RFantibody

1 Introduction

RFantibody¹ utilizes RFdiffusion² and RoseTTAFold2³ to fine-tune the structures of natural antibodies, specifically for antibody structure design and prediction, supporting the design of single-domain antibodies (VHH). It is capable of designing antibody structures with high binding affinity based on specified antigen epitopes. The design process is as follows:

• Given the antibody framework structure and the target antigen structure, binding hotspots can be specified.
• Using the diffusion model technique of RFdiffusion, the antibody structure is progressively "denoised" and optimized to design CDR loops that bind to the epitopes of the target antigen.
• CDR loop sequences are designed using ProteinMPNN4, achieving an amino acid recovery rate of 52.4%.
• The structure of the antibody-antigen complex is predicted and screened using the fine-tuned RoseTTAFold2.
  
  
  

图1. RFantibody图

2 Parameters

• Target PDB: Target antigen structure.
• Antibody PDB: Antibody framework structure.
• Antibody Heavy Chain: Specify the heavy chain framework of the antibody.
• Antibody Light Chain: Specify the light chain framework of the antibody.
• PPI Hotspot: Specify the binding hotspots on the target antigen.
• Number of Design: Number of designs to be generated.
• Design Loops: Allowed length range for CDRs (Chothia numbering)..
  • HCDR3: Default 5-13.
  • HCDR2: Default 6.
  • HCDR1: Default 7.
  • LCDR3: Default 9-11.
  • LCDR2: Default 7.
  • LCDR1: Default 8-13.

3 Results Explanation

• Generate a 3D structure file of the antibody, including the heavy chain (H), light chain (L), and target antigen chain (T).
• Scoring details:
  • interaction_pae: The average predicted aligned error (PAE) of the interactions. PAE measures the deviation of the relative positions between different parts of the protein structure predicted by the model compared to the actual structure. A smaller value indicates that the relative positions of different parts predicted by the model are closer to the true positions, which means the predicted interactions between different parts of the protein structure are more accurate.
  • pae: The expected value of the predicted PAE, which also reflects the model's accuracy in predicting the relative positions of residue pairs in the protein structure. A smaller value indicates that the predicted relative positions of residue pairs are closer to the true values.
    pred_lddt: The predicted local distance difference test (lDDT) score. lDDT is used to evaluate the accuracy of local structure predictions for each residue in the protein structure, measuring the differences in local distances between the predicted and actual structures. A higher score (typically ranging from 0 to 1) indicates more accurate local structure predictions.
  • target_aligned_antibody_rmsd: The root mean square deviation (RMSD) of the Cα atoms of the antibody after aligning with the target antigen structure. RMSD is a measure of the differences in atomic positions between two structures, calculated as the RMSD of the distances between Cα atoms. A smaller RMSD value indicates that the atomic positions of the predicted structure are closer to those of the reference structure, meaning the predicted structure is more accurate.
    target_aligned_cdr_rmsd: The RMSD of the Cα atoms of the antibody complementarity-determining regions (CDRs) after aligning with the target antigen.
  • framework_aligned_antibody_rmsd: The RMSD of the Cα atoms of the antibody after aligning the antibody framework.
  • framework_aligned_cdr_rmsd: The RMSD of the Cα atoms of the antibody CDRs after aligning the antibody framework.
  • framework_aligned_H1_rmsd: The RMSD of the Cα atoms of the antibody heavy chain CDR1 after aligning the antibody framework.
  • framework_aligned_H2_rmsd: The RMSD of the Cα atoms of the antibody heavy chain CDR2 after aligning the antibody framework.
  • framework_aligned_H3_rmsd: The RMSD of the Cα atoms of the antibody heavy chain CDR3 after aligning the antibody framework.
  • framework_aligned_L1_rmsd: The RMSD of the Cα atoms of the antibody light chain CDR1 after aligning the antibody framework.
  • framework_aligned_L2_rmsd: The RMSD of the Cα atoms of the antibody light chain CDR2 after aligning the antibody framework.
  • framework_aligned_L3_rmsd: The RMSD of the Cα atoms of the antibody light chain CDR3 after aligning the antibody framework.

4 Reference

[1] Bennett, N. R., Watson, J. L., Ragotte, R. J., Borst, A. J., See, D. L., Weidle, C., ... & Baker, D. (2024). Atomically accurate de novo design of antibodies with RFdiffusion. bioRxiv. https://doi.org/10.1101/2024.03.14.585103
[2] Watson, J.L., Juergens, D., Bennett, N.R. et al. De novo design of protein structure and function with RFdiffusion. Nature 620, 1089–1100 (2023). https://doi.org/10.1038/s41586-023-06415-8
[3] J. Baek, M., Anishchenko, I., Humphreys, I. R., Cong, Q., Baker, D., & DiMaio, F. (2023). Efficient and accurate prediction of protein structure using RoseTTAFold2. bioRxiv. https://doi.org/10.1101/2023.05.24.542179
[4] J. Dauparas et al. ,Robust deep learning–based protein sequence design using ProteinMPNN.Science378,49-56(2022). https://doi.org/10.1126/science.add2187