We quantified aggressive behavior of 40 wild-derived inbred lines, using a rapid and high-throughput behavioral assay 19. Variation in aggressive behavior was continuously distributed among these lines, as expected for a quantitative trait. There was significant genetic variation in aggression among lines (F_40,779 = 73.0168, P < 0.0001; Figure 1). Estimates of among line (σ_L²) and within line (σ_E²) variance components were σ_L² = 0.783 and σ_E² = 0.217, for a broad-sense heritability (H²) of aggressive behavior of H² = 0.78. Surprisingly, there was a 25-fold range of aggressive behavior in these lines: from an average of 3.3 to 76.9 aggressive encounters for 8 flies in a 2-minute observation period.

The variation among the inbred lines far exceeds that of lines selected for 21 generations for increased and decreased aggressive behavior, which only differ less than threefold (with a mean of 14.2 and 34.2 encounters in the high and low selection lines using the same assay) 18. Under a strictly additive model, we expect variation among fully inbred lines to be twice the additive genetic variation in the base population from which they were derived 16. Thus, under strict additivity, the estimate of the narrow sense heritability (h²) in this population would be h² = 0.64. This is much greater than the estimate of realized h² from response to selection (h² ≈ 0.09) 18, indicating that alleles affecting natural variation are recessive and/or interact epistatically.

Previously, we quantified variation in gene expression among these wild-derived inbred lines 21. A total of 7,508 transcripts were significantly variable among lines in males at a false discovery rate (FDR) of < 0.01 and 3,316 probes contained SFPs. We identified 133 QTTs (P < 0.01) associated with variation in aggressive behavior (Additional data file 1). In addition, 167 SFPs (P < 0.05) with a minor allele frequency of at least 10% were associated with variation in aggression; these represent 137 independent genes (Additional data file 2). Four of the QTTs were also implicated as candidates from the SFP analysis (CG1146, CG2556, CG31038 and methuselah-like 8). No gene ontology information is available for three of these genes (CG1146, CG2556, and CG31038). methuselah-like 8 encodes a predicted G protein coupled receptor that may affect the determination of life span 22.

In total, these analyses implicate 266 unique candidate genes associated with natural variation in aggressive behavior. These candidate genes are involved in a broad spectrum of biological processes, including vision, olfaction, learning and memory, and the development and function of the nervous system (Additional data files 1 to 4). However, the candidate genes are also involved in transcription, protein modification, mitosis and other basic cellular processes (Additional data files 1 to 4). More than half of the genes with annotations are involved in metabolism, nearly 60% have protein binding functions, and approximately 25% are implicated in development (Additional data file 5) 23.

Two categories of candidate genes are worthy of mention. We found a member of the Cytochrome P450 gene family associated with aggressive behavior, Cyp4p2. Members of this gene family have also been associated with aggressive behavior in previous studies 1718. Cytochrome P450s are generally involved in oxidation, metabolism, protection from xenobiotics, and possibly pheromone recognition 24. The repeated implication of this class of genes suggests that some or all of these functions, or yet unknown functions of this class of proteins, mediate aggressive behavior, although it remains unclear precisely how. We also found three genes that have been previously implicated in learning and/or memory to be associated with natural variation in aggression in this screen - nord, visgun, and klingon 25 - consistent with a previous report that Drosophila aggressive behavior is associated with learning and memory 26. Perhaps variation in these genes affects the fly's learning ability, which could subsequently influence the behavioral response to aggressive encounters. Assessment of these wild-derived lines in a learning and memory assay could inform our understanding of the relevance and variation of social memory in wild Drosophila.

