Corresponding author: S. M. Rollmann, Department of Biological Sciences, University of Cincinnati, Cincinnati, OH 45221-0006, USA. E-mail: firstname.lastname@example.org
A defining goal in the field of behavioural genetics is to identify the key genes or genetic networks that shape behaviour. A corollary to this goal is the goal of identifying genetic variants that are responsible for variation in the behaviour. These goals are achieved by measuring behavioural responses to controlled stimuli, in the present case the responses of Drosophila melanogaster to olfactory stimuli. We used a high-throughput behavioural assay system to test a panel of 157 Drosophila inbred lines derived from a natural population for both temporal and spatial dynamics of odour-guided behaviour. We observed significant variation in response to the odourant 2,3-butanedione, a volatile compound present in fermenting fruit. The recent whole genome sequencing of these inbred lines allowed us to then perform genome-wide association analyses in order to identify genetic polymorphisms underlying variation in responses. These analyses revealed numerous single nucleotide polymorphisms associated with variation in responses. Among the candidate genes identified were both novel and previously identified olfaction-related genes. Further, gene network analyses suggest that genes influencing variation in odour-guided behaviour are enriched for functions involving neural processing and that these genes form a pleiotropic interaction network. We examined several of these candidate genes that were highly connected in the protein- and genetic interaction networks using RNA interference. Our results showed that subtle changes influencing nervous system function can result in marked differences in behaviour.
Much of behaviour is generated by the integration of information transduced at the periphery, processed by circuitry in the central nervous system, and realized as bodily motion coordinated to respond to stimuli. This process is shaped by interactions among a large number of genes and a fundamental goal in behavioural genetics is to understand the mechanisms underlying the production of, and variation in, behaviour. This requires the identification of genes involved in the development and function of the neural circuitry that enables an organism to respond to environmental stimuli. It also requires a determination of how differences in those genes influence nervous system function to cause functional differences in behaviour.
In the case of odour-guided behaviour, significant advances have been made in uncovering the neural circuitry underlying the peripheral detection of chemical cues. The principles of odour coding are similar in both vertebrates and insects and thus Drosophila has emerged as a model for olfaction due to its quantitatively simpler olfactory system (Vosshall & Stocker 2007). In comparison to our current understanding of other sensory modalities, particularly vision, we lack detailed knowledge of how peripheral and central processing of olfactory signals results in behavioural output (Carey & Carlson 2011). This gap in knowledge applies particularly to one of the most intriguing features of any stimulus-sensation-response system, the nature and source of variation among individuals, populations or species. A promising way forward in closing this gap is to extend the traditional olfactory behavioural assay, to consider more nuanced spatial and temporal dynamics of odour-guided behaviour.
Behavioural performance in a T-maze assay system has been used extensively in the study of olfactory behaviour (e.g. Ai et al. 2010; Helfand & Carlson 1989; Störtkuhl et al. 2005; Suh et al. 2004). Typically, individuals are introduced into the centre of the apparatus and given a choice between two arms, one of which contains an airstream of diluted odourant and the other the vehicle control. A response index is then calculated based on the number of individuals in each arm after a set time period. However, because this traditional assay considers binary responses of organisms—either towards or away from an odour stimulus—it provides a simplistic assessment of the strength of response, and almost totally disregards the temporal dynamics of behaviour. It fails to measure more nuanced phenotypes of potential interest for their genetically based variability. We designed a behavioural paradigm that integrates an automated video tracking system with the robust T-maze assay system to record from multiple T-maze testing arenas simultaneously to obtain fine-scale resolution mapping of fly movement over time. These movements are deconstructed into different phenotypes, defined in space and/or time, and then used for genome-wide association (GWA) mapping to provide fine-scale resolution of the genetic basis of some new aspects of behavioural variation.
The recent whole genome sequencing of a panel of inbred Drosophila melanogaster strains (Drosophila Genetic Reference Panel, DGRP) derived from a natural outbred population presents a clear opportunity to assess the genetic basis of intraspecific variation in odour-guided behaviour (Mackay et al. 2012). These wild-derived lines represent genotypes that can be subjected to repeated behavioural measurements in a controlled environment. Furthermore, the limited population structure in the lines, an issue which in other systems has sometimes confounded interpretation of GWA results, makes them an excellent resource for GWA analyses (Mackay et al. 2012). Towards this end, we designed a high-throughput behavioural testing apparatus and monitored several spatial and temporal components of DGRP responses to the odourant 2,3-butanedione. This odourant is present in fermenting fruit, the natural food source for Drosophila (Mayr et al. 2003) and has been shown to modify CO2 behavioural responses in Drosophila and mosquitoes (Turner & Ray 2009; Turner et al. 2011). We then associated molecular polymorphisms with variation in these response variables to identify genes that contribute to odour-guided behaviour and subsequently conduct functional enrichment analyses to examine whether specific gene ontology (GO) categories were significantly enriched within each phenotype. Finally, we examined a subset of these candidate genes for their role in mediating odour-guided behaviour using RNA interference (RNAi).
