Fig 1.
A) Schematic of the LDM workflow that takes a GPCR model and a ligand as input, and outputs LDM models. Sampling steps are represented in blue while selection steps are represented in yellow. B) GPCRs are separated in regions in the LDM workflow that have increasing degree of sampling. The TM1 and cytoplasmic region are kept static (dark grey), the rest of the GPCR is flexible (light grey). The binding pocket region defines the docking area and undergoes further sampling (dotted circle). C) The LDM was evaluated with a benchmark divided into three scenarios where an origin X-ray crystal structure binding pocket is refined by the LDM using the ligand found in a destination X-ray crystal structure, referred to here as origin and destination, respectively. In self refinement, the origin and destination are the same structure, in chemotype switch, the origin and destination are bound by ligands of the same pharmacology (agonist or inhibitor) but with different chemotypes and in pharmacology switch, origin and destination are bound by ligands of different pharmacology (agonist to inhibitor or vice-versa).
Fig 2.
Comparison of VS performance for X-ray structures with all loops vs. ECL2-distal loop only.
VS performance in NSQ_AUC values of X-ray structures measuring A) recovery of known ligands against decoys and B) selectivity of agonists vs. inhibitors (or vice-versa). X-ray structures with all loops are represented in dashed bars and X-ray structures with ECL2-distal only are represented in black bars.
Fig 3.
Self refinement LDM experiment on the B2AR 4LDE-BI, using BI as a refinement ligand.
A) Dendrogram of the top 25 LDM models and the X-ray structures, a cut-off line identifies different LDM clusters and their representative LDM models are designated by a coloured dot. Representative LDM models are the highest scoring within the cluster based on the OPUS-ICM metric. B) PCA comparison of binding pocket conformation between the top 25 LDM models and the X-ray structures. LDM models are coloured based on their IFP Jaccard distance with the destination X-ray structure, 4LDE-BI. C) Binding poses of the representative LDM models and the destination X-ray structure with the X-ray ligand in black, LDM000 in red and LDM021 in blue. D) IFP of the representative LDM models and the X-ray structure. Interaction type is described for each residue of the binding pocket: hydrophobic interaction, hydrogen bond (H-bond) donor and acceptor, weak hydrogen bond (weak H-bond) donor and acceptor, ionic bond positive (+) and negative (-) and aromatic interaction. E, F, G) VS performance with the X-ray structure in black, LDM000 in red and LDM021 in blue. E) ROC curves visualising the recovery of B2AR agonists vs. decoys. F) ROC curves highlighting the selectivity of B2AR agonists over B2AR inhibitors. For E and F, the relative rank of the LDM refinement ligand is identified with a vertical dashed line. This vertical line is not visible as it overlays the axis indicating the refinement ligand is very highly ranked in all 3 cases. The ROC curve figure insets show NSQ_AUC values for each binding pocket. G) bar chart to visualise the EF for representative B2AR agonist chemotypes at EF1, EF5 and EF10. Chemotypes A, B ‘BI-like’, C ‘ADR-like’ and D ‘ISO-like’ represent only a subset of B2AR agonist ligands (S2C Fig). The EF bar chart inset shows the number of ligands for each of these chemotype clusters between parenthesis. X-ray structure chemotype EF shown in black bars, with the LDM models coloured based on their relative clusters identified in A.
Fig 4.
Self refinement LDM on the B2AR 2RH1-CAR, using CAR as the refinement ligand.
A) Dendrogram of the top 25 LDM models and the X-ray structure, a cutoff line identifies different LDM clusters. Representative LDM models are the highest scoring within the cluster based on the OPUS-ICM metric and are designated by a coloured dot. B) Comparison of binding pocket conformation between the top 25 LDM models and the X-ray structure. LDM models are colored based on their IFP Jaccard distance with the destination X-ray structure, 2RH1-CAR. C) Binding poses of the representative LDM models and the destination X-ray structure with the X-ray ligand in black and the LDM models coloured based on their representative clusters defined in A. D) IFP of the representative LDM models and the X-ray structure. Interaction type is described for each residue of the binding pocket: hydrophobic interaction, hydrogen bond (H-bond) donor and acceptor, weak hydrogen bond (weak H-bond) donor and acceptor, ionic bond positive (+) and negative (-) and aromatic interaction. VS performance is described with ROC curves to visualise E) the recovery of B2AR inhibitors vs. decoys and F) the selectivity of B2AR inhibitors over B2AR agonists. X-ray structures are in black, with the LDM models coloured based on their relative clusters identified in A. The relative rank of the LDM refinement ligand is identified with a vertical dashed line. In all cases, this vertical line is masked by the axis indicating that the refinement ligand was very highly ranked. The ROC curve figure inset shows NSQ_AUC values for each binding pocket. G) bar chart is used to visualise the EF for representative B2AR inhibitor chemotypes at EF1, EF5 and EF10. Chemotypes A ‘ICI-like’, B ‘KOL-like’, C ‘TIM-like’, D ‘CAR-like’ and E represent only a subset of B2AR inhibitor ligands (S2D Fig). The EF bar chart inset shows the number of ligands for each chemotype cluster between parenthesis. X-ray structure chemotype EF shown in black bars, with the LDM models coloured based on their relative clusters identified in A.
