2.1 Population Maintenance and Thermal Regimes

Two laboratory populations of Drosophila subobscura were founded from collections in Adraga (PT, Portugal) and Groningen (NL, the Netherlands) in 2013. These populations were threefold replicated shortly after (originating PT_1-3 and NL_1-3 populations) and maintained in with census sizes generally ranging between 500 and 1200 individuals (see Table S1), with discrete generations with synchronous 28-day cycle, 12L:12D photoperiod, reproduction at peak fecundity and constant 18°C temperature (Simões et al. 2017). In 2019, after 70 generations of laboratory adaptation, a new thermal regime was defined from each population: the global warming regime (W, specifically WPT and WNL) in which the temperature fluctuated daily (between 15°C and 21°C in the first generation) with increases in both daily mean and daily amplitude per generation (see Figure 1 and details in Santos et al. 2021). The increase in mean temperature in our experimental setup (0.18 per generation) aligns with the expected rate of temperature increase per decade (0.19°C–0.63°C [IPCC 2022]). In our thermal regime, the ratio of the increase in thermal amplitude relative to the mean (0.54/0.18) is also comparable with the IPCC prediction (2/1). The PT and NL populations were maintained in synchrony at a constant temperature of 18°C, serving as controls (C) in this experiment (i.e., assumed to represent the ancestral state of the warming populations; see Figure S1). From this point onwards, the two sets of populations derived from Portugal—PT and WPT—will be referred to the ‘low-latitude’ populations, whereas those derived from the Netherlands—NL and WNL—will be designated as the ‘high-latitude’ populations. This nomenclature is consistent with that employed in other studies (e.g., Santos et al. 2023). Given that only one population per latitude was sampled, and that the populations had already been in the laboratory for several generations when the thermal experiment began, any possible link between the patterns observed and the effects of latitudinal variation in nature must be treated with caution. Upon reaching a peak temperature of 30.2°C (generation 22), it was necessary to halt the increases in the thermal mean and amplitude of the global warming regime because of the significant decline in census size resulting from low viability (Table S1). At that time, the warming populations underwent a bottleneck (census size between 130 and 320 individuals), in contrast with previous generations (average census sizes between 830 and 940). Subsequently, the thermal cycle of the warming populations was reset to that of generation 20. Since that time, populations in the warming regime have been maintained within a temperature cycle with a mean temperature of 21.4°C, a lower extreme of 13.5°C and an upper extreme of 29.4°C (see Figure 1). This was the thermal cycle used for the warming environment in this experiment (see below). Importantly, the census size of our populations increased in the following generation in the warming environment, the assayed generation (see below). Although it was not a prolonged bottleneck in time, its magnitude may have had an impact at the genetic level, reducing the level of genetic variation to some extent and altering patterns of linkage disequilibrium in our populations (Frankham, Ballou, and Briscoe 2010; Schaper et al. 2012).

2.2 Experimental Design

After 23 generations in the warming (W) thermal regime, all populations were maintained for one complete cycle in a common garden environment (constant temperature of 18°C) prior to the assay. This step reduces maternal effects as all mothers of assayed individuals experienced similar environmental conditions. After common garden, the eggs laid by these females were placed in the two test environments, control and warming, to derive the sampled adults. A total of 24 samples were studied, corresponding to two thermal regimes (control vs. warming) × two histories (high latitude vs. low latitude) × three replicate populations (e.g., NL1-3) × two test environments (control vs. warming). Pools of 45 adult females from each population × environment were analysed in this experiment, with individuals being flash frozen (−80°C) on day eight of adult life. The age of the females was chosen as it corresponds to the median age of egg collection for the next generation in the maintenance protocol of both the control and warming regimes. Sex screening was done with CO₂ anaesthesia > 48 h before the flash freeze procedure. All sampled females were frozen in the morning of the same day.

