It's not all about CDRs: the surprising insights of ML-based antibody optimization

In this presentation, we will discuss unexpected findings from our work on ML-guided, lab-in-the-loop protein design targeting affinity, expression, and stability. Notably, we demonstrate that it is possible to achieve improved binding affinity without altering the complementarity-determining regions (CDRs). On the other hand, we show that even substantial modifications to CDRs, including the highly variable CDR3, can be applied while retaining strong binding properties, provided the changes are guided by a custom-conditioned protein language model (PLM). This opens an exciting path for active-learning-driven antibody optimization starting from an immunization campaign or even a single lead.