Materials and methods
Drosophila melanogaster isofemale lines from a single population in Raleigh, NC were each inbred by 20 generations of full-sib mating (Ayroles et al. 2009; Mackay et al. 2012). These lines comprise the DGRP and were obtained from the Drosophila Stock Center in Bloomington, IN. The olfactory mutant line Orco2 (23130) and its genetic control, w1118 (6326), were also obtained from the Drosophila Stock Center (Larsson et al. 2004). Transgenic RNAi flies were obtained from the Drosophila Transgenic RNAi Project (Harvard Medical School) and are as follows: UAS-lolaRNAi (26714), UAS-bnlRNAi (34572), UAS-Nedd4RNAi (34741), UAS-sifRNAi (25789), UAS-RfxRNAi (29355), UAS-svpRNAi (28689) and UAS-fzRNAi (34321) (Ni et al. 2008, 2009). Individual RNAi lines were each crossed to the following drivers: Orco-Gal4 (26818), elav-Gal4 (8765), OK107-Gal4 (845) and GH146-Gal4 (gift from F. Hamada, University of Cincinnati). RNAi mediated knockdown was thereby mediated in olfactory sensory neurons, pan-neurally, in the mushroom body or in projection neurons, respectively. Controls for genetic background were generated by crossing Gal-4 driver lines to the transgenic host strain, yv; attP2, y+, as appropriate. All flies were reared on standard cornmeal-agar-medium at 25°C under a 12-h light:12-h dark cycle, with lights on at 6:00.
Olfactory assay system
Odour-guided behaviour was measured using a modified T-maze design to allow for live video tracking and for high-throughput behavioural testing. The apparatus—the ‘olfactometer’—consists of six troughs arranged in parallel (L: ∼15 cm × W: 1.5 cm × H: 1.0 cm, Fig. 1). Similar to the traditional T-maze, flies were released from the centre of each trough, and had the option to move in one or the opposite direction. Six such experiments were performed and recorded simultaneously. The body of the olfactometer is made of odour-resistant polyethylene plastic, with opaque walls between troughs and the troughs are all covered by a single, clear acrylic lid. On each side of the olfactometer, light emitting capacitor panels were placed parallel to the troughs, providing diffuse, extended and non-directional (relative to the trough's long axes) light (CeeLite, Colmar, PA, USA). Air entered each trough through two portals, one at each end and exited through a single central floor vent under the point of animal release. Air entering all troughs originated from a single pure compressed source, drawn through active carbon filters. Air flow was regulated at the inputs to each trough by 12 air flow metres, and maintained at 500 ml/min. Air removal from the central vents was regulated by six flow metres, each set to 1000 ml/min, under vacuum provided by a single pump. Twelve glass bubblers, each unique to one end of a single trough, contained either odourant diluted in vehicle or vehicle only. Fly movement in all six troughs was captured with a video camera (Panasonic model wv-BP334; Panasonic Corporation, Kadoma, Osaka, Japan) and recorded using EthoVision software (Noldus Information Technology, Leesburg, VA, USA). Images were collected at a rate of 30 frames per second. Custom MATLAB 7.13 (MathWorks, Natick, MA, USA) code was used to perform these analyses.
Odour-guided behaviour was measured in 4–10 day-old adult mated flies from 157 wild-derived inbred lines comprising the DGRP. For each assay, 10 flies were aspirated into the centre of each of the six troughs and allowed to acclimate for 1 min. Upon release using a sliding gate, flies were allowed to walk down either arm of the trough. In experimental treatments, 0.1% 2,3-butanedione was delivered to one end of each trough, and vehicle (control:distilled water) was delivered to the other end. In control treatments, the vehicle alone was delivered to both arms. Dose response experiments had been previously performed on 10 randomly chosen DGRP lines to determine the optimal concentration for resolving variation among lines. Six replicate assays per sex and line were performed for a total of 3768 experiments. Experiments were conducted in the morning at 25°C. Replicates were arranged such that every line and sex was randomized with respect to date tested, trough used in the apparatus, side of the trough containing odourant and conditions in neighbouring troughs. For each trial, fly movement was recorded for 5 min. Experiments involving the Orco mutant or the RNAi silencing of candidate genes were conducted using the same behavioural assay protocols used for the measurement of odour-guided behaviour in the DGRP lines. Behavioural responses of the Orco −/− mutant and w1118 control were measured at 0.1%, 0.5% and 1% 2,3-butanedione. Additionally, the measurements for RNAi experiments were normalized to control for the effect of RNAi knockdown on general activity level. Namely, T' = Ax/Ac · T, where T is the original trait value, Ax is the activity level of the knockdown, Ac is the activity level of the control, and T' is the adjusted trait value.
