Using a combination of RF diffusion and AF_unmasked for de novo antibody and small binder design
Abstract
Recent advances in in-silico protein structure prediction and design, recognized with the 2024 Nobel Prize in Chemistry, enable the creation of small proteins that bind specific targets. This is particularly relevant for designing de novo monoclonal antibodies (mAbs) against defined epitopes of antigens such as viral or cellular proteins. mAbs are widely used to treat cancer and autoimmune diseases, but their discovery and isolation remain laborious. Generative machine-learning models like RF diffusion now allow the design of novel small binders like helical domains (Torres et al., 2023) and even antibody Fab domains targeting selected epitopes (Bennett et al., 2024). In our previous projects we could already show that mAbs designs derived from RFdiffusion can be validated using AF_unmasked by limiting the modelling to the parts relevant for the binding interface which should be computationally less expensive. In our previous tests we designed mAbs targeting epitopes on the corona virus spike protein and on auto-immunogenic human interferon (see activity reports for Berzelius 2025-55 and 2025-259). A main challenge remains efficient filtering of the outputs in order to reduce the number of candidate constructs to be tested in the wet-lab. Continuing this study, we will focus on exploring additional tools for assessing the quality of the designed interface. We will design small all-helical binders against epitopes on neurofibromin, an important tumor suppressor, which is presumably a less complex problem compared to designing the CDRs of mAbs. Promising mAb and small binder designs will be biochemically and structurally analyzed employing Bio Layer Interferometry (BLI) assays and cryo-EM. Finally, we will test the performance of the recently published RFantibody pipeline by the David Baker group by designing mAbs to a specific epitope on human transferrin receptor in parallel and in comparison to conventional wet-lab phage display methods. This will yield important feedback on how practicable current bioinformatical design pipelines really are when facing problems that are currently tackled by biochemical screening methods.
