Scaffold hopping from synthetic RXR modulators by virtual screening and de novo design

The concept of virtual screening and automated de novo design has been corroborated as a viable strategy for scaffold hopping from bioactive natural products to isofunctional, synthetically accessible mimetics.

Abstract

The lack of potent subtype-selective modulators of retinoid X receptors (RXRs) has hindered their full exploitation as promising drug targets. Using computational similarity searching, target prediction and automated de novo design, we identified novel RXR ligands exhibiting innovative molecular frameworks, pronounced receptor-subtype preference and suitable properties for hit-to-lead expansion.

The ligand-activated transcription factors retinoid X receptors (RXRs)1 fulfill a unique role among nuclear receptors by acting as universal heterodimer partners of numerous other members of this protein family. The RXR agonist bexarotene is approved for use in the treatment of certain cancers, but this drug is poorly tolerated.2,3 Evidence from animal models points to a considerable therapeutic value of RXR activation in neurodegenerative diseases,4 including multiple sclerosis5 and Alzheimer's disease.6 However, the lack of RXR modulators that selectively address the receptor subtypes RXRα, RXRβ and RXRγ, their often poor drug-likeness and their toxicity have prevented further studies and the expanded therapeutic use of RXR ligands in humans.7,8

The available RXR ligands beyond the putative endogenous agonist 9-cis retinoic acid9 (1a, Scheme 1) include several potent agonists and antagonists. However, their chemotype diversity is limited, and RXR agonists (e.g., 1a–c) and antagonists (e.g., 2a–c) share a small number of scaffolds (Fig. 1).10 90% of the known RXR modulators with a potency (EC₅₀ or IC₅₀) up to 50 μM described in ChEMBL11 contain only seven distinct reduced scaffolds.12,13 Moreover, most of them differ only in the attachment positions and lengths of the linkers (0 or 1 atoms), thus confirming the narrow chemical diversity of the known RXR ligands.

Scheme 1. RXR ligands: putative endogenous agonist 9-cis retinoic acid (1a), drug approved RXR agonist bexarotene (1b), the bexarotene analogue 1c and RXR antagonists 2a–c.

Fig. 1. Scaffold analysis of ChEMBL annotated RXR ligands in terms of most-frequently occurring scaffolds and reduced scaffolds. 90% of known RXR modulators (n = 521, agonists and antagonists) with a potency (EC₅₀ or IC₅₀) up to 50 μM share only seven different reduced scaffolds. The scaffolds display limited structural diversity, merely differing in linker length (0–1 atoms) and linker attachment positions.

Although the structure–activity relationship (SAR) of classical rexinoids has been extensively studied, no remarkable improvements in drug-likeness or subtype selectivity have been achieved to date. The discovery of novel chemotypes of RXR ligands will be mandatory for future drug development. To address this issue and identify RXR-targeting small molecules with novel molecular frameworks, we have employed virtual screening and computational de novo design as complementary approaches for scaffold hopping from known to new RXR modulators.

We computationally searched a collection of commercially available screening compounds (3.38 million compounds from selected providers) for potential RXR ligands by using two complementary techniques. First, the probability of RXR modulation was calculated for all compounds by the target prediction software SPiDER.14 This method is based on 2D pharmacophoric feature distributions and physicochemical properties. The compounds were then ranked according to the lowest p-values (false-positive potential) for RXR modulation. We used the chemically advanced template search (CATS)15 as the second in silico tool. CATS ranks all screening compounds by their 2D pharmacophore similarity to the known RXR ligands 1a–c. The SPiDER and CATS rankings were then merged, and from this consolidated list, we selected compounds 3–19 for in vitro characterization, considering their individual ranks and scaffolds.

Compounds 3–19 were analyzed in specific Gal4 hybrid reporter gene assays16,17 for the modulation of RXRα, RXRβ and RXRγ.‡ The compounds were tested for RXR activation and for RXR antagonism in competition with 1 μM bexarotene (1b). Single-point testing (30 μM) on RXRα, RXRβ and RXRγ confirmed six of the 17 candidates as RXR modulators, including partial agonists (12, 13, 19) and antagonists (5, 14, 15). For the RXR activating compounds, control experiments without the hybrid receptor showed no transactivation of the reporter gene, thus confirming that the observed effects are RXR-mediated (Fig. S1A†). Moreover, their RXR activating activity could be blocked with the RXR antagonist HX531 (2c, 1 μM) (Fig. S1B and C†). Full dose–response characterization revealed intermediate micromolar potency, despite the low transactivation efficacy for the partial agonists, which possess unique molecular frameworks that previously were not associated with RXR modulation (Fig. 2). Notably, the uncharged structure of 19 enables RXR modulator development without the acidic residue typically contained in RXR ligands.18

Fig. 2. Diversity of novel RXR modulators: (A) structures and Bemis–Murcko scaffolds of new RXR modulators. (B) A log P and A log S values26 of newly identified RXR modulators compared to ChEMBL-annotated RXR ligands.

