Unraveling the Multifaceted Actions of Neurosteroids: Machine Learning and In Vitro Screening for Novel Target Discovery

Availability of material, data, and code

This study did not generate any new large data sets. All screening related data sets are available in supplementary files, and small molecule library data sets are available open source for academic research use.

Raw data reported in this paper will be shared by the corresponding author upon reasonable request. Python analysis scripts were developed with the assistance of open-source resources and OpenAI tools to efficiently organize, process, and analyze data. Any additional information required to reanalyze the data reported in this paper is available from the corresponding author upon request.

Neurosteroids synthesis

AlloP, Prog, Etio and Andro were purchased from Sigma Aldrich and other NS and analogues used in screening were synthetized in the laboratory of Douglas F. Covey. Synthetic details are provided in references given in the Identifier/Reference column of the Key Resource Table. Reviews of synthetic methods have also been published (Jastrzębska and Covey, 2022; Covey et al., 2023).

Data collection and molecular similarity search

In this survey, we collected small molecule data from three openly accessible databases. Specifically, we obtained 6798 molecules including investigational drugs from the Drug Repurposing Hub, along with their respective target names available at http://www.broadinstitute.org/repurposing. Additionally, we downloaded 11926 small molecules and their associated target names from the DrugBank library (https://go.drugbank.com), as well as 3492 approved drugs along with their targets from the ChEMBL library, utilizing web resources through Python scripts (Davies et al., 2015). The canonical SMILES (Simplified Molecular Input Entry System) representations of these molecules were extracted using a Python program integrated with the open-source cheminformatics tool RDKit.

To identify neurosteroid-like molecules, we used t-distributed stochastic neighbor embedding (t-SNE) maps for a molecular similarity search in the chemical space of approved drugs and investigational small molecules. This method was previously used in several published articles (Janssen et al., 2019; Duran-Frigola et al., 2020; Li et al., 2020). AlloP, an FDA-approved NS, was used as a query molecule for a similarity search in chemical space. The subsequent step generated Morgan fingerprints for each molecule using RDKit’s “GetMorganFingerprintAsBitVect” function and the parameters radius=2 and nBits=4096. The radius parameter dictates the breadth of the fingerprint’s neighborhood, influencing the level of detail captured in the molecular representation. Meanwhile, the nBits parameter determines the length of the fingerprint bit vector, affecting the granularity of the molecular features encoded. These parameters strike a balance between capturing sufficient structural information and managing computational complexity. Following fingerprint generation, t-SNE was employed to project the high-dimensional molecular vector into a more manageable two-dimensional representation. The parameters, components=2 and random state=42, were specified to indicate the desired number of dimensions in the reduced space and ensure reproducibility across runs, respectively. By preserving local structure and revealing underlying patterns, t-SNE enabled intuitive exploration and interpretation of the molecular dataset. Subsequently, k-nearest neighbors (k-NN), a machine learning tool, was applied with parameter k=50 to determine the number of neighbors considered in the search. By computing distances between the query molecule and its nearest neighbors in the fingerprint space, this step facilitated the identification of molecules sharing structural similarities, thus offering valuable insights into potential drug candidates related to the query molecule. In parallel, K-means clustering was employed to partition the dataset into distinct groups based on the t-SNE-transformed features. To provide a visual reference for molecules within a certain distance from the query, a circle with a radius of 5 units was plotted around the query molecule in the t-SNE visualization. Additionally, extracted information for molecules within the circle surrounding the query molecule was obtained. This information typically includes details such as Molecule Name, Mechanism of Action (MOA), and Target names, and provides characteristics of molecules closely related to AlloP, aiding in further analysis and interpretation.

Drug-target network pharmacology

We conducted a similarity search to identify neurosteroid-like molecules and their targets with associated genes, followed by manual verification of target names using additional sources. In cases where target names were similar for multiple molecules, we ensured selection of each target only once for network pharmacology analysis. All potential target genes shared among the molecules were integrated into the STRING version 12.0 database (https://string-db.org) to construct a protein-protein interaction (PPI) network, with Homo sapiens selected as the organism. We opted for a medium confidence level, setting an interaction score threshold of >0.4 for the PPI network. This network consists of nodes representing target proteins and edges representing protein-protein interactions. The degree of each node indicates the number of direct connections it possesses, with higher degrees indicating greater importance.

