Part:BBa_K5083003

DPP-4 Protein Inhibitor

GLP-1 is rapidly degraded by DPP-4 in the gut, resulting in a half-life (t1/2) of only a few minutes. To address this, our team chose to use artificial intelligence to design a DPP-4 protein inhibitor to enhance the half-life of GLP-1 in the body, thereby improving its functionality. We utilized deep learning models such as RFdiffusion, ProteinMPNN, and ESMfold for protein design. Additionally, we validated the feasibility of the de novo designed proteins through computational methods like Amber and MM/PBSA.

Description

To address the issue of GLP-1's short half-life in the gut, we designed a DPP-4 protein inhibitor using deep learning models. This inhibitor can competitively bind to DPP-4, thereby increasing the half-life of GLP-1.

Usage and Biology

Our approach involved the following steps: 1.RFdiffusion: Generate the protein backbone structure. 2.ProteinMPNN: Design the side chains of the protein. 3.ESMfold: Screen for sequences with favorable properties. 4.Amber: Perform molecular dynamics simulations. 5.MM/PBSA: Compare the binding energies of GLP-1 and the inhibitor with DPP-4. 6.Trajectory Analysis: Evaluate the molecular dynamics simulation trajectories.

Fig 1.technical roadmap

Potential application directions

Figure 2. De novo design of DPP-4 protein inhibitors(A) Comparison of RMSD and pLDDT after modeling the artificially designed sequences(B) Affinity analysis of GLP-1 and DPP-4 protein inhibitors with DPP-4(C) RMSD of molecular dynamics simulations of the DPP-4 protein inhibitor(D) RMSF of molecular dynamics simulations of the DPP-4 protein inhibitor(E) Structural comparison of DPP-4 (orange) and its protein inhibitor (blue)

During the process of generating the backbone using RFdiffusion, we utilized the "Practical Considerations for Binder Design" module, selecting key residues E167, E168, and Y624 at the DPP-4 pocket center as hotspots. We generated 200 sequences of amino acids, each with a random length between 10 and 100 residues, and then selected the best scaffolds based on structural criteria. The selected scaffolds were used in ProteinMPNN’s FastRelax protocol for side-chain generation. After modeling 20,000 sequences with ESMfold, we filtered for those with RMSD < 15 and pLDDT > 80 compared to the protein backbone’s Cα, resulting in 18 sequences suitable for molecular dynamics simulations (Fig 2.A).

We performed extended molecular dynamics simulations on the selected sequences using the Amber program, with 100 ns of equilibration and 100 ns of production, totaling 200 ns. MM/PBSA calculations on the molecular dynamics results identified one sequence with better binding energy to DPP-4 than GLP-1. The DPP-4PI and DPP-4 complex had a system energy of -6522.0487 kcal/mol, compared to -6474.9481 kcal/mol for the GLP-1 and DPP-4 complex (Fig 2.B).

Finally, we analyzed the molecular dynamics simulation trajectories, and both RMSD and RMSF were within reasonable ranges (Fig 2.C and D). The structure from the stable frame of the molecular dynamics simulation showed the binding mode of DPP-4 with its protein inhibitor (Fig 2.E).

Sequence and Features

Assembly Compatibility:

• 10
  COMPATIBLE WITH RFC[10]
• 12
  COMPATIBLE WITH RFC[12]
• 21
  COMPATIBLE WITH RFC[21]
• 23
  COMPATIBLE WITH RFC[23]
• 25
  COMPATIBLE WITH RFC[25]
• 1000
  COMPATIBLE WITH RFC[1000]