Naturally segregating genetic variants contribute to thermal tolerance in a Drosophila melanogaster model system

Abstract

Thermal tolerance is a fundamental physiological complex trait for survival in many species. For example, everyday tasks such as foraging, finding a mate, and avoiding predation are highly dependent on how well an organism can tolerate extreme temperatures. Understanding the general architecture of the natural variants within the genes that control this trait is of high importance if we want to better comprehend thermal physiology. Here, we take a multipronged approach to further dissect the genetic architecture that controls thermal tolerance in natural populations using the Drosophila Synthetic Population Resource as a model system. First, we used quantitative genetics and Quantitative Trait Loci mapping to identify major effect regions within the genome that influences thermal tolerance, then integrated RNA-sequencing to identify differences in gene expression, and lastly, we used the RNAi system to (1) alter tissue-specific gene expression and (2) functionally validate our findings. This powerful integration of approaches not only allows for the identification of the genetic basis of thermal tolerance but also the physiology of thermal tolerance in a natural population, which ultimately elucidates thermal tolerance through a fitness-associated lens.

Introduction

Thermal tolerance is a complex physiological and behavioral trait that determines the survivability of many individuals and, ultimately populations. Temperature regulates and plays a critical role in a wide range of processes from cellular functions—action potentials and enzymatic reactions (Schulte et al. 2011; Yu et al. 2012; Dowd et al. 2015)—to more systematic evolutionary processes like migration and population growth (Frazier et al. 2006; Sunday et al. 2012) and ultimately fitness (Angilletta 2009). Populations experiencing different temperature environments have evolved different physiological adaptations, which often lead to different thermal limits among populations (Jørgensen et al. 2019). However, like many other complex traits, identifying the genetic variants driving thermal adaptation has been challenging. Therefore, a model system allowing for a comprehensive understanding of the genetic architecture underlying thermal tolerance would significantly advance our understanding of the proximate mechanisms determining thermal limits.

Many empirical studies—both in model and non-model systems—demonstrate the importance of understanding the mechanisms contributing to thermal tolerance, including in coral (Dixon et al. 2015), fish (Schulte et al. 2011; Anttila et al. 2013; Grinder et al. 2020), fruit flies (Overgaard et al. 2008; Bozinovic et al. 2016), birds (Nord and Nilsson 2011; DuRant et al. 2012, 2013), frogs (Pintanel et al. 2019; Díaz-Ricaurte et al. 2021), and lizards (Angilletta, Hill, et al. 2002). The ability to tolerate extreme temperatures is essential for ectotherms (Huey and Berrigan 2001; Angilletta, Niewiarowski, et al. 2002; Frazier et al. 2006; Hoffmann 2010; Sunday et al. 2011; Gunderson and Stillman 2015) because their body temperature tracks the ambient temperature to a greater degree than endotherms. Drosophila species have been used extensively as a research model because of their worldwide distribution and vastly different thermal environments and preferences (MacLean, Sørensen, et al. 2019). Drosophila melanogaster, in particular, has been used for decades to study these traits in laboratory and field settings (James et al. 1997; Hoffmann et al. 2003; Kristensen et al. 2008; Bergland et al. 2014; Ørsted et al. 2019; Machado et al. 2021; Rudman et al. 2022). For example, within a natural environment, Machado et al. (2021) demonstrated the effects of seasonal temperature shifts in allele frequencies of fruit flies, suggesting that short-term temperature changes can lead to rapid changes in allele frequencies. While these studies have been imperative in elucidating how ectotherms adapt to temperature changes and how the genome responds to such changes, there is still a lack of complete knowledge about the specific genetic mechanisms underpinning thermal tolerance.

Identifying the causative mechanisms at play in highly complex, multifaceted traits like thermal tolerance remains very challenging. Genetic mapping approaches, designed to identify loci associated with the naturally occurring variation in a trait continue to improve and have led to the discovery of many associations for thermal tolerance in fruit flies (Morgan and Mackay 2006; Vermeulen et al. 2008a, 2008b; Rand et al. 2010; Rolandi et al. 2018) and other systems including coral, fish, pigs, and cattle (Perry et al. 2001; Everett and Seeb 2014; Dixon et al. 2015; Kim et al. 2018; Matz et al. 2018; Cheruiyot et al. 2021). However, mapping approaches lack the resolution to identify single genes on their own (Mackay 2001). More recently researchers have been pairing mapping approaches with transcriptomics to identify candidate genes that influence complex traits (Cubillos et al. 2017; Highfill et al. 2017; Wen et al. 2019; Williams-Simon et al. 2019; Lecheta et al. 2020; Ng’oma et al. 2020), which was first introduced by Wayne and McIntyre (2002). While it is critically important to validate candidate genes that have been identified to be associated with complex traits, it has been a daunting task, and therefore few genetic studies of complex traits have definitively identified causal variants (Ding et al. 2016; Highfill et al. 2017; Gaspar et al. 2020). Approaches integrating several methods including multiparent mapping for high power and resolution, transcriptomics, and functional genomics provide a promising approach that moves beyond simply mapping loci toward identifying the key genes involved in complex traits. Here, we leverage the combined strengths of Quantitative Trait Loci (QTL) mapping, transcriptomics, and functional genetics to elucidate the genetic architecture of thermal tolerance using a large set of recombinant inbred lines (RILs) from the Drosophila Synthetic Population Resource (DSPR) multiparental population (King, Macdonald, et al. 2012; King, Merkes, et al. 2012). We measured over 700 RILs for thermal tolerance and identified multiple QTLs, including 1 large effect QTL. We then performed RNA-seq to identify several candidate genes that are located within our QTLs and performed validation studies using classical genetics. Through this multifaceted approach, we identified novel candidate genes associated with the critically important trait, thermal tolerance.

Methods

DSPR and fly husbandry

We used the DSPR, which is a multiparental mapping population for the D. melanogaster system (King, Macdonald, et al. 2012; King, Merkes, et al. 2012; Long et al. 2014); (http://FlyRILs.org). The DSPR consists of approximately 1,600 8-way RILs, with 2 populations (pA and pB), which consist of approximately 800 RILs each. Each population was created by crossing a set of 8 founder lines, with 7 unique founders to each population and 1 that is shared—AB8. Each RIL is a mosaic of the 15 founders (parental) lines, which have been fully re-sequenced, and the RILs have been genotyped at over 10,000 SNPs. This structure makes genomic analysis of this population more feasible, because of the ability to be able to track entire haplotypes (King, Macdonald, et al. 2012). To generate the RILs, the founders were mixed en masse for 50 generations to create a synthetic population, with 2 major goals—first to generate high genetic diversity, then for high mapping resolution. The population was then inbred for another 25 generations to create stable recombinant inbred lines (King, Merkes, et al. 2012). Within each of the pA and pB populations, 2 separate synthetic populations were maintained (e.g. pA1 and pA2) for the crossing phase before the creation of the RILs. We phenotyped 741 RILs in population A, scoring thermal tolerance for each line (Fig. 1a).

