LLMs Enable Large Scale Design of Nanobodies

In the coming decades, AI will be one major contributor to the offense-defense balance. If we want to steer the future towards a high-robustness state, it is important to develop in advance a portfolio of defensive technologies that directly benefit from economic growth and technological progress. What would a general defensive technology against biological threats look like? As a motivating example, we investigate whether present-day LLMs are capable of autonomous nanobody design. Using computational tools [2] as a harness for protein design, we enable a generalist LLM to design and test nanobodies. Examination of performance trends across LLMs and reasoning efforts reveals performance improvements in the designed antibodies, demonstrating a scaling trend relevant for safety. We argue that theoretical and empirical facts about LLMs make them exceptionally well suited for this type of hyperparameter search task [8][9]. Moreover, this type of evaluation enables both a new paradigm for capabilities assessment and potential for scaling defensive technologies.