Videotapes of all control and experimental trials were subjected to the following steps to extract the behavioural phenotypes. The video frames were first converted into 8-bit grayscale images, and a background image was subtracted from each frame in the video. The background image was generated by summing every tenth frame in the video and computing the mean image. The resulting ‘difference frames’ lacked any information that co-occurred in the background image and video frames, namely non-moving objects, and retained information that did not co-occur, namely objects (flies) that moved for more than ∼10% of video frames. These 8-bit difference frames were then filtered with a simple threshold to yield a binary image; pixels having grayscale values greater than 0.05 represented ‘fly’ and were given a value of 1, while all other pixels were given a value of zero. Rectangular areas in the binary images corresponding to the six individual troughs were extracted from each frame using coordinates obtained manually from a single video frame at the start of analysis. The pixel values in each of these six sub-images were summed down their columns to give bin heights corresponding to fly density along the ‘x-axis’ of the trough length. The behavioural phenotypes for each of the six replicates were then computed from spatiotemporal dynamics of the median location of fly density. For such dynamics to be tested for association with polymorphisms using standard statistical methods, the strong autocorrelation in the time-series data was removed by sub-sampling at an interval of 158 frames, or 5.3 seconds. This sampling interval was the median lag from autocorrelations performed on all control and experimental trials that yielded a correlation coefficient of 0.36 or less. Thus, the 9 × 103 data points from each trial (number of frames in a 5-min video) became 57 data points (see Results for further exploration of the effect of sampling interval).
The behavioural phenotypes consisted of the following five properties derived from the fly location data obtained from image analysis. In general phenotypes can be considered spatial if they are expressed in, e.g. cm, temporal if they are expressed in seconds, and spatiotemporal if they are expressed as the area under a curve in spatial-temporal coordinates (cm · second) (Fig. 1c).
Intensity of response: the farthest point achieved by the median fly position in the direction of the odourant source, measured from the median fly position of the control (units are cm, odour-side is positive, opposite side is negative). This phenotype is independent of time, and captures the maximum attraction to the odourant, at any time during the trial, however brief it may be.
Location ratio: the total fly density in the non-odour-half of the trough divided by that in the odour-half. Unlike the other phenotypes its value does not depend on controls or measures of centrality such as median. This phenotype captures the most basic aspect of general attraction to the odourant. Though this is a unitless number it reflects a purely spatial phenomenon.
Strength of response: the total area (cm · second) between the median fly position and the median of the control over time (odour-side positive).
4–5. Percentile(s) of strength: the times at which 30% and 90% of cumulative response strength was achieved. These phenotypes reflect response latency (30th percentile) and the pattern and extent to which the response was sustained, revealing potential habituation or adaptation to the odourant (90th percentile). Phenotypes from typical trials are shown in Fig. 1d. In the left panel 30th percentile occurs early, indicating a short latency, followed by an early 90th percentile, indicating relatively fast adaptation or habituation of the flies to the odour. The right panel has similar Strength and Intensity of response to the left, suggesting they share similar levels of attraction to the odour, but the 30th and 90th percentiles occur much later for longer latency and later adaptation/habituation, indicating a reduced ability or tendency to detect and/or respond to the odour (Fig. 1d). Units are in seconds and so these are technically temporal phenotypes, but note that they indicate the time at which a spatiotemporal process is completed.
General activity levels were also calculated by the digital equivalent of logging the frequency with which flies interrupt a light beam. We designated 35 temporal transects (digital light beams), evenly spaced across the width of the trough (would appear as vertical lines in Fig. 1 heat maps). Using the thresholded fly location data we determined how often the fly entered the transect, i.e. how many times the state of the pixels changed from ‘no fly’ to ‘fly’.
We found a subtle effect of activity level on Intensity (R = 0.0725, slope = 0.001) and 90th percentile (R = 0.115, slope = 0.112). We corrected for this affect by regressing the phenotype onto activity, and subtracting the resulting line from the phenotype values, i.e. by ‘de-trending’ the data.
Statistical analysis of behaviour
For each behavioural phenotype, comparisons among DGRP lines was made using the two-way anova model y = μ + L + S + L × S + E, where sex (S) is the fixed effect of sex, line (L) and L × S are the random effects of line and line by sex, respectively, and E indicates error. If no significant difference between sexes was observed, the data were pooled. Correlations between each of the five phenotypes were assessed by calculating Pearson product–moment correlation coefficients and using sequential Bonferroni correction for multiple testing (Rice 1989).
