Identification of Nonpeptide Oxytocin Receptor Ligands by Receptor‐Ligand Fingerprint Similarity Search

Abstract

Oxytocin (OT) is a peptide hormone secreted by the pituitary gland that binds to a specific G protein-coupled receptor (OTR) to mediate both peripheral and central actions.[1] It notably induces uterine contraction during labor and stimulates lactation in breast-feeding mothers. Centrally, OT is implicated in numerous social behaviors [2][3][4] including pair bonding, maternal behavior, sexual arousal, learning, selfconfidence and is often referred as the hormone of love. [5] As most peptides, OT displays very poor pharmacokinetic properties and does not cross the blood-brain barrier [6] . Non-peptide orally-available and selective OTR agonists would be of invaluable help to better investigate the still poorly known central effects of OT and enable the treatment of pathologies (e.g. male erectile dysfunction) where OT function is impaired. Only one OTR partial agonist (compound 1, Figure 1), discovered from a gene reporter highthroughput screen [7] is known to date.Despite a high sequence similarity between OT and vasopressin; [8] OTR and vasopressin receptors (V1AR, V1BR, V2R) [9] and the availability of non-peptide vasopressin receptor agonists, [10,11] identifying OTR non-peptide agonists has proven to be extremely difficult. We notably showed that systematic variations of known V1AR and V2R agonists only lead to OTR antagonists but not to agonists.[12] Likewise, truncation of the terminal homopiperazine ring leads to a functional antagonist (compound 2, Figure 1), illustrating very subtle structure-activity relationships. [12] As part as our global initiative for identifying novel nonpeptide OTR agonists, we herewith report the in silico screening of an in-house collection of commercially-available drug-like compounds (Bioinfo-DB).[13] Since the X-ray structure of OTR is currently unavailable and docking-based screening of GPCR ligands require preliminary knowledge (known ligands, site-directed mutagenesis data) [14] that were missing, we restricted the panel of possible approaches to ligand-centric methods. Three similarity search methods (topomer matching, [15] shape overlap, [16] circular fingerprint similarity [17] ) were applied to find 91 hits similar to reference compound 1. In addition, a recently described chemogenomic screening method (PLFP) based on concatenated protein-ligand fingerprints [18] was utilized. Chemogenomic [19] (or proteochemometric [20] ) methods circumvent known limitations of pure ligand-based quantitative structure-activity (QSAR) approaches by considering both ligand and target descriptors [20][21][22] to predict target-ligand associations (binding affinities, [23,24] complex formation [18] ). Since similar receptors are supposed to bind to similar ligands, [25] novel interactions can thus be inferred from known data on similar ligands and/or similar targets. In the current study, PLFP screening provided 37 additional virtual hits.The above described 128 compounds were tested in duplicate at two concentrations (1 and 10 mM) in a functional fluorescence-based assay on HEK cells o...

Red Fluorescent Turn‐On Ligands for Imaging and Quantifying G Protein‐Coupled Receptors in Living Cells

“…Interestingly, both NR‐PEG‐CBT and NR‐CBT conjugates showed high affinity for OTR ( K i =0.54 and 0.55 n M , respectively). In addition, the activities of NR‐PEG‐CBT and NR‐CBT were evaluated by measuring intracellular calcium release18 (Figure S2). NR‐CBT inhibited oxytocin‐induced calcium release in a dose‐dependent manner, thus highlighting its antagonistic character (IC 50 =74 n M ); however, NR‐PEG‐CBT induced calcium release in a dose‐dependent manner, thus showing its agonistic character (EC 50 =244 n M ).…”

Unsupervised Representation Learning for Proteochemometric Modeling

“…With an expressive protein descriptor, a model could leverage similarities and differences between proteins directly to model their binding behaviors rather than merely using correlations found among the compounds they bind to, as in multi-task modeling, or ignoring protein relationships altogether as in single-task modeling [ 1 , 2 ]. Consequently, PCM models have found success in a variety of protein targets and using many different machine learning methods, including random forests (RF) and support vector machines (SVM) [ 11 , 12 , 13 , 14 , 15 ].…”

Structure versus function—The impact of computational methods on the discovery of specific GPCR–ligands

“…The extension of chemical space to ligands of closely related binding sites is a promising approach in sparse data scenarios. Weill et al [39] used a protein-ligand fingerprint (PLFP) screening to search for chemical similarity to any known ligand of the oxytocin receptor, but also of closely related receptors (vasopressin receptors: V 1A R, V 1B R, V 2 R). MD simulations can sample receptor conformations representing the key interacting points of the receptor, which are less biased towards one specific chemotype.…”