The RXR antagonists 5, 14 and 15 displayed low micromolar IC₅₀ values (assay agonist: 1 μM bexarotene (1b)) and possessed unique scaffolds including the non-acidic structure of 5. Interestingly, 14 and 15 revealed opposite subtype preferences, with 14 favoring RXRα and 15 being more potent for RXRβ and RXRγ. Since 14 and 15 possess similar scaffolds but differ in their substitution patterns, their SAR may provide an opportunity for subtype-selective RXR ligand development (cf. Fig. S2†).

Among the newly identified RXR antagonists, compound 14 showed the highest efficacy, reducing the bexarotene-induced RXR activation to nearly baseline levels. We rescreened the collection of screening compounds for analogues with similar molecular frameworks to obtain a preliminary SAR of this ligand class. 249 compounds that retained the central 1,2,3-substituted five-ring scaffold of 14 were retrieved and computationally ranked for their similarity to the known RXR antagonists 2a–c using Morgan fingerprints (1024 bit, RDKit20), CATS and LIQUID21 descriptors. The combination of these three computational metrics covers fundamental aspects of the molecular structure as well as the 2D and 3D pharmacophores. Six compounds, including the query antagonist 14, appeared two or more times among the top-5 of these similarity rankings, thus confirming the validity of this approach. Of the remaining five compounds, four (20–23) were readily available and were characterized in vitro (Table 1).

Table 1. In vitro activity of RXR modulators obtained by virtual screening (mean ± SEM, n ≥ 4 independent experiments).

ID	RXRα	RXRβ	RXRγ
5	IC50 = 7 ± 2 μM	IC50 = 10.9 ± 0.3 μM	IC50 = 8.2 ± 0.8 μM
12	EC50 = 21 ± 1 μM (2.9 ± 0.1-fold act.)	Inactive (30 μM)	Inactive (30 μM)
13	EC50 = 21 ± 3 μM (3.0 ± 0.5-fold act.)	EC50 = 42 ± 1 μM (8.2 ± 0.1-fold act.)	EC50 = 51 ± 9 μM (12 ± 4-fold act.)
14	IC50 = 4.8 ± 0.7 μM	IC50 = 12.3 ± 0.4 μM	IC50 = 18.1 ± 0.1 μM
15	IC50 = 18 ± 4 μM	IC50 = 6 ± 1 μM	IC50 = 4.9 ± 0.5 μM
19	EC50 = 27 ± 3 μM (6.2 ± 1.0-fold act.)	EC50 = 37 ± 1 μM (3.3 ± 0.1-fold act.)	EC50 = 24 ± 6 μM (2.3 ± 0.1-fold act.)
20	Inactive (30 μM)	Inactive (30 μM)	Inactive (30 μM)
21	IC50 = 2.0 ± 0.8 μM	IC50 = 18 ± 2 μM	IC50 = 1.8 ± 0.1 μM
22	IC50 = 12 ± 7 μM	IC50 = 14 ± 4 μM	IC50 = 17 ± 8 μM
23	IC50 = 12 ± 4 μM	IC50 = 11 ± 2 μM	IC50 = 18 ± 1 μM
1b 19	EC50 = 33 ± 2 nM	EC50 = 24 ± 4 nM	EC50 = 25 ± 2 nM

Compounds 21–23 showed RXR antagonistic activity in the reporter gene assays, whereas compound 20 was inactive. The most linear analogue 21 possessed the highest potency (IC₅₀ = 2.0 ± 0.8 μM on RXRα). The close analogues 22 and 23, differing from 14 only in their aromatic ring substituents, were slightly less potent and displayed IC₅₀ values of 10–20 μM on all receptor subtypes. Analogue 20, with a 4-chlorophenyl residue comprising the largest 4-substituent at the 5-ring core, was inactive. Together, these results draw a preliminary SAR of 1,2,4-triazole based RXR antagonists, suggesting that small moieties in 4-position are preferred and that a linear extension of the lipophilic 5-substituent might enhance RXR antagonistic potency.

