Rewarding compounds identified from the medicinal plant Rhodiola rosea

Rhodiola rosea L. (Sedum roseum) is a folk-medicine plant of the family Crassulaceae. In accordance with historical reports of the use of root material from this plant to treat headaches and ‘hysteria’ (Linnaeus, 1749), R. rosea products have been used for stress relief and the focusing of attention (Wiegant et al., 2009; Panossian et al., 2010). Contemporary scientific literature has reported increases in life span as a result of R. rosea food supplementation in the worm Caenorhabditis elegans and the fly Drosophila melanogaster (Jafari et al., 2007; Wiegant et al., 2009; Gospodaryov et al., 2013; Arabit et al., 2018), as well as effects related to nervous system function. For example, food supplementation with R. rosea root material or extract enhances odor–reward associative memory in Drosophila larvae and can partially compensate for age-related memory decline in adult flies (Michels et al., 2018). In bees, acute food supplementation with R. rosea extract can enhance the acquisition and the consolidation of such associative memories (Michels et al., 2018). In neither case was behavior towards the odor or the reward affected. Memory-enhancing effects have also been reported in rodents and humans (Petkov et al., 1986; Panossian et al., 2010; Michels et al., 2018; but see Ishaque et al., 2012). However, whether one and the same chemical constituent of R. rosea mediates all these effects has remained unclear (for more details, please see Discussion). Indeed, identification of the respective bioactive compound(s) is critical not only for drug development but also for quality control during the production of R. rosea preparations, and for the dosage of R. rosea products. Possibly as a result of variation in their effect associated with variation in the concentration of the effective compounds, which cannot be determined as long as the effective compounds themselves have not been identified, the European Medicine Agency (EMA) and the National Institutes of Health of the USA (NIH) have issued skeptical reports (European Medicines Agency, 2012; https://nccih.nih.gov/health/rhodiola, accessed 27 July 2020). We have therefore endeavored to identify bioactive compounds from R. rosea, using for a bioassay the rewarding effect of its roots, which was discovered in the course of the present study.

So far, about 140 compounds have been detected in the roots of R. rosea, including phenylethanoids and phenylpropanoids, monoterpene alcohols, cyanogenic glycosides, flavonoids, proanthocyanidins and gallic acid derivatives (Panossian et al., 2010, and references therein). The observed effects have been considered to be mainly based on phenylethane and phenylpropane derivatives such as tyrosol and cinnamic alcohol and derived glycosides such as salidroside and rosavin. Thus, several commercial R. rosea preparations are standardized for their salidroside and rosavin content (Elameen et al., 2010). High rosavin levels seem to be characteristic for R. rosea and are used for differentiation from related species (Panossian et al., 2010). The main chemical classes of volatile compounds are monoterpene hydrocarbons, monoterpene alcohols and straight-chain aliphatic alcohols (Rohloff, 2002).

The starting point of the present study was the peculiar rose-like smell of ground R. rosea root material. Indeed, the common name of R. rosea in English is rose root. The monoterpene geraniol, its formate and acetate derivatives, as well as phenylethyl alcohol contribute to this rose-like odor (Rohloff, 2002). We report that for Drosophila larvae, a genetically tractable and placebo-clear model system, root material from R. rosea has an attractive smell as well as an attractive taste, and that R. rosea has a strong rewarding effect in these animals. We followed up on this latter observation and initiated a bioassay-guided fractionation approach supported by correlation analyses to identify the rewarding compound(s) in R. rosea (Degenhardt et al., 2014; Hielscher-Michael et al., 2016). We will briefly discuss whether these rewarding effects indicate a habit-forming or addictive potential, and weigh these effects against the recently reported memory-enhancing effect of both R. rosea and at least one of its compounds that are here identified as rewarding (Michels et al., 2018).

Crudely ground R. rosea root material was bought from Eveline24.de (Maardu, Estonia; Ch./ Lot: 17884). Voucher samples (QGB011; corresponding to the accession Rhodiola⁴) (Michels et al., 2018) are deposited at the IPB (Halle, Germany) and the LIN (Magdeburg, Germany). The traditional medicinal species R. rosea L. (Crassulaceae) is at present treated as a synonym of the accepted species Sedum roseum (L.) Scop. (accessed 1 January 2013: http://www.theplantlist.org/).

¹H nuclear magnetic resonance (NMR) and ¹³C NMR spectra were recorded in solutions on a Varian Mercury 400 spectrometer at 400 MHz and 101 MHz, respectively. Chemical shifts (δ) are reported in ppm and were referenced to TMS (¹H NMR) or to the solvent signal (¹³C NMR). Coupling patterns are designated as s(inglet), d(oublet), t(riplet), q(uartet), m(ultiplet), ps(eudo) and br(oad). The low-resolution electrospray (ESI) mass spectra were recorded on a SCIEX API-3200 instrument (Applied Biosystems, Concord, ON, Canada) combined with an HTC-XT autosampler. GC-EI-MS measurements were obtained using a QP-2010 Ultra system (Shimadzu Corporation, Kyoto, Japan) equipped with a Phenomenex column (ZB-5MS, 30 m×0.25 µm) after trimethylsilylation of the compounds. For headspace gas chromatography mass spectrometry (GCMS) without derivatization, the following conditions were applied: incubation temperature 40°C, incubation time 30 min, syringe temperature 45°C, split injection (1:5), detected mass-to-charge ratio (m/z) 40–500, injector port temperature 220°C, source temperature 220°C, interface temperature 300°C, carrier gas helium, column flow 1.02 ml min⁻¹, column temperature program: 40°C for 1 min, then raised to 300°C at a rate of 10°C min⁻¹ and held on 300°C for 5 min. Before measurement, the sample was wetted with water. Compounds were identified by comparing their electron ionization-mass spectrometry (EI-MS) spectra with the NIST 08 or NIST 11 database. The relative abundance of components was calculated based on the peak areas of the total ion chromatogram. Ultra-high performance liquid chromatography-high-resolution mass spectrometry (UHPLC-HRMS) measurements were performed as previously described (Coors et al., 2019). TLC was performed on silica gel 60 F₂₅₄ plates (Merck). Spots were visualized using UV light and vanillin-H₂SO₄ detection reagent. For preparative column chromatography (CC), silica gel 60 (63–200 µm and 40–63 µm) was used.