GO and KEGG enrichment analysis

Gene Ontology (GO) and KEGG (Kyoto Encyclopedia of Genes and Genomes) pathways analyses are essential components of drug-target identification and development, playing integral roles in understanding disease mechanisms, identifying therapeutic targets, and advancing drug discovery efforts. The STRING database also offers functionalities for functional enrichment analysis, including KEGG and GO analysis. In this analysis we used similar data sets as used in network pharmacology: the KEGG pathway and the functional enrichment of GO pathway, including biological processes. Molecular function and cellular component results were downloaded and further analyzed with Python script for visualization purposes. The top 50 enriched KEGG pathways and GO-terms are plotted based on the observed-to-background gene count ratio. Bars are color-coded by the false discovery rate (FDR) value.

Target selection

Following enrichment analysis, we meticulously scrutinized each pathway and GO-term, weighing numerous factors to choose targets for experimental validation. Our primary criterion was to prioritize targets aligned with our research focus, especially those yet unexplored for our NS molecules. We identified targets significantly enriched in pertinent GO-terms or KEGG pathways target genes, indicating their potential pivotal roles in the biological system with therapeutic implications. Moreover, we analyzed the interactions among these targets, recognizing highly connected ones as potential hubs or key regulators within biological networks. We evaluated the functional significance of the targets based on their roles in biological pathways, gene ontology categories, or protein-protein interaction networks. Although some targets lacked multiple connections, they were novel to our study. We conducted a literature review to gather further evidence supporting the involvement of selected targets in relevant biological processes or diseases. Additionally, we considered the feasibility of experimental validation, considering factors such as experimental techniques and required resources.

In vitro screening

Broad in vitro biological screenings described below were conducted by Eurofins DiscoverX Corporation (11180 Roselle Street, Suite D, San Diego, CA 92121). We outsourced all in vitro screening, a panel of 47 targets (table S5) including ion channels, GPCRs, nuclear receptors, transporters, non-kinase enzymes, and some kinases using SAFETYscan E/IC₅₀ ELECT service from Eurofins. The experimenters were blinded to any hypotheses about the activity of the analogues tested. Most of the cell-based assays were performed using cell lines, including CHO-K1, CHO, and HEK-293, as maintained by Eurofins DiscoverX. Sample sizes were determined a priori and represent a single replicate at each of 10 concentrations for biological screening performed by the contract research organization; the results of the in vitro screening should thus be considered hypothesis generating. All 19 compounds, including NS and their synthetic analogues, were tested at 1.5 nM, 4.5 nM, 13.7 nM, 41.2 nM, 123.5 nM, 370.4 nM, 1.1 μM, 3.3 μM, 10 μM and 30 μM. We also screened 10 μM AlloP, Prog, and their enantiomers for kinase inhibition activity across 481 kinases using SelectScreen Kinase Profiling Services from Thermo Fisher (www.thermofisher.com/selectscreen).

Ion channel assays

For ion channel activity assays, including ionotropic receptors, we used CHO-K1 or HEK-293 cell lines. In brief, cells were seeded into Poly-D-lysine-coated, black-walled, clear-bottom 384-well microplates, followed by incubation at 37°C. Dye loading used a 1X Dye loading buffer with 2.5 mM Probenecid, incubated for 30–60 min at 37°C. For agonist determination, cells were incubated with the test sample, while antagonist determination involved pre-incubation with the sample in EC₈₀ agonist. Signal detection was performed on a FLIPR Tetra, and data analysis used the CBIS suite to calculate percentage activity or inhibition based on relative luminescence units (RLU).

GPCR cAMP assay

Cyclic adenosine monophosphate (cAMP) modulation was evaluated using the DiscoverX HitHunter cAMP XS+ assay in stably transfected cells. Cells were seeded into a 384-well microplate and incubated at 37°C. The media was replaced with HBSS/HEPES and cAMP XS+ Ab reagent, followed by incubation. In the G_s agonist format, cells were treated with the sample and incubated at 37°C or room temperature for 30–60 min to induce a response. For the G_i agonist format, cells were exposed to the sample in the presence of EC₈₀ forskolin for 30–60 min, with subsequent steps similar to the G_s agonist protocol. In the antagonist format, cells were first pre-incubated with the sample for 30 min, then challenged with agonist at EC₈₀ concentration and further incubated for 30 min. In the case of G_i coupled GPCRs, EC₈₀ forskolin was included. After appropriate sample incubation, signal detection was generated using cAMP XS+ ED lysis cocktail reagents for one hour followed by incubation with cAMP XS+ EA reagent for three hours at room temperature, chemiluminescent signal detected via a PerkinElmer Envision^™ instrument. Data analysis was completed using the CBIS data analysis suite, applying specific formulas to determine percentage activity or inhibition depending on the assay mode.

