We utilize Mosaic for nanobody design. Mosaic is a method that hallucinates structures using a custom loss on prediction models, and demonstrated high success for de novo binder design during the Nipah competition. We tailor this framework specifically for nanobody generation by integrating AbLang and AbMPNN into the loss function. Furthermore, we implemented a targeted loss constraint designed to reward CDR-specific interactions while penalizing framework-mediated binding. This ensures the resulting binders adhere to biologically relevant interfaces. From an initial library of 1,000 generated candidates, the final selections were determined by ranking the top performers according to their AlphaFold3 ipTM scores.

For each designed nanobody, we also create a "grafted" version based on a known nanobody with a similar sequence, where we keep the designed CDRs and replace the framework.

Building on the Mosaic framework's proven success in the Nipah competition, our high-confidence structural predictions suggest that this methodology may be as robust for specialized nanobody design as it is for general protein binders.

If only taking parts of these submissions, we recommend the top 10 models that got the highest confidence score by AF3:

(non-grafted) a69b10, a88b6, a51b7, a30b4, a54b0

(grafted) a69b10_grafted1, a88b6_grafted2, a20b1_grafted2, a51b7_grafted3, a52b15_grafted2