High-resolution electrospray ionization (ESI) mass spectra were obtained from a Bruker Apex III Fourier transform ion cyclotron resonance (FTICR) mass spectrometer (Bruker Daltonics, Billerica, MA, USA) equipped with an Infinity™ cell, a 7.0 T superconducting magnet (Bruker, Karlsruhe, Germany), a radio-frequency-only hexapole ion guide and an external APOLLO electrospray ion source (Agilent, off-axis spray, voltages: negative ion mode: endplate, 3700 V; capillary, 4200 V; capillary exit, −100 V; skimmer 1, −15.0 V; skimmer 2, −6.0 V; offset −3.75 V; positive ion mode: endplate, −3700 V; capillary, −4.200 V; capillary exit, 100 V; skimmer 1, 15.0 V; skimmer 2, 12.0 V; offset 3.75 V). Nitrogen was used as a drying gas at 150°C. The sample solutions were introduced continuously via a syringe pump with a flow rate of 120 µl h⁻¹. The instrument was externally calibrated using the ES tuning mix (Agilent G2422A). The data were acquired with 512 k data points and zero-filled to 2048 k by averaging 12 scans (ion accumulation: 1 s). The mass spectra were evaluated using the Bruker software XMASS 7.0.8.

To relate the metabolite profiles obtained by direct-infusion ESI-FTICR-MS to the rewarding activity of the respective plant-derived materials, reverse metabolomics with the AcorA algorithm was used (Degenhardt et al., 2014; Hielscher-Michael et al., 2016). In brief, from the fractions resulting from liquid–liquid partitioning of the R. rosea crude extract and the sub-fractions E1–E5 and H1–H4, 1 ml of a solution in methanol (10 mg ml⁻¹) was purified using a Chromabond^® RP18 SPE cartridge (Macherey-Nagel). The resulting samples were adjusted to a concentration of 1 mg ml⁻¹ and 100 µl of the solution was diluted with 500 µl methanol, centrifuged and then subjected to triplicate high-resolution electrospray ionization mass spectrometry (HRESIMS) measurements. The ESI-FTICR-MS files were converted to the mzData file format. After peak-picking and aligning the data from the positive ionization mode using the R package XCMS (Smith et al., 2006), 936 different m/z features were detected. To determine the correlation between the rewarding activity and the mass signal intensities, the Spearman rank correlation coefficient was calculated after performing a permutation test with 500 repeats. This AcorA resulted in the identification of 51 positively correlating features (Table 1; Table S1). The elemental composition of the correlating features was calculated based on the ESI-FTICR-MS results. The ESI-FTICR-MS data from the negative ionization mode could not be correlated because of an unsuccessful R-data conversion and the limitations of the Bruker software.

Finely chopped R. rosea root (1.565 kg) was exhaustively extracted with 80% aqueous un-denatured ethanol (3×6 l) at room temperature. The extracts were combined, filtered and the organic solvent evaporated under reduced pressure until the aqueous residue remained. This crude extract was successively partitioned with 3×500 ml n-heptane, ethyl acetate and n-butanol, resulting, after removal of the solvents, in fractions of different polarity: n-heptane (Rho_heptane, 8.72 g), ethyl acetate (Rho_{ethyl acetate}, 25.66 g), n-butanol (Rho_butanol, 92.42 g) and water (Rho_water, 312.01 g). Following assessment of their influence as a reward in associative learning in larval Drosophila, the active fractions were further separated according to polarity using silica gel column chromatography.

A 3.40 g sample of the Rho_{ethyl acetate} fraction was sub-fractionated on a short column, eluting with a gradient system: 250 ml n-hexane, 300 ml chloroform (combined E1, 0.38 g), 300 ml ethyl acetate (E2, 1.40 g), 200 ml ethyl acetate/methanol (8:2; E3, 0.51 g), 300 ml methanol (E4, 0.34 mg) and 300 ml methanol containing 2 ml 2 mol l⁻¹ HCl (E5, 1.15 g).

A part of the active n-heptane fraction (4.90 g Rho_heptane) was sub-fractionated using n-heptane (H1, 1.22 g), CHCl₃ (H2, 0.66 g), ethyl acetate (H3, 2.06 g) and methanol (H4, 0.60 g), 500 ml in each case, as eluting solvents to give the respective sub-fractions.

The resulting sub-fractions E1–E5 and H1–H4 were then assayed for their activity as a reward as described below.

The most active fractions and compounds associated with the best-correlating peaks from AcorA (see above) were separated and characterized as detailed in the following.

The most active sub-fraction (E5) was further separated by column chromatography on silica gel, eluting with a chloroform/methanol gradient to obtain β-sitosterol-β-d-glucoside (32.7 mg) as the major constituent.

β-Sitosterol-β-d-glucoside (BSSG): ¹H NMR (400 MHz, DMSO) δ [ppm]: 5.329 (brs, 1H, H-6); 4.220 (d, J=7.9 Hz, 1H, H-1′); 0.957 (s, 3H, H-19); 0.902 (d, J=6.6 Hz, 3H, H-21); 0.651 (s, 3H, H-18). ¹³C NMR (100 MHz, DMSO) δ [ppm]: 140.5 (C5), 121.2 (C6), 100.8 (C1′), 77.0 (C3), 76.7 (C3′), 76.7 (C5′), 73.4 (C2′), 70.0 (C4′), 61.0 (C6′), 56.2 (C14), 55.4 (C17), 49.6 (C9), 45.2 (C24), 41.9 (C13), 39.3 (C12, under DMSO signal), 38.3 (C4), 36.8 (C1), 36.2 (C10), 35.5 (C20), 33.4 (C22), 31.44 (C8), 31.4 (C7), 29.3 (C2), 28.7 (C25), 27.8 (C16), 25.5 (C23), 23.9 (C15), 22.6 (C28), 20.6 (C11), 19.7 (C27), 19.1 (C19), 18.9 (C26), 18.6 (C21), 11.8 (C29), 11.7 (C18). The NMR data are in accordance with Faizi et al. (2001).

Also from the E5 sub-fraction, ferulic acid eicosyl ester was isolated, ESI-MS guided by the occurrence of the ion at m/z 497 [M+Na]⁺ of the correlating feature 3 (Table 1) in positive ionization mode, as previously described (Michels et al., 2018). Therefore, an aliquot of the ethyl acetate extract (Rho_{ethyl acetate}, 7.8 g) was fractionated in analogy to the above-described sub-fractionation procedure on a short silica gel column, eluting with 400 ml chloroform (3.79 g), 300 ml ethyl acetate (2.50 g) and 200 ml methanol (1.30 g). The fraction eluted with chloroform was subjected to repeated column chromatography on silica gel using a gradient system (n-hexane:CHCl₃ 8:2 gradient to 0:1, followed by ethyl acetate and methanol), followed by column chromatography with CHCl₃ affording 15.5 mg ferulic acid eicosyl ester in a 3:1 mixture as trans and cis (interconvertible) diastereomers, 13.3 mg fatty alcohol and 2.2 mg stigmast-4-en-3-one.

