Researchers from a leading biotechnology corporation developed the Lab-in-the-loop (LitL) approach to integrate machine learning (ML) tools into a closed-loop, design-test-learn system for therapeutic antibody design. Therapeutic antibody design presents challenges due to factors such as optimizing one pharmacologically relevant property, such as affinity, can compromise another, like solubility, requiring an iterative search through a vast sequence space.

While ML is promising, models trained on limited datasets often optimize for single properties and ignore therapeutic constraints. LitL overcomes this by using generative ML and property prediction, followed by experimental validation and adaptive model retraining. Twist's high-accuracy synthesis of linear DNA was essential for encoding the custom Complementarity-Determining Regions (CDRs), safeguarding the integrity of the system by ensuring what the ML model designed was accurately represented in the wet lab.

Over four rounds, the LitL system generated over 1,800 unique variants and identified antibodies with binding affinity improvements ranging from 3X to 100X against four antigen targets.