To systematically ascertain the function of signalling pathway genes and their interactions with Hippo signalling in the development of the D. melanogaster ovary, we first curated a list of all known and predicted signalling genes (Gramates et al., 2017; Kanehisa et al., 2010; Mbodj et al., 2013). We identified 475 genes belonging to the 14 developmental signalling pathways characterised in D. melanogaster (Table 1; Supplementary file 1), and obtained UAS:RNAi lines for 463 of these genes from the Vienna Drosophila RNAi centre (VDRC) or the TRiP collections at the Bloomington Drosophila Stock centre (BDSC) (all D. melanogaster genetic lines used are listed in Materials and methods).

We previously showed that reducing the levels of hpo in the somatic cells of the larval ovary using traffic jam Gal4 (tj:Gal4) driving hpo[RNAi] increased both ovariole number and egg laying of adult female flies (Sarikaya and Extavour, 2015). To identify genes that modify these phenotypes, we used tj:Gal4 to drive simultaneous hpo[RNAi] and RNAi against a signalling candidate gene, and quantified the phenotypic change (Figure 1a–d). We observed that on driving two copies of hpo[RNAi] using tj:Gal4, we obtained a further increase in both egg laying and ovariole number (Figure 1e). This indicates that ovaries have further potential to increase ovariole number and egg laying beyond the increase induced by tj:Gal4 driving one copy of hpo[RNAi], and that tj:Gal4 can drive the expression of two RNAi constructs, indicating that our screen could identify both enhancers and suppressors of the tj:Gal4 >hpo[RNAi] phenotype.

We proceeded to identify modifiers of the tj:Gal4 >hpo[RNAi] phenotype by crossing males of each of the 463 candidate genes RNAis individually with tj:Gal4 >hpo[RNAi] females, and performing three phenotypic screens on the offspring. In the first screen (Figure 1a), we measured egg laying of three F1 female offspring (tj:Gal4 >hpo[RNAi], signalling candidate[RNAi]) over five days. To address batch variation (Figure 1—figure supplement 1), we standardised egg laying measurements by calculating the Z scores (Z_gene = number of standard deviations from the mean) for each candidate line relative to its batch controls. 190 genes had an egg laying |Z_gene| below 1. Previous studies have shown that the egg laying of newly eclosed adult mated females correlates with ovariole number during the first five days (Klepsatel et al., 2013b). We therefore eliminated these 190 genes from subsequent screening, because the change in egg laying was so modest that we considered these candidates were unlikely to show changes in ovariole number when compared to controls.

In the second screen (Figure 1b), we measured egg laying in a wild-type background (tj >signalling candidate[RNAi]) for the 273 remaining candidate genes. For the third screen (Figure 1c), we quantified the ovariole number of tj:Gal4 >hpo[RNAi], signalling candidate[RNAi] F1 adult females for the same 273 candidate genes. To choose candidates from the second and third screens for further study, we wished to account for the fact that the two screens had different effective numbers of data points. This was because egg laying data were obtained from individual vials of three females over five days, while ovariole numbers were obtained from 20 ovaries from ten females (see Materials and methods). We therefore selected the 67 genes with a |Z_gene| above two for ovariole number (Figure 1c and d; Table 2), and the 49 genes with a more conservative |Z_gene| above five for egg laying (Figure 1a, b and d; Table 2), for a total of 116 positive candidates for subsequent analyses.

Standardisation of the results from the three screens using Z scores allowed us to compare the effects of individual genes on one or both of egg laying and ovariole number. We performed a pairwise comparison of the Z_gene values for all combinations of screens, and considered genes with |Z_gene| values that were above the thresholds set for the phenotype in each screen (above two for ovariole number, above five for egg laying; green dots in Figure 2a–c). Across all three screens, loss of function of our positive candidates yielded reductions in ovariole number and egg laying more commonly than increases (Figure 2a–c). Comparing the |Z_gene| values of egg laying and ovariole number of tj:Gal4 >hpo[RNAi], signalling candidate[RNAi] adult females revealed that genes that caused a change in egg laying did not always similarly affect ovariole number, and vice versa (Figure 2a). We therefore hypothesise that egg laying and ovariole number may be regulated by genetically separable mechanisms. This hypothesis notwithstanding, we observed a weak but statistically significant correlation between egg laying and ovariole number (p=1e10⁻⁵; Figure 2d), and this correlation was most significant in adult females that had a drastic reduction in both phenotypes (Figure 2a).