Ferulic acid eicosyl ester (FAE-20): ESI-MS, 473.3647 [M-H]⁻ calculated for C₃₀H₄₉O₄⁻ 473.3636 accompanied by homologs bearing a C₁₈ or C₂₂ fatty alcohol chain: 445 [M-H]⁻ in accordance with C₂₈H₄₅O₄⁻ (FAE-18); and 501.3959 [M-H]⁻ calc. for C₃₂H₅₃O₄⁻ 501.3949 (FAE-22). ¹H NMR (400 MHz, CDCl₃) δ [ppm]: trans: 7.608 (d, J=16 Hz, 1H, H-7); 7.075 (dd, J=8, 2, 2H, H-6); 7.037 (d, J=2 Hz, 1H, H-2); 6.913 (d, J=8 Hz, 1H, H-5); 6.291 (d, J=16 Hz, 1H, H-8); 5.993 (s, 1H, OH); 4.187 (t, J=7 Hz, 2H, H1′); 3.930 (s, 3H, OMe); 1.638 (m, 2H, H-2′); 1.45–1.21 (m, 10H); 0.879 (t, J=7 Hz, 3H, Me-20′). cis: 7.756 (d, J=2 Hz, 1H, H-2); 7.103 (dd, J=8, 2, 1H, H-6); 6.878 (d, J=8, 1H, H-5); 6.795 (d, J=12.7 Hz, 1H, H-7); 5.948 (brs, 1H, OH); 5.815 (d, J=12.5 Hz, 1H, H-8); 4.117 (t, J=7 Hz, 2H, H-1′); 3.930 (s, 3H, OMe). ¹³C NMR (100 MHz, CDCl₃) δ [ppm]: trans: 167.4 (C9); 147.9 (C3); 146.7 (C4); 144.6 (C7); 127.0 (C1); 123.0 (C6); 115.6 (C8); 114.7 (C5); 109.3 (C2); 64.6 (C1′); 55.9 (OMe); 31.9 (C2′); 29.7; 29.63; 29.58; 29.5; 29.34; 29.28; 28.8; 26.0; 22.7 (C3′-C19′); 14.1 (Me). cis: 166.7 (C9); 147.0 (C4); 143.7 (C7); 127.0 (C1); 125.5 (C6); 116.9 (C5); 113.8 (C8); 112.7 (C2); 64.4 (C1′); 55.9 (OMe). The NMR data are in agreement with Hennig et al. (2011) and Baldé et al. (1991).

Fatty alcohols: gas chromatography electron ionization mass spectrometry (GC-EI-MS): octadecanol 1.6%, nonadecanol 0.4%, eicosanol 69.2%, heneicosol 1.3%, docosanol 27.0%, tetracosanol 0.5%. ¹H NMR (400 MHz, CDCl₃) δ [ppm]: 3.639 (t, J=7.0 Hz, 2H); 1.566 (m); 1.360–1.255 (m); 0.880 (t, J=7.0, 3H).

Stigmast-4-en-3-one: ESI-MS, m/z: 413 [M+H]⁺, 435 [M+Na]⁺, 825 [2M+H]⁺. ¹H NMR (400 MHz, CDCl₃) δ [ppm]: 5.722 (s, 1H, H-4); 1.181 (s, 3H, H-19); 0.926–0.782 (m, H-21, H-26, H-27, H-29); 0.710 (s, 3H, H-18).

A part of the sub-fraction H3 was purified by column chromatography on silica gel, eluting with CHCl₃ (400 ml), followed by CHCl₃:methanol 9:1 (300 ml) and methanol (300 ml), to obtain an unsaturated fatty acid fraction (FA_isolated) consisting mainly of linoleic acid (51.2 mg).

Linoleic acid: ESI-MS, m/z: 279 [M-H]⁻, 303 [M+Na]⁺. ¹H NMR (400 MHz, CDCl₃) δ [ppm]: 5.417–5.301 (m, -CH=CH-); 2.823–2.756 (m, =CH-CH₂-CH=); 2.347 (t, -CH₂-COOH); 2.098–1.989 (m, -CH₂-CH=); 1.651–1.597 (m, -CH₂-CH₂-COOH); 1.393–1.256 (m, -CH₂-); 0.890 (t, CH₃).

β-Sitosterol-β-d-glucoside and trans-ferulic acid eicosyl ester were synthesized as previously described (Michels et al., 2018).

Wild-type D. melanogaster larvae were used throughout, unless mentioned otherwise. They were of the Canton-Special (CS) strain and aged 5 days after egg laying, still in their feeding stage. Animals were kept in mass culture, maintained at 25°C, 60–70% relative humidity and a 12 h:12 h light:dark cycle.

The anosmic Orco¹ mutant strain (Bloomington Stock Center no. 23129) carries a loss-of-function allele of the Orco gene and has been described previously (Larsson et al., 2004; Vosshall and Hansson, 2011). The mushroom body driver strain OK107-Gal4 (Connolly et al., 1996), the split-Gal4 strains APL-Gal4 (SS01671), DAN-h1-Gal4 (SS01696) (Saumweber et al., 2018) and DAN-h1, DAN-i1, DAN- j1, DAN- k1-Gal4 (with DAN-j1 included only stochastically; SS01948) (Eschbach et al., 2020) were crossed to UAS-Kir2.1::GFP (Kir^GFP) (Saumweber et al., 2018) as the effector for neuronal silencing in the experimental genotypes. To obtain the driver controls, the Gal4 driver strains were crossed to the CS wild-type strain, as this was the genetic background of the effector strain that was omitted from the cross (Saumweber et al., 2018). As effector controls, a strain homozygous for both the attP40 and attP2 landing sites, but without a Gal4 inserted, was the genetic background of the split-Gal4 driver strains (‘empty’ Gal4; Pfeiffer et al., 2010) and was therefore used for generating the effector control when the experimental genotypes featured a split-Gal4 driver element; for the experiment featuring OK107-Gal4, the genetic background of the driver strain was w¹¹¹⁸ and this strain was used for generating the effector control; in both cases, the crossings yielding the effector control were carried out with UAS-Kir2.1::GFP. Experimenters were blind to genotype.

To verify effector expression from UAS-Kir2.1::GFP, larval brains of experimental genotypes were dissected in Ringer solution and fixed in 4% paraformaldehyde dissolved in PBS for 30 min. After three washes (each 10 min) in PBST, the brains were treated in blocking solution containing 3% normal goat serum (Dianova) in PBS for 2 h. Tissue was then incubated overnight with the primary antibodies rabbit anti-GFP (1:1000; Invitrogen A11122) and mouse anti-FASII (1:50; DSHB). Six washing steps in PBS (each 10 min) were followed by incubation with Alexa Fluor 488 goat anti-rabbit (1:200; Invitrogen A11034) and Cy3 donkey anti-mouse (1:200; Jackson ImmunoResearch Laboratories 715-165-150) as secondary antibodies. After final washing steps with PBS, samples were mounted in Vectashield (Linaris) and scanned under a confocal microscope. Projections of the stacks were accomplished with Fiji ImageJ software.