Calcium flux assay

Stable cells were seeded into Poly-D-lysine coated 384-well microplates and incubated before testing. Assays were conducted in a 1X dye loading buffer containing dye, Additive A, and Probenecid. Cells were loaded with dye, incubated, and then tested for agonist or antagonist activity using FLIPR Tetra. Agonist activity was measured by adding a compound and monitoring calcium mobilization, while antagonist activity was measured after pre-incubation with a sample and subsequent agonist challenge. Data analysis involves calculating the area under the curve and determining the percentage of activity or inhibition using specific formulas.

Nuclear translocation assay for nuclear receptor

PathHunter nuclear hormone receptor (NHR) cell lines (stably transfected CHO-K1 or HEK293 cells) were seeded into white-walled 384-well microplates containing charcoal-dextran filtered serum. To determine agonist activity, cells were treated with a 5X sample and incubated for 3–16 hours, allowing for concentration-dependent translocation kinetics (Cutress et al., 2008; Trebble et al., 2013; Lv et al., 2021). For antagonist determination, cells were pre-incubated with an antagonist, then challenged with an EC₈₀ agonist and incubated. Signal detection was achieved by adding the PathHunter Detection reagent cocktail and measuring the chemiluminescent signal using a PerkinElmer Envision^™ instrument. Data analysis involved calculating percentage activity for agonists and percentage inhibition or inverse agonist activity for antagonists, with results capped at 0% or 100% as necessary

KINOMEscan binding assays

To assess kinase activity, Eurofins used the KINOMEscan binding assay as previously described (Fabian et al., 2005; Karaman et al., 2008). In brief for most of the assays, kinase-tagged T7 phage was grown in E. coli, followed by infection with T7 phage and incubation at 32°C until lysis (90–150 min). The lysates were centrifuged (6000*g) and filtered. Remaining kinases were produced in HEK-293 cells and tagged for qPCR detection. Streptavidin-coated beads were treated with biotinylated ligands for 30 min to create affinity resins and blocked with excess biotin and washed to remove unbound ligands. Binding reactions were started by combining kinases, liganded beads, and test compounds in 384-well plates. After one hour incubation with shaking at room temperature, affinity beads were washed and resuspended in elution buffer for 30 min at room temperature. The kinase concentrations in eluates were measured by qPCR. Percent response and binding constants (Kds) were calculated using a dose-response curve.

Transporter assays

CHO cells were seeded into Poly-D-lysine coated 384-well microplates and incubated at 37°C prior to testing. After cell plating and incubation, the media was removed, and 25 μL of 1X compound in HBSS/0.1% BSA was added, followed by a 30-minute incubation at 37°C. Assays were performed in 1X Dye Loading Buffer, with 25 μL of 1X dye added to the wells, and cells were incubated for 30–60 min at 37°C. Fluorescence signal detection was then conducted using a PerkinElmer Envision^™ instrument, and compound activity was analyzed using the CBIS data analysis suite, where percentage inhibition for blocker mode assays was calculated, and percent response for primary screens was capped at 0% or 100% based on the results.

Enzymatic assays

After preparing enzymes, including acetylcholinesterase (AChE), cyclooxygenase- 1 and −2 (COX1 and COX2), monoamine oxidase A (MAOA), phosphodiesterase-3A and −4D2 (PDE3A and PDE4D2), enzymatic activity assays were conducted to measure substrate consumption or product formation over time using various detection methods. For AChE, the enzyme and test compounds were preincubated before substrate addition, with the signal detected by absorbance at 405 nm. For COX1 and COX2, the enzymes were equilibrated with compounds, and the reactions were measured on a fluorimeter. For MAOA, reactions were initiated with kynuramine and terminated with NaOH, with signal detection via spectrofluorimetry (Matsumoto et al., 1985). For PDE3A and PDE4D2, the reactions involved cAMP substrate and were terminated by 3-isobutyl-1-methylxanthine (IBMX), with detection using a cAMP detection kit. Signal detection was performed using a PerkinElmer Envision^™ instrument, and data were analyzed using the CBIS data suite to calculate percentage inhibition and percent response in primary screens.