Fig. 1.

Schematic of the experimental design. a) Phenotyping and QTL mapping using the DSPR RILs, b) RNA-seq and differential expression analysis, and c) using RNAi lines for functional validation of candidate genes within the QTL interval.

We raised stock RILs at 18°C and 60% relative humidity on a standard fruit fly diet (Supplementary Table 1) with live yeast added to the top of the food after it cooled. Maintenance of our stock lines occurred at 18°C because this made it more feasible to maintain the high number of RILs we needed for this experiment. Two weeks before phenotyping for thermal tolerance, approximately 10 female and 6 male flies were placed into a flask (Fisher Scientific, cat. no.: AS-355) of food and maintained at 25°C for 1 week to allow for mating and egg laying. After the first week (7 days, post-oviposition), we removed all the adults (parental generation) from the flasks first to ensure that there was no mixing of generations and second to control for density while making sure we had sufficient adults for phenotyping. After 2 weeks (14 days post-oviposition), we anesthetized adult flies on ice for no more than 10 min and selected 60–80 female flies for the thermal tolerance assay. We then divided flies equally between 2 separate vials of food and allowed them to recover for at least 24 h before the phenotyping assay. The age range for all individuals we assayed was 15–22 days post-oviposition (5–13 days post-eclosion). This age range allowed us to obtain and assay a large number of flies per line. Throughout this experiment, we only measured mated female flies to keep the experiment at a reasonable size. Successfully mapping a complex trait such as thermal tolerance requires assaying a massive number of individuals and measuring both sexes would effectively double the size of the experiment. Thus, while we anticipate sex effects, we were unable to include both sexes at the scale that would be needed.

Measuring thermal tolerance in the DSPR

We measured thermal tolerance using a thermally-controlled and highly sensitive apparatus known as the “heat box” (Wustmann et al. 1996). Within the heat box, there are 16 individual chambers, which allows for concurrent testing of flies. In each chamber, temperatures are set at 24°C for the first 30 s for acclimation, then immediately shift to 41°C for 9.5 min for a total of 10 min (Wustmann et al. 1996; Zars 2001). Every chamber has a temperature sensor that reports the temperature of that chamber in real time. The temperature change within each chamber is rapid and takes 4 s to reflect 41°C (Wustmann et al. 1996). Additionally, we independently confirmed that the chambers reflected the temperature at the given time by inserting a thermocouple in each chamber periodically. While 10 min is a short period of time, the vast majority of RILs are incapacitated within this timeframe with fewer than 1% of RILs remaining active for the full 10 min. We assayed the RILs in a randomized order, limiting the number of flies assayed on a single day from a single RIL up to 27 individuals total (2 groups in the heat box), to ensure each RIL was assayed over multiple days, however, we ran several RILs on any given day. For each RIL, we assayed a minimum of 30 individuals. We also measured the founder lines for thermal tolerance, following the same experimental design as we did for the RILs, except for AB8, which did not thrive in our hands, and thus, we do not have an estimate for that line. We measured a total of 424 individuals from the founding populations, with an average of 61 individuals per founder (range: 48–67).

We define thermal tolerance as the amount of time it takes before a fly becomes incapacitated for at least 60 s. We used a custom R script to automate scoring incapacitation based on the criteria defined above (i.e. no activity for 60 s). The heat box contains an infrared system for monitoring each fly's position within the chamber. We used a sliding window to assess the variance in position in 60 s intervals and scored a fly as incapacitated when this variance was below a threshold value indicating no substantial movement within that interval. This approach has the advantage of applying purely objective, defined criteria to scoring incapacitation. However, it is important to confirm the program produced the correct results. We validated our script by comparing over 5,000 visually hand-scored incapacitation scores to our automated-scored values (Supplementary Fig. 1). These scores are highly correlated (r = 0.88), and manual inspection of discrepancies between the hand-scored and automated-scored values revealed the automated values to be more accurate and consistent for our defined criteria for incapacitation.

QTL mapping and heritability

We performed a genome scan on 741 RILs using a custom R script as described in Williams-Simon et al. (2019). We used the average incapacitation times from each RIL as phenotype and applied the Haley–Knott regression (Haley and Knott 1992). This statistical approach regresses each phenotypic average on the 8 founder haplotype probabilities (Broman and Sen 2009; King, Macdonald, et al. 2012) by fitting the following model:

$$y_i={\displaystyle \sum _{j=1}^7}{p}_{ij}{b}_{ij}+e_i,$$

where $y_i$ is the phenotype (i.e. thermal tolerance) of the ith RIL, $p_{ij}$ is the probability the ith RIL has the jth haplotype at the locus, $b_{ij}$ is the vector of effects for the jth haplotype, and $e_i$ is the vector of residuals (Broman and Sen 2009). Prior to performing this genome scan, we transformed the raw phenotypic data, using a square root transformation, which improved the normality distribution of the data to meet the assumptions of our statistical test. The DSPR RILs contain 2 subpopulations within each set of RILs, stemming from the synthetic populations being maintained en masse in 2 sets before the creation of the RILs (King, Macdonald, et al. 2012). Including subpopulations (pA1 and pA2) in the model did not affect the mapping results (Supplementary Fig. 2a). We also performed a second genome scan after correcting for our most significant QTL (Q5) and did not obtain any additional peaks (Supplementary Fig. 2b). We also performed local association mapping for our most significant QTL (Q5) by fitting a similar model with SNP probabilities rather than haplotype probabilities. We calculated these probabilities from the haplotype probabilities for each RIL at each position and the SNP genotypes in the founder lines.

To identify statistically significant QTLs, we used permutations to calculate both the false discovery rate (FDR) and the family-wise error rate (FWER). First, we permuted thermal tolerance along with place learning and place memory phenotypes (Williams-Simon et al. 2019), because we assayed all 3 of these phenotypes on the same individual fly during heat box the experiment. For context, we define place learning and memory as the ability of an individual fly to associate a side of the chamber between 2 different temperatures (aversive—41°C or preferred 24°C), and the remembering to avoid the side associated with the aversive temperature (Ostrowski et al. 2018), respectively. Second, we computed the number of false positive QTLs at a range of different significance thresholds. We removed any QTL that were within 2 cM of a more significant QTL to identify distinct QTLs. Then, we calculated the empirical false discovery rate (FDR, expected false positives/total positives) for each threshold and determined the threshold corresponding to a 5% FDR. We also calculated the threshold corresponding to a 5% family-wise error rate (FWER, the probability of 1 or more false positives experiment-wide) by determining the lowest P-value for each set of genome scans from each permutation and calculated the 95% quantile of the resulting set of P-values (Churchill and Doerge 1994). For each identified peak, we calculated the estimated effects of each haplotype at each QTL, the percent variance explained (PVE) by the QTL and the Bayesian credible interval (BCI), following the methods described previously for the DSPR (King, Macdonald, et al. 2012). If the BCI of 2 significant QTLs did not overlap, we considered these distinct QTLs.

