De Novo Binder Design for a Single-Residue Modification: Asparagine Deamidation

Engineering proteins capable of discriminating between single-residue modifications, such as asparagine deamidation, is a frontier challenge in structural biology and protein design. In this work, we leverage generative machine learning tools alongside a custom $\textit{in vivo}$ split-spGFP biosensor to both design and validate protein binders sensitive to a $\sim$1 Da electrostatic shift. To target the deamidated state, our computational pipeline used a post-hoc $\Delta\Delta$G evaluation to select for an aspartic acid target. Surprisingly, empirical screening revealed a binder, pSD009, that preferred the unmodified asparagine, demonstrating inverted specificity to the $\textit{in silico}$ prediction. To investigate this discrepancy, we conducted a cross-architecture analysis, finding that while both Rosetta $\Delta\Delta$G and Chai-1 pAE\_Interaction failed to anticipate this inversion, AlphaFold 3 predicted the correct $\textit{in vivo}$ specificity using that same confidence metric. This selectivity towards the native state suggests a novel biological application: leveraging custom binders to filter out non-deamidated species, thereby differentiating stable proteins from their degraded variants. In addition, this work demonstrates the emerging power of SoTA generative models to achieve single-residue precision, highlighting a new strategy for highly sensitive binder design.