Analyses

All activity data were analyzed using the CBIS data analysis suite (ChemInnovation, CA) to determine IC₅₀/EC₅₀ values via non-linear regression curve fit analysis. For GPCRs, ion channels, and nuclear receptors, agonist mode assays measured percentage activity based on the relative increase in luminescence compared to vehicle (basal) and maximum control, while antagonist mode assays calculated percentage inhibition relative to the EC₈₀ control. Additional modes of action included G_i agonist and antagonist assessments in GPCR assays, in which inhibition was quantified relative to forskolin control, and inverse agonist activity in nuclear receptor assays, where reduction in receptor activity was measured against the vehicle and maximum control. In calcium mobilization and transporter assays, agonist mode activity was determined by comparing relative fluorescence units (RFU) to control ligands, while antagonist or blocker assays measured inhibition relative to EC₈₀ or positive control.

KINOMEscan binding assays calculated percent response using a normalized formula based on the difference between test compound signal, positive control, and negative control signals. Percent control values were converted into percent response by subtracting from 100, ensuring a standardized output. Binding constants (K_d) were derived using the Hill equation, with a Hill slope set to −1, and curve fitting was performed using a non-linear least squares approach with the Levenberg-Marquardt algorithm. In enzymatic assays, inhibition of enzyme activity was determined based on the luminescence reduction relative to vehicle and positive controls.

For all screenings, the percentage response was capped at 0% or 100% where the calculated percent response returned a negative value or value exceeded than 100%, respectively. Percent responses were used to visualize curve fit through GraphPad Prism 10. We used Python-based Matplotlib for visualization of data, such as heatmaps, t-SNE plots, and gene enrichment analyses.

Molecular similarity of neurosteroids with investigational & approved drugs

To explore novel therapeutic targets of NS, we started with a cheminformatics approach. We compared chemical space similarity of NS with other bioactive small molecules from open-source, small-molecule databases, using AlloP as the query molecule because it is a clinically approved prototype. In the resulting t-distributed stochastic neighbor embedding (t-SNE) map with Drug Repurposing Hub molecules (Figure 1A), each point is a molecule, colored according to its assigned cluster. The query molecule (AlloP) is highlighted as a red cross in the plot, facilitating visual inspection of its neighborhood in the reduced-dimensional space. Unsurprisingly, many near-neighbors had a familiar steroid-like backbone (Figure 1B).

Upon performing the t-SNE based chemical similarity search from Drug Repurposing Hub molecules (6798 molecules), we investigated 37 neighboring molecules located within a 5-unit radius (red dotted circle) of chemical space around AlloP (Figure 1A and 1B). All neighboring molecules are listed in Table 1 along with varied modes of action of the drugs. These molecules showed a wide range of pharmacological activities, including androgen receptor modulation, anticancer activity, inhibition of the hedgehog pathway, and modulation of neurotransmitter receptors. Neurotransmitter receptor targets include GABA-A receptors, glutamate receptors, and both muscarinic and nicotinic acetylcholine receptors. Other targets included an ionotropic serotonin receptor (HTR3A), chemokine receptor (CXCR3), smoothened GPCR, and kinase associated proteins (Src, FAK, Erk, PI3K, mTOR, GSK3, CDKL5, SHIP1). The queried molecules have a steroidal backbone and generally, are known for their significant activity towards nuclear receptors including estrogen receptors (ESR1, ESR2), androgen receptor (AR) and glucocorticoid receptor (GR). The molecules are also involved in the activity of cytochrome P450 enzymes (CYP19A1, CYP17A1), and some show NS-like modulation of GABA-A receptors (GABRA1–6, GABRB1–3, GABRD-RE) and glutamate receptors (GRIN1, GRIN2A-B, GRIN3A-B).

In addition to the Drug Repurposing Hub database, we utilized the DrugBank and ChEMBL datasets for chemical similarity searches. Analysis of the DrugBank dataset through t-SNE embedding (Figure S1) revealed a similar list of molecules (Table S1) to those in the Drug Repurposing Hub data set. From the ChEMBL dataset, we used approved drug molecules, which retrieved only four molecules including the query molecule allopregnanolone (Figure S2 & Table S2). This analysis demonstrates that AlloP has few structurally similar clinically approved drugs. Notably, the neighboring molecules of AlloP from the Drug Repurposing Hub dataset encompass targets across all classes found in DrugBank and ChEMBL. Hence, we restricted our further analysis to the targets within the Drug Repurposing Hub dataset.