We calculated broad-sense heritability (H² = V_G/V_p) of thermal tolerance by estimating the genetic and phenotypic variance and covariance from a linear mixed model lme and VarCorr functions in the nlme package (Pinheiro and Bates 2000; Pinheiro et al. 2023) followed by a jackknife, which allowed us to obtain standard errors of our estimate (Roff and Preziosi 1994; Roff 2008). The jackknife removes each observation once, the model is then fit, and the quantitative genetic parameter (i.e. heritability) is estimated and a pseudo value is calculated. Prior to fitting the model, we quantile-transformed the raw data to ensure normality. Thus, for our dataset of 741 RILs, each RIL is deleted, 1 at a time, to produce 741 pseudo values. Because the lme function uses restricted maximum likelihood, it is possible for the model to not converge. This occurred for 72 cases, thus, these pseudo values were not included in the calculation of our estimates. We also performed a model comparison using a likelihood ratio test to determine the significance of RIL (i.e. genotype). We compared the model fitting only the intercept to the model including RIL, fitting both via maximum likelihood such that model comparisons would be valid.

Structural variants analysis

Structural variants (SVs), for example, insertions, deletions, and transposable elements, contribute to the phenotypic variation in complex traits (Frazer et al. 2009; Eichler et al. 2010; Chakraborty et al. 2018). Chakraborty et al. (2019) used the DSPR to identify genome-wide SVs in both populations A and B, which we utilized to determine if there were any SVs located within the BCI of our most significant QTL (Q5). To ensure that we captured all the SVs located in the QTL region, we included any genes that had any portion within the BCI, even if the entire gene did not fall within the QTL BCI.

RNA-sequencing

We performed RNA-seq on mated female heads to determine differentially expressed genes between the top and bottom cohorts of thermal tolerance (Fig. 1b). It is important to note that in our design, we only used heads which allowed us to only focus on gene expression within a specific tissue type. There is physiological evidence that incapacitation in high temperatures results from failure in the central nervous system (Jørgensen et al. 2020), thus, we chose the head as the most logical tissue to focus on.

First, we selected 14 RILs from the high (n = 7) and low (n = 7) 5% cohorts after phenotyping the first 140 RILs and followed the same set-up protocol (see phenotyping section above). In total, we had 6 pooled samples consisting of 5 individual female heads from each of the 7 lines in the cohort. Thus, for each of the 6 samples, we pooled a total of 35 female heads in a single sample, with 5 heads from each RIL contributing to each sample. Adult female flies 14 days post-oviposition were flash-frozen in liquid nitrogen and stored at −80°C, until further processing. We used a narrower age range here, choosing flies at the same age that females were set up for phenotyping. In addition, we also took a different set of individuals from the same RILs and cycled them through an “incubation assay” for 1 min intervals at 41°C and 24°C (room temp) for a total of 6 min, and left control samples on the benchtop at a constant 24°C also for 6 min. Then, all the vials–from both experimental and control groups–were left on the benchtop at room temp for 3 h. To ensure: (1) flies were just incapacitated and not dead from heat stress, and (2) the gene expression changes had time to occur but also that there was minimal time to allow for gene expression turnover. Immediately after the 3 h, flies were flash-frozen in liquid nitrogen and stored at −80°C. For the incubator assay, all RILs were tested between 8 Am and 10 Am , to help reduce possible confounding effects that would affect gene expression. The samples were then freeze-dried overnight to prevent the degradation of RNA, using a lyophilizer (Labconco, cat. no.: 77550-00). Second, the samples were shipped to Rapid Genomics (Gainesville, FL) for transcriptome processing. The total RNA was extracted with Dynabeads mRNA direct kit from Life Technologies, mRNA was then fragmented and converted into double-stranded cDNA, followed by the standard proprietary library prep for 1 lane of Illumina HiSeq 3000 instrument to generate paired-end (PE) reads. The first 16 samples were PE 150 bp, while the remaining samples were PE 100 bp. Third, We analyzed the RNA-seq data using a bioinformatic pipeline, as described previously in Williams-Simon et al. (2019). Briefly, we trimmed the samples using the software cutadapt (Martin 2011), then ran a quality control test using fastqc (Leggett et al. 2010), which did not show any significant issues with the reads. We then used parts of the Tuxedo pipeline (Pertea et al. 2016) to align the reads against the D. melanogaster reference transcriptome (Ensembl Release Version: 84) using HISAT2 (Kim et al. 2015) and StringTie (Pertea et al. 2015, 2016) assembled and quantified transcripts. To calculate the read count for each gene we used a prepDE Python script provided for this purpose (http://ccb.jhu.edu/software/stringtie/index.shtml?t=manual#deseq). The resulting dataset provided the input needed to use the DESeq2 package (Love et al. 2014) to identify differentially expressed genes. We filtered out any zero count genes and performed surrogate variables analysis (SVA) on the normalized count data (Leek et al. 2012, 2019) to estimate any unknown batch effects. We identified 2 significant surrogate variables, which we included as terms in the models we fit along with our known batch effect (i.e. sampling day) in all following bioinformatics analyses. We then used the DESeq2 package (Love et al. 2014) to test for overall treatment effects and performed contrasts between pool (high and low cohorts), condition (temperature: 41°C or room temperature), and the interaction between pool and condition, identifying significantly differentially expressed genes for each. To estimate the differences in expression in each pool and condition, we used the function lfcShrink, which performs shrinkage on log₂(fold changes). All log₂(fold changes) reported here are the shrinkage estimated values using the “normal” estimator. We then used the contrasts to identify candidate genes that were differentially expressed between the different groups localized within the BCIs of the mapped QTLs (Q1–Q7).

RNAi lines and crossing design

Because the Q5 QTL indicated a very strong association with thermal tolerance and the BCI was fairly narrow, we further investigated the genes within this QTL using RNAi lines. We sourced the UAS-RNAi transgenic fly lines from the Transgenic RNAi Project (TRiP) from the Bloomington Drosophila Stock Center (BDSC; Perkins et al. 2015). BDSC reported that the TRiP RNAi lines might have a low dose of Wolbachia pipientis, therefore we treated the flies with tetracycline (Sigma-Aldrich, cat. no.: T4062-5G, batch: SLCF9393). We followed a tetracycline treatment protocol from Hoffmann et al. (1986). We incorporated 0.3 mg/mL final concentration of tetracycline into an agar base fly food recipe and allowed the flies to remain on that food for 2 generations flipping them onto new agar food after 2 weeks. One of the lines did not survive during the tetracycline treatment (BDSC stock no.: 63713). After the treatment, we switched the population back to our “normal” fly food recipe (Supplementary Table 1) and allowed the lines to re-acclimate to the food for 5 generations before beginning experiments. The TRiP collection did not have RNAi lines for all 21 genes in the Q5 interval. However, we obtained RNAi lines for 14 genes from the BDSC, some of which had more than 1 available RNAi line (Supplementary Table 2). We checked each UAS line for reports of off-target effects and did not identify any reported off-target effects (https://fgr.hms.harvard.edu/up-torr). In total, we screened 20 UAS-RNAi lines using 2 different promoters: (1) ubiquitously expressed using an actin (Act5C-Gal4) driver (BDSC–3954: y1 w*; P{Act5C-GAL4}17bFO1/TM6B, Tb1), and (2) pan neuronally expressed using embryonic lethal abnormal vision (elav) driver (BDSC–25750: P{w[+mW.hs]=GawB}elav[C155]w[1118]; P{w[+mC]=UAS-Dcr-2.D}2). All the UAS-RNAi lines were in either P2 (Chr 3) or P40 (Chr 2) background, which we used as controls for comparing thermal tolerance.