A total of 26 of the 266 candidate genes identified in this study overlapped with the candidate genes implicated from the correlated response of the transcriptome to selection for divergent level of aggressive behavior 18, from a different sample of the same base population as the one from which the inbred lines were derived (Additional data file 6). This is no more overlap than expected by chance (χ₁² = 0.36, P > 0.05). There are several possible - and not mutually exclusive - reasons why the degree of overlap between the two experiments is not more extensive. First, the observation that there is no more overlap between the two experiments than expected by chance could mean that there are many rare alleles affecting aggressive behavior segregating in nature, such that two independent samples captured different subsets of alleles. Second, the flies from the selection lines were not mated, and had been starved for 90 minutes prior to RNA extraction, in contrast to the mated, fully fed flies for which transcript profiles were obtained in this experiment. Third, the control line was the most extreme for many of the transcripts that were divergent among the selection lines; this type of transcript-phenotype association will not be detected in a linear regression. Fourth, selection causes linkage disequilibrium between the selected locus and linked unselected loci; changes in transcript abundance among these linked loci between the selection lines are false positive associations. In contrast, the rapid decay of linkage disequilibrium in regions of normal recombination in unselected Drosophila 2728 minimizes false positive associations of transcript abundance of linked loci in the unselected inbred lines. Fifth, a greater fraction of the genetic variation among the inbred lines than the selection lines is due to dominance and epistasis. The transcriptional signature of a homozygous recessive allele in the inbred lines is likely to be different from the same allele as a heterozygote in the selection lines. Thus, the overlap of genes between the two studies may be enriched for loci with additive effects that causally affect natural variation in aggressive behavior.

To evaluate whether the candidate genes suggested from these analyses potentially affect aggressive behavior, we assessed aggression levels of P-element insertional mutations in 12 of the candidate genes, and their co-isogenic control lines. Nine of the mutant alleles were associated with significantly different aggression levels from the control (Figure 2). This high 'success' rate shows that expression profiling of wild-derived genetically divergent lines is an efficient method for identifying candidate genes affecting complex traits, as has been observed previously 1718293031.

Flies with mutations in CG11448, CG13760, CG2556, CG31038, CG32425, late bloomer and skuld are all more aggressive than their controls, while flies with mutations in GTPase-activating protein 1 (Gap1) and schizo are less aggressive than the control strain. No gene ontology information is available for the predicted genes tested; however, CG11448 is homologous to the amyloid beta A4 precursor protein, which is implicated in Alzheimer's disease. late bloomer has a role in nervous system development and synapse biogenesis. It is homologous to TSPAN7, a tetraspanin protein implicated in mental retardation 32. skuld is involved in numerous transcription-related processes, and also has roles in metabolism and development. Gap1 has roles in the cell cycle, and is also involved in signal transduction and numerous developmental processes, such as axis specification and sensory organ development. Finally, schizo is involved in several signal transduction pathways, the development of the central nervous system, and muscle development. It is homologous to the human protein ADP-ribosylation factor guanine nucleotide exchange factor 2, dysfunctions of which are associated with microcephaly 33.

The transcriptome is highly genetically inter-correlated 21. This correlation structure can be used to infer modules of genetically correlated transcripts associated with aggressive behavior, after removing the correlations among the transcripts attributable to their association with aggression itself. The number and contents of modules are determined such that the average correlation of probe sets within a module is maximized, while the average correlation among probe sets in different modules is minimized. The 133 QTTs grouped into 9 modules, ranging in size from 2 to 54 probe sets (Figure 3a; Additional data file 7). The correlated transcript modules associated with aggressive behavior can also be represented as an interaction network, with edges between transcripts in the network determined by genetic correlations in transcript abundance exceeding a threshold value (Figure 3b represents |r| ≥ 0.7). Note that these are, at present, undirected networks. We do not know which transcripts are causally associated with variation in aggression, due to functional polymorphisms in cis-regulatory regions, and which transcripts are trans-regulated and change expression as a consequence of cis-regulatory variation at another locus 34.

