Hemoglobinopathies, such as sickle-cell disease (SCD) and transfusion-dependent β-thalassemia (TDT), affect 360,000 individuals yearly and is a significant economic burden. Recent research identify CHD4 as a promising target for SCD and TDT treatment by reactivating production of fetal γ-globin. However, the lack of chemical CHD4 modulators hinders drug development efforts.
Recent advancements in deep-learning diffusion algorithms have demonstrated remarkable promise in the rapid design of entirely new proteins, including those with potential binding capabilities. When given the structure of a target protein, RoseTTAFold diffusion (RFdiffusion) can produce candidate binding proteins of any size, whilst boasting an experimentally verified success rate of approximately 20%.
In our initial investigations, we generated 20 candidates (binders) ranging from 50–100 residues for the CHD4 HMG domain. The top-ranked binders exhibit favorable geometry, and independent AlphaFold2 multimer analyses predict them to bind to the HMG domain as designed. By using NMR spectroscopy, isothermal titration calorimetry, and real-time fluorescence assays, we have identified several binders that can bind the HMG domain with high affinity and inhibit the remodeling activity of CHD4. Our work further demonstrates that RFdiffusion is a powerful deep-learning method for the de novo design of proteins.