For each line, we placed adults from the parental stock in fresh vials and allowed them to mate. On day 3, we cleared the vials of all adults, then on days 9–12, we collected 3 males (Gal4) and 5 females (UAS-RNAi) to set up the cross to obtain Gal4-UAS-RNAi F₁ progeny. The Act5C-Gal4 line has a tubby (TM6B, tb1) phenotypic marker, which we selected against at the pupae stage of development in the F1 because tubby is a pupae marker. Therefore, to make a fair comparison, we also transferred all lines on days 19–22 to control for transferring the tubby lines. On days 20–23, we anesthetized the flies using CO₂, collected 10 F₁ females (0–1 day post-oviposition), placed them in a new vial with food, and allowed at least 12 h of recovery time before measuring thermal tolerance.

Measuring thermal tolerance in RNAi lines

We measured the time it takes for individuals to become incapacitated in our RNAi crosses (Fig. 1c). Flies were placed individually into small clear cylinder tubes (5 mm diameter × 65 mm length; monitor tubes; TriKinetics, Inc., USA), which were plugged on both ends to prevent the flies from escaping. We placed the tubes onto a custom-milled aluminum plate (6.11 mm thick) with 34 individual slots (2.1 mm thick spaced 1.6 mm apart) matching the diameter of the fly tubes, allowing even heating throughout the tube. The aluminum plate is mounted to a Peltier-element based heating plate (PELT-PLATE; Sable Systems International, Inc., North Las Vegas, NV), which is connected to a controller, maintaining the plate at a constant temperature of 41°C. A temperature probe was also placed into a monitor tube for each run to monitor the accuracy of the temperature within the fly tubes relative to the PELT-PLATE.

We measured each cross in 3 batches of ∼10 individuals per run for a total of 10 min at 41°C. We deliberately designed systems for identifying when human errors occurred during testing, such as the plate not being aligned with the heat source, or the camera being out of alignment and dropped any samples where we identified an issue to ensure consistent measurements. Following this process, we had measurements for between 21 and 60 individuals for each cross. Before each trial, we confirmed that the aluminum plate was at ambient temperature before placing the monitor tubes into the wells. We randomized the assay order for the lines for the first replicate and kept that order for the second and third batches.

During each 10-min thermal tolerance trial, the entire set of flies was recorded using a Raspberry Pi High Quality Camera (12.3 MP) at 1 frame per second using a Raspberry Pi 4 Model B and custom Python code. Simultaneously, this script also recorded the timestamp, temperature reported by the PELT-PLATE, and temperature in the monitor tube measured via a thermocouple. Because the Raspberry Pi camera produces a highly distorted “fisheye” image, we first de-warped the individual images via checkerboard calibration using 15 images covering the entire plate surface (Supplementary Fig. 3). The images were subsequently combined into a single movie for later analysis, both using routines in the OpenCV library (Bradski 2000).

Movies of the thermal tolerance trials were analyzed using DeepLabCut (version 2.2.3; Mathis et al. 2018; Nath et al. 2019) and custom Python scripts to enable batch processing of trials. Because each trial recorded the activity of many flies simultaneously, we needed to first split each composite video into a movie of a single fly's movements. The rigid nature of the aluminum thermal plate allowed for the automated extraction of video for individual flies. First, we located the 4 corners of the plate using a trained DeepLabCut model (Supplementary Fig. 4a). With coordinates of the 4 corners, we applied a second dewarping step, rotated the image so that the edges were normal to the frame, and cropped the image to the 4 corners using routines in the OpenCV library (Bradski 2000). Second, we leveraged the fixed coordinates of the fly tubes within the plate to subset each video into 34 separate videos, each with a single fly, no fly, or the thermocouple (Supplementary Fig. 4b). These videos were then processed using a second DeepLabCut model, which had been trained on 978 manually labeled frames from individual fly movies (1,000,000 training iterations; train error = 1.64 pixels; test error = 1.69 pixels; Supplementary Fig. 4c). This automated system was able to track 1,350 flies across ∼800,000 individual frames, yielding Cartesian (x, y) coordinates for each fly at each 1 s timestep. With these time and position data, we used an R script to automate identifying when individual flies were incapacitated using the same criteria described for the heat box.

To determine the effect of each RNAi cross on the incapacitation phenotype, we compared the UAS-Gal4-RNAi crosses and the background control cross by fitting separate ANOVAs for each RNAi line with the cross id as the predictor and incapacitation time as the response variable. We followed this model with planned post hoc comparisons using the glht function in the multcomp package (Hothorn et al. 2008), comparing each RNAi cross with the relevant matching background cross.

Validation of thermal tolerance phenotyping

Our mapping experiment used the heat box (see Measuring thermal tolerance in the DSPR section above) to track activity in individual flies in chambers lined individually with Peltier elements. We developed an alternative method for measuring incapacitation following this experiment for 2 reasons: (1) the heat box showed more variability than desired in the target temperature reached, and we identified instances where individual heat box chambers did not reach the desired temperature (these data were discarded from our dataset, see above), and (2) the heat box can measure, at most, 16 individuals at 1 time, limiting the speed we can perform assays. The heat plate method we developed, described above, resolved both issues, increasing the throughput of the assay and allowing for consistent assay temperatures with continuous monitoring. To confirm that our phenotyping methods are comparable, we measured a subpopulation of 39 RILs from those that were assayed in the heat box. We assayed these RILs in 2 biological replicates, assaying ∼60 individuals each time (range: 27 individuals to 72 individuals) for a set of RILs distributed across the range of thermal tolerance scores using the heat plate apparatus and methodology (Supplementary Fig. 5a). We observed shorter incapacitation times overall on the heat plate, which is expected given that the heat plate showed better consistency in reaching and maintaining the target temperature of 41°C than the heat box, but confirmed that the scores are correlated between the 2 methods (r = 0.36). We acknowledge that the correlation between the assays is not particularly high, and that there are differences between the 2 assays. For assaying thermal tolerance, the heat plate is more consistent, which is why we have moved to this method. In the heat box measurements, the same individual was first assayed for place learning and memory, resulting in individuals previously exposed to 41°C at varying durations as they learn (for more details, see section above and Williams-Simon et al. 2019). In addition, in the heat box, each chamber has a separate heating element, leading to small variations in the temperature in each, whereas the heating element in the heat plate assay is 1 block, which means the onset of the 41°C is uniform. Thus, while both assays give a proxy measurement for thermal tolerance, we believe these differences in the assays are driving the differences we see some differences between the 2 assays. As with the heat box, we validated our measurements by hand-scoring a set of actual videos of flies moving for incapacitation and examined a set of traced movies to confirm both the tracking performance and the programmatic scoring of incapacitation. We compared the human scores to the computer-generated scores, which had an overall positive correlation (Supplementary Fig. 5b; r = 0.47).