We evaluated the biological plausibility of the modules by querying whether genes in the modules are enriched for shared gene ontology categories, tissue-specific expression patterns, or DNA sequence motifs (the latter using the Multiple EM for Motif Elicitation (MEME) tool). Approximately one-third of the transcripts in module 6 affect ion binding, relative to approximately 2% of the probe sets in the genomic background; this is a significant enrichment (P < 0.01). Nearly 50% of the annotated genes in module 6 are involved in establishment of localization, compared to approximately 13% of the background (P < 0.001); 25 to 30% of the genes in modules 6 (P < 0.05) and 7 (P < 0.01) are involved in cell communication, whereas only 13% of the background falls into that category (Figure 4). Module 7 is enriched for several categories related to development (Figure 4). Transcripts in modules 6 and 7 are enriched in the brain, head, and thoracicoabdominal ganglion (Figure 5), indicating that these genes function primarily in central nervous system functions. However, the fact that they fall into distinct modules suggests that their specific functions differ, or that they are differentially regulated in a temporally or spatially specific manner.

Additional support for the hypothesis that genes in a module are co-regulated is generated by shared MEMEs among members of a module 35 (Figure 6). The P-value for each gene containing the consensus sequence represents the probability of a random sequence having the same match score or higher. Of 35 genes in module 6, 29 share a motif with a 20-bp consensus sequence, and the significance values for genes containing this motif range from P = 2.68 × 10^-4 to P = 1.82 × 10^-10 (Figure 6a). Of 54 genes in module 7, 18 share a 14-bp motif, with P-values ranging from P = 9.32 × 10^-6 to P = 4.73 × 10^-9 (Figure 6b).

Although many of the QTTs lack annotation, we can infer potential functions based on the characterized genes that fall into the same correlated module. Three of the four QTTs in module 1 belong to a large transcriptional module enriched for male biased transcripts 21, and these genes are highly expressed in the testis 36; perhaps this module is related specifically to male reproductive functions. Of the five QTTs in module 4, three are involved in visual perception. Their correlated expression implies that the others, CG13928 and CG6403, might share a similar function. The fact that all of these transcripts are highly expressed in the head supports this possibility (Figure 5). Three of the four annotated genes in module 8 are involved in metabolic functions, suggesting a similar role for the uncharacterized genes in that module. Additional tests can help us tease apart the relationships among genes within a module. For example, manipulation of a single gene and assessment of the effects on other genes within the same module can elucidate causality and direction of effects.

The wild-derived inbred lines have been assessed for variation in other complex traits: longevity, starvation stress resistance, chill coma recovery time, locomotor reactivity (a startle response), copulation latency, competitive fitness and sleep traits 2137. At the level of organismal phenotype, only locomotor reactivity was significantly genetically correlated with aggressive behavior (r_G = 0.49, P < 0.001). However, organismal genetic correlations can only be significant if alleles affecting both traits have largely similar positive or negative effects on the traits 16. There can be substantial pleiotropy in the absence of genetic correlation if alleles at many loci affect both traits, but the sign of the effects is not correlated. This motivated us to ask whether particular modules of transcripts associated with aggressive behavior were associated with modules of transcripts associated with the other traits (Additional data file 8). Many of the probe sets implicated in multiple traits correspond to predicted genes about which little is known. However, transcript abundance of synaptogyrin, which is involved in synaptic vesicle exocytosis 38, is associated with variation in starvation resistance and fitness 21. Rab9 is associated with chill coma recovery 21 and sleep 37. GRHRII, which encodes a predicted G-protein coupled receptor 39 and gonadotropin-releasing hormone receptor 40, is associated with starvation resistance 21 and sleep 37.

In addition to examining genetic correlations between QTTs affecting aggressive behavior, we can ask which of the genes affecting aggressive behavior are most highly correlated (r ≥ 0.70) to the transcriptome (Figure 3b). Three QTTs stand out as being highly connected. miple transcript abundance is highly correlated with 22 other transcripts. It is highly pleiotropic, and is thought to affect locomotor behavior, muscle development, ATP binding, synapse biogenesis, and response to stimulus 22. VAChT expression is correlated with 21 other transcripts. It is described as an acetylcholine transporter, and is also involved in the response to a chemical stimulus 22. Another 'hub' gene is unc-104, which falls into many of the Gene Ontology categories described for miple; it is also involved in nucleotide binding. Mutations in human homologues have been implicated in spastic paraplegia and Charcot-Marie-Tooth disease 22. Additional highly connected genes are the computationally predicted genes CG2790, CG13928, CG14853, and CG6156, about which little annotation information is available, although CG2790 and CG13928 are reportedly involved in zinc ion and protein binding.