2.3 RNA Extraction, Sequencing and Preprocessing of Reads

Female pools were disrupted using the TissueRuptor from Qiagen, and total RNA was extracted using the RNeasy Plus Mini kit (also from Qiagen). Nucleic acid quantification was performed using Nanodrop, and RNA integrity was verified by running the samples through electrophoresis on an agarose gel. Sequencing was carried out with Illumina SBS technology using rRNA removal protocol and strand-specific libraries, and aiming at 50 million reads (15 GB) per sample, with paired-end reads of 150 bp. After sequencing, the quality of the paired-end raw reads was evaluated with FastQC v0.10.1 (Andrews 2010). Reads were pre-processed using fastp version 0.21.1 (Chen et al. 2018) to trim/remove those of poor quality, setting an average quality value of 20 and a minimum read length of 120 base pairs (bp) which corresponds to 80% of read size length. The default value of 5 was used for the n_base_limit parameter. A total of 2,726,550,002 raw reads (pair1 and pair2) were generated (average of 112,783,075 per sample) with a length of 150 bp. After trimming low-quality bases with fastp, 99% of the reads were retained for downstream analyses.

2.4 Mapping of the Reads Against the Reference Genome and Features Assignment

The pre-processed reads from the 24 samples were mapped against the Drosophila subobscura reference genome—assembly UCBerk_Dsub_1.0 (Bracewell et al. 2019) using STAR software version 2.7.9a (Dobin et al. 2013). STAR was run in a 2-pass mapping with default parameters. The assignment of the reads to features (genes) was done with FeatureCounts version 2.0.0 (Liao, Smyth, and Shi 2014) with the following parameters: –p –T 9 –F GFF –t gene –g ID –s 2 –C –a GCF_008121235.1_UCBerk_Dsub_1.0_genomic.gff. The ‘–s’ option was set to 2 as the sequencing protocol used a dUTP second strand marking RNA.

A total of 2,522,844,484 reads (93%) mapped to the reference genome (on average 105,118,520 per sample), with 2,338,445,150 (86%) mapping only once (on average 97,435,215 per sample), hereafter referred to as unique mapped reads (UMR). Mapping results are shown in Tables S2 and S3. A total of 14,306 genes (99% of the genes in the reference genome) were obtained after gene count, and 12,823 (89% of the genes in the reference genome) were used in the overall expression analysis (Figure S2 and Table S4 show additional information per sample).

2.5 Differential Expression Analysis

Read count files were normalised with DESeq2's median of ratios normalisation method using DESeq2 (Love, Huber, and Anders 2014) inside galaxy (Afgan et al. 2018, Galaxy Version 2.11.40.7 + galaxy2). This method enables the estimation of size factors for each sample that are used to normalise samples to account for bias in library sizes and RNA composition across samples (Chatterjee et al. 2018).

A principal component analysis (PCA) of gene expression variation was computed on the normalised counts for all genes and samples from the different thermal regimes, histories and environments. To further scrutinise evolutionary changes between thermal regimes in the warming environment, additional PCA analysis were performed for samples tested in that environment, using three different datasets: (1) all genes, (2) genes with differential expression between thermal regimes in the high-latitude populations and (3) genes with differential expression between thermal regimes in the low-latitude populations (obtained from models explained below). We also implemented a principal variance component analysis (PVCA) to our data that allow to complement each PCA analysis by estimating the proportion of variation attributed to each effect under study.

Generalised linear mixed effects (GLMMs, Bolker et al. 2009) were applied using the normalised expression counts of each transcript (gene) as raw data to test for the effects of different fixed factors: selection, history and environment allowing to address specific questions (see below). By ‘selection’ here we mean evolutionary differences between thermal regimes, not that they are entirely due to selective forces. Nevertheless, as is common in experimental evolution studies, consistent differences between thermal regimes across replicate populations are tentatively interpreted as related to selective forces. These models incorporated the variation between replicate populations, taking into account the pairing due to their ancestry across regimes and/or environments as a random effect. Specifically, in models including the factor History, such random effect was defined as the ancestral population nested in History (AP:History) to account for the pairing of replicate populations as a function of their ancestry (e.g., PT₁ is the ancestral of WPT₁, NL₁ is the ancestral of WNL₁ etc.)—see Table 1, Row 1. To select the most appropriate model and distribution to detect differential expression of normalised read counts, we performed several tests (script available at https://github.com/marta-antunes/WithInteractions) with combinations of different models (with different interaction terms with the random factor tested in a stepwise manner—see Table S5) and distributions (Poisson and negative binomial distributions). For each gene, the best combination was the one with the lowest value of Akaike information criterion (AIC—function AICtab, Burnham and Anderson 2002, Table S1). The combination used for all genes was the best for the majority of them (Figure S3). On this basis, we used the models in Table 1 (without interactions with the random factor) and a negative binomial distribution to perform several analyses.