All data analyses were performed in R 4.2 and lower (Posit team 2023; R Core Team 2023), Python 3.9.0 (Van Rossum and Drake 2009), and DeepLabCut 2.2.3 (Mathis et al. 2018).

Results

Phenotypic distributions

We first quantified thermal tolerance in 741 RILs (39,392 individuals) in the DSPR using the heat box (Fig. 1a). We found a wide range of phenotypic variation with approximately a 15-fold range from 24 to 361 s between RILs with low and high average incapacitation times (Fig. 2). We expected thermal tolerance to show high variability among RILs because the original founding parents of the RILs were collected from different continental locations (King, Merkes, et al. 2012) with varying adaptation to different temperatures and it is also a characteristic complex trait. We also phenotyped the DSPR founder lines (A1, A2, …, A7) directly for thermal tolerance (Fig. 3b).

Fig. 2.

Incapacitation scores for all individuals (>48,000) for the 741 DSPR RILs measured. Each point represents an individual fly's phenotype with points plotted with semi-transparency so overlap can be visualized as darker areas. The RILs are sorted on the x axis by their mean incapacitation score. The mean score ± 1 SE for each RIL is plotted from min to max. The ids of the 14 RILs used for RNA-seq are labeled on the x axis.

Fig. 3.

Genome-wide scans. a) Genome scan for thermal tolerance in the DSPR RILs showing the −log₁₀(P-value) vs position in the genome with different chromosome arms denoted by shaded boxes. Each QTL that reached at least the FDR threshold is labeled at the corresponding QTL peak. The FDR (solid line) and FWER (dotted line) thresholds are shown with horizontal lines. b) The raw incapacitation scores measured in the inbred founder lines. Each point represents an individual fly's phenotype with points plotted with semi-transparency so overlap can be visualized as darker areas. The mean score ± 1 SE for each founder line is plotted. c) The average incapacitation scores in the RILs with RILs separated by the founder haplotype they harbor at the Q6 location. Each point represents a RIL mean with points plotted with semi-transparency so overlap can be visualized as darker areas. The mean score ± 1 SE for the set of RILs harboring each founder haplotype at Q6.

We estimated broad-sense heritability (H²) for thermal tolerance in the RILs, which was H² = 0.24, 95% CI = 0.22–0.26. In addition, the effect of RIL ID on thermal tolerance was highly significant ($x_1^2=11,232.13$, P < 0.0001), indicating that there is a genetic basis for thermal tolerance in the DSPR RILs.

QTL

We identified 7 QTLs (Q1–Q7; Fig. 3a; Table 1) at the 5% FDR and higher. Of those 7 QTLs, 5 reached or passed the FWER threshold, which is the more stringent threshold. Any significant QTLs are potential regions of the genome that influence thermal tolerance, with Q5 having the strongest association. For each significant QTL, we determined the region most likely to contain causative variants by obtaining a Bayesian credible interval (BCI; Manichaikul et al. 2006; Broman and Sen 2009). For Q5, the most significant QTL, the BCI is 20.93–21.09 on chromosome 3R. Therefore, for each of our significant QTL, we can estimate the founder (A1, A2, A3, A4, A5, A6, A7, and AB8) haplotypes’ effect at the specific QTL location for thermal tolerance (Q5: Fig. 3c; all QTL: Supplementary Fig. 6). For example, at Q5, the set of RILs harboring the A5 haplotype at that location have the highest average incapacitation score, while those harboring the A4 haplotype have the lowest, though the means do not form 2 distinct groups, suggesting the potential for multiple causative variants in this interval.

Table 1.

Details for all QTLs.

ID	Arm	Position (Mbp)	Lower (Mbp)	Upper (Mbp)	−log10(P-value)	PVE (%)
Q1	2L	20.11	2L:22.07	2R:8.69	7.57	6.4
Q2	3L	9.39	8.82	9.42	3.74	3.8
Q3	3R	8.02	5.53	8.28	5.30	4.9
Q4	3R	15.40	15.26	17.37	4.93	4.6
Q5	3R	20.98	20.93	21.09	21.79	14.9
Q6	3R	26.33	26.00	26.63	5.84	5.2
Q7	3R	28.94	26.97	30.40	3.41	3.6

We also performed local association mapping within the Q5 peak interval to identify which DSPR variants showed the strongest association with thermal tolerance. Given that we have complete genome sequences for all founder haplotypes and have inferred the haplotype identity for each RIL at each position, we can compute the probability a given RIL harbors a given SNP at each position. We can then use these probabilities in a linear model rather than the founder haplotype probabilities. This approach is most useful to identify the set of “in phase” SNPs that might correspond to a clear high vs low group, which would be suggestive of a single causative SNP underlying the QTL. However, in this case, and most cases in the DSPR, there is evidence for multiallelism (more than 1 causative SNP) with haplotypes not forming a clear high and low grouping (Supplementary Fig. 7), thus interpretation is more challenging. In addition, because haplotype probabilities do not change very much across the BCI region, the association mapping results will be similar for SNPs that share the same haplotype groupings (e.g. the set of SNPs unique to 1 founder will all have the same predictive power). As a result, there are significant SNPs located near all genes in the interval and thus this analysis does not narrow down our set of potential candidate genes.

RNA-seq and genome-wide expression patterns

We performed RNA-seq on a pooled set of mated female heads to identify differentially expressed genes between low and high cohorts of thermal tolerance performance, assaying expression at both room temperature and following exposure to high temperature. Thus, we obtained genome-wide expression profiles (Fig. 4) for 4 separate groups: thermal tolerance high performers at room temperature (H_RT), thermal tolerance low performers at room temperature (L_RT), thermal tolerance high performers after incubation (H_Incu), and thermal tolerance low performers after incubation (L_Incu).

Fig. 4.

Differential gene expression. a,b) MA plots showing log₂(fold change) for the high vs low pools a) and the incubated vs room temperature pools b) vs the normalized mean expression value for each gene. Normalized mean expression is shown on a log₁₀ scale. Genes are colored by P-value with nonsignificantly differentially expressed genes in pink and increasingly significant genes in darker shades of blue. c,d) log₂(fold change) for the high vs low pools c) and the incubated vs room temperature pools d) vs −log₁₀(adjusted P-value). The gradient of the points shows the normalized mean expression value of each gene. The Hsp genes that fall within Q2 are labeled on each plot.

