High-throughput screening for selective appetite modulators: A multibehavioral and translational drug discovery strategy

Abstract

How appetite is modulated by physiological, contextual, or pharmacological influence is still unclear. Specifically, the discovery of appetite modulators is compromised by the abundance of side effects that usually limit in vivo drug action. We set out to identify neuroactive drugs that trigger only their intended single behavioral change, which would provide great therapeutic advantages. To identify these ideal bioactive small molecules, we quantified the impact of more than 10,000 compounds on an extended series of different larval zebrafish behaviors using an in vivo imaging strategy. Known appetite-modulating drugs altered feeding and a pleiotropy of behaviors. Using this multibehavioral strategy as an active filter for behavioral side effects, we identified previously unidentified compounds that selectively increased or reduced food intake by more than 50%. The general applicability of this strategy is shown by validation in mice. Mechanistically, most candidate compounds were independent of the main neurotransmitter systems. In addition, we identified compounds with multibehavioral impact, and correlational comparison of these profiles with those of known drugs allowed for the prediction of their mechanism of action. Our results illustrate an unbiased and translational drug discovery strategy for ideal psychoactive compounds and identified selective appetite modulators in two vertebrate species.

Fig. 1. Homeostatic state modulates feeding behavior but not spontaneous activity, arousal to visual or acoustic signals, or habituation in zebrafish larvae.

(A) Zebrafish larvae were raised under controlled conditions until 7 dpf, loaded into a well of a 96-well plate, and fasted for different time periods before assessment of distinct behaviors. (B) A custom-built imaging platform can quantify larval feeding behavior by measuring intestinal food content if combined with feeding fluorescently labeled live prey, larval locomotion by tracking swimming behavior, larval arousal by presenting visual and acoustic signals, and larval habituation (nonassociative learning) by repeatedly presenting inconsequential stimuli. The two exemplary images show data quality used for larval activity tracking and larval intestinal food content quantification using two distinct invisible light sources. All measurements were acquired under daylight conditions. Scale bars, 1 mm. (C) After different fasting periods, larvae were given access to fluorescently labeled paramecia, and their intestinal food content was quantified for 2 hours. Using a biological inspired curve-fitting algorithm, we extracted total food intake (D), initial intake rate (E), and digestion rate (F) (15) (n = 48; one-way ANOVA, Dunnett’s post test, ***P < 0.001). (G) Simultaneously, larval locomotor activity was tracked for 2 hours (n = 48; lc, line crosses). (H) Larval zebrafish react to dark flashes with increased motor output (50). After different fasting periods, larvae were exposed to eight dark flashes [750-ms duration, 30-s interstimulus interval (ISI)] and their locomotion was quantified. The triggered average is shown (n = 48). (I) An acoustic stimulus triggers larval locomotion (51). Identical experiment as in (H), but here a single tap was presented instead of a dark flash (n = 48). (J) Zebrafish larvae habituate to repetitive, inconsequential stimuli (52). Here, we presented 30 sequential taps with a short ISI (2 s) to differently fasted larvae and quantified larvae’s locomotor response. (K) The habituation metric is the activity ratio of the initial and last three taps indicated by the black bars (n = 48; one-way ANOVA, Dunnett’s post test, P > 0.05).

Fig. 2. Available appetite modulators trigger a pleiotropy of behaviors.

(A) Schematic of the multibehavior protocol used to test drug impact. Briefly, 2-hour-fasted larvae were pre-exposed to a drug for 30 min before multibehavioral profiling. Here, we tested many appetite-modulating drugs used in humans (4, 5). Each drug’s behavioral impact was condensed into a single, quantitative behavioral barcode, which is illustrated here for two classic anorectic drugs—nicotine and rimonabant—in detail (n ≥ 12). (B) After providing access to live prey, we quantified larvae’s intestinal food content for 2 hours and condensed the fluorescent traces into two feeding periods (F1 and F2). (C to E) Other feeding metrics can be extracted using post hoc, one-way ANOVA, Dunnett’s post test (**P < 0.01 and ***P < 0.001). (F) Simultaneously, larval locomotor activity was tracked and summarized in a single metric (S, spontaneous activity). (G) Following these 2 hours, we presented eight consecutive dark flashes to larvae and measured their activity. The triggered average response was condensed into two time points: visual periods 1 (V1, first 2 s) and 2 (V2, the following 26 s). (H) Next, we presented eight consecutive taps and measured larval kinematic response. Acoustic period 1 (A1) reflects the first 2 s of the triggered average, while A2 reflects the following 26 s. (I) Subsequently, a 3-min rest period preceded a sequence of 30 nonsequential taps with short ISI (2 s) used to quantify habituation (H). (J) Last, four dark flashes were presented identically as before in (G) to determine lethargy (L1 and L2), which is the activity ratio of the initial and post-dark flash triggered average. (K) All 10 behavioral metrics (F1, F2, S, V1, V2, A1, A2, H, L1, and L2) were acquired on a single well basis and normalized to on-plate vehicle controls (n ≥ 24). Each compound was tested in multiple animals (n ≥ 12), and their behavioral effect size and reproducibility were condensed using strictly standardized median difference (SSMD). Each square represents an SSMD value (red, higher than control; blue, lower than control) for a single behavioral metric, and jointly, they form a compound’s behavioral barcode. The black box indicates the barcodes for nicotine and rimonabant, shown in detail in (B) to (J); the orange box denotes feeding behavior; and the green box indicates habituation. *Barcodes shown for drugs tested at 100 μM. (L) Overview of all compounds tested in the high-throughput screen using the above protocol. Compounds with unknown pharmacology represent chemical structures with unknown biological activity. (M) Overview of all animals used in the high-throughput screen. Each compound was tested in multiple animals (n ≥ 6), and effects were quantified relative to on-plate vehicle controls (n ≥ 24). Sixty-two compounds were excluded in the absence of detectable activity in the last 2 min of the experiment (L1 + L2 = 0), potentially based on the strong anesthetic or lethal impact of the tested compound. Behavioral metrics: F, feeding; S, spontaneous activity; V, visual response; A, acoustic response; H, habituation; L, lethargy; 1 or 2, different time periods.