For the behavioral experiments, the R. rosea root was finely ground using a Precellys 24 homogeniser (VWR, Erlangen, Germany). For simplicity, the resulting root powder will be called Rhodiola throughout the paper. For the behavioral experiments, Petri dishes (Sarstedt, Nümbrecht, Germany) with an inner diameter of 60 mm were filled with 1% agarose (electrophoresis grade; Roth, Karlsruhe, Germany), allowed to solidify, covered with their lids, and then left untreated at 4°C until the next day. To serve as a potential reinforcer, these could be supplemented with 0.1, 1.0 or 10.0 mg ml⁻¹ Rhodiola mixed into a 1% agarose solution shortly after boiling the agarose solution.

All the behavioral experiments were performed under natural light at 22–25°C. Animals were randomly collected from the food vial by removing a spoonful of food medium containing larvae and were transferred to a tap-water droplet on a plain plastic Petri dish. From this droplet, cohorts of 15 animals were collected, briefly washed in tap water and transferred for the respective experiment.

The odor–reward associative learning experiments follow the procedures introduced by Scherer et al. (2003) and Neuser et al. (2005) for the association of odor with fructose as the reward in the two-group, one-odor version introduced by Saumweber et al. (2011) (for a detailed manual, see Michels et al., 2017) – with the modification that instead of fructose, Rhodiola was used. In brief, an odor was presented to separate groups of animals, either paired or unpaired with Rhodiola during training, followed by an odor preference test. A higher test preference for the odor after paired versus unpaired training indicates appetitive associative memory; that is, it indicates that Rhodiola was effective as a reward.

Odor was applied by adding 10 µl of n-amyl acetate (AM; CAS: 628-63-7; purity: 98.5%, diluted 1:50 in paraffin oil; Merck, Darmstadt, Germany) to the Teflon containers described above. These were placed into Petri dishes with either plain agarose as the substrate or agarose supplemented with Rhodiola at the amounts mentioned in Results.

For the learning experiments, the regular lids of the Petri dishes were replaced by lids perforated in the center by 15 holes of 1 mm diameter to improve aeration. Then, a cohort of 15 larvae was placed into a Petri dish featuring, for example, AM as the odor and Rhodiola as the reward. The Petri dish was closed with its lid and the larvae were allowed to move freely for 5 min. Then the larvae were transferred to a Petri dish with an empty odor container (Em) and agarose as the substrate. Such paired training (AM+/Em) was repeated two more times. In half of the cases, the sequence of training trials was as indicated in this example; in the other half it was the reverse (Em/AM+). Unpaired training consisted of separate presentations of odor and Rhodiola (AM/Em+ or when using the reverse sequence of trials Em+/AM). After either paired or unpaired training, the larvae were tested for their odor preference by allowing them to choose between two sides of a test Petri dish, one with an AM-filled odor container and the other with an empty odor container (the test Petri dish featured agarose alone as the substrate). After 3 min, the number of animals on the odor side and no-odor side and in the neutral zone (5 mm) was counted and the odor preference determined, with due adjustments, as in Eqn 1.

Again, scores ranged between −1 and 1. Given that larvae of both groups were equated for handling and stimulus exposure and thus for the effects of non-associative learning, positive PI scores indicate appetitive associative memory, i.e. a rewarding effect exerted by Rhodiola. Negative scores indicate aversive associative memory, i.e. a punishing effect exerted by Rhodiola.

The Petri dishes were coded such that experimenters were blind with respect to whether the Petri dishes featured Rhodiola or not. This was decoded only after the experiment.

For the control group, the procedure was identical to the above (including mock-codes of the Petri dishes), except that Rhodiola was omitted. Data for the Rhodiola and the control groups were gathered alternately. This is essential for meaningful statistical comparisons, given the variability of memory scores in the larvae (and indeed in Drosophila and insects in general).

The one-trial version of the learning paradigm followed the same principle, without repeating the training cycle. To match our procedures to a recent report of one-trial learning with various other tastant reinforcers (Weiglein et al., 2019), these experiments featured 30 larvae, Petri dishes of 90 mm diameter, AM at a 1:20 dilution as the odor, and training trials that lasted 4 min, unless mentioned otherwise.

Exactly the same learning protocols as described above were used whenever either fructose (2 mol l⁻¹; purity: 99%; Roth, Karlsruhe, Germany) or Rhodiola-related candidate rewarding substances were used (e.g. those with a positive AcorA value). If these substances were added to the agarose in dissolved form, using an appropriate solvent, we also included solvent groups in our experiments. In other words, the only difference between the experimental groups, trained with the candidate rewarding compound, and the corresponding solvent group was whether or not the candidate compound was present. Except for the fractions (Rho_heptane, Rho_butanol, Rho_ethylacetate, Rho_water), all substances were dissolved in 100% ethanol. The candidate substances tested for their reward capacity, at the concentrations mentioned in Results and in the figures, were as follows (for details of how these were obtained, please see sections above): Rhodiola crude extract (Rho_extract); fractions of Rho_extract (Rho_water, Rho_butanol, Rho_ethylacetate, Rho_heptane) dissolved in water (Rho_water), 50% aqueous ethanol (Rho_butanol) or 80% aqueous ethanol (Rho_{ethyl acetate} and Rho_heptane), with the respective solvents being used for the solvent groups; sub-fractions of the ethyl acetate (E1-5) and the heptane fractions (H1-4); isolated and synthesized β-sitosterol-β-d-glucoside (BSSG_isolated and BSSG_synthesized); isolated and synthesized ferulic acid ester (FAE-20_isolated and FAE-20_synthesized); isolated fatty acid fraction (FA_isolated); trans-ferulic acid (purity 99%, Aldrich, Darmstadt, Germany); 1-eicosanol (purity 96%, Alfa Aesar, Karlsruhe, Germany); and stigmasterol (purity 95%, ACROS Organics, Morris Plains, NJ, USA). Experimenters were blind to the substance used.

Statistical analyses of behavioral data were performed with Statistica (version 8.0, StatSoft, Inc., Tulsa, OK, USA) on a PC. In a conservative approach, non-parametric tests were used throughout. For comparisons with chance levels (i.e. with zero), one-sample sign tests (OSS) were applied (http://www.R-project.org/) and for multiple-group comparisons, Kruskal–Wallis (H) tests were used. In the case of significance, follow-up pairwise comparisons using Mann–Whitney U-tests were conducted. The significance level used was 5%, and maintained at that level in the case of multiple comparisons within one experiment by using a Bonferroni correction (P<0.05 divided by the respective number of pairwise tests). Data are displayed as box plots representing the median as the middle line, 25% and 75% quantiles as box boundaries, and 10% and 90% quantiles as whiskers.