First, to determine the level of genome-wide expression patterns in the different cohorts, we performed principal components analysis (PCA) on all samples using the expression matrix (Supplementary Fig. 8). PC1 explains 29% of the variance, while PC2 explains 22% of the variance. There is a clear separation between the samples incubated at high temperatures and those measured at room temperature. Both high- and low-performance pools following high-temperature exposure (Incu) cluster together, whereas there is a substantial separation between high and low pools (H and L) at room temperature (RT).

Next, we determined differences in gene expression patterns in the different cohorts (Fig. 4; Supplementary Fig. 8b). There were 2,642 genes that were differentially expressed between high vs low (pool) cohorts for thermal tolerance genome-wide. In this comparison, 1,215 genes had higher expression in the high pool and 1,427 had higher expression in the low pool (Fig. 4). Of those genes, 191 were located between our QTL's BCI (Q2 = 18; Q3 = 49; Q4 = 35; Q5 = 8; Q6 = 15; Q7 = 66; Supplementary Fig. 9), obviously, higher numbers of genes are located under the QTLs with wider BCIs. Q1 is excluded here because the BCI spans a centromere. In the incubated vs room temperature (condition) comparison, 2,862 genome-wide genes were differentially expressed between high and low cohorts for thermal tolerance. In this comparison, 1,608 genes had higher expression in the incubated condition and 1,254 had higher expression at room temperature. There were 256 genes differentially expressed within our QTL intervals (Q2 = 29; Q3 = 66; Q4 = 59; Q5 = 8; Q6 = 12; Q7 = 82; Supplementary Fig. 10).

We then wanted to examine if any Heat Shock Protein (HSP) genes within our QTL intervals were differentially expressed within the cohorts (pool or condition) because Hsps are characterized as genes important for thermal tolerance in fruit flies and other species (Krishnan et al. 1989; Sørensen et al. 2001; Pörtner 2002; Fangue et al. 2006; Lockwood et al. 2017). Within Q2, which is located on chromosome 3L, there were 6 differentially expressed Hsp genes (Hsp22, Hsp23, Hsp26, Hsp27, Hsp67Ba, and Hsp67Bc) in the condition cohort, which has some consistency with work done by Freda et al. (2022), whereas only Hsp23 was significant in the pool cohort.

Candidate genes in the Q5 region

Among our identified QTLs, Q5 was much more significant than the other peaks (Fig. 3a), indicating that this locus has a larger effect on thermal tolerance, and therefore we consider Q5 a QTL of high interest. Q5 is located on chromosome 3R there are 21 genes located within that BCI region (Supplementary Figs. 9d and 10d). Eight of the 21 genes were significantly differentially expressed between high and low pools at room temperature: annexin B9 (anxB9), cortactin (cortactin), dynein heavy chain (dhc93AB), malvolio (mvl), Na pump alpha subunit (α), rho GTPase activating protein (RhoGAP93B), CG31207, and CG7079. We focused on differential expression based on standing expression in the low and high pools at room temperature because our thermal tolerance assay is short (10 min) and changes in gene expression in response to temperature would require longer timescales to impact the phenotype.

RNAi

In order to investigate the effect of differences in gene expression for genes within the Q5 interval on thermal tolerance we used the TRiP lines with knockdown expression of a particular gene. We quantified thermal tolerance in 20 UAS-RNAi mutant lines (1,969 individuals), representing 14 genes (Supplementary Table 2), which we crossed to Act-Gal4 (all cells) and elav-Gal4 (all neurons). We observed significant phenotypic effects on several of these crosses (Fig. 5; Supplementary Fig. 11; Supplementary Table 2), with the most common effect being a decrease in incapacitation time compared to the background cross. Adding this RNAi evidence to our existing data, we selected 5 genes (anxB9, cortactin, Dhc93AB, Mvl, and CG31207) as possible candidates, based on several lines of evidence.

Fig. 5.

Validation through functional genetics. The effect of RNAi knockdown for each assayed gene within the Q6 BCI. Plotted points show the difference between average incapacitation scores between each RNAi cross and the corresponding background cross for each gene within the Q6 BCI assayed. Negative numbers indicate the RNAi cross had a lower average incapacitation score than the matching background cross. Different shapes show drivers targeting different sets of tissues with circles representing the driver targeting all cells (ubiquitously expressed) and triangles showing the driver specific to neurons (pan-neuronal). Some genes had more than 1 RNAi line available so for those cases, there are two or more triangle or circle to represent the different RNAi lines. In some cases, all crosses for a particular driver—RNAi line combination was not successful and thus those are not plotted. Genes are ordered on the x axis by the degree of differential expression between pools of lines showing high vs low incapacitation. The log₂(fold change) for each gene is displayed at the top of the plot with significantly differentially expressed genes shown in red (dark red = 0.01 < P < 0.05; red = P < 0.01).

First, we started with a list of all the genes that were located within the BCI for Q5 and that were also significantly differentially expressed in our RNA-seq dataset (Fig. 5; Supplementary Fig. 12). Second, we combined this list with the RNAi data, where we measured the changes in incapacitation compared to the background line when that particular gene was knocked down. We considered a gene worthy of further consideration if the direction of the RNAi effect was consistent with our RNA-seq results (e.g. a gene with higher expression in the high thermal tolerance group should show a decrease in thermal tolerance in the RNAi cross), this effect was significant for at least 1 cross for that focal gene, and there was overall consistency in the direction of the effect in the set of crosses for the gene. The 1 exception to this criterion is Dhc93AB, which we include because it was one of our most significantly differentially expressed genes and the phenotypic effect of the RNAi line was consistent with the expected effect, though this effect was not significant. The following genes met these criteria: anxB9, cortactin, Dhc93AB, Mvl, and CG31207, and the effects and significance for these genes are shown in Table 2. While not every RNAi cross associated with a given gene showed a statistically significant effect, there is general consistency in the direction of the effect for the genes in our candidate list across crosses.

Table 2.

Effect estimates and P-values from post hoc tests for each RNAi cross in our candidate list.

Gene	RNAi ID	Cross type	Effect estimate	P-value
AnxB9	38523	All cells	−34.6	0.42
	38523	Neurons	−28.2	0.72
	58111	All cells	−59.2	0.003
	58111	Neurons	−50.3	0.01
Cortactin	32871	All cells	−25.5	0.74
	32871	Neurons	5.9	1
	44425	All cells	−52.0	0.02
	44425	Neurons	−58.4	0.002
Dhc93AB	51511	All cells	−7.5	1
	51511	Neurons	−34.5	0.40
Mvl	55316	All cells	−21.1	0.98
	55316	Neurons	−96.3	1.24 × 10−9
CG31207	60074	All cells	−25.0	0.89
	60074	Neurons	−70.1	8.42 × 10−5

Furthermore, we investigated if there were any SVs located within the DSPR founders with haplotypes located at the Q5 region, following the protocol of (Chakraborty et al. 2019). SVs are hidden mutations within the genome that are known to shape variation in complex traits (Fraser and Xie 2009; Eichler et al. 2010; Chakraborty et al. 2019), and therefore identifying them is a critical step in dissecting complex traits. We found 4 SVs that fall within Q5 (Supplementary Table 4), including an insertion within anxB9, that occurs in the founder A3 haplotype. Interestingly, the A3 founder has one of the lower incapacitation scores (Fig. 3c) at Q5. Within all 7 QTLs, we found a total of 413 SVs (Supplementary Table 3). Lastly, we performed a manual search on FlyBase (Gramates et al. 2022) to investigate previous findings of these genes being implicated with thermal tolerance. Of the 5 genes we present, only very recently anxB9 has been shown to be slightly differently expressed in D. melanogaster larva after exposure to cold stress (Freda et al. 2022). The other 4 genes have not been shown to have direct implications for thermal tolerance previously. Thus, our results showing both differential expression in low and high-performing cohorts and a consistent effect following RNAi knockdown of these same genes within a large effect QTL are the first suggestion that these genes may be involved in thermal tolerance.

