Numerical discrimination in Drosophila melanogaster

Abstract

Sensitivity to numbers is a crucial cognitive ability. The lack of experimental models amenable to systematic genetic and neural manipulation has precluded discovering neural circuits required for numerical cognition. Here, we demonstrate that Drosophila flies spontaneously prefer sets containing larger numbers of objects. This preference is determined by the ratio between the two numerical quantities tested, a characteristic signature of numerical cognition across species. Individual flies maintained their numerical choice over consecutive days. Using a numerical visual conditioning paradigm, we found that flies are capable of associating sucrose with numerical quantities and can be trained to reverse their spontaneous preference for large quantities. Finally, we show that silencing lobula columnar neurons (LC11) reduces the preference for more objects, thus identifying a neuronal substrate for numerical cognition in invertebrates. This discovery paves the way for the systematic analysis of the behavioral and neural mechanisms underlying the evolutionary conserved sensitivity to numerosity.

Keywords: CP: Neuroscience; Drosophila melanogaster, numerosity, insect cognition, approximate number system, associative learning, LC11 neurons.

Figure 1.. Drosophila spontaneously prefers larger numbers of items in a 1vs3 choice test

(A) Schematic of the experimental setup. (B) Representative example. Red dashed lines illustrate y-limits of the preference zone (red area) for each stimulus. Kernel density plots on the top and right of the platform denote the position permanence of the fly along x and y axes respectively. (C and D) At population level, flies show a significant spontaneous preference for three stripes over one stripe. (C) Left: heatmap illustrates the relative frequency of the fly location at each position of the platform (red denotes high-frequency permanence, while blue denotes low frequency). Middle: kernel population y-density plot. Right: mean population PI is significantly different from chance preference (n = 60, PI = 0.42 ± 0.38, p = 5.9e–09, Wilcoxon signed rank test). D. Control for overall area occupied (n = 56, PI = 0.29 ± 0.40, t₍₅₅₎ = 5.41, p = 1.4e–06, one-sample t test). (E–G) Flies kept their preference for more objects when tested with arrays of squares. (E) Flies’ performance in a 1vs3 squares contrast (n = 29, PI = 0.45 ± 0.31, t₍₂₈₎ = 7.92, p = 1.27e–08, one-sample t test). (F) Control for overall area occupied (n = 60, PI = 0.14 ± 0.38, p = 8.7e–03, Wilcoxon signed rank test). (G) Control for total dark area (n = 59, PI = 0.18 ± 0.34, t₍₅₈₎ = 4.00, p = 1.8e–04, one-sample t test).

Figure 2.. Flies spontaneously present a preference for more units in a 2vs4 numerical discrimination test, irrespective of non-numerical visual cues

(A) Flies preferred the four-squares stimulus (n = 60, PI = 0.27 ± 0.33, t₍₅₉₎ = 6.30, p = 4.13e–08, one-sample t test). (B) Table illustrating the non-numerical features controlled in successive experiments. (C–E) Series of experiments showing that the preference for numerically larger arrays is preserved when controlling for non-numerical cues. (C) Control for overall area occupied (n = 60, PI = 0.49 ± 0.36 p = 4.62e–10, Wilcoxon signed rank test). (D) Control for total dark area and horizontal extension of the numerical sets (n = 60, PI = 0.12 ± 0.34, t₍₅₉₎ = 2.74, p = 8.1e–03, one-sample t test). (E) Control for spatial distribution (n = 60, PI = 0.10 ± 0.38, t₍₅₉₎ = 2.09, p = 0.04, one-sample t test).