Expression of all of these integral genes is highly enriched in the brain, head, and thoracicoabdominal ganglion. Furthermore, the male accessory glands exhibit enrichment of unc-104, and CG6156 is up-regulated in the crop, tubule, larval tubule, and larval fat body 36. Their high degree of connectivity implies that these genes might be central to networks involved in aggressive behavior. The range of biological processes and molecular functions in which they are involved makes it difficult to isolate which are relevant to aggression, but their high expression levels in the head and nervous system unsurprisingly implicate those tissues in the modulation of aggression. We can also use these data to develop hypotheses about the highly connected yet uncharacterized genes CG2790, CG13928, CG14853, and CG6156.

Aggression is clearly a highly complex trait - we have identified 266 candidate genes associated with natural variation in aggressive behavior, none of which have been previously implicated to affect aggression. Follow-up functional validation shows that 75% of P-element insertional mutations tested in these candidate genes indeed affect aggression. The candidate genes embrace a wide range of biological functions with plausible connections to aggressive behavior (sensory perception and chemosensation, function and development of the nervous system), as well as other general functions with less obvious relationships to aggression per se (metabolism, protein modification, mitosis). Analysis of natural variants affecting complex traits that have survived the sieve of natural selection thus gives insights about the genetic basis of complex behaviors that are not possible from analysis of mutations of large effect. That none of the genes previously implicated to affect aggression was detected in this screen is somewhat surprising. There are several possible explanations. The known candidate genes may not be genetically variable at the level of transcription; we could not detect genetically variable transcripts at these loci because they are expressed at low levels or at a different developmental stage; our SFP map detects only a small fraction of polymorphic variants; and the candidate genes may not tolerate functional variation due to strong purifying selection. For example, variation in fruitless was not associated with variation in aggressive behavior in this study or previous studies 1718. Only one of the seven probe sets on the array that target fruitless was genetically variable, and variation in fruitless expression for this probe set was not associated with variation in aggressive behavior.

The QTTs associated with natural variation in aggressive behavior group into genetically correlated modules with shared functional annotations, sequence motifs, and tissue-specific expression. These modules are, in turn, correlated with other traits, providing insights about the molecular basis of pleiotropy between aggression and other behavioral and fitness-related traits. These results provide the foundation for a systems genetics analysis of natural variation in aggressive behavior. The future availability of whole genome DNA sequence variation for these lines will enable us to discriminate cis- from trans-acting polymorphisms, and infer the direction of the flow of information through the network. The entire suite of 266 candidate genes provides a focal point for linkage analysis of segregating populations derived from the inbred lines. Further, the inbred lines can be characterized for other quantitative traits, including components of metabolism, which will enable us to interpret the balance of selective forces maintaining variation for aggressive behavior in natural populations on a genome wide scale. Extension of these analyses to a larger sample of inbred lines will increase the power of network analyses, and provide a more representative sample of allelic diversity associated with aggressive behavior. Finally, it is not inconceivable that our understanding the genetic underpinnings of variation in aggressive behavior in Drosophila could be used to develop novel pharmacological therapies for treatment of pathological aggression in humans and domestic animals.

The 40 inbred lines were derived by 20 generations of full-sib mating from isofemale lines that were collected from the Raleigh, NC farmer's market in 2003 21. Flies were reared under standard culture conditions on cornmeal-molasses-agar medium at 25°C, 60 to 75% relative humidity, on a 12-h light-dark cycle. P-element insertional mutations and their co-isogenic control lines were obtained from Bloomington Drosophila Stock Center, Bloomington, Indiana, USA.