We found that at least some genes from all tested signalling pathways could affect both egg laying and ovariole number (Figure 3). To determine if some pathway(s) appeared to play a more important role than others in these processes, we asked whether any of our screens were enriched for genes from a specific signalling pathway. To measure enrichment, we compared the distribution of individual pathway genes among the positive candidates in each screen, to a randomly sampled null distribution of pathway genes among a group of the same number of genes randomly selected from our curated list of 463 signalling genes (Figure 3a). Involvement of a pathway in the regulation of a phenotype would be reflected in a difference between the representation of pathway genes in an experimentally derived list and a randomly selected group of signalling genes. We found that rather than only one or a few pathways showing functional evidence of regulating ovariole number or egg laying, nearly all pathways affected both phenotypes (Figure 3a). We further tested this result by calculating the hypergeometric p-value for the enrichment of each signalling pathway, in each of the three groups of genes. Consistent with the results of the random sampling approach (Figure 3a), we found that most pathway members were not significantly enriched for egg laying or ovariole number phenotypes (Figure 3b). The absence of significant enrichment of any specific pathway is not simply attributable to the pool of genes that were screened, because our experimental manipulations of ovariole number and egg laying did cause a change in the distribution of signalling pathway members (Figure 3—figure supplement 1). Instead, both phenotypes appeared to be regulated by members of most or all signalling pathways (Figure 3). The only two exceptions to this trend were a greater than twofold enrichment of (1) genes from the Notch signalling pathway in the regulation of ovariole number (p-value<0.05, pink bar in Figure 3a,b), and (2) members of the Hedgehog (Hh) signalling pathway in the regulation of Hippo-dependent egg laying (p-value<0.05, brown bar in Figure 3a and b; Figure 3—figure supplement 2). In summation, our analyses of the enrichment of signalling pathways within the different screens indicated that both ovariole number and egg laying are regulated by genes from nearly all described animal signalling pathways (Figure 3a), rather than being dominated by any single pathway.

Comparing the results of the Egg Laying screens performed in a wild type background (Figure 1b) or in a hpo[RNAi] background (Figure 1a), revealed that most of the genes that met a threshold of |Z_gene| > 5 in one screen, did not meet that threshold in the other screen (Figure 2c). This result suggests the existence of both Hippo-dependent and Hippo-independent mechanisms of regulation of egg laying. The interpretations of separable Hippo-dependent and -independent regulation of egg laying, and of the separable regulation of ovariole number and egg laying, were supported by the results of the network analysis described in the following section.

The finding that these reproductive traits were regulated by the genes of all signalling pathways led us to consider the broader topology of putative gene regulatory networks in the analysis of our data. Previously characterised genes in the ovary are often pleiotropic and can regulate both ovariole number and egg laying (Gilboa, 2015; Sarikaya and Extavour, 2015). As with proteins in a linear pathway, proteins in a protein-protein interaction network (PIN) are more likely to function in conjunction with genes that are connected to them within the network (e.g. Ideker and Sharan, 2008; Jeong et al., 2001). Centrality is one measure of the connectedness of a gene in the PIN and can be used to identify the most important functional centres within a protein network (Hahn and Kern, 2005; Ma'ayan, 2011). Most centrality measures use path length, which is a measure of the number of other proteins required to link any two proteins in the network. Here, we used four commonly used metrics to quantify gene centrality, each measuring slightly different properties (Jalili et al., 2016; Koschützki and Schreiber, 2008). (1) Degree centrality is proportional to the number of proteins that a given protein directly interacts with. (2) Betweenness centrality measures the number of shortest paths amongst all the shortest paths between all pairs of proteins that require passing through a particular protein. (3) Closeness centrality measures the average shortest path that connects a given protein to all other proteins in the network. (4) Eigenvector centrality is a measure of the closeness of a given protein to other highly connected proteins within the network.