All experiments and analyses comply with applicable ethical regulations and law. The data underlying the presented figures and used for statistical analyses are documented in the analytical and behavior data files (Table S2 and Table S3). Assuming that effect sizes for the materials used as rewards in the present study would be equal to or moderately less than those in previous studies from our laboratory using ‘canonical’ rewards, sample sizes were limited to a range equal to or slightly higher than in our previous publications. The same applies for the olfactory and gustatory preference assays.

The first time we used freshly ground, dried R. rosea root material, our attention was caught by a pleasant, mild, fruity, rose-like smell combined with a peculiar note reminiscent of rubber. This prompted us to ask whether our experimental animals too might find Rhodiola to have an attractive smell. Wild-type CS larvae showed attraction to Rhodiola in an olfactory preference test; this attraction was abolished in mutant larvae carrying a loss-of-function mutation in the Orco gene (Fig. 1A). This gene codes for an olfactory co-receptor required for sensory function in all known larval olfactory sensory neurons, its loss rendering the animals unresponsive to odors but leaving taste processing unaffected. We conclude that the net effect of the volatile compounds from Rhodiola (Table 2) is attractive to larval Drosophila, and requires Orco function.

We found Rhodiola to have a delayed yet strong, unpleasant and ‘dry’, astringent, bitter taste, so we wondered whether Rhodiola might have an aversive taste to the larvae, too. This was not the case. Rather, wild-type larvae showed attraction in a combined odor/taste preference test; the fact that the anosmic Orco¹ mutant larvae likewise showed such attraction (Fig. 1B) suggests that in these mutants the taste of Rhodiola is attractive.

Given the sensory attractiveness of Rhodiola to our experimental animals, we wondered whether it could also serve as a reward in a Pavlovian conditioning experiment. To test for such a rewarding effect, the odor AM was presented to separate groups of larvae, either paired or unpaired with Rhodiola. Then, the preference for the AM odor was measured, and the difference in preference between paired-trained and unpaired-trained groups of larvae was quantified through the associative PI. This revealed that Rhodiola had a rewarding effect at all three doses tested (Fig. 1C). Even the lowest dose used (0.1 mg ml⁻¹) was sufficient to establish memory that remained significant until 80 min after training (Fig. 1D). This rewarding effect and in particular the low dose at which it was observed were striking, because it was of about the same strength as fructose at a 2 mol l⁻¹ concentration (360 mg ml⁻¹) (Fig. S1). Reducing the number of training trials from three to one confirmed the rewarding effect of Rhodiola, though as expected at lower levels and with less stable memory over time (Fig. 1E,F).

These strong rewarding effects of Rhodiola prompted us to venture into a bioassay-guided fractionation approach to identify the compounds that mediate this effect. Before doing so, however, we wanted to see whether an extract, i.e. soluble constituents of Rhodiola, might also be effective as a reward, and whether the neurogenetic ‘fingerprint’ of learning with such an extract corresponds to learning with more canonically used taste rewards.

From a 4 kg sample of R. rosea root material, an exhaustive extraction with 80% ethanol was performed. As shown by NMR, the resulting Rho_extract was dominated by sugars (Fig. 2A–C). Beside sucrose and glucose, a heptulose was also detected in MS measurements by ions at m/z 209 [M-H]⁻, 245 [M+Cl]⁻, 233 [M+Na]⁺ and 249 [M+K]⁺. The occurrence of sedoheptulose is a characteristic feature of Crassulaceae (Nordal, 1940). Given that the rewarding effect of Rhodiola was observed at doses much lower than those effective for sugars (Fig. 1; Schipanski et al., 2008; Rohwedder et al., 2012), it appeared very unlikely that these sugars mediate the rewarding effects of Rhodiola. We therefore performed a detailed MS analysis of the Rho_extract. This revealed the above-mentioned sugars and the occurrence of previously described characteristic, specialized metabolites such as phenylethanoid and phenylpropanoid glycosides like salidroside, rosavin and phenylethyl glycoside (Coors et al., 2019), monoterpene glycosides such as rosiridin, rhodioloside E and sacranoside A, cyanogenic glycosides such as lotaustralin, rhodiocyanosides A and B as well as gallic acid and derivatives (Fig. 2B,C and Table 3). The glycosidic moieties of the observed monoglycosides (e.g. salidroside, ferulic acid glucoside) are most likely glucosides whereas diglycosides are suggested to contain 6-O-α-l-arabinopyranosyl-α-d-glucopyranoside, also found in the marker compound rosavin (Table 3). Interestingly, several of the major volatile compounds (phenylethanol, octanol, hexanol and myrtenol; Table 2) were also detected as their corresponding non-volatile glycosides. The broad band of background absorptions visible in the photodiode array (PDA) profile (Fig. 2B) is typical for R. rosea and might be due to epigallocatechin gallate oligomers (Panossian et al., 2010).

Behaviorally, presenting an odor together with Rho_extract revealed a strong and dose-dependent rewarding effect in larval Drosophila (Fig. 2D), indicating the extractability of rewarding compounds from Rhodiola. Therefore, subsequent analyses of this rewarding effect were performed with this Rho_extract.

Using the Rho_extract as the reward, we proceeded to investigate whether mushroom body function is required for Rho_extract learning (Fig. 3A). The mushroom bodies are situated at a side branch of the olfactory sensory-motor loop. They are largely dispensable for innate olfactory and gustatory behavior and feature a notably sparse multimodal representation of the sensory environment, including olfactory stimuli. Further, they receive intersecting input from distinct sets of modulatory neurons mediating punishment and reward information, respectively (Berck et al., 2016; Rohwedder et al., 2016; Eichler et al., 2017). A simplified working hypothesis of how odor–reward learning comes about is as follows. Upon odor-induced activation of the mushroom body neurons coincident with, for example, dopaminergic modulatory neurons mediating sugar reward information, an associative memory trace is formed. This takes place at the synapse from the odor-activated mushroom body neurons to their downstream partners, the mushroom body output neurons; signaling from these mushroom body output neurons eventually causes learned modulation of behavior (Pauls et al., 2010; Michels et al., 2011; Aso et al., 2014; Hige et al., 2015; Perisse et al., 2016; Rohwedder et al., 2016; Gerber and Aso, 2017; Saumweber et al., 2018). To see whether this working hypothesis is in principle also valid for odor–Rho_extract associative learning, the mushroom body neurons were silenced by transgenic expression of a hyperpolarizing GFP-tagged potassium channel (MB^OK107>Kir^GFP). This abolished odor–Rho_extract associative memory scores (Fig. 3B,C). In turn, silencing the mushroom body intrinsic GABAergic APL neurons, which arguably lifts inhibition of the mushroom body neurons and thus distorts the sparseness of the sensory representation, reduced memory scores by about a third (Fig. 3D,E). Whereas for silencing of the mushroom body neurons, impairment of innate odor preference might contribute to the reduced memory scores, this is not the case for silencing of the APL neurons (Fig. S2A,B). Of note is that the genetic controls did not show innate preference for the Rho_extract in the taste assay (Fig. S2C,D). On the one hand, this is unfortunate because it leaves tests of Rho_extract taste preference after silencing candidate neurons inconclusive. On the other hand, this lack of preference is informative. It suggests that the extraction protocol did not yield taste preference-promoting compounds (genetic controls in Fig. S2C,D), but did yield rewarding compounds (genetic controls in Fig. 3C,E,G,I), and thus suggests that different compounds mediate these two behavioral effects. We next sought to identify the rewarding compounds of the Rho_extract through bioassay-guided fractionation and AcorA (see above). This seemed particularly interesting because the silencing of those dopaminergic modulatory neurons that are required for sugar reward learning in the larva (Rohwedder et al., 2016; Saumweber et al., 2018) was without effect on memory through the Rho_extract reward (Fig. 3F–I). This suggests that the Rho_extract engages a ‘non-canonical’ reward pathway, potentially unrelated to sugar reward. If so, only the non-polar fractions of the Rho_extract, lacking most if not all sugars, should have a rewarding effect. The next experiment tested whether this was indeed the case.

