We want to build technology that protects us from the biggest threats we will face as a society in the coming decades. To achieve this with high robustness it is important to maintain a balanced portfolio of defensive technologies that both leverage scaling factors like compute, power, and industrial capacity, while maintaining uncorrelated components that function independently of those general improvements.
Resilience against threats of biological origin is especially relevant due to direct impact on public health. This has been made apparent with the recent concerns about AI-enabled bio-threats [6][7]. But, what if we could use the power of AI for the benefit of defense?
To investigate this question, we focus on potential application of AI for Bio Defense, the automated design of pharmacological compounds. Concretely we focus on a type of protein: the nanobodies. Nanobodies (also known as Single Domain Antibodies) are a class of proteins with the ability to bind to specific antigens, much like antibodies, but substantially smaller and simpler.
Nanobodies offer potential advantages that are highly relevant for situations in which a fast reaction is necessary (like a natural or man-made pandemic). For example, nanobodies have very short development times and don't require specialized equipment to be developed, this is specially relevant when dealing with fastly evolving strains. They also have many drawbacks, which would make mass deployment very challenging.
There is growing interest in using nanobodies for pharmaceutical applications [5][10][11].
In recent years, the tools available to protein engineers have been dramatically expanded by ML-based biological models [12][13]. This is also true for nanobody design [2]. However despite the ease of use of these tools, many aspects of protein design are still poorly understood and rely on diffuse and tacit knowledge that only researchers have.
This prompts our central question: "Can we leverage the biological knowledge present in state-of-the-art LLMs to automate the design of nanobodies?"
To answer this question we evaluate several LLMs on the autonomous design of nanobodies for Nipah virus using RFantibody as a harness. Several in-silico metrics [4] will be used to assess the quality of the designed nanobodies.