We hypothesised that if the candidate genes we identified in our screen as playing roles in ovarian function worked together as a PIN, then the degree of centrality of a gene might be an indicator of function. To test this hypothesis, we first compiled a PIN consisting of all described interactions between D. melanogaster proteins, from the combination of publicly available protein-protein interaction (PPI) studies in the DroID database (see Materials and methods). We then calculated the four centrality measures described above for all genes within the D. melanogaster PIN (Supplementary file 1). We rank ordered only the genes tested in each screen by their score for each centrality measure, and asked whether their rank order correlated with the results of the screen, plotting these results as a receiver operating characteristic (ROC) curve. Positive correlations between centrality (a continuous variable) and phenotype (a binary variable: above or below the |Z_gene| threshold) are reflected in an area under the curve (AUC) of more than 0.5. We found that the higher the centrality score, the greater the likelihood that a gene had |Z_gene| values above our threshold for effects on ovariole number and egg laying (Figure 4a; Figure 4—source data 1). This supports the premise that the positive candidates identified in our screen function together as a network in the regulation of either ovariole number or egg laying. Interestingly, while the centrality of genes did predict whether a gene would affect our phenotypes of interest, it could only weakly predict the strength of that effect (p-value<0.05 in Figure 4—figure supplement 1).

The centrality analyses above suggested that the genes implicated in ovariole number and egg laying displayed characteristics of a functional network. PINs can often be further sorted into a collection of sub-networks. A sub-network is a smaller selection of proteins from the PIN. Examples of such sub-networks could be proteins within the same subcellular organelle (Foster et al., 2006) or genes that are expressed at the same time (Spellman et al., 1998), thus making them likely to function together (Srinivasan et al., 2007). A putative module is a sub-network that can perform regulatory functions as a unit, independent of other sub-networks, and has key measurable features (Barabási and Oltvai, 2004; Hartwell et al., 1999; Ravasz et al., 2002; Yook et al., 2004). Genes and interactions between genes are not mutually exclusive to such putative modules and can be shared between putative modules. We therefore asked if our sub-networks, consisting of genes that showed similar mutant phenotypes, might display features of modularity. To determine whether genes that were implicated in regulation of ovariole number and egg laying interacted with each other in specific groups more than would be expected by chance, we created four lists of genes, called ‘seed’ lists, based on their individual phenotypic effects based on our screen results: (1) the core seed list, including genes positive in all three screens (Figure 4b); (2) the egg laying seed list, including genes positive in the wild type background egg laying screen (Figure 1b; Figure 4c); (3) the hpo[RNAi] egg laying seed list, including genes positive in the hpo[RNAi] background egg laying screen (Figure 1a; Figure 4c); and (4) the hpo[RNAi] ovariole seed list, including genes positive in the hpo[RNAi] background ovariole number screen (Figure 1c; Figure 4c). Interestingly, the core seed list, comprising genes that affected all three measured phenotypes, only consisted of genes that caused a reduction in both ovariole number and egg laying (Figure 4b).

We then asked whether the members of these four seed lists were more connected than would be expected by chance. In other words, we formally tested them for modularity as defined above. Meeting our criteria for modularity would suggest that the genes in these phenotypically separated seed lists might operate together as putative functional modules within the Drosophila PIN. We performed our modularity test using four commonly measured network metrics: (1) Largest Connected Component (LCC) (the number of proteins or nodes connected together by at least one interaction), (2) network density (the relative number of edges as compared to the theoretical maximum), (3) total number of edges, and (4) average shortest path (average of the minimum distances connecting any two proteins). We considered a sub-network to show modular features if they showed most of the following properties: higher LCC, higher network density, more edges, and shorter average shortest path length, when compared to a similarly sized, randomly sampled selection of genes from the PIN.