For the Orco mutant and RNA interference experiments, differences between the control and experimental lines were made using the anova model in which y = μ + G + S + G × S + E, where genotype (G) and sex (S) are fixed effects. In both experiments, if no significant difference was observed between the sexes, data was pooled. All data was analyzed using jmp 8.0 software (SAS Institute, Cary, NC, USA).
Genome-wide associations were computed for each of the five phenotypes using single nucleotide polymorphism (SNP) data collected and made publically available by Mackay et al. 2012 (http://dgrp.gnets.ncsu.edu; DGRP Freeze 1.0). The same 157 DGRP lines used to assess natural variation in odour-guided behaviour were used for these analyses. Associations were computed for SNPs with the minor polymorphism present in 4 or more DGRP lines and for which mean sequence coverage was between × 2 and × 30. Statistical significance of the associations between individual polymorphic sites and phenotypic variation was determined by the anova model y = μ + M + S + M × S + L(M) + E, where μ is the mean, M represents the polymorphic marker, S is the sex, and L represents the random DGRP line effect. In analyses for each individual sex, the model y = μ + M + E was used. A significance threshold of 10−5 was used for all analyses. Linkage disequilibrium (LD) between significant SNPs was examined using MMC clustering (Stone & Ayroles 2009). Estimated effect sizes were calculated as one half the difference in mean phenotypic trait value for the major and minor allele classes.
Gene set functional enrichment analyses and network visualization
To understand global interactions of the genes associated with significant SNPs within each phenotypic category, overrepresentation analyses for biological processes (BP) and molecular functions (MF) were performed. Such analyses typically apply hypergeometric tests to estimate whether a biological process is represented in a given data set more than expected by chance. Using the FlyMine (Lyne et al. 2007) and WebGestalt (Zhang et al. 2005) servers, which query GO annotations, functional enrichment analyses were performed separately for each of the five groups. We restricted our analyses of GWA candidate genes to genes with SNPs within 1 kb of the candidate gene(s) and at a threshold of 10−5. For network representation of specific and shared enriched BP and MF among different phenotypic groups of genes, we used Cytoscape (Shannon et al. 2003) and Gephi (Bastian & Jacomy 2009) applications. Network representation of gene distribution across different phenotypes was also done using Cytoscape and Gephi. Additionally, we mined Drosophila genetic and protein–protein interactions to see if any of the genes within each of the phenotypes have any known interactions. We used the DroID (Yu et al. 2008) and NCBI's Entrez Gene (Maglott et al. 2011) for mining the Drosophila genetic and protein–protein interactions.
Measurement of odour-guided behaviour
We designed a behavioural paradigm that integrates an automated video tracking system with the robust T-maze assay system to record from multiple T-maze testing arenas simultaneously to obtain fine-scale resolution mapping of fly movement over time (Fig. 1a,b). Using this assay, variation in spatial and temporal dynamics of behaviour was quantified among 157 DGRP lines in response to 2,3-butanedione, a compound present in Drosophila feeding substrates and a by-product of fermentation (Mayr et al. 2003; Nishimura et al. 1989; Takeoka et al. 1990). Specifically, we quantified five odour-evoked behavioural phenotypes (Fig. 1c) that were defined to assess the latency and degree of attraction to an odourant as well as the attenuation of the behavioural response.
We began by addressing an issue inherent to analyses of measurements of animal movements, i.e. the fact that sequential measurements of the same animal are highly correlated, and thus cannot be subject to standard statistical procedures. A standard way of alleviating this issue is to subsample the data at a longer time interval; the longer the interval, the lower the autocorrelation. The level of autocorrelation that is acceptable is arbitrary, but is kept to a minimum in order to meet the assumptions underlying subsequent statistical tests. We tested the effect of sampling interval on phenotype values by computing the phenotypes at sampling intervals of 50, 100, 150, 200, 250 and 300 frames. Correlations between the resulting phenotypic values (i.e. for Location ratio between 50 frames and 100, 150, 200, 250 or 300 frames, and so on for all interval combinations; equaling a total of 75 comparisons) were high (R = 0.853–0.999), and all were highly significant (all P ≪ 0.001). In 73 out of 75 comparisons smaller sampling intervals produced phenotypes with larger values, on average, as evidenced by the less-than-unity slopes when phenotypes computed from higher sampling intervals were regressed onto those from lower sampling intervals. This is to be expected, given the increased likelihood of sampling any given phenotypic value, including large ones. Higher sampling intervals also reduce the large incongruities seen with lower sampling rates in the 30th and 90th percentile phenotypes, but which were virtually absent in Location ratio and Strength. Together, these analyses indicate that, within the range tested, the choice of a particular sampling interval is arbitrary with respect to consistency in phenotype values. Therefore, we selected a sampling interval (158 frames) that yielded a low level of autocorrelation (R = 0.36) as well as accurately representative and normally distributed data.