Discussion

We identified 7 genomic regions that influence thermal tolerance in the DSPR, including 1 large effect locus on chromosome 3R. We also identified multiple differentially expressed genes between cohorts that differ in high vs low thermal tolerance. For our most significant QTL, we integrated our RNA-seq findings and RNAi results to identify several candidate genes with multiple lines of evidence for their involvement in thermal tolerance. Overall, we found evidence for multiple potential causative variants within the large effect locus on Q5, supporting the hypothesis that in some cases, mapped QTLs do not represent a single causative variant, but instead a collection of variants acting in concert.

Thermal tolerance and its genetic basis have been investigated extensively previously, especially in the D. melanogaster model, providing the opportunity to compare our results to previously obtained results. However, a caveat is that there is no single method for assaying thermal tolerance, which both limits direct comparison in some cases, and highlights the fact that a different assay would very likely yield different results. The most common methods for assaying thermal tolerance include single extreme temperature assays (e.g. knockdown time assays) and ramping temperature assays (Sørensen et al. 2013). These are employed both for naive individuals, and for those exposed previously to variable temperatures for variable times or at different stages, with some studies finding evidence for a different genetic basis for thermal tolerance at different life stages (e.g. Freda et al. 2017, 2019, 2022). In addition, most often survival is assayed but other phenotypes such as fertility (e.g. van Heerwaarden and Sgró 2021) have also been assayed. In addition, different rearing conditions used in different laboratories may also cause differences in results and limit the ability to make direct comparisons. For example, in our own experiment we maintained stock lines at 18°C before moving stocks to 25°C for testing in the F1 generation. This cold exposure could have intergenerational effects that are specific to our study. Nevertheless, it is essential to observe if there might be common loci identified in these different studies. Our results expand on previous studies that took genome-wide approaches to investigate the genetic basis of thermal tolerance in D. melanogaster (Norry et al. 2004, Norry, Gomez, et al. 2007; Norry, Sambucetti, et al. 2007, Norry et al. 2008; Morgan and Mackay 2006; Rand et al. 2010; Rolandi et al. 2018). We first compared our QTL dataset to Norry et al. (2008), because they characterized thermal tolerance in 2 divergent populations that were measured for thermal tolerance in both males and females. Norry et al. (2008) used heat-hardening measurements (exposure to high temps repeatedly) with an exposure temperature similar to our measurements. We found that our Q2 had an overlap with the data that was presented in that study, however, when we compared to the list of candidate genes only Hsp23 was common among the 2 groups. This is not surprising because Frydenberg et al. (2003) performed genomic analysis on isofemale lines with a specific focus on Hsp23, Hsp26, and Hsp27, and found that at Hsp23, allele frequency differs across latitude clines in D. melanogaster. Interestingly, both the founding parents from DSPR—which we used in this study—and the RILs used in Norry et al. (2008) were collected in different latitudinal climatic locations, and represent different sets of genotypes, which could also contribute to the differences in the studies. We also compared our study to the findings of Rolandi et al. (2018). This group studied the genetic basis of thermal tolerance by measuring the upper thermal limits of 34 lines from the DGRP and performed GWAS to identify QTLs. The DGRP—like the DSPR—is one of the multiparental populations in fruit flies. Briefly, they reported 8 major SNPs that are associated with upper thermal limits in both males and females. We did not find any overlap between the SNPs they identified as candidates and our QTLs, nor did they correspond with genes that were significantly expressed within our QTLs. We also cross-referenced the candidate genes found in Norry et al. (2008) and Rolandi et al. (2018) to our list of genes that were significant between high and low thermal tolerance pools and found only 1 overlap—Hsp23 (Norry et al. 2008). We have not directly compared effect estimates for the loci identified in these previous studies and our own, thus, loci we did not identify in this study may still be consistent with those findings and not identified in our study due to power. Thus, while this comparison to previous work does not allow us to make conclusions about differences in the genetic basis of thermal tolerance in these different study populations, this comparison does show that we have identified a set of novel QTLs and candidate genes that potentially influence thermal tolerance in the D. melanogaster, adding to our knowledge about the genetic basis of this critical trait.

Our findings emphasize the natural phenotypic and genetic variations in thermal tolerance (Sørensen et al. 2001, 2019; Norry et al. 2004, 2008; Overgaard and Sørensen 2008; Overgaard et al. 2011; Rolandi et al. 2018; MacLean, Sørensen, et al. 2019, MacLean, Overgaard, et al. 2019). This approach contrasts with a mutagenesis approach, which, while powerful, does not focus on natural genetic variation that is segregating. For example, we did not identify any Hsp70 genes in our integrated approach comparing high and low thermal tolerance groups. This gene family has been well-established as having a role in heat tolerance, which stems from much of the earlier work done on elucidating mechanisms of thermal tolerance (Welte et al. 1993; Feder et al. 1996). However, more recent work by Jensen et al. (2010) suggests that Hsp70 is not related to heat tolerance variation in adult D. melanogaster. Additionally, a GWAS study by Lecheta et al. (2020) also did not identify any hsp genes in their findings. The only gene from the Hsp family that was significantly differentially expressed within our design is Hsp23, within our pool treatment. Another potential reason for the lack of Hsp70 genes in our study could be that most of the studies that identified Hsp70 using D. melanogaster had a longer exposure or repeated exposure (Norry et al. 2008; Rampino et al. 2009; Cleves et al. 2020), whereas our measurement of thermal tolerance was a total of 10 min. Our focus on natural genetic variation is one of the advantages of the DSPR. In the DSPR system, the fly genome is unmodified (natural), however, we can rear them in the lab within a controlled environment, therefore reducing possible uncontrollable confounding effects. This approach allows for unbiased genomic scans to uncover natural segregating variants that would likely not be identified using single-gene approaches. Because of this systematic approach, we can identify multiple genome loci that influence complex traits, like thermal tolerance.