Network pharmacology and enrichment analysis

The targets identified from the Drug Repurposing Hub dataset offer insights into drug action, safety profile, and therapeutic potential, and contribute to the identification of new targets for NS. To explore the most important targets among the set, we employed protein-protein interaction (PPI) network analysis and enrichment analysis using the STRING database (Figure 2). This analysis shows connections weighted by the degree of published exploration, so thin connections may reveal important but relatively under-investigated pathways. Our network analysis elucidated lists of pivotal targets such as GABA and glutamate receptors, steroid synthesis and signaling, but it also revealed other novel targets such as acetylcholine muscarinic GPCRs (CHRM2 and CHMR3), nicotinic acetylcholine receptors, serotonin receptors, and kinases (Figure 2A). These novel targets offer a future investigation to explore the bioactivity of NS beyond the well-studied GABA-A receptors.

Enrichment analysis helps identify overrepresented biological terms or pathways among the target proteins, elucidating their functional significance. In this study, we conducted gene ontology (GO) enrichment analysis to explore the molecular functions of AlloP neighbor targets (Figure 2B). Among the top fifty significantly enriched molecular functions are GABA receptor activity, glutamate receptor activity and nuclear steroid receptor activity, which are well-known for their involvement with steroid-like molecules (Figure 2B). The analysis revealed other important molecular functions covering multiple neurotransmitter receptors and signaling systems. Many of these have been explored less than GABA-A receptors for activity of NS, and we selected several for our study of the polypharmacology of NS molecules.

Biological activity of neurosteroid analogues

Based on network and molecular function analyses, we selected 47 targets (Table S5) to evaluate the in vitro activity of natural and synthetic NS as shown in Figure 3. In addition, we compared the list of neighboring targets submitted for network analysis (Table S6A) with the set of commercially available targets (Table S6B) included in the screening package. This comparison led to the identification of ten overlapping gene targets, including GABA-A and NMDA receptor subunits, androgen receptor (AR), glucocorticoid receptor (GR), muscarinic receptors CHRM2 and CHRM3, and serotonin receptor HTR3A (Figure S11). We evaluated NS activity at ionotropic glutamate receptor (NMDAR1A/2B), neurotransmitter receptors including GPCRs (muscarinic: CHRM1–3, serotonergic: HTR1–2, dopaminergic: DRD1–2, histaminergic: HRH1–2, adrenergic: ADRA1–2, ADRB1–2), other transmembrane receptor (adenosine: ADOA2A, arginine vasopressin: AVPR1A, cholecystokinin: CCKAR, cannabinoid: CNR1–2, opioids: OPRD1, OPRK1, OPRM1 and insulin: INSR), ligand gated ion channel (serotonergic: HTR3A and nicotine acetylcholine: nAChR(α4/β2), calcium ion channel (Ca_v1.2), and nuclear receptors (AR and GR). We also included neurotransmitter associated transporters; serotonin (SERT), dopamine (DAT) and norepinephrine (NET) and enzymes; acetylcholinesterase (AChE) and monoamine oxidase (MAOA) that were highly involved in neurotransmitter activity functions. Despite kinase activity being situated near the bottom of our enrichment analysis (Figure 2B), we selected kinase targets for activity evaluation due to their novelty and potential functional relevance; these were tested with a subset of NS analogues at a single concentration (10μM). Some targets were also included that did not appear in the enrichment analysis but have strong relevance to side effects or were part of a commercial screening package; examples include KvLQT1, hERG, Cox1-2, EDNRA, PDE3A and PDE4D2.

All 19 NS and analogues (Figure 3) were tested for activity in a concentration dependent manner (ten concentrations, 1.5 nM to 30 μM) to determine EC₅₀/IC₅₀ values at chosen targets. To present results in a compact form, we represented all tested NS activity in heatmaps (Figure 4, 5 and 6), EC₅₀/IC₅₀ values in Table 2 and Table S3, EC₅₀/IC₅₀ value of reference ligands listed in Table S5. In more detail, we will discuss the activity results in the following sections.