Then, we employed the computational de novo design software design of genuine structures (DOGS)22 to obtain new chemotypes. The software constructs new molecules from a collection of commercially available building blocks and a set of synthetic reaction schemes. DOGS has previously been proven to generate synthetically accessible, isofunctional mimetics of the given template.23 We used seven known RXR ligands from the literature and our in-house library of nuclear receptor modulators as templates for DOGS and retrieved a total of 2308 de novo designs. The designs were then ranked in silico using the SPiDER and CATS software. Additionally, the frequency of their generation from different templates was analyzed. From the consolidated ranking, we selected designs 24–26 for synthesis by considering their individual ranks and building block availability.

Compounds 24–26 were accessible in single-step procedures (Scheme 2). Suzuki reaction of 1-bromo-3,5-di-tert-butylbenzene (28) with 4-boronobenzoic acid (29) and of 3-(5-(4-chlorophenyl)oxazol-2-yl)propanoic acid (30) with phenylboronic acid (31) yielded 24 and 25, respectively. 26 was available through Paal–Knorr reaction of 5-amino-1-(3-chlorophenyl)-1H-pyrazole-4-carbonitrile (32) with hexane-2,5-dione (33).

Scheme 2. Synthesis of de novo designs 24–26: reagents & conditions: (a) Pd(PPh₃)₄, Cs₂CO₃, toluene/EtOH, reflux, 16 h, 41%; (b) C₆H₅-B(OH)₂ (31), Pd(dppf)Cl₂, Cs₂CO₃, toluene, 80 °C, 16 h, 31%; (c) hexane-2,5-dione (33), montmorillonite K10, μW, 90 °C, 30 min, 89%.

The computationally generated compounds 24–26 were then characterized in reporter gene assays, which confirmed designs 24 and 26 as RXR agonists, while 25 partially antagonized the RXR transactivation (Table 2). However, a scaffold similar to 25 has been associated with luciferase inhibition;24 therefore, the activities of 25 should be interpreted cautiously. Agonist 24 displayed low micromolar potency for all three subtypes but revealed remarkable subtype preference for RXRβ and RXRγ in terms of transactivation efficacy. It behaved as a moderate partial agonist for RXRα but a strong full agonist for RXRβ and RXRγ. By exploiting this particular selectivity profile, 24 can serve as a starting point for the development of subtype-specific RXR modulators.

Table 2. In vitro activity of de novo designs 24–26 (results are mean ± SEM, n ≥ 4 independent experiments).

ID	RXRα	RXRβ	RXRγ
24	EC50 = 11.8 ± 0.7 μM (5.7 ± 0.4-fold act.)	EC50 = 11.7 ± 0.1 μM (70.6 ± 0.2-fold act.)	EC50 = 14.1 ± 0.1 μM (58.3 ± 0.2-fold act.)
25	IC50 = 11 ± 8 μM	IC50 = 22.4 ± 0.4 μM	IC50 = 6 ± 1 μM
26	EC50 = 19 ± 5 μM (3.6 ± 0.4-fold act.)	EC50 = 16 ± 2 μM (11.1 ± 0.3-fold act.)	EC50 = 17 ± 1 μM (8.1 ± 0.2-fold act.)
1b 19	EC50 = 33 ± 2 nM	EC50 = 24 ± 4 nM	EC50 = 25 ± 2 nM

Using virtual screening and an automated de novo design, we have identified a total of twelve novel RXR modulators comprising ten different scaffolds. Although the query/template compounds contained only a few molecular frameworks, we retrieved novel RXR modulator chemotypes by both in silico approaches. The new RXR modulators are markedly different from the ChEMBL-annotated RXR agonists and antagonists (IC/EC₅₀ < 50 μM, N = 521) in terms of their scaffolds, with the average Jaccard–Tanimoto similarity (extended connectivity fingerprints,25 radius = 0 to 4 bonds, 1024 bit) below 0.20 ± 0.02 (Fig. S3†). In vitro characterization indicated that several of the new ligands possess subtype preference, thus facilitating the development of subtype-selective RXR modulators. Most of the new RXR ligands appear superior to the ChEMBL-annotated compounds in terms of both lipophilicity and solubility, as indicated by their lower Alog P and higher Alog S values (Fig. 2B).

Our computational scaffold hopping approach yielded several new lead compounds that expand the SAR of RXR modulators. These chemical entities provide a variety of innovative molecular frameworks to address key challenges in RXR targeting by enabling improved drug-like properties and greater receptor subtype selectivity. The concept of combined computational similarity searching and de novo design has been corroborated as a viable strategy for chemical scaffold hopping.

Conflicts of interest

G. S. declares a potential financial conflict of interest in his role as life-science industry consultant and cofounder of inSili.com GmbH, Zurich.