Genome scans are often paired with RNA-seq to further dissect the genetic basis of complex traits. We integrated RNA-seq to identify differentially expressed genes between high and low thermal tolerance cohorts within our 7 QTL regions (Wayne and McIntyre 2002). This integration continues to be a powerful approach used to narrow down potential candidate genes within QTLs that affect complex traits. We compared both low and high (baseline at room temp) groups and incubated vs room temp for genome-wide expression. This approach allowed us to identify a large set of genes whose expression differs constitutively among high and low thermal tolerant lines and the set of genes with expression changes following exposure to high temperatures. This rich dataset complements our QTL mapping results, giving additional information about the connection between gene expression and thermal tolerance. Often, the resolution of mapping approaches presents a challenge, with tens to hundreds of genes located within an associated QTL. For example, there are 472 genes located within the BCI of Q7 for the high and low cohorts, however, through RNA-seq, we were able to narrow down this list to 66 candidate genes. Notably, we might be excluding potential coding variants with this approach, however, initially focusing on regulatory variants was the most feasible approach in this case. An important limitation of our study is that we selected the 7 high and 7 low RILs for RNA-seq after initially measuring 140 RILs. Unfortunately, after measuring an additional 601 RILs and improving the accuracy and objectivity of scoring incapacitation, RILs in the high and low cohorts do not completely represent the tails of the thermal tolerance distribution. In particular, a single RIL that was originally in the high pool has a final mean incapacitation score closer to the RILs in the low pool. One of the reasons for developing a code-based scoring of incapacitation based on our activity traces was because we noticed human subjectivity in scoring the traces by hand. For example, in some of the early scoring of individuals, incapacitation was scored at timepoints after 60 s of inactivity due to signals of later activity, which can represent seizure rather than controlled movement by the fly. We addressed this issue and created objective criteria and implemented all of our scoring via code after validating this method (see Methods). However, because we chose the RILs for our RNA-seq pools from the earliest scoring, the line means were different from the final dataset with improved scoring. Despite this limitation, the high and low pools remain enriched for high vs low thermal tolerance lines (Fig. 1a). Furthermore, work done by Nielsen et al. (2006) suggests that phototransduction genes—mainly expressed in the eyes of flies—are upregulated in flies selected for increased heat tolerance, and because we used whole heads for the RNA-seq experiment, we wanted to rule out any possible phototransduction genes. We briefly compared our gene list from Q5 to their findings and found no commonality between the lists, which suggests that most of the genes we present to be candidate genes of thermal tolerance are not upregulated due to direct phototransduction activity, as described by Nielsen et al. (2006).

To functionally validate the QTL, we used the TRIP resource (Perkins et al. 2015), which is a powerful resource that provides the ability to screen more than 71% of the genes in the fly genome, because there is a mutant line for almost all the genes. We took advantage of this resource and ordered lines with gene disruption for the genes that fall within our Q5 region. In total, we screened 20 lines that we crossed to 2 pan-neuronal (elavGAL4) fly lines and 1 ubiquitously expressed line (act-Gal4). We then measured thermal tolerance in those lines together with background controls to identify how the reduction in the expression of our candidate genes influenced thermal tolerance. As hypothesized, we were not able to identify a single causal gene for thermal tolerance, but instead, our results are consistent with several genes contributing to the phenotype (Fig. 5). In fact, recent work by Herrmann and Yampolsky (2021) suggests that approximately 2,000 genes (1/8 of the Drosophila genome) either directly or indirectly contribute to temperature adaptation, suggesting a highly polygenic architecture with most variants having small effects and likely involving complex epistatic interactions (reviewed in Mackay 2015). This high complexity poses an obvious challenge and potential limit to our ability to validate all genes involved in thermal tolerance, and other highly complex traits.

Here, we present evidence for some of the genes we found to be influencing thermal tolerance, though the other genes within the QTL interval should not be discounted at this point. We highlight 5 lines that we consider to show interesting patterns. Those lines had a knockdown expression of anxB9, cortactin, Dhc93AB, Mvl, and CG31207. Based on a FlyBase and brief literature search, none of these 5 genes have been previously shown to directly influence thermal tolerance in adult D. melanogaster. In fact, relatively little is known about these genes. Briefly, anxB9 is shown to be involved in the biological processes described in transportation, localization, and development, which is concluded based on work done mostly in mutants (Bejarano et al. 2008; Tjota et al. 2011). Cortactin is involved in cell organization and reproduction, in addition, to transport and development (Somogyi and Rørth 2004; Quinones et al. 2010; O’Connell et al. 2019), while dhc93AB is involved in nervous system processes, specifically sensory perception and movement (Gaudet et al. 2011; Senthilan et al. 2012). The gene mvl is shown to be involved in other nervous system processes and respond to stimuli, which were also inferred from D. melanogaster mutants (Orgad et al. 1998; Southon et al. 2008; Panda et al. 2013). Very little is known about CG31207, and thus our work presents some of the first phenotypic data to be associated with this gene.

Our study shows the challenges of identifying the genetic basis of highly polygenic traits even with a high powered design, an advanced QTL mapping panel, and multiple informative datasets. However, we have also demonstrated how we can gain insightful information from a multipronged approach, even if we do not identify “a single gene” associated with a particular QTL. Instead, taken together, our results are consistent with the hypothesis that there are multiple contributing variants within the large effect QTL we identify and additional contributing loci across the genome. Overall, our findings suggest a highly polygenic architecture for thermal tolerance, which may be characteristic of many complex traits, as has been suggested previously (Rockman 2012). This model of the genetic architecture of complex traits has a long history, from the infinitesimal model proposed by Fisher (1918), to the more recent omnigenic model developed from this foundation (Boyle et al. 2017). Future approaches to the investigation of the genetic mechanisms underlying thermal tolerance should recognize this inherent complexity and consider pulling in additional intermediate phenotypes and methods for assaying heat tolerance.

Data availability

Raw phenotypic data for measurements of thermal tolerance both in the heat box and heat plate, including the DeepLabCut models, are available from Zenodo (https://zenodo.org/record/7767713). RNA-seq data are available from the NCBI Short Read Archive (Leinonen et al. 2011) under SRA accession number: (https://www.ncbi.nlm.nih.gov/sra/SRR24828389). Raw re-sequencing data of the DSPR founder lines are deposited in the NCBI SRA under accession number SRA051316 and the RIL RAD genotyping data are available under accession number SRA051306. Founder genotype assignments from the hidden Markov model are available as 2 data packages in R (http://FlyRILs.org/). Additionally, see King, Merkes, et al. (2012) and Long et al. (2014) for more details on the DSPR datasets. The complete code used to perform all analyses is available on GitHub (https://github.com/EGKingLab/LearnMemTol).

Supplemental material available at GENETICS online.

Funding

This material is based upon work supported by the National Science Foundation under grant no. IOS-1654866 (TZ and EGK), National Institutes of Health R01GM117135 (EGK), Univerisyt of Missouri – MU Research Council grant URC-16-007 (TZ), National Institute of Health IMSD R25GM056901 (PAW-S), and Howard Hughes Medical Institute Gilliam Fellowship (PAW-S).