Different models were applied separately for high- and low-latitude populations to address specific questions (see Rows 2, 3 and 4 of Table 1): (i) the plasticity of the control high-latitude and low-latitude populations, assessed with model (2) comparing the expression in warming and control environments (factor Environment), with a total of six samples per analysis (e.g., three low-latitude replicates tested in the warming environment + three low-latitude replicates tested in the control environment); (ii) the evolutionary changes due to selection in the warming test environment, with the model (3) comparing the gene expression of populations evolving in different thermal regimes (control and warming) and tested in the warming environment; and (iii) the evolution of plasticity, with model (4) with two thermal regimes (control and warming) and two test environments (control and warming), with 12 samples tested per analysis. Additional models were applied to test for the effects of history in the control (ancestral) populations. First, a model was fitted using data from both (high- and low-latitude) populations and the two test environments to directly address the effects of history on the plasticity of the control populations including factors Environment and History as well as their interaction (model (5), Table 1). Furthermore, differences between control populations of distinct history were also assessed at each test environment separately (models with only History as a fixed factor—see model (6), Table 1). Finally, a model (model (7), see Table 1) was applied to directly test for a change in the magnitude of differences between latitudinal populations (e.g., reduction if involving convergence) as a result of thermal selection in the warming environment, by analysing variation in paired differences (absolute values) between replicate populations of distinct latitudes for both control (ancestral) and warming thermal regimes (e.g., NL_1–3—PT_1–3 vs. WNL_1–3—WPT_1–3). This was done separately for the genes that showed a differential expression between thermal regimes in low-latitude and high-latitude populations (obtained from model [3]). We also performed the same analyses excluding common genes that showed significant differences between thermal regimes in both low- and high-latitude populations. The model used in these analyses included the fixed factor Selection and the random effect Block (i.e., defined as the set of same numbered replicate populations, e.g., Block 1 with data from NL₁-PT₁ and WNL₁-WPT₁ comparisons).

All differential expression analyses were carried out in four steps: (1) selection of the read count files to be used in the analysis; (2) normalisation of the read counts using DESeq2 (Love, Huber, and Anders 2014) within galaxy (Afgan et al. 2018); (3) selection of genes that have data in at least three of the samples (> 11,000 genes in each analysis—see Table S7); and (4) detection of differential expression using glmmTMB R package (Brooks et al. 2017) assuming the negative binomial distribution (scripts generated can be found at https://github.com/marta-antunes/WithInteractions). Considering the need to account for the multiple testing involved in our analysis, p values were corrected with false discovery rate (FDR for alpha = 0.01) calculated as follows: $$$ \frac{0.01}{\sum_{i=1}^n\frac{1}{i}} $$$ where n is the gene count (Benjamini and Yekutieli 2001, theorem 1.3). Veen diagrams were created using InteractiVenn (Heberle et al. 2015).

2.6 Gene Ontology (GO) Analysis

Given the paucity of information available on the D. subobscura genome in comparison to the D. melanogaster genome, we used Proteinortho (Lechner et al. 2011) within galaxy to detect D. melanogaster orthologs for D. subobscura proteins. The input files used were protein sequence files from D. melanogaster (Release 6 plus ISO1 MT) and D. subobscura download from ncbi. All the parameters were set to default values. Among all these proteins, we selected only those corresponding to our candidate genes—obtained from the GFF file of D. subobscura genome (see Github). Around 8% of the proteins of D. melanogaster did not have correspondence in D. subobscura. The gene symbols for each locus accession were obtained by selecting the ‘Gene’ option in Batch Entrez tool from NCBI (https://www.ncbi.nlm.nih.gov/sites/batchentrez). We then used D. melanogaster gene symbols as input to BiNGO (Maere, Heymans, and Kuiper 2005) within cytoscape 3.9.1 to perform an enrichment analysis. The reference list used consists of the complete annotation of the D. melanogaster genome provided by BiNGO. The parameters used were the default ones with two exceptions: Ontology file that was set to GO_Biological_Process, GO_Molecular_Function or GO_Cellular_Component and organism/annotation was set to D. melanogaster.