Cholinergic receptor activity

Acetylcholine receptors have two classes, ionotropic nicotinic receptors (nAChR) and metabotropic muscarinic receptors (Picciotto et al., 2012). We evaluated NS activity at nicotinic acetylcholine α4β2 receptors (nAChRα4β2) using the FLIPR Tetra System in opener and blocker format at ten concentrations (Figure 4A and B). Among the nineteen NS analogues, twelve exhibited blocker activity at nicotinic α4β2 receptors in a dose dependent manner (Figure 4B and S4A), and six did not show activity (Figure 4A). AlloP exhibited an IC₅₀ value of 6.38 μM, similar to Prog (IC₅₀ 6.48 μM) and the sulphated NS analogue MQ221 (IC₅₀ 6.86 μM), while the unnatural enantiomers of AlloP and Prog did not show any activity (Figure 4B). B260, with a carbonitrile group at carbon 17 (C17), had the highest potency (IC₅₀=3.13 μM), and an unnatural synthetic enantiomer of Etio, with a ketone group at C17, exhibited similar potency (IC₅₀=3.23 μm). By contrast, the 5α diastereomer of B260 (B372) and natural Etio had similar but higher IC₅₀ values of 8.4 μm and 9.95 μM respectively. Other NS analogues including Andro, ent-Andro, alfaxalone, ECN, and MQ34 exhibited low potency (14 to 27 μM).

At muscarinic acetylcholine receptors, the calcium flux assay for CHRM1/3 (Gαq coupled) and cAMP assay for CHRM2 (Gαi/o coupled) were employed in agonist and antagonist modes (Figure 5A and B). Among the molecules tested, only 3α5βPC showed potent agonist activity (EC₅₀=0.56 μM) at CHRM2 in the cAMP assay in a concentration dependent manner (Figure S4C). Conversely, analogue MQ221 only showed antagonistic activity toward CHRM2 (IC₅₀= 9.05 μM), while Prog and ent-Prog exhibited antagonism with IC50 values ranging from 12 μM to 22 μM at CHRM1 and CHRM2 (Figure S4B, and S4D). Finally, ent-Prog also exhibited low-potency antagonism of CHRM3 with IC50 value 23.99 μM (Figure 4B; Table 2).

Serotonergic receptor activity

Next, we examined the NS analogues at the serotonin ligand-gated ion channel HRT3A in both blocker and opener modes using the FLIPR Tetra system (Figure 4). The activity dose response curves in figure S5A demonstrate that eight analogues inhibited HRT3A activity with IC₅₀ values ranging from 6.15 μM to 18.21 μM (Table 2). The closely related NS Andro and Etio showed similar inhibitor potency, with IC₅₀ values of 7.43 μM and 7.35 μM, respectively. Additionally, the enantiomer of Etio (ent-Etio) exhibited similar potency (IC₅₀=6.15 μM), whereas ent-Andro showed low potency (IC₅₀=19.31 μM) compared to Andro (Table 2). Other NS, including ECN (IC₅₀=13.25 μm), B260 (IC₅₀=18.21 μM), MQ34 (IC₅₀=17.55 μM) and the B260 diastereomer (B372), AlloP, and MQ34 showed a clustered IC₅₀ range (~12.33–16.68 μM), somewhat higher than Andro. We also tested serotonin GPCRs including HRT1A and HRT1B (Gαi/o coupled) by cAMP assay and HRT2A and HRT2B (Gαq coupled) by calcium flux assay in agonist and antagonist mode (Figure 5). These receptors are highly involved in various physiological processes, including mood and cognition. The concentration-response relationship suggested that only B260 showed antagonistic activity at HRT1A (IC₅₀=5.14 μM) and HRT1B (IC₅₀=14.58 μM) (Figure S5B and S5C).

Adrenergic receptor activity

We evaluated the activity of analogues at adrenergic receptors including ADRA1A, ADRA2A, and ADRB1–2 (Figure 5A, B). KEGG pathway enrichment analysis for drug targets suggested that the hsa04261 pathway (adrenergic signaling in cardiomyocytes) was significantly enriched but not in the top 50. This analysis outcome was mirrored in the screening results, most analogues showed activity toward ADRB1–2 (Figure 5B, S6A and S6B), receptors that are predominantly expressed in cardiomyocytes and regulate cardiac function (Madamanchi, 2007). Prog, 3α5βPC and B372 showed antagonism with IC₅₀ values of ~5–6 μM at ADRB1, good potency compared to other analogues (Table 2, Figure S6A). Weaker antagonism was observed for ent-Prog (IC₅₀ 13.26 μM), ECN (IC₅₀14.58 μM), and YX23 (IC₅₀ 8.23 μM), ent-AlloP (IC₅₀ 13.23 μM), B260 (IC₅₀ 25.13 μM), and MQ34 (IC₅₀ 15.23 μM) Table 2 (Figure S6A). 3α5βPC and B372 also showed similar antagonistic potency at ADRB2 (4.57 μM and 6.04 μM IC₅₀ respectively), while others including MQ35 (IC₅₀=12.01 μM), YX23 (10.73 μM), Prog (IC₅₀=11.92 μM) and ent-Prog (IC₅₀=14.26 μM) showed modest antagonistic activity (Figure S6B). Andro, ECN and B260 showed poor potency toward ADRB2 (>20 μM IC₅₀). Only ECN (6.82 μM) and MQ221 (1.99 μM) antagonized α adrenergic receptor 2A (ADRA2A) (Table 2; Figure S6C), and none of the steroids showed agonist activity.

