BiasNet: A Model to Predict Ligand Bias Toward GPCR Signaling

Abstract

Signaling bias is a feature of many G protein-coupled receptor (GPCR) targeting drugs with potential clinical implications. Whether it is therapeutically advantageous for a drug to be G protein biased or β-arrestin biased depends on the context of the signaling pathway. Here, we explored GPCR ligands that exhibit biased signaling to gain insights into scaffolds and pharmacophores that lead to bias. More specifically, we considered BiasDB, a database containing information about GPCR biased ligands, and focused our analysis on ligands which show either a G protein or β-arrestin bias. Five different machine learning models were trained on these ligands using 15 different sets of features. Molecular fragments which were important for training the models were analyzed. Two of these fragments (number of secondary amines and number of aromatic amines) were more prevalent in β-arrestin biased ligands. After training a random forest model on HierS scaffolds, we found five scaffolds, which demonstrated G protein or β-arrestin bias. We also conducted t-SNE clustering, observing correspondence between unsupervised and supervised machine learning methods. To increase the applicability of our work, we developed a web implementation of our models, which can predict bias based on user-provided SMILES, drug names, or PubChem CID. Our web implementation is available at: drugdiscovery.utep.edu/biasnet.

Figure 1.

Scatter plot showing the performances of the RF, XGB, MLP, and D-MPNN trained and tested on long ECFPs.

Figure 2.

Violin plot comparing molecular fragments for G protein biased ligands and β-arrestin biased ligands. (a) G protein biased ligands tend to have less secondary amines than β-arrestin biased ligands. (b) G protein biased ligands tend to have less aromatic amines than β-arrestin biased ligands.

Figure 3.

Isoproterenol and BI-167107 interacting with side chain residues from the β-2 adrenergic receptor. Isoproterenol was docked against 6KR8. BI-167107 was co-crystallized with its target receptor in 3P0G. Panels (a) and (c) show two-dimensional representations of ligands. Hydrogen bond interactions are shown as violet arrows, and π–π interactions are shown as green line segments. Panels (b) and (c) show three-dimensional molecular docking poses of ligands. Key residues, hydrogen bond distances, and hydrogen bond angles are labeled. (a) Isoproterenol forms a hydrogen bond contact with SER 203 on TM5 (b) Isoproterenol in the binding pocket of the β-2 adrenergic receptor. (c) BI-167107 does not form any hydrogen bond contacts with SER 203, SER 204, or SER 207 on TM5. (d) BI-167107 co-crystallized with the β-2 adrenergic receptor.

Figure 4.

Biased scaffolds.

Figure 5.

Five clusters are identified by t-SNE unsupervised learning. Clusters are marked by different colored rectangles. Blue circles correspond to G protein biased ligands. Red circles correspond to β-arrestin biased ligands. Cluster 1 (red), cluster 2 (yellow), cluster 3 (green), cluster 4 (blue), and cluster 5 (black).

Figure 6.

Common structures of ligands from each of the five clusters. R6 on structure 3 denotes a possible oxygen atom, methylamine, or ethylamine substitution, and R7 denotes a possible methyl substitution or a hydrogen atom. Atoms “X” on structure 5 denote either a halogen or hydrogen atom. For specific ligands, see Supporting Information Figures S2–S5.

Figure 7.

Biased drug molecules.