To determine whether these criteria would correctly identify signalling genes, which are known to function together as a module, we measured these four parameters in the original set of genes (all signalling genes) used in this study (Table 1). We found that the signalling genes display features of modularity when compared both to a randomly selected set of genes, as well as to a degree-controlled list of genes (Figure 4—figure supplement 2a). We then used this approach to test the modularity of the four phenotypic sub-networks, when compared to two different ‘control sub-networks’ consisting of a group of the same number of genes as contained in the sub-network, one chosen randomly from among the candidate genes from our initial screen list (Table 1), and the second chosen from a degree-controlled list of genes selected from the entire PIN (see Materials and methods: Building degree-controlled randomised networks). We found that the four predicted phenotypic sub-networks showed higher LCC, higher network density, and more edges (Figure 4—figure supplement 2b), compared to both ‘control sub-networks’. This result suggests that these sub-networks display many features of modularity (although their average shortest path length is higher than controls, rather than lower) and may function as putative modules within the PIN to regulate one or both of ovariole number or egg laying.

Based on published molecular interactions, in addition to the four criteria described above, further evidence for putative functional modules of genes can also be obtained by applying algorithms that use either the shortest path method (Bromberg et al., 2008) or the Steiner Tree approach (Huang and Fraenkel, 2009). Such methods identify and predict functional connections between the seed proteins, as well as additional nodes (proteins or genes) that have not been experimentally tested within the given parameters, but are known to interact with the seed genes in the PIN (Albert and Albert, 2004; Yu et al., 2006). This process can provide evidence for or against the existence of a predicted functional module, and subsequent experimental testing of this predicted module can confirm its functionality. Given its recent success in predicting gene modules, we applied the previously published Seed Connector Algorithm (SCA), a member of the Steiner Tree algorithm family (Wang et al., 2017; Wang and Loscalzo, 2018), to the groups of genes that had similar phenotypic effects in our screens (seed genes; Figure 4b and c). The SCA connects seed genes and previously untested novel genes (connectors) to each other using a known PIN, producing the largest possible connected putative module given the data. Using the PIN and the aforementioned four lists of seed genes, we applied a custom python implementation of the SCA (Materials and methods: 04_Seed-Connector.ipynb) to build and extract the largest possible (given our PIN) connected putative modules that regulate egg laying and ovariole number.

This SCA method yielded four putative modules, one for each seed list, which we initially referred to as the Core Module (Figure 5b), hpo[RNAi] Egg Laying module (Figure 5—figure supplement 1), Egg Laying Module (Figure 5—figure supplement 2), and hpo[RNAi] Ovariole Number Module (Figure 5—figure supplement 3) respectively. Each of these four putative modules contained seed genes, which had been functionally evaluated in our screens (green and red circles in Figure 5), as well as connector genes, which were genes newly predicted as regulators of these phenotypes (green and red triangles in Figure 5). Of the four putative modules generated by the SCA, we found that the Core module had higher centrality measures than the other three putative modules (Figure 5—figure supplement 4). We interpret this to mean that the genes regulating these ‘Core phenotypes’ are more strongly connected to each other.

We found, however, that these four groups of genes produced by the SCA did not have increased LCC values, increased network density, more edges nor decreased average shortest path (Figure 5—figure supplement 5), compared to our ‘control sub-networks’. This result shows that the SCA in this instance does not provide evidence for putative functional modules from the four phenotypic sub-networks in our system, above and beyond the evidence provided by the application of the four network metrics discussed above. To be conservative in our description of these results, we therefore henceforth refer to these four groups of genes united by phenotype and with strong predicted interactions, as sub-networks rather than as modules. We noted that each of these four sub-networks contains genes from most, if not all, known signalling pathways, rather than only genes from a single pathway (Figure 5—figure supplement 6).