Cannabinoid and other GPCRs receptor activity

Several lines of evidence suggest that retrograde endocannabinoid signaling suppresses GABAergic synapses (Alger, 2002; Barna et al., 2024). Consistent with this, the retrograde endocannabinoid signaling pathway (hsa04723) was identified in our KEGG pathway enrichment analysis, as illustrated in figure S3A. Despite a few prior studies of NS activity at cannabinoid receptors (CNR) (Vallée et al., 2014; Haney et al., 2023), we assessed the interaction of NS analogues with CNR1 and CNR2 using a cAMP assay. Interestingly, five compounds exhibited an agonist response with CNR2, while two showed an antagonist response with CNR1 (Figure 5). Specifically, the enantiomer of Etio and 3α5βPC demonstrated CNR2 agonist dose responses with EC₅₀ values of 2.86 μM and 2.75 μM, respectively (Figure S7A). Additionally, the endogenous NS Andro exhibited an EC₅₀ of 4.67 μM, whereas its enantiomer displayed a substantially lower agonist potency with an EC₅₀ value of 19.84 μM (Figure S7A). The unnatural enantiomer MQ35 demonstrated a modest agonist response with EC₅₀ value 7.82 μM at CNR2, while the natural enantiomer MQ34 exhibited no activity. Only two 5β-reduced compounds, ECN and B260, demonstrated CNR1 antagonism with IC₅₀ values of 9.55 μM and 4.89 μM, respectively (Figure S7B).

We tested analogues against several other important families of GPCRs that respond to neurotransmitters/neuromodulators: opioids, histamine, cholecystokinin, and arginine vasopressin (Figure 5). However, with in vitro screening, we did not observe notable activity except for the dopamine receptor 1 (DRD1), histamine receptor 2 (HRH2), and delta opioid receptor (OPRD1). Three analogues including 3α5βPC, B372 and MQ221 exhibited DRD1 antagonist activity with similar range of IC₅₀ values (5.67–8.25 μM; Figure S7C), while ent-Prog had very weak potency (Table 2). Among the tested analogues, only three molecules, including 3α5βPC and Prog, along with its enantiomer, exhibited weak HRH2 antagonistic activity, with IC₅₀ values of 8.89 μM, 25.37 μM, and 18.4 μM, respectively (Table 2; Figure S7D). Only B372 exhibited OPRD1 antagonism with 4.25 μM IC₅₀ (Figure S7E).

Ion Channel Activity

Enrichment analysis and beta-adrenergic receptor activity of NS suggested that NS related compounds have some association with regulation of cardiac function. This notion is also supported by several previous studies (Shiyue et al., 2017; Head et al., 2019; Pan et al., 2024). When we examined our enrichment GO analysis (Figure S3B), we found ion channel complex (GO:0034702) containing ion channels such as NaV1.5, Ca_v1.2, KvLQT1 and hERG highly associated with cardiac functions (Grant, 2009). We tested our NS analogues at all four ion channels (Figure 4). Eight of nineteen analogues inhibited Na_v1.5 channel activity in a dose dependent manner (Figure S8A). The enantiomer of Prog (ent-Prog) and the carbon 3 diastereomer of B260, B372, exhibited ~10 μM IC₅₀ and others including ent-AlloP, Prog, Andro, ECN, MQ34 and MQ35 have IC₅₀’s ranging from 16–26 μM (Table 2). Note that many of these values are considerably lower than the reference compound lidocaine, which is a low-potency blocker (Table S5). Only B260 exhibited inhibitory activity at Ca_v1.2, with an IC₅₀ of 10.71 μM (Table 2; Figure S8B). All nineteen NS also were tested for KvLQT1 and hERG activity and we did not find any response up to 30 μM (Figure 4). In all enrichment analyses, including GO-term and KEGG, glutamate receptor function was strongly represented (Figure 2B, S3A, S3B, and S3C). Nevertheless, none of the NS showed blocker or opener activity at N-methyl-D-aspartate receptors NMDAR (1A/2B) (Figure 4A and B). We note that our compound list contains at least two NS with previously documented inhibitory activity at neuronal NMDARs (Zhu and Vicini, 1997; Mennerick et al., 2001). We address caveats of the screening assays in the Discussion.