We observed broad variability in the five phenotypes (Figs. 1d; S1a–e) and narrow-sense heritability estimates for these phenotypes ranged from 2% to 14%. In pairwise comparisons these phenotypes were correlated to varying degrees, with the exception of Intensity by 30th Percentile—meaning that the temporal latency to respond had little bearing on the maximum odour-side distance eventually attained (Table S1). The fact that the phenotypes were correlated was not surprising since they all comprise a somewhat unified olfactory behavioural response. However, the predominance of low correlation coefficients led us to predict that variation in these correlated phenotypes will be associated with partially non-overlapping genetic variants. Our results support this prediction, with the degree of overlap scaling linearly with the degree of phenotypic correlation (R2 = 0.498, F1,19 = 17.82, P = 0.0005). The spatial and temporal behavioural measurements were therefore pursued as informative distinct phenotypes.
Moreover, to further confirm that these behavioural responses are odour-guided, we examined behavioural responses of Orco mutant flies to 2,3-butanedione. Orco is implicated in the localization of the tuning odourant receptor proteins to the sensory dendrites (Benton et al. 2006; Larsson et al. 2004). Thus Orco mutant flies lack the majority of tuning odourant receptors, but do have derived ionotropic glutamate receptors that function in detection of acid and amine odours (Benton et al. 2009; Silbering et al. 2011). We observed a dose dependent behavioural response to 2,3-butanedione in wild-type flies, with attraction at an intermediate odourant concentration and repulsion at a higher concentration. Orco mutant flies showed a lack of significant behavioural responses to odour, regardless of concentration (Fig. 2). At intermediate concentrations, there were significant differences between Orco and wild-type flies for all five phenotypes (Table S2A). Differences were also observed in response to 1% 2,3-butanedione for Location ratio. As expected, no significant differences were observed for Intensity, 30th percentile and 90th percentile, as non-zero values occur only in the case of attraction. Wild-type flies are repulsed at 1%, producing for example, 30th percentile = 0, which is indistinguishable from the control. These results are as predicted for an assay that tests odour-mediated behaviour.
Spatial dynamics of odour-guided behaviour
Two spatial properties of fly responses to 2,3-butanedione were measured, Intensity and Location ratio, to assess the degree of attraction to the odourant. Significant phenotypic variation among the 157 lines was observed for both phenotypes (Table S2B). In general, flies exhibited positive behavioural responses to 2,3-butanedione, moving towards the odour source with a mean Intensity of 2.42 cm ± 0.03 (mean ± standard error; Fig. S1a) and range from 0.83 cm to 3.83 cm. Sexual dimorphism was observed in measurements of Location ratio. Females tended to be more biased towards to odour-half of the trough [0.77 ± 0.28, range = 0.22–2.22 (females) and 0.86 ± 0.28, range = 0.19–1.73 (males), respectively; Fig. S1b]. The fact that a small subset of lines showed inverted behavioural responses illustrates that qualitative, not only quantitative, differences in odour sensitivity exist among the wild-derived lines and is consistent with previous findings (Richgels & Rollmann 2011).
Temporal dynamics of odour-guided behaviour
The temporal and spatiotemporal dynamics revealed several additional behavioural features that are incorporated into the following measurements: the Strength of response, and its 30th and 90th percentiles, which are the times at which the cumulative Strength of response was complete to these degrees, respectively (Fig. 1c). Significant phenotypic variation was seen for all three of these spatiotemporal behavioural measurements (Table S2B; Fig. S1c–e). In general, 30% of the strength of response occurred just over a minute after the start of the introduction of odour (mean = 84.3 seconds; Fig. S1c). This value reflects a composite of factors that influence the timing and onset of response, including sensitivity and preference. Variation in the 30th percentile phenotype was observed among lines—the range was 13–184 seconds, with a standard error of 22. Times marking completion of ninety percent of the strength of response similarly showed strong variability, with a mean and standard error of 216 seconds ± 36 (Fig. S1d) and a range of 26–283 seconds. This phenotype reflects processes such as sensory adaption and habituation, which influence the prolongation and cessation of response. Clearly, some lines initiate and attenuate their responses rapidly while other lines respond in a more sustained manner. Overall these values indicate that the response was much more likely to be strong at the start of the trial, since it was not distributed evenly throughout the 300 seconds, in which case the 30th and 90th percentiles would occur at 90 and 270 seconds, respectively. Indeed, the majority of fly lines showed a pattern of movement consistent with sensory adaptation or habituation, in that they had very strong initial responses towards the odour, and subsequently moved back and forth between the non-odour and odour sides of the arena. Finally, significant variation was observed in Strength of response (mean and standard error of 20.5 cm · second ± 158). Strength of responses ranged from −376 to 556 cm · second, demonstrating broad intra-specific behavioural variability. Thus, broad temporal and spatiotemporal variability was observed among the DGRP lines.