Fig. 3. Diversity of drug-induced behavioral changes.

(A) Hierarchical clustering groups hit compounds with similar behavioral barcodes based on their correlation. Dendrogram color code localizes selective clusters shown at higher magnification. (B) tSNE groups hit compounds with similar barcodes using nonlinear probability distributions and preserve local distance metrics. All 4801 hits are shown in a 2D space. The tSNE map is false colored, with the primary behavior modulated for a given compound. The magnification of the rectangle is shown in (C). (D) tSNE map location of selected hierarchical clusters magnified in (A). (E) Histogram of the pairwise Pearson correlation (R) for compound barcodes sharing one or more targets, or compound barcodes with unknown target (two-tailed, two-sample Kolmogorov-Smirnov test; line depicts the median). (F) The chemical structure of two novel compounds with unknown target. PubChem compound identifier (CID): unknown 1, 1206526; unknown 2, 2836278. (G and H) Unknown compound 1 behavioral barcode showed high pairwise Pearson correlation with drugs targeting histamine receptor 1 (H1, R = 0.70 ± 0.13; mean ± SD), and unknown compound 2 showed high pairwise Pearson correlation with drugs targeting muscarinic acetylcholine receptor 3 (M3, R = 0.72 ± 0.15). All the relevant barcodes are shown for illustration, and the box shows the mean barcode for all H1 or M3 drugs relevant for correlation. The asterisk (*) depicts trans-tripolidine, l-hyoscyamine, and ipratropium bromide. (I and J) Competitive in vitro human receptor binding assays for compound unknown 1 and H1 receptor and for compound unknown 2 and M3 receptor with positive control, respectively (n = 4, mean ± SD). (K) Physiochemical properties of all tested compounds. Lines indicate the boundaries of Lipinski’s rule of five (26). (L) Cumulative histogram of pairwise structural similarities (2D Tanimoto coefficient, T) for compounds within behavioral clusters (fig. S3) or shuffle control (two-tailed, two-sample Kolmogorov-Smirnov test). (M) The tSNE map is falsely labeled with the pairwise structural similarity coefficient (T) for the illustrated example compound. Each dot represents a different compound (red, similar structures; blue, dissimilar structures). A Tanimoto coefficient of > 0.85 reflects very similar structures.

Fig. 4. Identification of novel and selective appetite modulators.

(A) Barcodes of all hits modulating feeding behavior selectively based on the SSMD threshold sorted by their effect size on feeding period 1 (F1). (B) Twenty-seven structurally novel compounds were selected based on the displayed behavioral barcodes and unique chemical structures (see fig. S5). All compounds were validated, and the complete validation data are shown in Fig. 5 and fig. S6. (C) Heatmap illustration of in vitro target binding profiles for all 27 compounds across 43 different human targets. All binding assays were validated with a positive control. Targets: adrenergic receptors (α-1A, α-1B, α-1D, α-2A, α-2B, β-1, and β-2); dopamine receptors (D1 to D5); γ-aminobutyric acid (GABA)–A receptor, peripheral benzodiazepine receptor (PBR), and allosteric benzodiazepine binding site on rat brain slices (BZP); histamine receptors (H1, H3, and H4); muscarinic receptors (M1 to M5); serotonin receptors (5-hydroxytryptamine, 5-HT1A, 5-HT1B, 5-HT1D, 5-HT1E, 5-HT2A, 5-HT2B, 5-HT2C, 5-HT3, 5-HT5A, 5-HT6, and 5-HT7); sigma receptors 1 and 2; opioid receptors, δ-opioid receptor (DOR), κ-opioid receptor (KOR), and μ-opioid receptors (MOR); and biogenic amine transporters dopamine transporter (DAT), norepinephrine transporter (NET), and serotonin transporter (SERT).

