Empowering AlphaFold2 for protein conformation selective drug discovery with AlphaFold2-RAVE
- Institute for Physical Science and Technology, University of Maryland, United States
- University of Maryland Institute for Health Computing, United States
- Biophysics Program, University of Maryland, United States
- Department of Chemistry and Biochemistry, University of Maryland, United States
Figures
Figure 1
A schematic of the AF2RAVE-Glide workflow.
A schematic of the AF2RAVE-Glide workflow: (i) decoy structures generated by reduced multiple sequence alignment (MSA) AlphaFold2 (AF2), (ii) regular space clustering and unbiased molecular dynamics …
Figure 2 with 1 supplement
DFG-in and DFG-out conformation adopted by DDR1 kinase and their relation with Type I and Type II inhibitors.
(A) The NMR (nuclear magnetic resonance) structures of Abl1 kinase overlay, comparing the activation loop (A-loop) in the active DFG-in state (red, Protein Data Bank [PDB]: 6XR6) with the classical …
Figure 2—figure supplement 1
The distributions of ligand RMSDs for Glide XP docking poses of DDR1 and type I/type II inhibitors (upper/lower panel).
Results from cross-docking against four crystal holo structures, docking against the AF structure, and docking against 15 classical DFG-out structure in reduced multiple sequence alignment (rMSA) …
Figure 3 with 6 supplements
Reduced MSA AF2 is capable of generating crystal-like DFG-out state for DDR1 kinase.
(A) Upper panel: distribution of A-loop location for the reduced multiple sequence alignment (MSA) AlphaFold2 (AF2) structures of Abl1 kinase. 4 out of 1198 structures are A-loop folded. Lower …
Figure 3—figure supplement 1
The plot illustrates the number of gaps in the multiple sequence alignment (MSA) generated by mmseq2 (using ColabFold; Mirdita et al., 2022) for different kinases.
The non-gap count describes the coverage of each position in the MSA. The presence of residue positions with gap counts higher than 40% of the total sequence in DDR1 implies that it has fewer …
Figure 3—figure supplement 2
The AlphaFold2 (AF2) pLDDT rank is plotted against the CA RMSDs from the AF2 structure (the one with the highest pLDDT) for each structure in the reduced multiple sequence alignment (rMSA) AF2 ensemble for Abl1, DDR1, or Src kinase.
An RMSD cutoff of 7 Å (dashed black line) is applied to filter out unphysical structures with large RMSD from the native structure. Each rMSA AF2 ensemble consists of 1280 structures, 640 for MSAs …
Figure 3—figure supplement 3
The AlphaFold2 (AF2) pLDDT rank is plotted against the CA RMSDs from the AF2 structure for each structure in the AF2-cluster ensemble for Abl1, DDR1, or SrcK.
An RMSD cutoff of 10 Å (dashed black line) is applied to filter out unphysical structures with large RMSD from the native structure. After the RMSD filter, 197 out of 362 structures remain for Abl1, …
Figure 3—figure supplement 4
Ligand RMSDs are plotted against the docking scores for the Induced Fit docking (IFD) poses of type II inhibitors (ponatinib and imatinib) against AlphaFold2 (AF2) structure (blue) or classical DFG-out structures in reduced multiple sequence alignment (rMSA) AF2 ensembles (red).
(A) IFD results for Abl1. (B) IFD results for DDR1. The pose with the lowest ligand RMSD from each input structure is marked by hexagon.
Figure 3—figure supplement 5
Docking studies on holo-model from rMSA AF2 for DDR1 kinase.
(A) Comparison of the DFG motif for DDR1 in its co-crystalized structure with imatinib (Protein Data Bank [PDB] 4BKJ), its ‘holo-model’ structure, and its AlphaFold2 (AF2) structure. (B and C) In …
Figure 3—figure supplement 6
Ligand RMSDs are plotted against the DiffDock confidence scores for the DiffDock poses of type II inhibitors (ponatinib and imatinib) against DDR1 AlphaFold2 (AF2) structure (blue) or the classical DFG-out structures in DDR1 reduced multiple sequence alignment (rMSA) AF2 ensemble (red).
The pose with the lowest ligand RMSD from each input structure is marked by hexagon.
Figure 4 with 6 supplements
Ranking the DDR1 structures using AF2RAVE protocol on the learnt latent space.
(A or B) The unbiased molecular dynamics (MD) trajectories of DDR1 are projected onto the learnt state predictive information bottleneck model (SPIB) latent space. In plot (A), the colors of sample …
Figure 4—figure supplement 1
Reduced MSA AF2 generated DDR1 kinase structures on the Dunbrack space.
(A) The reduced multiple sequence alignment (rMSA) AlphaFold2 (AF2) ensemble for DDR1 is projected in the Dunbrack space. Sample points are color-coded based on the CA RMSD from the AF2 structure …
Figure 4—figure supplement 2
Umbrella sampling for DDR1 kinase.
(A) Distributions from different umbrella sampling windows in the latent space. (B) The distribution overlap graph for all the umbrella sampling windows. The mean value of each distribution is shown …
Figure 4—figure supplement 3
DDR1 potential of mean force (PMF) calculated with all the umbrella sampling windows.