Genome wide association analyses
Genome-wide association analyses were performed in order to dissect the genetic architecture underlying natural variation in odour-mediated behaviour. We computed the statistical association between 2 425 403 SNPs, made publically available by Mackay et al. 2012, and each of the five odour-guided behavioural phenotypes (Tables S3 & S4–S8). These analyses identified 155 SNPs to be significantly associated with Intensity, 161 with 30th percentile, 585 with 90th percentile, 73 with Strength of response, and 396 with Location ratio (Table S9) at a significance threshold of P < 10−5. This threshold corresponds to a false positive discovery rate (FDR) of 16% for Intensity, 33% for Strength, 15% for 30th percentile, 4% for 90th percentile and 6% for Location ratio, though precise FDR calculations are not feasible given the non-independence among SNPs (see below). At an FDR of 5%, 834 SNPs are associated with 90th percentile and 319 SNPs with Location ratio. No SNPs met the 5% FDR threshold for the remaining three phenotypes. The distribution of SNPs associated with all phenotypes was typically complex and broad, with multiple SNPs within and between chromosomal regions associated with variation in odour-guided behaviour. Of the significantly associated SNPs, some were significantly associated for both sexes, only one sex, or in the SNP by sex interaction term (Fig. 3; Figs. S2–S6; Table S9). In the case of Location ratio, where responses differed between the sexes, analyses identified 144 SNPs for sexes pooled, 200 for females, 41 for males, and 12 for the SNP by sex interaction. Finally, we assessed the extent of non-random associations (LD) among polymorphic markers. Significant LD was typically found between polymorphisms in close physical proximity, suggesting that the patterns of genotype–phenotype associations are not solely attributable to the high LD among SNPs and that multiple genes contribute to variation in odour-guided behaviour (Figs. S2–S6). The occasional long-range LD was found. For example, long-range LD of ∼1.4 kb (restricted to three SNPs associated with Strength), to 6.7 kb (six SNPs associated with 30th percentile) to 28 kb (two SNPs associated with 90th percentile) was observed.
Quantile–quantile (q–q) plots of the P-values indicate varying degrees of deviation from the null hypothesis (no indication of causal variants), from little in the case of Intensity to strong in the case of Location ratio (Fig. S7). Interpretation of the results for each phenotype should be tempered by their degree of adherence to the null. For example, there was strong support for the presence of specific genetic variants affecting variation in 90th percentile and Location ratio, but we proceed with caution in the case of Intensity. The q–q plots also show little evidence of population stratification, in concurrence with initial estimates (Mackay et al. 2012).
Candidate gene network analyses and enrichment of GO categories
A systems biology analysis of the significantly associated genes further illustrates the genetic architecture of odour-guided behaviour. A network representation of gene distribution across the different phenotypes was created, and the Drosophila genetic and protein interaction database was mined to see if there were any known interactions between genes associated with a given phenotype. Overall, we found that the set of polymorphic genes associated with each of the behavioural phenotypes was largely non-overlapping (Fig. S8a). Several GWA candidate genes could, however, be clustered based on known genetic and protein–protein interactions (Fig. S8b) and a large number of the GWA candidates have been previously implicated in olfactory functions. Genes highlighted by our GWA analyses that have been shown in previous studies to play a role in olfaction include: bicoid-interacting protein (bin3), calmodulin (Cam), frizzled (fz), ftz transcription factor 1 (ftz-f1), genderblind (gb), longitudinals lacking (lola), Pinocchio (Pino), cAMP-dependent protein kinase R2 (Pka-R2), Regulatory factor X (rfx), seven up (svp), scribbled (scrib) and tachykinin (Tk), as well as odourant receptors, Ir76b, Or33b, Or67a, Or67c and Or98b (Fig. S8; Table S9; (Abuin et al. 2011; Anholt et al. 1996; Anholt & Williams 2010; Dubruille et al. 2002; Grosjean et al. 2008; Ignell et al. 2009; Park et al. 2000; Rollmann et al. 2005; Sambandan et al. 2006; Sen et al. 2003; Spletter et al. 2007; Swarup et al. 2013). Significant epistatic interactions between Cam and scrib and independently between Pino and scrib have been demonstrated previously (Anholt et al. 2003; Fedorowicz et al. 1998; Ganguly et al. 2003). Or98b is thought to be co-expressed with Or85b and single unit electrophysiological recordings revealed robust responses to 2,3-butanedione (Couto et al. 2005; Hallem et al. 2004). We also found a few GWA candidates that are involved in olfactory learning and memory [Mob2 and visgun (vsg)] (Akalal et al. 2011; Dubnau et al. 2003). These results are to be expected if our GWA analyses have in fact served to identify real links between genes and olfactory behaviour.