Our network analysis identified four highly connected sub-networks of genes that regulate two distinct developmental processes, together with or independently of Hippo signalling activity: ovariole number determination, which occurs primarily during larval development, and egg laying, which takes place in adult life (Figure 5). We wished to assess the degree to which there might be any shared genetic components between these four phenotypic sub-networks, and whether the addition of connector genes by the SCA had any impact on this. To understand potential interactions between the phenotypic sub-networks in the regulation of both ovariole number and egg laying, we constructed a composite network of all genes in each of the four phenotypic sub-networks (Figure 5b; Figure 5—figure supplement 1; Figure 5—figure supplement 2; Figure 5—figure supplement 3), which we refer to as the ‘meta network’ (Figure 6a). We then grouped the genes of the meta network into seven bins based on their phenotypic effects as measured in the three screens, resulting in sub-groups I through VII shown in Figure 6a. To ask whether the genes in these phenotypic groupings showed any notable interaction patterns, we compared the connectivity between genes assigned to the same phenotypic group, to the connectivity of a group of the same size randomly assembled from the genes of the meta network (Figure 6—figure supplement 1). As a measure of connectivity, we used an edge density map, which reflects the number of interactions between the genes within a group and between groups. We quantified the deviation between the edge density of each of groups I through VII, and their corresponding randomly assigned groups of the same size, by computing their respective Z score. When we included only seed genes in each of groups I through VII, we found that the edge density values of these groups were somewhat lower (Z score <−1.5) than those of the randomised groups (Figure 6—figure supplement 1a). The single exception to this was group IV, whose members shared more edges with each other than did the members of its randomised comparison group of the same size (Figure 6—figure supplement 1a). In other words, these groups of phenotypically binned seed genes were not notably more connected to each other then we would expect by chance.

In contrast, expanding each of the seven phenotypic sub-groups to include both seed genes and the connector genes predicted by the SCA changed the edge densities of these groups relative to their randomised control groups. Specifically, edge densities were much lower between groups I, II and III (Z score < −3), and much higher within group IV (Z score >3) (Figure 6—figure supplement 1). This shows that applying the SCA to these phenotypically binned groups increased the non-random differences in connectivity between them that were already present within the seed genes (Figure 6—figure supplement 1a), thus clarifying the internal structure of the meta network.

We then asked if these seven sub-groups were as connected to each other, as were the genes within each of the sub-groups, again using the edge density assessment as described above (Figure 6b). This analysis yielded three principal findings. First, edge densities between the three groups corresponding to the three scored phenotypes (I, II and III in Figure 6a) were very low (Figure 6b).This implies that the genes in each of the groups that regulate only one phenotype (I, II and III in Figure 6a) share more interactions with themselves than with genes in the other two groups, suggesting that each of these initially scored phenotype can be largely regulated by an independent, non-interacting set of genes. Second, the core group (IV in Figure 6a) displayed a higher edge density with the other three groups (I, II and III in Figure 6a) than any of those three groups did with each other (Figure 6b). Consistent with the definition of core genes as regulating all three scored reproductive phenotypes, this result suggests that the core genes, in contrast to those from the other three groups, may share substantial functional interactions with genes of the other groups. Finally, three small additional groups emerged from this analysis (V, VI and VII in Figure 6a), suggesting small networks of genes that might work together to regulate two of the three scored phenotypes. In sum, this meta network analysis supports the hypothesis of three potentially largely non-interacting genetic networks that regulate Hippo-dependent ovariole number, Hippo-dependent egg laying, and Hippo-independent egg laying, respectively. The presence of smaller sub-networks (V, VI and VII in Figure 6a) that interact with each other further supports the observation that the putative modules predicted by the SCA – which we refer to as sub-networks – could include genes that function within more than one such sub-network (Figure 5—figure supplement 4). Moreover, each of these genetically separable sub-networks included genes in multiple signalling pathways (Figure 6c).

The four predicted phenotypic sub-networks produced by the SCA approach included connector genes that were not included in our original screen, and thus had not been tested for possible effects on our phenotypes of interest (triangles in Figure 5b; Figure 5—figure supplements 1–3). Given that prior work in human disease models showed that predicted disease modules can correctly identify gene involvement in the relevant diseases (Chen et al., 2006; Gonzalez et al., 2007; Wang et al., 2017; Wang and Loscalzo, 2018), we asked whether our deployment of the SCA had likewise successfully predicted novel, functionally important genes in each sub-network. To do this, we used UAS:RNAi lines for each connector, driven by tj:Gal4 to measure the effects of knocking down each of the connector genes (triangles in Figure 5b and Figure 5—figure supplement 1; Figure 5—figure supplement 2; Figure 5—figure supplement 3) both on phenotypes within the sub-network where they were predicted (Figure 7a–b, Figure 7—source data 1), and on either of the other two tested phenotypes (Figure 7c, Figure 7—source data 2).

