Our client, a leading biopharmaceutical company, aspired to design antibodies with enhanced stability, binding affinity, selectivity, and manufacturability. Traditional methods require synthesizing and testing numerous variants, which are costly and slow. Arrayo was tasked with developing predictive computational models that pre-screen antibody sequence designs for desired properties leveraging numeric embeddings from large language models (LLMs) and structure-derived descriptors.  Historically, creating accurate models has been challenging due to limited data availability.

Delivery

The approach involves converting antibody sequences into high-dimensional embeddings. These embeddings encode complex structural and functional information. From these data representations, machine learning models are trained to predict desired antibody properties. Even with current data limitations, these models enable preliminary screening, highlighting candidates that are most likely to possess optimal properties, before committing resources to synthesis and lab testing.

Value

This computational strategy significantly accelerates the early-stage antibody design process by reducing the number of candidates requiring synthesis, thus saving time and costs. As more data to build models become available, the models are expected to become more predictive, ultimately enabling fully automated antibody designs. This will facilitate faster development of optimized antibodies with improved therapeutic performance and a streamlined path from discovery to clinical application.