Fig 5.
Chemotype switch LDM experiment on the B2AR from 4LDO-ADR to 4LDL-ISO, using ISO as a refinement ligand.
A) Dendrogram of the top 25 LDM models and X-ray structures, a cutoff line identifies different LDM clusters and their representative LDM models are designated by a colored dot. Representative LDM models are the highest scoring within the cluster based on the OPUS-ICM metric. B) Comparison of binding pocket conformation between the top 25 LDM models and X-ray structures. LDM models are colored based on their IFP Jaccard distance with the destination X-ray structure, 4LDL-ISO. C) Binding poses of the representative LDM models and the destination X-ray structure with the X-ray ligand in black and the LDM models coloured based on their representative clusters defined in A. D) IFP of the representative LDM models and the X-ray structures. Interaction type is described for each residue of the binding pocket: hydrophobic interaction, hydrogen bond (H-bond) donor and acceptor, weak hydrogen bond (weak H-bond) donor and acceptor, ionic bond positive (+) and negative (-) and aromatic interaction. VS performance is described with ROC curves to visualise E) the recovery of B2AR agonists vs. decoys and F) the selectivity of B2AR agonists over B2AR inhibitors. X-ray structures are in black, with the LDM models coloured based on their relative clusters identified in A. The relative rank of the LDM refinement ligand is identified with a vertical dashed line. In all cases, this vertical line is masked by the axis indicating that the refinement ligand was very highly ranked. The ROC curve figure inset shows NSQ_AUC values for each binding pocket. G) bar chart is used to visualise the EF for representative B2AR agonist chemotypes at EF1, EF5 and EF10. Chemotypes A, B ‘BI-like’, C ‘ADR-like’ and D ‘ISO-like’ represent only a subset of B2AR agonist ligands (S2C Fig). The EF bar chart inset shows the number of ligands for each chemotype cluster between parenthesis.
Fig 6.
Chemotype switch LDM experiment on the AA2AR from 3PWH-ZM to 3RFM-CAF, using CAF as a refinement ligand.
A) Dendrogram of the top 25 LDM models and X-ray structures, a cutoff line identifies different LDM clusters and their representative LDM models are designated by a colored dot. Representative LDM models are the highest scoring within the cluster based on the OPUS-ICM metric. B) Comparison of binding pocket conformation between the top 25 LDM models and X-ray structures. LDM models are colored based on their IFP Jaccard distance with the destination X-ray structure, 3RFM-CAF. C) Binding poses of the representative LDM models and the destination X-ray structure with the X-ray ligand in black and the LDM models coloured based on their representative clusters defined in A. D) IFP of the representative LDM models and the X-ray structures. Interaction type is described for each residue of the binding pocket: hydrophobic interaction, hydrogen bond (H-bond) donor and acceptor, weak hydrogen bond (weak H-bond) donor and acceptor, ionic bond positive (+) and negative (-) and aromatic interaction. VS performance is described with ROC curves to visualise E) the recovery of AA2AR inhibitors vs. decoys and F) the selectivity of AA2AR inhibitors over AA2AR agonists. X-ray structures are in black, with the LDM models coloured based on their relative clusters identified in A. The relative rank of the LDM refinement ligand is identified with a vertical dashed line. This vertical line for some models may be masked by others if the ligand is very highly ranked. The ROC curve figure inset shows NSQ_AUC values for each binding pocket. G) bar chart is used to visualise the EF for representative AA2AR inhibitor chemotypes at EF1, EF5 and EF10. Chemotypes A ‘ZM-like’, B ‘T4G-like’, C ‘XAC-like’ and D ‘CAF-like’ represent only a subset of AA2AR inhibitors ligands (S2B Fig). The EF bar chart inset shows the number of ligands for each chemotype cluster between parenthesis. Origin and destination X-ray structure chemotype EF shown in grey and black bars, respectively, with the LDM models coloured based on their relative clusters identified in A.
Fig 7.
Pharmacology switch LDM experiment on the B2AR from 2RH1-CAR to 4LDE-BI, using BI as a refinement ligand.