Of the ten predicted novel connectors within the Core sub-network, loss of function of several of these had significant effects on at least one of the three scored phenotypes. Five affected ovariole number, two affected Hpo-dependent egg laying and one affected Hpo-independent egg laying. However, only one of them significantly altered all three scored phenotypes (Figure 7a; Figure 7—source data 1).

The predicted connector genes from two of the other three phenotypic sub-networks showed high positive prediction rates for novel genes within the sub-networks. RNAi against seven out of 18 of the hpo[RNAi] Egg Laying connectors, three out of 11 of the hpo[RNAi] Ovariole Number connectors, and none of the 11 Egg Laying connectors, significantly affected the sub-network phenotype (Figure 7—source data 1). Thus, although the Egg Laying connectors failed to impact this phenotype in our assay, 41.1% and 27.2% of the connectors from the other two sub-networks were correctly predicted (Figure 7b; Figure 7—source data 1).

In sum, taken across all sub-networks, this methodology correctly identified genes regulating at least one of the scored reproductive phenotypes, at significantly higher rates than those obtained in the original screen of 463 members of all known signalling pathways (Figure 7c; Figure 7—source data 2). By this measure, testing network-predicted novel genes derived from experimentally obtained data was even more successful than testing signalling pathways as a means of identifying novel genes that regulate ovariole number and egg laying.

The D. melanogaster ovary is a commonly studied model for organogenesis (Chen et al., 2001; Godt and Laski, 1995; Lobell et al., 2017; Sarikaya and Extavour, 2015), stem cell maintenance (Gilboa, 2015) and interactions of development and ecology (Cohet and David, 1978; Hodin and Riddiford, 2000a; Klepsatel et al., 2013a; Sarikaya et al., 2019). Nevertheless, our understanding of the genetic mechanisms that regulate these processes remains fragmentary. In this paper, we have identified four distinct protein interaction sub-networks that regulate ovariole number and egg laying in the D. melanogaster ovary. These sub-networks consist of both novel and previously characterised genes that regulate either ovariole number or egg laying or both, thus enhancing our understanding of the genetic underpinnings of this reproductive system.

Of the four sub-networks, the Core sub-network affects both ovariole number and egg laying. The Core sub-network contains numerous housekeeping genes, including regulators of transcription, translation and cell division such as polo (Llamazares et al., 1991), cyclin K (Edwards et al., 1998), nucleosome assembly protein 1 (Ito et al., 1996) and eukaryotic translation release factor 1 (Chao et al., 2003). While polo and eukaryotic translation release factor one are members of signalling pathways, cyclin K and nucleosome assembly protein one are genes predicted by the SCA. Given that the Core sub-network largely consists of genes whose loss of function decreases both of these parameters, we hypothesise that these are essential genes for the basic structure and function of the ovaries. Essential genes are more interconnected in a PIN with higher centrality measures (Jeong et al., 2001; but see Yu et al., 2008) and interestingly, we find that the genes in the Core sub-network also have higher connectivity than those in the other three sub-networks (Figure 5—figure supplement 4).

In addition to genes that regulate basic cellular processes, the Core sub-network is enriched for the central components of the Hh signalling cascade, namely patched (ptc), smoothened (smo) and costa (cos) (Lee et al., 2016). However, we find that the loss of Hh ligand, which is expressed in the TF cells in the developing larval ovary (Lai et al., 2017), does not significantly affect either ovariole number or egg laying. Though surprising, ligand-independent activation of Hedgehog signalling has been observed before. For example, in the Drosophila eye, loss of either ptc or cos in clones leads to non-cell autonomous proliferation in wild type cells, as well as growth disadvantages in the mutant tissue (Christiansen et al., 2012). In another example, sufficient intracellular smo levels can also activate downstream transcription of Hh pathway targets, showing that Hh itself is not always required to activate the cascade (Jiang et al., 2018).