Hanson et al., 2019, found the A-loop folded DFG-out state to be more stable than the A-loop folded DFG-in/inter state for DDR1; Vani et al., 2024, reported that the A-loop extended DFG-out state is …
Figure 4—figure supplement 4
Potential of mean force (PMF) values and Boltzmann ranks of candidate structures fluctuate with the selection of the umbrella sampling windows and the simulation length of umbrella sampling trajectories, demonstrated with the DDR1 system.
Figure 4—figure supplement 5
Free energy profile for DDR1 in the latent space, calculated from unbiased molecular dynamics (MD) simulations.
The 15 DDR1 classical DFG-out structures in reduced multiple sequence alignment (rMSA) AlphaFold2 (AF2) are shown as red cross and circles (top 5 structures ranked by free energy values). The …
Figure 4—figure supplement 6
Unbiased simulation coverage study on DDR1 and Abl1 kinase.
(A) The distribution overlap graph for all the unbiased molecular dynamics (MD) trajectories starting from 15 classical DFG-out structures in DDR1 reduced multiple sequence alignment (rMSA) …
Figure 5 with 7 supplements
Transferrable learning of holo-like structure for Abl1 kinase.
(A) The reduced multiple sequence alignment (MSA) AlphaFold2 (AF2) structures of Abl1 are projected onto the latent space. Sample points are color-coded based on the A-loop location. Light green …
Figure 5—figure supplement 1
Ligand RMSDs are plotted against the docking scores for the Induced Fit docking (IFD) poses of type II inhibitors (ponatinib and imatinib) against Abl1 tAF2 structures.
The pose with the lowest ligand RMSD from each input structure is marked by hexagon.
Figure 5—figure supplement 2
The projection of A-loop folded structures from the reduced multiple sequence alignment (rMSA) AlphaFold2 (AF2) ensemble or the AF2-cluster ensemble (cAF2) on the AF2RAVE potential of mean force (PMF) for Abl1 or DDR1.
Figure 5—figure supplement 3
Structural visualization of tAF2 generated classical DFG-out Src kinase.
(A) The AlphaFold2 (AF2)-template structure for Src kinase is superimposed with its template structure (classical DFG-out in DDR1 reduced multiple sequence alignment [rMSA] AF2 ensemble, …
Figure 5—figure supplement 4
Comparison between performance of rMSA AF2 and AF2-cluster on Src kinase.
The projection of (A) the reduced multiple sequence alignment (rMSA) AlphaFold2 (AF2) ensemble or (B) the AF2-cluster ensemble on the AF2RAVE latent space for SrcK. The classical DFG-out SrcK …
Figure 5—figure supplement 5
Ligand RMSDs are plotted against the docking scores for the Induced Fit docking (IFD)/IFD-trim docking poses of type II inhibitors (ponatinib and imatinib) against the SrcK tAF2 structure.
The pose with the lowest ligand RMSD from each input structure is marked by hexagon.
Figure 5—figure supplement 6
Accounting for broken αC helix during umbrella sampling of DDR1 and Abl1 kinase.
(A) One representative frame with αC helix broken in Abl1 umbrella sampling trajectories. The backbone of the αC helix is shown with cyan sticks, while the DFG motif is shown as orange sticks. (B or …
Figure 5—figure supplement 7
Abl1 potential of mean force (PMF) calculated from umbrella sampling after discarding windows with αC helix broken.
The four holo-like structures (‘holo-models’) are enriched to the top six based on PMF values.
Tables
Table 1
Comparing the Induced Fit docking (IFD) performance of various structure generation methods for docking type II kinase inhibitors.
| Source-protein (# of structures) | Lowest imatinib ligand RMSD (Å) | Lowest ponatinib ligand RMSD (Å) | Ratio of structs. w. ligand RMSD <3 Å |
|---|---|---|---|
| AF2-Abl1 (1) | 9.22 | 9.4 | 0/1 |
| AF2-DDR1 (1) | 9.24 | 9.33 | 0/1 |
| rMSA AF2-Abl1 (2) | 10.14 | 9.11 | 0/2 |
| rMSA AF2-DDR1 (15) | 1.04* | 0.89 | 1/15 |
| tAF2-Abl1 (30) | 2.74 | 0.78 | 4/30 |
| AF2RAVE-DDR1 (2) | 1.04* | 0.89 | 1/2 |
| AF2RAVE-Abl1 (8) | 2.74 | 0.78 | 4/8 |
-
*
Result from IFD-trim.
Additional files
-
MDAR checklist
- https://cdn.elifesciences.org/articles/99702/elife-99702-mdarchecklist1-v1.pdf
-
Supplementary file 1
Comparison between AF2RAVE ranks and DiffDock confidence scores.
Confidence score for the DiffDock pose aligns with AF2RAVE potential of mean force (PMF) values. The DiffDock confidence score of the pose with the lowest ligand RMSD (marked in red/bold) from each classical DFG-out structure in DDR1 reduced multiple sequence alignment (rMSA) AlphaFold2 (AF2) ensemble is compared with the AF2RAVE PMF value for corresponding structure (marked in red/bold).
- https://cdn.elifesciences.org/articles/99702/elife-99702-supp1-v1.xlsx