2.7 Additional Analyses

To test whether the selected and non-selected genes are equally plastic, we used a Fisher's exact test with a 2 × 2 contingency table for each latitudinal population (i.e., number of selected or not selected genes versus being plastic or not plastic; see Table S8A,B; https://github.com/marta-antunes/WithInteractions).

To assess replicate heterogeneity, we calculated the mean Jaccard index of the three possible replicate combinations (e.g., PT₁-PT₂, PT₂-PT₃ and PT₁-PT₃) for low- and high-latitude populations for the candidate genes of the high latitude that were not candidates in low latitude. The Jaccard index for each combination of two replicate populations expression data was calculated by the formula $$$ \frac{\sum \limits_{k=1}^n\min \left( Xk, Yk\right)}{\sum \limits_{k=1}^n\max \left( Xk, Yk\right)} $$$ where X and Y are two vectors with length n (Charikar 2002, code used is available at https://github.com/marta-antunes/WithInteractions). Jaccard index of zero denotes no similarity between the combination and Jaccard index of one indicates maximum similarity.

2.8 Analysis of Adaptive Versus Non-Adaptive Plasticity

The genes used in this analysis were the ones that showed both significant ancestral plasticity (e.g., comparing expression of NL populations in warming and control environments) and effects of selection (e.g., comparing expression of NL vs. WNL populations in the warming environment) in each latitudinal population. By analysing only genes with signs of selection, it was possible to ascertain whether the direction of plasticity was ‘adaptive’ or ‘non-adaptive’. We assume adaptive initial plasticity when the differential expression between environments (i.e., plasticity) in the control populations is in the same direction (up or down-regulation) as the evolution of expression between the warming populations and their controls tested in the warming environment.

3.1 Overall Gene Expression Analysis

To obtain a measure of the overall variation in gene expression across different thermal regimes, test environments and histories (latitudinal populations) a PCA was performed on the normalised counts of all samples (PCA; Figure 2A). A combined reading of the two PCA axes highlights clear differences between test environments (i.e., plasticity) as the major source of gene expression variation between samples. This is corroborated by a PVCA that allows to assess the effects of different factors on gene expression variation. In fact, this analysis showed that 40.6% of the variation can be attributed to the test environments, whereas only 17.3% and 1.4% can be attributed to differences between latitudinal populations (history) and thermal regimes, respectively (see Figure 2A).

Applying a generalised linear model, we found more differentially expressed genes between samples from the same populations tested in different environments (effect of plasticity) than between populations with a different History or Thermal regime across environments (3741 genes vs. 1426 and 700 genes respectively, FDR < 0.01—see Tables S9 and S10 and Figure S4). Furthermore, from the 700 differentially expressed genes between thermal regimes (detected across the two test environments), 45% (313) were plastic and 33% (233) showed historical differences (see Figure S4). Below, we focus on the detection of candidate genes for selection—those that were significantly differentially expressed between thermal regimes—in populations tested in the warming environment, which is the new environment to which populations are adapting.

3.2 Evolutionary Changes in the Warming Environment

Given the high percentage of genes showing significant interaction between selection and history in the overall analysis (17%, see Table S9), we conducted a separate analysis of the response to the thermal regimes in the two latitudinal populations. Specifically, for each latitude of origin, we analysed gene expression differences between the warming populations and their respective controls tested in the warming environment.

A total of 221 differentially expressed genes were identified as being common to both low- and high-latitude populations (Figure S5), with a high proportion (64%) showing changes in the same direction (142 genes; Table S11). We found a higher proportion of differentially expressed genes between thermal regimes in high-latitude populations than in low-latitude ones (35% vs. 8%; Figure S5 and details below). These genes were evenly distributed along the genome in both populations (Figure 3).