The hpo[RNAi] Ovariole Number sub-network is composed of genes that affect the Hippo signalling activity-dependent determination of ovariole number during development. Establishment of ovariole number occurs largely during the third instar stage of larval development in D. melanogaster (Godt and Laski, 1995; Hodin and Riddiford, 1998; King et al., 1968; Sahut-Barnola et al., 1996). During this period, the TFCs are specified in the anterior of the ovary and undergo rearrangement into stacks of cells called TFs, each of which gives rise to an ovariole (Godt and Laski, 1995; Sahut-Barnola et al., 1995). TF specification requires the expression of engrailed (En) (Bolívar et al., 2006) and the transcription factors Bab1 and Bab2, encoded by the bric-à-brac locus (Couderc et al., 2002; Godt et al., 1993). A third transcription factor, Lmx1a, was recently found to be necessary for the specification of the TFCs (Allbee et al., 2018). Our hpo[RNAi] Ovariole Number sub-network identifies numerous additional novel transcription factors including bunched (bun) and retinoblastoma-family protein (rbf), which we hypothesise could also be involved in the specification of ovariole number. bun and rbf have been implicated in the migration (Dobens et al., 2005) and endoreplication (Cayirlioglu et al., 2003) of the follicle cells during oogenesis, but have not, to our knowledge, been previously identified as playing a role in the context of larval ovary development.

The TFCs specified in the larval ovary undergo a process of convergent extension to form TFs. This process of convergent extension requires cell intercalation, and the actin depolymerising factor Cofilin, encoded by the gene twinstar, is essential to this process (Chen et al., 2001). During intercalation, the cells also dynamically modify their actin cytoskeleton and their expression of E-cadherin (Godt and Laski, 1995). Our hpo[RNAi] Ovariole Number sub-network further identifies Rho1 (Barrett et al., 1997) and Rho kinase (Rok) (Mizuno et al., 1999) as necessary for correct ovariole number. During the extension of the D. melanogaster embryonic germ band, a commonly studied model of convergent extension, the localised activation of the actin-myosin network facilitated by Rho1 and Rok is necessary for cell intercalation (Kasza et al., 2014). Given the known roles of Rho1 and Rok as regulators of the actin cytoskeleton (Ridley, 2006), we propose that TF assembly in the ovary requires both these proteins for correct cell intercalation. A third actin cytoskeleton regulator, misshapen (msn), was also identified by our hpo[RNAi] Ovariole Number sub-network. msn encodes a MAP kinase previously shown to regulate the polarisation of the actin cytoskeleton during oogenesis (Lewellyn et al., 2013), but has not, to our knowledge, been studied to date in the context of larval ovarian development.

We propose that the polarity of the somatic cells in the ovary is also necessary for correct larval ovary development, given the presence of the lateral membrane proteins discs large 1 (dlg1) and prickle (pk) in the ovariole sub-network. During the maturation of the TFs during larval development, the TFCs undergo significant cell shape changes, coincident with localised expression of beta-catenin and actin to the lateral edges of the TFCs (Godt and Laski, 1995). Restriction of the E-cadherin domain in epithelia requires establishment of the basolateral domain (Harris and Peifer, 2004) and we propose that testing a similar requirement for dlg1 and pk in the larval ovary would be a fruitful avenue for future studies.

Using a systems biology approach to analyse RNAi screening data has proven fruitful, providing us with new insights into the development and function of the D. melanogaster ovary by identifying novel and previously understudied genes that regulate this process. Systematic analysis of the function of single genes in development has been a historical convention and has provided valuable and precise genetic interaction information (Jansen et al., 2002; von Mering et al., 2002). With the advent of genome-wide analysis, however, we can use data from a larger number of genes to predict the identity of additional functionally significant genes with relative ease (Yu et al., 2006). We note that the novel gene prediction rate within each individual sub-network ranged from as high as 41.1% from the hpo[RNAi] Ovariole Number sub-network to as low as 0% from the Egg Laying sub-network (Figure 7b; Figure 7—source data 1). We suggest that this may be due to multiple factors. Firstly, the possible incompleteness of the PIN is expected to lead to some areas of the network being sparse or non-existent (von Mering et al., 2002). If the sub-network of interest happens to fall in such regions of the PIN, prediction algorithms will fail. Secondly, the initial restriction of tested genes to signalling pathway members might have provided a seed list too sparse to usefully predict functional connectors. Finally, it could be the case that ‘Egg Laying’ is such a complex phenotype that its gene regulation cannot be adequately captured within a highly connected network of the type suited for identification by the analyses we have used here.