Fig. 5. “Ideal” candidate compounds’ appetite modulators identified in zebrafish.

(A) Chemical structure of four compounds, two orexigenic candidates (oC) and two anorectic candidates (aC), with unknown biological activity to date. All structures fulfill Lipinski’s rule of five (26). Two-hour-fasted individual larval zebrafish (7 dpf) were pre-exposed for 30 min to candidate compounds before giving access to labeled food. MW, molecular weight. Subsequently, we quantified larval feeding behavior (B to E), arousal to visual (F) and acoustic (G) stimuli, spontaneous activity (H), habituation (I), and lethargy (J) (n = 48; one-way ANOVA, Dunnett’s post test, ***P < 0.001). ns, not significant. (K and L) Competitive in vitro human receptor binding assays for aC D and 5-HT2B or 5-HT2C receptor with positive control (n = 4, mean ± SD). cpm, counts per minute. (M and N) Developmental assay: Embryos were continuously exposed to drugs starting at 1 hpf, and survival rates were recorded until 6 dpf. Body length was recorded on day 5. EtOH (350 mM) served as positive control (n = 27; one-way ANOVA, Dunnett’s post test, ***P < 0.001). (O) Example locomotion trace of an individual larval zebrafish performing thigmotaxis (wall loving) in daylight and, to a lesser extent, in darkness. The time in the outer circle versus the inner circle was quantified as a thigmotaxic index in the presence of candidate compounds (n = 21; one-way ANOVA, Dunnett’s post test, ***P < 0.001). Scale bar, 10 mm. (P) The swimming path of individual larvae in the phototaxis assay during daylight and darkness in which one-half of the well is covered by a visible light–proof shelter. Phototaxis was quantified as the time in the open space versus that under the shelter with the treatment of the candidate compounds (n = 21; one-way ANOVA, Dunnett’s post test, ***P < 0.001). Scale bar, 10 mm. (Q) Turn angles triggered by perpendicular to the larvae’s body axes moving black stripes—the optomotor response—under the influence of candidate compounds. If larvae turned in the direction of the stimulus, this was quantified as a correct turn (n = 16; two-tailed, two-sample Kolmogorov-Smirnov test, P > 0.05). (R) Example swim traces of an individual larval zebrafish in an aversion experimental paradigm, where a noxious stimulus is located at one end of the swimming chamber. The time in either half of the tank was quantified as a preference index with treatment of the candidate compounds in daylight or darkness (n = 16; one-way ANOVA, Dunnett’s post test, ***P < 0.001). Scale bar, 20 mm. (S) Larval zebrafish swim in bouts, short bursts of motor activity, and these characteristic swim kinetics were quantified under the influence of the candidate compounds in a large arena (n = 20; one-way ANOVA, Dunnett’s post test, P > 0.05). Scale bar, 20 mm.

Fig. 6. Orexigenic and anorectic candidate compounds selectively modulated food intake in mice.

(A and B) Ad libitum fed or 16-hour-fasted mice received an intraperitoneal injection of candidate compounds (30 mg/kg), and their subsequent food intake was monitored continuously. Their cumulative food intake is shown relative to vehicle control in (B) (n = 10). ip, intraperitoneally. Summary characteristics of food intake and food-seeking behavior during the dark phase in ad libitum fed (C) or 16-hour-fasted (D) mice (n = 10; one-way ANOVA, Dunnett’s post test, *P < 0.05, **P < 0.01, and ***P < 0.001). Anamorelin (Ana) and rimonabant (Rim) served as positive controls. (E) Blood glucose levels measured 30 min after compound application (n = 10; one-way ANOVA, Dunnett’s post test, ***P < 0.001). Insulin (Ins) served as positive control. (F) Spontaneous locomotion of mice quantified 30 min after compound application (n = 10; one-way ANOVA, Dunnett’s post test, *P < 0.05). (G) Triggered average displacement response of mice to eight light flashes of 5-s duration and the corresponding light flash index after treatment with candidate compounds (n = 10). (H) Triggered average displacement response of mice to eight acoustic taps and the corresponding tap response index after treatment with candidate compounds (n = 10; one-way ANOVA, Dunnett’s post test, *P < 0.05). (I) Example locomotion trace of individual mice performing thigmotaxis (anxiolytic behavior) treated with vehicle (V) or rimonabant (rim). The time in the outer versus the inner square was quantified as a thigmotaxic index after treatment with candidate compounds (n = 10; one-way ANOVA, Dunnett’s post test, **P < 0.01). Scale bar, 10 cm. (J) Fluid intake and saccharine preference of mice measured on the test day of a conditioned taste aversion test. LiCl served as positive control (n = 8; one-way ANOVA, Dunnett’s post test, ***P < 0.001).