3.3 Evolution of Plasticity

The plasticity of the control populations was analysed as a baseline for studying the evolution of plasticity. This was achieved by comparing the gene expression of each control population (e.g., NL1-3) between the two test environments. A total of 2427 differentially expressed genes were identified between environments (hereafter referred to as ‘plastic genes’) in the control low-latitude populations (PT1-3, Table S15A) and 1567 in the control high-latitude populations (NL1-3, Table S15B). These findings are presented in Table 2. Of these plastic genes, 606 were common to both origins (Figure S7). Interestingly, most of these genes responded in the same direction in both high-latitude and low-latitude populations (593 out of 606 = 98%). When we directly tested how the plastic response varied between control latitudinal populations (the interaction between history and environment, see model 5 Table 1), we found 334 (out of 12,453) genes showing significant differences in plasticity between the two populations (see Table S16A). The number of genes showing overall historical differences (across environments) in the control populations was much higher (3353, see Table S16B). Interestingly, when comparing the control populations in each environment separately (model 6 Table 1), we observed a clearly higher number of genes showing differences between latitudinal populations in the warming environment (3344) than in the control environment (1994)—see Table S17A,B respectively.

For the warming populations, we found 2385 plastic genes in the low-latitude populations (WPT_1-3, see Table 2 and Table S18A), whereas in the high-latitude populations, we found 3096 plastic genes (WNL_1-3, see Table 2 and Table S18B).

Below, for each history, we test whether plasticity has evolved through selection (meaning consistent evolutionary changes in plasticity between thermal regimes across replicate populations) by comparing the plasticity of populations that evolved in the warming regime with that of their respective control regime.

3.3.1 Low-Latitude Populations

The generalised mixed-effects model applied to test the evolution of plasticity (model 4 in Table 1) revealed 198 significant genes (FDR < 0.01) for the interaction between selection and environment, indicating differential evolution of plasticity due to the imposition of the thermal regime (Table 2 and Table S19A).

A total of 88 genes out of the 198 analysed (44%) changed their expression in opposite direction between the two environments when comparing the warming versus the control regimes: 54 (61%) genes were down-regulated in PT (lower expression in the warming environment compared with the control environment) and up-regulated in WPT (see left side of Figure 4), whereas 34 (39%) showed the opposite pattern (the few lighter dots below the zero line on the right side of the plot). In general, there is a pattern of up-regulation in WPT with higher expression in the warming environment when compared with expression in the control environment.

3.4 Plasticity Versus Evolution

To grasp better the connection between ancestral (control) plasticity and evolution, we examined the overlap of genes associated with control plasticity and evolutionary changes (candidate genes for selection) for both high-latitude and low-latitude populations (Figure 5A,B). This study allows us to understand the extent to which plastic genes are recruited by selection. We found that, in high-latitude populations (Figure 5B), 17% of the genes recruited by selection were plastic (723/(723 + 3472)), which is significantly higher than the 11% (844/(844 + 6861)) of plastic genes that were not recruited by selection. (Fisher's exact test with Table S5B, p value < 2.2e-16). In low-latitude populations (see Figure 5A), 34% of the recruited genes (321/(321 + 628)) were plastic, again significantly higher than the 19% of plastic genes not recruited (2106/(2106 + 8845)) (Fisher's exact test, with Table S8A, p value < 2.2e-16). These results indicate a greater recruitment of plastic genes by selection in low-latitude populations compared with high-latitude populations, even though this impact is significant in both historical populations.

To understand whether candidate genes for selection exhibited ancestral (control) adaptive plasticity, we estimated the number of genes showing expression values with the same sign in both ancestral plasticity and selection (by either consistent up or down-regulation, see Figure 5C). We found that 80% of the selected genes showed ancestral adaptive plasticity in low-latitude populations but only 17% in high-latitude populations (Figure 5C).

3.5 Comparison With Candidate Genes From Other Thermal E&R Studies

We have compiled a list of candidate genes for heat selection from several studies of thermal experimental evolution in Drosophila (Laayouni et al. 2007; Mallard et al. 2018; Michalak et al. 2019; Sørensen et al. 2020). Those genes that have a match in the D. subobscura reference genome are listed in Table 3, along with an indication of a significant effect for the term Selection and direction of change (up or down-regulation) in our warming populations of low and high latitude. It is important to note that this list is not exhaustive, but aims to provide a representative sample of the genes previously identified and highlighted as relevant in the studies mentioned above.