Functional enrichment analyses also revealed that specific GO categories were significantly enriched within each of the five phenotypes (Fig. 4; Fig. S9; Table S10). For instance, there was significant enrichment of GO terms involved in neural function for four of the five phenotypes: 90th percentile, Intensity, Location ratio and Strength. Specific GO terms include axon guidance, generation of neurons, neuron differentiation, neurogenesis and neuron recognition. There was also significant enrichment of GO terms associated with response to a chemical stimulus or chemotaxis for all phenotypes, with the exception of 30th percentile. Some of the GO terms were composed of largely overlapping subsets of GWA candidate genes. Such was the case for analyses of Location ratio candidates and GO enriched terms neuron differentiation and generation of neurons. On the other hand, functional enrichment analyses of GWA candidate genes associated with Intensity, Location ratio and Strength independently resulted in the identification of the same significantly enriched GO term, generation of neurons, but the GWA candidate genes within this term largely differed by phenotype (Table S10). Moreover, in the case of Location ratio, GO analyses of sex-specific GWA candidates revealed the significant enrichment of GO terms to be female specific.
To summarize, our GWA analyses identify SNPs associated with multiple known olfactory genes as well as genes previously unknown to mediate odour-guided behaviour. Our network analyses also point to many polymorphic genes contributing to odour-guided behaviour and show that these candidate genes are enriched for functions involved in neural processing. Several GWA candidate genes were associated with more than one phenotype. This overlap may constitute additional support for these genes in mediating odour-guided behaviour or simply reflect the correlation between the phenotypes. There are indications that, despite this overlap, different spatial and temporal aspects of the behaviour are in fact largely distinct phenotypes. Thus, GWA analyses using several, albeit related, components of a behaviour can identify some candidate genes that underlie one component, and some that underlie several or all components.
Candidate genes associated with odour-guided behaviour: RNA interference
Polymorphisms in both known and novel candidate genes were identified as contributing to variation in odour-guided behaviour. To determine whether candidate genes for which natural genetic variation was associated with variation in olfactory behaviour are integral to the manifestation of olfactory behaviour we conducted a targeted transgenic RNAi knockdown experiment using the GAL4/UAS system (Brand & Perrimon 1993; Ni et al. 2008, 2009). Gene selection for the RNAi experiment was based on the level of statistical significance of the genotype–phenotype association, involvement in nervous system function, and/or level of network interconnectivity. We examined the potential role of GWA candidate genes in olfactory neural circuitry, by knocking down the expression of individual genes in different neuronal populations, and measured behavioural responses to 2,3-butanedione. Differences in odour-guided behavioural responses relative to the control were observed for 86% of genes tested (Fig. 5; Table S11). The effects on the five odour guided phenotypes depended on which gene was disrupted as well as its neuroanatomically specific effect. For GWA candidate genes previously implicated in odour detection, processing, and/or sensory organ development, RNAi knockdown of these genes resulted in effects on spatial and/or temporal behavioural responses. For instance, seven up (Svp) affects progenitor cell identity during antennal development (Sen et al. 2003) and RNAi knockdown in olfactory sensory neurons resulted in increased Location ratio responses. Rfx is a transcription factor implicated in sensory neuron differentiation and in Rfx mutant flies, larvae had abnormal chemosensory responses (Dubruille et al. 2002). Here, we show that RNAi knockdown in adults, resulted in reduced Intensity responses. For genes previously not known to play a role in olfaction, however, there is a question as to their role in directly influencing olfactory behaviour. Still life (sif) is a gene involved in the regulation of axonogenesis (Sone et al. 2000) and knockdown of sif in projection neurons resulted in generalized changes in odour-guided behaviour. The genes fz and Nedd4 also had effects on multiple odour-guided phenotypes and the phenotypes affected depended on the location of fz or Nedd4 knockdown. Fz and Nedd4 play roles in the conserved Notch and/or Wnt receptor signalling pathways. Notch signalling has been shown to affect olfactory receptor neuron diversification (Endo et al. 2007; Lieber et al. 2011; Reddy et al. 1997) and Wnt receptor signalling, through glycogen synthase kinase-3β, to influence processes such as olfactory habituation and maintenance of olfactory sensory neuron activity (Chiang et al. 2009; Sakurai et al. 2009; Wolf et al. 2007; Yao et al. 2007). Finally, we note that the extent to which our RNAi experimental results paralleled our GWA results varied. For example, our GWA analyses revealed that polymorphisms within or near Nedd4 were significantly associated with the phenotypes Location ratio and 90th percentile, while RNAi mediated knockdown of Nedd4 resulted in observed effects on these two phenotypes as well as Intensity. Future work will be needed to assess the mechanisms by which these polymorphic genes affect behaviour and their impact on the evolution of adaptive trait variation in nature.