Ovariole number in D. melanogaster is the outcome of a discrete developmental process with a clear beginning and end, comprising a specific series of cellular behaviours that take place in the confines of one organ (Godt and Laski, 1995; Hodin and Riddiford, 2000a; Sahut-Barnola et al., 1996). Once established during larval life, ovariole number in Drosophila remains unaltered through to and during adulthood, even if oogenesis within those ovarioles suffers congenital or age-related defects (King, 1970). Because previous work suggested that ovariole number in Drosophila could have at least some predictive relation to egg laying (Cohet and David, 1978; Klepsatel et al., 2013b; Sarikaya and Extavour, 2015), we reasoned that scoring the latter phenotype in a primary screen (Figure 1a) could be an effective way to uncover ovariole number regulators (Figure 1c). While our results showed that this was true in many cases, it was also clear that these two traits can vary independently (Figure 2), highlighting the fact that ovariole number is not the only determinant of egg laying. Egg-laying dynamics, even during the limited five-day assay used in our study, are likely influenced not just by a single anatomical parameter such as ovariole number, but rather by many biological, biomechanical, hormonal and behavioural processes. Consequently, the sub-network we were able to extract from the results of this screen (Figure 1a and b) might be too coarse to extract novel genes that participate in the potentially complex gene interactions regulating egg laying. Furthermore, genes predicted within each of the sub-networks are unlikely to function exclusively within just one sub-network. This conclusion is supported by our observation that genes predicted to function in any of the sub-networks, also function in at least one of the four sub-networks at a higher rate than genes selected for screening by their presence in a signalling pathway (Figure 7c). We also observe that although substantial regions of the meta network do not share interactions with genes in the other sub-networks (Figure 6b), we do find smaller sub-networks where there is some overlap, further indicating pleiotropy of some genes in both egg laying and ovariole number regulation.

The predictive rates of the approach we have used here, although encouraging, are likely limited by the degree of noise in the high-throughput data used to generate the PIN (Li et al., 2010), the sparseness of the PIN, and the degree of misidentification of protein interactions (Zhang et al., 2017). Addressing one or more of these parameters could improve the outcomes of future network predictions from developmental genetics data. For example, the problem of sparseness, which is a paucity of high confidence detectable interactions relative to all biologically relevant interactions, has been addressed in other studies by using an ‘Interolog PIN’ (Matthews et al., 2001) in place of an organism-specific PIN. An Interolog PIN combines known interactions from multiple organisms and has been used successfully to identify, for example, gene modules relevant in squamous carcinoma, based on a starting dataset of microarray data on differentially expressed genes between cancer cells and the surrounding tissue (Wachi et al., 2005). Future studies applying such an Interolog PIN to the outcomes of genetic screens for developmental processes of interest could potentially overcome the problem of sparseness, as well as the biases towards proteins that are more heavily studied and thus better represented in organism-specific PINs.

This study did not generate new unique reagents. This study generated new python3 code available on GitHub: https://github.com/extavourlab/hpo_ova_eggL_screen; Kumar, 2020 (copy archived at https://github.com/elifesciences-publications/hpo_ova_eggL_screen).

Further information and requests for resources and reagents should be directed to and will be fulfilled by the Lead Contact, Cassandra G. Extavour (extavour@oeb.harvard.edu).

Wild type and mutant lines of Drosophila melanogaster were obtained from publicly accessible stock centers and maintained as described in ‘Fly Stocks’ below. Genotypes and provenance are provided in the Key Resource Table. Candidate genes were randomly assigned to batches for screening (see the Supplementary file 1 for which genes were in each batch). F1 animals from the same cross were randomly assigned to experimental groups for phenotyping in all screens.

This study generated a series of python3 Ipython notebook files that perform the entire analysis presented in this study. All the results presented in this paper, including the figures with the exception of the network visualisations, which were created using Cytoscape3 (https://cytoscape.org/) can be reproduced by running the aforementioned python3 code. The raw data, calculations made with these data, and code used for calculations and analyses (Ipython notebooks) are available as supplementary information. For ease of access, legibility and reproducibility, the code and datasets have been deposited in a GitHub repository available at https://github.com/extavourlab/hpo_ova_eggL_screen.