From the Rho_extract we performed successive liquid–liquid partitioning with solvents of different polarity. This resulted in four fractions (Rho_water, Rho_butanol, Rho_{ethyl acetate}, Rho_heptane) comprising substances from the Rho_extract separated by polarity (Fig. 4A,B) and partially overlapping with ‘neighboring’ fractions. The observed ¹H NMR spectra (Fig. 4B) confirm the predominant occurrence of free sugars in the most polar water fraction (Rho_water). The neighboring Rho_butanol fraction contains in addition glycosidic compounds, gallic acid and galloyl derivatives, whereas the non-polar Rho_{ethyl acetate} and Rho_heptane fractions comprise aromatic compounds, phytosterols, fatty acids and fatty alcohols. Of these four fractions, only the Rho_{ethyl acetate} and the Rho_heptane fractions, i.e. the two non-polar ones, turned out to have a rewarding effect at low concentrations that exceeded the solvent control (Fig. 4C–F). The lack of a low-concentration rewarding effect in the Rho_water and Rho_butanol fractions specifically suggested that sugars, at the concentrations at which they are present in these fractions, are not sufficient as a reward. This is consistent with the previously reported dose–effect relationships for the rewarding effects of sugars (Schipanski et al., 2008; Rohwedder et al., 2012), and with the observation that silencing the dopamine neurons mediating sugar reward was without effect on the reward capacity of the Rho_extract (Fig. 3G,I).

Notably, only the most non-polar Rho_heptane fraction was effective as a reward at the lowest of the concentrations employed. This fraction does not contain any of the taste rewards known to be effective in Drosophila larvae (sugars, amino acids, sodium chloride) (Niewalda et al., 2008; Schipanski et al., 2008; Rohwedder et al., 2012; Schleyer et al., 2015). We were therefore curious whether further sub-fractionation of the two neighboring non-polar fractions would allow us to identify candidate substances hitherto not suspected of being rewarding.

From the Rho_{ethyl acetate} and Rho_heptane fractions, another round of separation resulted in five and four sub-fractions (E1–E5 and H1–H4), respectively, which included partially overlapping sets of increasingly non-polar compounds (Fig. 4G). Using these nine sub-fractions at the same single concentration (3.35 µg ml⁻¹) in odor–reward associative learning experiments revealed that sub-fractions E5, H1 and H2 maintained a rewarding effect beyond solvent control levels. More importantly, these results yielded sufficient variance in memory scores between sub-fractions to relate them statistically to analytical data. To this end, we looked for correlations of memory scores with the intensity of each analytical signal in all samples (AcorA). Specifically, the metabolite profiles of all fractions and sub-fractions were determined by direct-infusion ESI-FTICR-MS in positive ionization mode. Each analytical signal was characterized by a specific m/z value and intensity. After peak-picking and aligning these data, 936 different m/z features were detected across all samples (Table S2). To determine the correlation between memory scores and the mass signal intensities, the Spearman rank correlation coefficient was calculated, resulting in the identification of 51 positively correlating features (Table 1; Table S1). Based on these, the elemental composition of candidate compounds was calculated, which, through comparison with known natural products, allowed structure proposals for candidate compounds to be derived (Table 1; Table S1). Indeed, as the rewarding effects we had observed so far were not obviously getting successively weaker when Rhodiola, Rho_extract, the extract's non-polar fractions or the sub-fractions derived from these were used as the reward (Figs 1C, 2D and 4C–G), we felt encouraged to pick the most promising of these to see whether any of them might be rewarding in isolation, i.e. as a pure compound.

The correlating features from Table 1 and Table S1 suggest a high prevalence of mostly oxygenated fatty acid derivatives, sterol derivatives and terpenoids. In addition to the statistical correlation, we therefore compared the Rho_extract and all the above-mentioned fractions by TLC, HPLC, low- and high-resolution MS, and NMR (Fig. 4B). This indicated that all the active samples did indeed contain fatty acid, fatty alcohol or phytosterol derivatives. These compound classes were therefore considered to be particularly promising candidates for exerting a rewarding effect. We should note that the Rhodiola volatiles detected by GCMS (Table 2), and in particular phenylethanol as a major component, do not ionize under ESI conditions and could therefore not be included in the above ESI-MS analyses. NMR experiments, however, suggested the occurrence of phenylethanol and its derivatives in the non-polar fractions.

To determine the actual structure and verify the biological activity of the members of the candidate compound classes that correspond to the correlating structure proposals, isolation of the top candidates was attempted (Table 1; Table S1). Candidate 1 (m/z 485.36057; C₂₉H₅₀O₄Na⁺) and candidate 2 (m/z 365.26635; C₂₀H₃₈O₄Na⁺), which yielded the highest correlation coefficients, could not be obtained as pure compounds or enriched fractions. In contrast, it was possible to isolate candidate 3 (497.36086; C₃₀H₅₀O₄Na⁺) from the Rho_{ethyl acetate} fraction, and the structure was determined as ferulic acid eicosyl ester (FAE-20) (Michels et al., 2018). The isolated sample contained FAE-20 as a natural mixture of trans and cis isomers (3:1), accompanied by homologs bearing a shorter (C₁₈) or longer (C₂₂) even-numbered carbon alcohol chain.