Studies of odour-guided behaviour in Drosophila can serve as a model for understanding the genomic architecture underlying behaviour. In our study of both the spatial and temporal dynamics of odour-guided behaviour to 2,3-butanedione we identified many molecular polymorphisms that were associated with variation in behaviour, with some contributing to variation in a single sex. This result is not unexpected. Previous studies have shown epistatic networks of loci underlying odour-guided behaviour and in some cases different loci affected behavioural variation in males and females (Anholt et al. 2003; Mackay et al. 1996). Importantly, however, classic studies of olfactory behaviour typically did not dissect the spatial and temporal responses to odour and our GWA analyses revealed largely non-overlapping polymorphic loci underlying these behavioural phenotypes. Our results highlight the benefits of dissecting behavioural responses in both space and time because doing so may identify candidate loci contributing to variation in more subtle aspects of what is broadly called ‘olfactory behaviour’.
Genetic variation in odourant receptor and odourant binding protein loci have been previously linked to differences in olfactory behaviour (Richgels & Rollmann 2011; Rollmann et al. 2010; Wang et al. 2007, 2010). Interestingly, however, in this study the preponderance of polymorphisms associated with behavioural responses to 2,3-butanedione were in genes involved in neural development and central processing. This suggests that prime contributors to differences in olfactory behaviour may be at the level of changes to central brain circuitry, and more than changes in simply detection and sensitivity. This may be the reason we found qualitative, and not only quantitative, differences between lines, i.e. responses ranged from highly attracted to slightly repelled.
We speculated that our GWA analyses, if resulting in numerous true positives, would identify genes with known interactions in documented networks. We used a systems biology approach to examine each phenotype, and found enrichment of several neural-specific GO terms. We also found evidence for genetic and protein–protein interactions between genes. The significant enrichment of neural GO terms is in line with Swarup et al. 2013 that found neurogenetic networks underlying avoidance responses to benzaldehyde. Furthermore, the examination of a subset of GWA candidate genes using targeted RNAi experiments resulted in several aberrant olfactory responses, supporting our assertion that our GWA analyses can identify candidate genes important in mediating olfactory behaviour. These directed disruptions to specific populations of cells not only implicated genes previously not known to mediate olfactory behaviour, but also allowed us to gain insights into the mechanism of action of these genes. Finally, in many cases GWA results paralleled RNAi results in that the same olfactory phenotypes were affected. However, there were some differences and we speculate that these differences reflect, potentially, the nature of the change (protein coding vs. non-coding) or the restriction of RNAi knockdown to specific cell types. Regardless, the combined work sets the stage for future work on detailed mechanisms by which allelic variation in these loci affect behaviour.
Genome wide association studies also have limitations that should be acknowledged. The statistical power to detect a significant association increases with sample size. Hence, larger sample sizes could potentially reveal additional polymorphisms influencing variation in responses to 2,3-butanedione and additional candidate genes. In addition, examination of other populations could identify polymorphisms in additional genes that contribute to olfactory behaviour. These limitations are not restricted to our study. In our study we are fortunate to be able to take advantage of a population of inbred lines derived from nature that have limited population structure (Mackay et al. 2012) and overall low LD, with the exception of polymorphisms in relatively close physical proximity. These genotypes also allow for repeated behavioural measurements in a controlled environment, reducing environmental noise. Examination of genome wide associations in other populations has been proposed as a method of testing GWA results (Liu et al. 2008) and future analyses of alternative populations (e.g. Flyland or Drosophila Synthetic Population Resource Panel; Huang et al. 2012; King et al. 2012) could reveal additional insights into the genetic architecture of spatial and temporal aspects of olfactory behaviour. However, different environmental conditions have been shown to alter epistatic networks underlying olfactory behaviour in Drosophila and epistasis has been shown to be an important contributor to quantitative trait variation (Huang et al. 2012; Jhaveri et al. 2003; Sambandan et al. 2008). Therefore, testing for associations in a different population may also risk dismissing genetic variants that show genotype by environment interactions (Chan et al. 2011).
In conclusion, our study has identified polymorphisms in new genes influencing olfactory behaviour and suggests that subtle changes influencing nervous system function can result in marked differences in behaviour. This work highlights the benefits of considering spatial and temporal information in analyses of the genetic basis of behaviour.
We would like to acknowledge Tiffany Cook, Jim Deddens, Mike Magwire, and Ken Petren for help with statistical analyses and/or helpful discussions. We also thank Sarah Smith, Nina Mecca and Richard Kirby for technical assistance and the TRiP at Harvard Medical School (NIH/NIGMS R01-GM084947) for transgenic RNAi stocks. Fumika Hamada kindly provided select Gal-4 driver lines. This work was supported by the National Institutes of Health (GM080592 to S.M.R.).