Behavioral assays were performed as described previously 18 on socially experienced, 3- to 7-day-old males. Flies were not exposed to anesthesia for at least 24 h prior to the assay. A total of 20 replicate assays were performed for each line, with one replicate per line per day for a total of 20 days. Each replicate consisted of a group of eight 3- to 7-day-old flies of the same genotype. The flies were placed in a vial without food for 90 minutes, after which they were transferred (without anesthesia) to a test arena containing a droplet of food and allowed to acclimate for 2 minutes. After the acclimation period, the flies were observed for 2 minutes; the aggression score of each replicate was the total number of aggressive interactions observed among all eight flies in the 2-minute observation period. Behavioral assays were conducted in a behavioral chamber (25°C, 70% humidity) between 8 a.m. and 11 a.m.

The gene expression analysis has been described previously 21. Briefly, RNA was extracted from two independent pools of 25 3- to 5-day-old mated whole flies/sex/line that were frozen at the same time of day, labeled, and hybridized to Affymetrix Drosophila 2.0 arrays, using a strictly randomized experimental design. The raw array data were normalized using a median standardization. The measure of expression was the median log2 signal intensity of the probes in the perfect match probe sets, after removing probes containing SFPs between the wild-derived lines and the reference strain sequence used to design the array. Negative control probes were used to estimate the level of background intensity; probe sets with expression levels below this threshold were considered to be not expressed.

The analysis of variance (ANOVA) model Y = μ + L + ε was used to partition variation in male aggressive behavior and transcript abundance between lines (L, random) and the variation within lines (ε). A FDR of < 0.01 41 was used to assess significance of the L term in the analyses of natural variation in gene expression, to account for multiple testing. Broad sense heritabilities (H²) were estimated as:

- where σ_L² and σ_E² are the among line and within line variance components, respectively. Estimate of cross-trait genetic correlations were:

- where cov_ij is the covariance of line means between trait i and trait j, and σ_i and σ_j are the square roots of the among line variance components for the two traits. Differences in aggressive behavior between P-element insert lines and their co-isogenic controls were assessed by t-tests, with significance levels based on Bonferroni-corrected P-values. Simple linear regressions were used to identify QTTs significantly associated (P < 0.01) with variation in aggressive behavior across the 40 lines. Similarly, ANOVA models (Y = μ + M + ε, where M denotes SFP presence or absence) were used to identify SFPs significantly associated (P < 0.05) with variation in aggressive behavior.

The genetic correlations between all transcripts significantly associated with aggressive behavior were computed after removing the correlation between these transcripts and the phenotype. This was achieved by fitting the model Y = μ + E + ε(Y is the phenotype and E is the covariate median log2 expression level) and extracting the residuals to compute the genetic correlations for module construction 21. Modules of transcripts associated with aggressive behavior with coordinated patterns of expression across the 40 lines were then quantified as described previously 21 by transforming the pairwise genetic correlations among transcripts into Euclidean-like distances, which were used to construct an affinity matrix. The transcripts were partitioned into modules using a graph-theoretical approach that envisions the transcripts as nodes in an undirected graph whose edges are weighted by the entries of the affinity matrix. Transcriptional modules common to aggressive behavior and other phenotypes measured on the 40 wild-derived inbred lines 2137 were identified by comparing the transcripts in each aggression module to the transcripts in each module from the other phenotypes, and determining whether the overlap between the modules exceed what is expected by chance using a Fisher's exact test 21.

Statistical analyses were performed using JMP 7.0 software (SAS, Cary, NC, USA). Functional annotations of genes are based on FlyBase 38 annotations; additional information was obtained using FlyMine v12.0 22 and Babelomics v2 and v3 23. Categories that were represented by fewer than 5% of the gene list queried were excluded. Statistically significant over- or under-representation was determined by the online software used when available; otherwise, a chi-square test was performed, using the appropriate genomic background to determine the expected values.