The molecular formula of candidate 1 (C₂₉H₅₀O₄) differs from that of candidate 3 (C₃₀H₅₀O₄) by only one carbon, suggesting a structural similarity; however, in MS/MS experiments, the two compounds did not fragment in an analog manner. Whereas FAE-20 in the negative mode showed the predominant loss of a methyl moiety followed by loss of an extended series of neutral radicals differing by 14 u from the long alkyl chain (Schmidt et al., 2019), candidate 1 exhibited the loss of an acetyl function. The fragmentation in the positive mode was dominated by the loss of water (MS2), followed by the elimination of a C₁₈ alkyl chain (MS3) and an acetyl moiety (MS4). This fragmentation did not allow an unambiguous structure proposal for candidate 1, but excluded ferulic acid as a compound moiety. We note that nevertheless candidate 1 might belong to the group of alkyl hydroxycinnamates.

Additionally, further separation of the most active E5 sub-fraction, using column chromatography, yielded BSSG as its major constituent. Its [M-H₂O+Na]⁺ peak at m/z 581.4182 appeared as correlating signal 46 (Table S1). Furthermore, fatty acids, fatty alcohols including eicosanol as the major compound, and stigmast-4-en-3-one were obtained as constituents of active sub-fractions. Because BSSG and FAE-20 isolated from Rhodiola exhibited a rewarding effect in low micromolar or even nanomolar concentrations, these two compounds were synthesized from commercially available starting material (Michels et al., 2018) and retested for their rewarding effect, which was indeed confirmed for both BSSG and FAE-20 (Fig. 5A–D).

Under acidic conditions, which are likely to occur along the alimentary canal upon ingestion and possibly also in the rotting-fruit habitat of larval Drosophila, esters might partly hydrolyze into an acid and an alcohol. We therefore wondered whether ferulic acid and 1-eicosanol, the products of the hydrolysis of FAE-20, would be effective as rewards. 1-Eicosanol was also found to be present in the fatty alcohol fraction isolated from an active sub-fraction (see Materials and Methods). The highest tested concentration of ferulic acid (7.06 µmol l⁻¹) was revealed to be more rewarding than the solvent control (Fig. 5E), whereas 1-eicosanol was without any such effect (Fig. 5F).

Similarly, sterol glycosides might be hydrolyzed into the aglycon and the sugar moiety. As mentioned above, sterol derivatives were regarded as potentially effective compounds. As representative compounds of this kind, BSSG and stigmast-4-en-3-one were isolated. The rewarding potential of commercially available stigmasterol, representing a common plant sterol structurally closely related to sitosterol and stigmast-4-en-3-one, was therefore also evaluated; however, this did not show a rewarding effect (Fig. 5G).

To ascertain whether fatty acids can be considered candidate compounds that contribute to the observed rewarding effect, an isolated fatty acid fraction was also tested (Fig. 5H). Despite apparent trends, this did not have a rewarding effect beyond that of the solvent control at any of the concentrations used. According to NMR measurements, this fraction contains a mixture of unsaturated fatty acids with linoleic acid as the main compound (correlating signal 38; Table S1). Linoleic acid had previously been described as a major fatty acid in R. rosea (Pooja et al., 2006). However, MS experiments on this isolated fraction showed additional signals referring instead to oxygenated fatty acids that were also determined as correlating features (Table 1; e.g. 7: m/z 319 [M+Na]⁺, linoleic acid+1 oxygen; 8: m/z 335 [M+Na]⁺, linoleic acid+2 oxygens) (Fig. S3). Some of these might be lipid oxidation products arising in traces during storage of extracts. Their MS/MS fragmentation behavior is in accordance with hydroxylated fatty acids (e.g. Fig. S3B). We were not able to determine conclusive amounts or concentrations of these hydroxylated fatty acids in any of the fractions or sub-fractions by NMR experiments, or to enrich them during isolation. Taken together, the most plausible explanation for these results is that, because of their excellent ionization behavior in the electrospray ionization mass-spectrometric measurements (such as MS and MS/MS), even trace amounts of these substances might lead to false positive correlation results. Thus, the compound causing correlating signal 2 (m/z 365.26635 [M+Na]⁺ C₂₀H₃₈O₄Na⁺), which is in accordance with an oxygenated fatty acid, is also unlikely to be a promising candidate compound.

In conclusion, there is no evidence that fatty alcohols or unsaturated fatty acids, or oxygenated derivatives therefrom, exert a measurable rewarding effect in larval Drosophila. Rather, FAE-20, and to a lesser extent BSSG and ferulic acid are, as single compounds, rewarding to larval Drosophila.

The volatile and non-volatile compounds from R. rosea root material identified here are consistent with previous reports. Specifically, to detect the volatile compounds, we used headspace GCMS focusing on major peaks, yielding results in good agreement with those of Rohloff (2002), who used headspace or steam distillate analysis. We did not identify decanol, though, which with related aliphatic alcohols may be responsible for the rubber-like background scent of R. rosea root material, whereas the remaining volatile substances, including the previously unreported linaloyl oxide and eucalyptol, might underlie its floral flavor. The non-volatile compounds detected here include phenylethanoid and phenylpropanoid glycosides, monoterpene glycosides, cyanogenic glycosides, gallic acid derivatives, sugars, sterols and fatty acids as previously reported (Brown et al., 2002; Panossian et al., 2010; Lee et al., 2016; Marchev et al., 2017). Notable is the high sedoheptulose content, a sugar that in its free form is a storage compound in Crassulaceae (Hegnauer, 1964; Ceusters et al., 2013).

To the best of our knowledge, ferulic acid or its derivatives have not so far been isolated from R. rosea although other phenolic acids such as benzoic, cinnamic, coumaric or caffeic acid have been reported (Panossian et al., 2010). However, ferulic acid has been shown to occur in Rhodiola heterodonta (Krasnov et al., 1976), and an ester of ferulic acid and hexanoic acid has been obtained from Rhodiola wallichiana (Song et al., 2018). Importantly, we recently described FAE-20 (Michels et al., 2018) as a bioactive constituent of R. rosea. This compound has not been detected in Crassulaceae before but has been found in the rhizomes of Zingiber ottensii (Zingiberaceae) (Akiyama et al., 2006), in the stem bark of Pavetta owariensis (Rubiaceae) (Baldé et al., 1991) and in Baccharis genistelloides (Asteraceae) (Hennig et al., 2011). Related alkyl hydroxycinnamates, typically consisting of the phenylpropanoids coumaric, caffeic or ferulic acid esterified with fatty alcohols of different chain length, are often found in association with plant lipid barriers such as cuticles and periderm in plant bark and roots, suggesting an involvement in suberin formation (Domergue and Kosma, 2017). Therefore, their occurrence might be expected in other folk-medicine plant roots with similar effects, such as ginseng (Panax ginseng), taiga root (Eleutherococcus senticosus) or maca root (Lepidium meyenii).