Neurotransmitter metabolic enzymes & transporters

Given the actions of steroids at some neurotransmitter metabolic enzymes, we evaluated all nineteen NS analogues at AChE and MAOA. Only Prog was active, inhibiting AChE with IC₅₀ values of 12.48 μM (Figure 6 and S10C). Neurotransmitter transporters are essential for regulating the concentration of neurotransmitters in the synaptic cleft and ensuring the precise control of neurotransmitter signaling (Torres et al., 2003). We evaluated NS analogues activity at multiple neurotransmitter transporters including SERT, DAT, and NET. None of the compounds had effects on any of the transporters (Figure 6).

Nuclear hormone receptor activity

Unsurprisingly, network pharmacology drug-target enrichment data, along with previous work, showed that steroid-like molecules are associated with the GO-term GO:0003707 (nuclear steroid receptor activity; Figure 2B). We focused on nuclear receptor GR and AR for our investigation of NS analogues molecules activity because of limited exploration with NS analogues. Understanding the direct activity of GR and AR is crucial for elucidating their potential effects on various physiological processes. We investigated the activity of the 19 analogues at GR and AR using nuclear protein interaction and nuclear translocation assays, respectively in agonist and antagonist mode (Figure 5). Dose response results suggested that only Prog showed GR antagonist activity with an IC₅₀ value of 3.61 μM (Table 2; Figure S10B). Concentration response of AR activity suggested that all compounds antagonized AR activity except sulphated compound MQ221, and both enantiomers of Etio (Figure S10A), which highlights the differential activity of NS and their analogues at androgen receptors. Notably, the molecules B260, B372, ECN, MQ34, Prog, and YX23 exhibited the most potent antagonistic activity, with IC₅₀ values of 0.53 μM, 0.75 μM, 0.16 μM, 0.12 μM, 0.03 μM, and 0.60 μM respectively (Table 2). Interestingly, the unnatural enantiomers of MQ34 (i.e., MQ35) and Prog (ent-Prog) exhibited IC₅₀ values of 4.27 μM and 2.92 μM respectively; the natural enantiomers were ~35–100 times more potent (Table 2). Other analogues such as CDNC24, AlloP, and Andro also demonstrated moderate antagonistic effects, with IC₅₀ values ranging from 2.52 μM to 4.82 μM (Table 2). Additionally, unnatural enantiomers of AlloP and Andro (e.g., ent-AlloP and ent-Andro) displayed activity similar to the natural molecule, with IC₅₀ values ranging from 4.12 μM to 4.82 μM. However, 3α5βPC exhibited weaker antagonistic activity at AR, with an IC₅₀ value of 26.95 μM.

Kinases and non-kinase enzymes

Some NS-like molecules have shown association with kinase signaling activity, as indicated by our KEGG pathway analysis and previously published work (Adams et al., 2015). Nineteen analogues were evaluated for kinase activity at LCK, VEGFR2, ROCK1, and INSR; no alterations in kinase activity relative to baseline were observed (Figure 6). Additionally, we assessed AlloP and Prog, along with their enantiomers at single concentration (10 μM), for their kinase inhibition activity at 481 kinases. However, we did not observe notable kinase activity inhibition by these NS (Table S3). Furthermore, we investigated the inhibitory effects of all NS on phosphodiesterases (PDEs) and cyclooxygenases (Cox). The results are illustrated in Figure 6. Interestingly, the enantiomer of Prog inhibited the activity of PDE3A and PDE4D2 in a dose dependent manner with IC₅₀ values of 3.45 μM and 9.76 μM, respectively while sulphated NS MQ221 inhibited PDE3A response with 7.05 μM IC₅₀ (Table 2; Figure S9 A and B). None of the NS analogues showed inhibitory activity against Cox-1 and Cox-2 enzymes.