Figure 3.. Flies use ANS to discriminate among numerosities

(A) Flies tested in a 3vs4 squares contrast showed no numerical preference (n = 50, PI = 0.05 ± 0.32, t₍₄₉₎ = 1.11, p = 0.27, one-sample t test). (B) Control for overall area occupied (n = 58, PI = 0.08 ± 0.38, t₍₅₇₎ = 1.63, p = 0.11, one-sample t test). (C) Control for total dark area and horizontal extension of the numerical sets (n = 60, PI = −0.04 ± 0.41, t₍₅₉₎ = 0.93, p = 0.36, one-sample t test). (D) Flies preferred three squares in a 2vs3 contrast with the same dark area as 3vs4 (n = 60, PI = 0.25 ± 0.33, t₍₅₉₎ = 5.88, p = 1.96e–07, one-sample t test). (E) Flies preferred six squares in a 2vs6 contrast (n = 60, PI = 0.16 ± 0.40, t₍₅₉₎ = 3.15, p = 0.003, one-sample t test). (F) Flies preferred eight squares in a 4vs8 contrast (n = 60, PI = 0.13 ± 0.36, t₍₅₉₎ = 2.82, p = 0.007, one-sample t test). (G and H) Flies preferred eight squares in a 6vs8 contrast. (G) n = 60, PI = 0.14 ± 0.37, t₍₅₉₎ = 3.01, p = 0.004, one-sample t test; (H) n = 60, PI = 0.12 ± 0.37, t₍₅₉₎ = 2.49, p = 0.01, one-sample t test. (I) Flies preferred 12 squares in a 9vs12 contrast (n = 60, PI = 0.11 ± 0.40, t₍₅₉₎ = 2.09, p = 0.04, one-sample t test). (J) Discriminative performance across numerical ratio (p < 2.2e–16, Kruskal-Wallis rank test. 0.25, n = 60, mean ± SD = 0.58 ± 0.31, p = 4.5e–11; 0.33, n = 384, mean ± SD = 0.29 ± 0.39, p = 1.6e–31; 0.50, n = 360, mean ± SD = 0.24 ± 0.38, p = 8.57e–25; 0.67, n = 399, mean ± SD = 0.20 ± 0.36, p = 5.86e–26; 0.75, n = 348, mean ± SD = 0.08 ± 0.38, p = 1.05e–04; 1.00, n = 48, mean ± SD = 0.06 ± 0.42, p = 0.38, Wilcoxon signed rank test). (K) Spearman correlation shows that discrimination accuracy decreases as the numerical ratio between quantities becomes closer to 1.0.

Figure 4.. Flies display three main stable behavioral patterns of numerical preference

(A) Mean trajectory for each pattern of behavior. Color code indicates the recording time (%). (i) Flies with stable preference for more items along all recording time; (ii) hesitant flies that finally choose more items; (iii) Flies that show no preference at the beginning but then decide for fewer items. (B) Y position over time for the three categories of behavioral preference. (C) Population heatmap corresponding to each day. (D) For each particular day, flies preferred four squares (day 1, n = 59, PI = 0.26 ± 0.44, t₍₅₈₎ = 4.48, p = 3.61e–05, one-sample t test; day 2, n = 59, PI = 0.24 ± 0.42, t₍₅₈₎ = 4.45, p = 3.97e–05, one-sample t test). The performance is maintained between consecutive days (p = 0.71; t₍₅₈₎ = 0.37, paired t test). (E) Pearson correlation plot. Gray shadow indicates confidence interval (95%). (F) Matrix heatmap showing the relationship between patterns of behaviors in day 1 and day 2. The relative and absolute (between parentheses) number of animals with the specific relationship are indicated in each box.

Figure 5.. Flies’ numerical preference can be modified by associative conditioning

