Engineering antigen-dependent stability in nanobodies: A bioinformatics tool for tuning intracellular nanobody stability

Nanobodies, derived from camelid antibodies, are highly prized reagents due to their ability to bind their antigen with high affinity, specificity and due to their ease of genetic expression inside cells. Conditionally stable nanobodies, nanobody variants containing mutations in their sequence, are stable only upon antigen binding, making them ideal biosensors. However, the generalizability of conditionally stable nanobodies depends on contextual interactions between nanobody sequences as well as with their fusion protein sequences. To account for contextual effects and thereby creating an individualized way of engineering nanobody for desired stability modulations, we developed a novel bioinformatics tool based on large language model (LLM) that takes an input nanobody sequence and makes predictions about its intracellular stability. The predictions integrate information about nanobody binding to avoid disruptive mutations, guided by a dataset of 389 nanobody-target interfaces. The LLM was first pre-trained on 1.4 million camelid antibody sequences, then fine-tuned to perform binary classification of intracellular stability, producing an F1 score of 0.8. Generative models were developed using a form of Masked Language Modeling (MLM). Models' predictions were validated using reporter assays and western blot in human cell culture. The tool integrates concepts from LLM, nanobody bidirectional stability modulation, nanobody interface classifications, and web integration of modalities to create a tool that can facilitate biomedical and cancer endeavors in the creation of biosensors against any target intracellular proteins for which nanobodies can be selected. This tool has the potential to greatly speed up biosensor engineering, promoting diverse applications in biomedicine and cancer research.