A) Dendrogram of the top 25 LDM models and X-ray structure(s), a cutoff line identifies different LDM clusters and their representative LDM models are designated by a colored dot. Representative LDM models are the highest scoring within the cluster based on the OPUS-ICM metric. B) Comparison of binding pocket conformation between the top 25 LDM models and X-ray structures. LDM models are colored based on their IFP Jaccard distance with the destination X-ray structure, 4LDE-BI. C) Binding poses of the representative LDM models and the destination X-ray structure with the X-ray ligand in black and the LDM models coloured based on their representative clusters defined in A. D) IFP of the representative LDM models and the X-ray structures. Interaction type is described for each residue of the binding pocket: hydrophobic interaction, hydrogen bond (H-bond) donor and acceptor, weak hydrogen bond (weak H-bond) donor and acceptor, ionic bond positive (+) and negative (-) and aromatic interaction. VS performance is described with ROC curves to visualise E) the recovery of B2AR agonists vs. decoys and F) the selectivity of B2AR agonists over B2AR inhibitors. The relative rank of the LDM refinement ligand is identified with a vertical dashed line. This vertical line may be masked by other curves if the ligand is very highly ranked. The ROC curve figure inset shows NSQ_AUC values for each binding pocket. Finally, a G) bar chart is used to visualise the EF for representative B2AR agonist chemotypes at EF1, EF5 and EF10. Chemotypes A, B ‘BI-like’, C ‘ADR-like’ and D ‘ISO-like’ represent only a subset of B2AR agonist ligands (S2C Fig). The EF bar chart inset shows the number of ligands for each chemotype cluster between parenthesis. Origin and destination X-ray structure chemotype EF shown in grey and black bars, respectively, with the LDM models coloured based on their relative clusters identified in A.
Fig 8.
Pharmacology switch LDM experiment on the B2AR from 3P0G-BI to 2RH1-CAR, using CAR as a refinement ligand.
A) Dendrogram of the top 25 LDM models and X-ray structures, a cutoff line identifies different LDM clusters and their representative LDM models are designated by a colored dot. Representative LDM models are the highest scoring within the cluster based on the OPUS-ICM metric. B) Comparison of binding pocket conformation between the top 25 LDM models and X-ray structures. LDM models are colored based on their IFP Jaccard distance with the destination X-ray structure, 2RH1-CAR. C) Binding poses of the representative LDM models and the destination X-ray structure with the X-ray ligand in black and the LDM models coloured based on their representative clusters defined in A. D) IFP of the representative LDM models and the X-ray structures. Interaction type is described for each residue of the binding pocket: hydrophobic interaction, hydrogen bond (H-bond) donor and acceptor, weak hydrogen bond (weak H-bond) donor and acceptor, ionic bond positive (+) and negative (-) and aromatic interaction. VS performance is described with ROC curves to visualise E) the recovery of B2AR inhibitors vs. decoys and F) the selectivity of B2AR inhibitors over B2AR agonists. The relative rank of the LDM refinement ligand is identified with a vertical dashed line. This vertical line may be masked by other curves if the ligand is very highly ranked. The ROC curve figure inset shows NSQ_AUC values for each binding pocket. Finally, a G) bar chart is used to visualise the EF for representative B2AR inhibitor chemotypes at EF1, EF5 and EF10. Chemotypes A ‘ICI-like’, B ‘KOL-like’, C ‘TIM-like’, D ‘CAR-like’ and E represent only a subset of B2AR inhibitor ligands (S2B Fig). The EF bar chart inset shows the number of ligands for each chemotype cluster between parenthesis. Origin and destination X-ray structure chemotype EF shown in grey and black bars, respectively, with the LDM models coloured based on their relative clusters identified in A.
Table 1.
Summary of the LDM benchmark outcomes.
The results are organised by scenario: self refinement, chemotype switch and pharmacology switch. Each line describes an LDM experiment including the origin and destination X-ray structures as well as the overall VS performance outcome. The best LDM model is compared to the origin X-ray structure in recovery of known ligands vs. decoys, selectivity of agonists over inhibitors (or vice-versa) and chemotype enrichment. VS recovery and selectivity performance is improved ↑, similar → or worse ↓ using NSQ_AUC values. VS chemotype enrichment evaluates the LDM refinement ligand chemotype enrichment by comparing the EF1 values of LDM models with their origin X-ray structure. It is “improved” when an LDM model outperforms the origin X-ray structure, and it is “narrowed” when the same performance is observed and EF1 value for other chemotypes is worse for the LDM model. LDM experiments were assigned to two groups based on the similarity of their LDM model binding pose to that of the destination X-ray structure and their improvement in VS performance over the origin X-ray structure. Group A includes LDM models with similar binding poses and improved performance and group B includes LDM models with a different binding pose and improved performance or a similar binding pose and no performance improvement. CCR5 agonist ligands were not available from the GLL/GDD, hence VS selectivity for the self refinement CCR5-MRV experiment was not calculated and is marked N/A.