pareto front that satisfies these constraints.