The current study ascertained rewarding effects exerted by FAE-20, BSSG and ferulic acid. This does not rule out the possibility that further compounds from the non-polar fractions or their mixtures might also have rewarding effects, in particular at higher concentrations.

Ferulic acid and related compounds constitute dietary polyphenolic antioxidants present in fruits that can counteract oxidative stress in wild Drosophila populations (Dweck et al., 2015). Flies and larvae are unable to smell ferulic acid directly, but are attracted to antioxidant-containing food by yeast-produced volatiles, including ethylphenols, particularly ferulic acid-derived 4-ethyl guaiacol. The yeast bouquet also contains phenylethyl alcohol found in the headspace of R. rosea, probably contributing to the attraction shown by the larvae in the Rhodiola odor preference test. Furthermore, the esterification of ferulic acid with eicosanol might increase the volatility of the compound, potentially influencing its attractiveness to Drosophila.

BSSG, the glucoside of the most prominent phytosterol sitosterol, is a common plant constituent (for R. rosea: Kurkin et al., 1985; Kurkin et al., 1991). Its rewarding effect in Drosophila is surprising as BSSG can be found in other plants that so far have not been flagged as particularly attractive or rewarding to fruit flies. In fact, BSSG is known as a bioavailability improver, so its effect might be indirect (Bin Sayeed et al., 2016). In humans, phytosterol consumption is associated with cholesterol-lowering effects (Brufau et al., 2008). Further, phytosterol glycosides structurally resemble corticosteroids involved in the protective inactivation of the stress system (Panossian and Wikman, 2010). Notably, the long-term uptake of high amounts of BSSG as found in cycad seeds (Cycas micronesica) has been reported to exert neurotoxic effects (Tabata et al., 2008).

Compared with the sugars known to be rewarding in larval Drosophila, BSSG, ferulic acid and FAE-20 are effective at remarkably low concentrations. Whereas all three identified substances act in the low or sub-micromolar range, the sugars sucrose and fructose show their rewarding potency only at low molar concentrations (0.02 and 0.2 mol l⁻¹, respectively) (Schipanski et al., 2008; Rohwedder et al., 2012). The sugars glucose, xylose, arabinose and maltodextrin are effective as a reward in the same molar range, with maltodextrin and arabinose being effective at a slightly higher concentration of 1 and 2 mol l⁻¹, respectively (Rohwedder et al., 2012).

For salt (NaCl), rewarding effects have been observed for the low molar range, whereas increasing salt concentration changes it to a punishing effect (Niewalda et al., 2008), consistent with our own cooking experiences.

Furthermore, most if not all amino acids can have a rewarding effect (Kudow et al., 2017), but only for aspartic acid is a dose–effect function available, indicating a rewarding effect in the millimolar range (Schleyer et al., 2015).

Thus, the effective rewarding concentrations of BSSG, ferulic acid and FAE-20 are notably lower than for all previously known taste rewards. These known taste rewards are all water soluble and would all be contained in the polar fractions of the Rho_extract. Moreover, they are all in themselves of value to the larvae for energy production (sugars), growth (amino acids) and homeostasis (salt). Neither the low effective dose nor the chemical structures of BSSG, ferulic acid and FAE-20 suggest any such direct biological value. Rather, their rewarding effect may come about pharmacologically, more akin to other secondary (specialized) plant metabolites such as nicotine, caffeine or cocaine. While the molecular targets of Rhodiola, BSSG, ferulic acid and FAE-20 remain unknown, we might expect membrane interactions in the case of BSSG and FAE-20. Both compounds show an amphiphilic structure consisting of a hydrophilic side (glucose moiety or ferulic acid, respectively) and a hydrophobic part (sterol skeleton or alkyl chain). Phytosterols and their glycosides are typical constituents of plant plasma membranes that regulate membrane fluidity and permeability (Brufau et al., 2008; Ferrer et al., 2017). The sugar part is probably located in the plane of the polar head groups of the membrane whereas the sterol moiety is integrated into the hydrophobic core of the lipid bilayer (Ferrer et al., 2017).

Interestingly, the neuronal requirements for the rewarding effect of the Rho_extract only partially overlap with those for odor–sugar reward learning (Fig. 3A). In both cases, intact mushroom body Kenyon cells as well as a functional APL neuron are required (Michels et al., 2011; Saumweber et al., 2018; Thum and Gerber, 2018; Lyutova et al., 2019; Zhou et al., 2019) but different neurons are apparently required to carry reward information to the mushroom body Kenyon cells. That is, silencing one of the dopaminergic input neurons to the medial lobe of the mushroom body, namely DAN-h1, impairs odor–fructose associative memory scores (Saumweber et al., 2018), but leaves the rewarding effect of the Rho_extract intact (Fig. 3G). Indeed, the silencing of all medial lobe DANs was inconsequential when Rho_extract was used as the reward (Fig. 3I), whereas this abolishes odor–fructose as well as odor–arabinose and odor–sorbitol memory scores (Rohwedder et al., 2016). Thus, given the above results, and given that DANs outside the medial lobe are apparently involved in aversive rather than appetitive learning (Eschbach et al., 2020), the remaining candidates are octopaminergic/tyraminergic mushroom body input neurons (see also Hammer, 1993; Schroll et al., 2006; Eichler et al., 2017; Saumweber et al., 2018) and mushroom body input neurons with as yet unknown transmitters (MBIN-b1 and -b2, MBIN-c1, MBIN-l1, MBIN-e2).

Food supplementation with Rhodiola or the Rho_extract has an enhancing effect on odor–sugar associative memory in larval and aged Drosophila, and in bees (Michels et al., 2018). Furthermore, one of the compounds identified as rewarding in the present study, FAE-20, also has such a memory-enhancing effect in larval and aged adult Drosophila, and in aged mice as well (Michels et al., 2018). Thus, on the one hand Rhodiola, the Rho_extract and FAE-20 have a memory-enhancing effect that in humans would be welcome in cases of impaired mnemonic function. On the other hand, the present study shows that in larval Drosophila at least, they also have a strong rewarding effect. Should this be the case in humans, too, derivatives of FAE-20 should be sought that maintain its memory-enhancing effect but are not rewarding and are thus less likely to induce habitual use or addiction. In the search for such derivatives, the genetic tractability of Drosophila may allow for an accelerated analysis of the differential mechanisms underlying memory enhancement and the rewarding effect of FAE-20.

We thank S. Lenuweit, M. Wurm, A. Sibinski, V. Menne and F. Gerstner for experimental help, T. Saumweber for discussions, R. D. V. Glasgow for language editing, B. Westermann and A. Schaks for their contribution to chemical syntheses, and J. Schmidt for MS experiments.