Flies were trained to associate a non-preferred set of squares (CS+) to sugar (US). (A) Flies trained in a 1vs3 squares contrast showed no preference for three squares in the testing session. Left: testing session kernel density plot for each group (pink, control group; yellow, trained group; orange, overlap of the two curves). Control group (CT) preferred the three squares while the trained group (TR) showed no numerical preference. Right: unlike the control group (n = 38, PI = 0.54 ± 0.36, p = 3.25e–09, Wilcoxon signed rank test), trained flies showed no preference for the set of three squares (n = 37, PI = 0.08 ± 0.36, t₍₃₆₎ = 1.31, p = 0.2, one-sample t test; comparison between groups, p = 1.8e–06, Wilcoxon rank-sum test). (B) Flies trained in a 2vs3 squares contrast showed inverse numerical preference during the testing session. Trained flies significantly preferred the smaller set of squares (n = 45, PI = −0.20 ± 0.34, t₍₄₄₎ = −3.93, p = 2.9e–04, one-sample t test), opposite to the control group (n = 42, PI = 0.39 ± 0.36, t₍₄₁₎ = 7.19, p = 8.87e–09, one-sample t test; comparison between groups, t₍₈₄₎ = 7.9, p = 7.94e–12, Welch two sample t test). (C and D) Flies trained with a numerical contrast they could not discriminate (i.e., 3vs4) did not show a change in their preference during the testing session. (C) Flies trained to sugar associate three squares did not show a numerical preference (n = 41, PI = 0.008 ± 0.44, t₍₄₀₎ = 0.11, p = 0.91, one-sample t test), same as the control group (n = 40, PI = 0.08 ± 0.64, p = 0.33, Wilcoxon signed rank test; comparison between groups, p = 0.47, Wilcoxon rank-sum test). (D) Flies trained to associate four squares to sugar did not show a numerical preference in the testing session (n = 45, PI = 0.08 ± 0.47, t₍₄₄₎ = 1.18, p = 0.24, one-sample t test), same as the control group (n = 46, PI = 0.03 ± 0.60, p = 0.71, Wilcoxon signed rank test; comparison between groups, p = 0.25, Wilcoxon rank-sum test).

Figure 6.. Silencing LC11 neurons diminishes numerical discrimination

(A) Anatomy of LC11 neurons. (B) Performance of flies with silenced LC11 tested in a 1vs3 stripes contrast. LC11>TNT: n = 60, PI = 0.11 ± 0.30, p = 0.002, Wilcoxon signed rank test. LC11/+: n = 60, PI = 0.27 ± 0.28, p = 3.74e–08, Wilcoxon signed rank test. +/TNT: n = 60, PI = 0.30 ± 0.34, p = 3.43e–07. (C) LC11-silenced flies tested in a 2vs3 stripes contrast were unable to discriminate in opposition to the control group. LC11>TNT: n = 53, t₍₅₂₎ = −0.47, p = 0.64, PI = −0.02 ± 0.29, one-sample t test. LC11/+: n = 55, t₍₅₄₎ = 3.53, p = 0.0009, PI = 0.14 ± 0.30, one-sample t test. +/UAS: n = 56, t₍₅₅₎ = 3.27, p = 0.002, PI = 0.11 ± 0.25, one-sample t test. (D) Anatomy of LC10a neurons. (E) Numerical discrimination performance of flies with silenced LC10a neurons in a 1vs3 stripes contrast. LC10a>TNT, n = 58, t₍₅₇₎ = 5.12, p = 3.93e–06, PI = 0.19 ± 0.28, one-sample t test. LC10a/+: n = 58, p = 6.63e–09, PI = 0.39 ± 0.33, Wilcoxon signed rank test. +/TNT: n = 56, p = 1.67e–07, PI = 0.32 ± 0.33, Wilcoxon signed rank test. (F) Numerical discrimination performance of LC10a-silenced flies in a 2vs3 stripes contrast. LC10a>TNT: n = 56, p = 0.02, PI = 0.09 ± 0.29, Wilcoxon signed rank test. LC10a/+: n = 57, t₍₅₆₎ = 3.39, p = 0.001, PI = 0.14 ± 0.31, one-sample t test. +/TNT: n = 57, t₍₅₆₎ = 2.10, p = 0.04, PI = 0.10 ± 0.36, one-sample t test. (G and H) Flies with silenced LC11 neurons trained by pairing two squares with sugar in a 2vs3 contrast. (G) Kernel density plot for each group (pink, control group; yellow: trained group. (H) LC11-silenced flies showed no preference during the testing session (LC11>TNT_CT vs LC11>TNT_TR: p = 0.8, Wilcoxon rank-sum test). In contrast, both control groups significantly diminished the magnitude of their preference after learning (LC11/+_CT vs LC11/+_TR: p = 0.009, Wilcoxon rank-sum test. +/UAS_CT vs +/UAS_TR: p = 0.008, Wilcoxon rank-sum test). Scale in (A) and (B), 50 μm.