Among this repertoire of genes, we found that only 17% (6 out of 35) showed signs of selection in our low-latitude populations, whereas a much higher percentage (44%—15 out of 34) was obtained for the high-latitude populations (Table 3). Also, only two genes show significant signs of selection in both populations. One of them, gene GD25858, coding for ornithine decarboxylase antizyme, shows patterns of up-regulation in both populations. Sestrin and SNF4Agamma genes, which were highlighted by Mallard et al. (2018) as major candidate genes for thermal adaptation in their genome analysis, also show evidence of thermal selection in our study. However, Sestrin only showed a significant sign of selection in the low-latitude populations, whereas SNF4Agamma was only significant in the high-latitude populations (see Table 3). The catalytic subunit α of SNF4A, which was not mentioned by Mallard et al. (2018), was also found to be down-regulated in our high-latitude populations showing the same pattern as the gamma subunit.

4.1 Ample Gene Expression Response to Selection

Given that we did not find an evolutionary response for life-history traits in our populations by generation 22 (Santos et al. 2023), we anticipated that only a relatively low proportion of the transcriptome would exhibit signs of selection. However, this expectation was not met. In fact, we identified hundreds to thousands of genes differentially expressed between thermal selective regimes. It is likely that the bottleneck suffered by our warming populations two generations before this analysis caused depletion of genetic variation. However, our estimates of the total number of variable sites derived from our transcriptomic data do not show evidence of the expected decline in the variability of the warming populations (see Supporting information). This suggests that the bottleneck did not have a substantial impact on the overall genetic variability of our populations. Two factors may have contributed to this: firstly, the bottleneck was not too drastic (the lowest census size was 130), and secondly, census sizes recovered rapidly. Nevertheless, it is possible that the bottleneck may have resulted in the loss of rare alleles and the generation of some linkage disequilibrium patterns (Frankham, Ballou, and Briscoe 2010; Schaper et al. 2012).

It is crucial to acknowledge that changes in gene expression in some of these candidate genes may occur as a correlated response also leading to consistent evolutionary changes in all replicate warming populations, rather than being the direct targets of selection. Pleiotropy, linkage disequilibrium and the side effects of changes in regulatory genes may contribute to the observed variation. Moreover, we cannot exclude effects of consistent drift between replicated warming populations given spurious indications of selective response. The decoupling between life-history traits and gene expression patterns is noteworthy and might be associated with changes in other relevant traits—such as adult survival or thermal tolerance (Rogell et al. 2014; Schou et al. 2014; Manenti, Loeschcke, and Sørensen 2018) not studied directly by Santos et al. (2023). Nevertheless, that study focused on reproductive output, expected to be a most relevant trait, as a surrogate for fitness. Our whole transcriptome approach covers gene expression changes impacting on a wide range of phenotypic functions and traits, including those related to metabolism, physiology and morphology. These might have altogether given higher power to be detected, as correlated responses to the ultimate target of selection, which is fitness (Falconer and Mackay 1996). In any case, the decoupling is likely to reflect a different timing of thermal evolution between the various biological levels. To rephrase, there may be a temporal discrepancy between the evolution of gene expression patterns and fitness-related traits, with a comparatively slower response (and/or capacity to detect it) in the latter relative to the former. Indeed, we have observed the evolution of life-history traits after 39 generations of thermal evolution in our warming populations, particularly in the low-latitude populations (Santos et al. 2023). This slow response serves to reinforce the role of initially adaptive plasticity as a primary mechanism for coping with a warming environment. Furthermore, it also highlights the importance of long-term studies.

We anticipated that the control populations were already adapted to the laboratory environment at the time the warming selective regime was imposed. Our team has followed the dynamics of laboratory adaptation and observed a quick response during the first 20 generations in the laboratory followed by a reduction in the evolutionary rate (Simões et al. 2007, 2019). This suggests that our control populations were already very close to the fitness optimum when the experiment started—after 70 generations of adaptation to the laboratory environment. However, the underlying genes may continue to change frequencies even after the trait optimum has been reached (Franssen et al. 2017; Seabra et al. 2018). According to Franssen et al. (2017), adaptation to the trait optimum involves (1) directional allele frequency changes followed by (2) a plateauing and (3) further allele frequency changes eventually leading to fixation or loss of alleles after the trait optimum was reached. Considering this, it is possible that some changes in gene expression might still occur in the control populations in our study, though of a low magnitude. Having said that, changes in allele frequencies in response to the general laboratory conditions may also occur in the warming populations. However, more pronounced gene expression changes are expected to occur in the warming populations as a result of thermal adaptation. Our stringent criteria to detect candidate genes reinforces this expectation.

A relevant finding is that the numerous candidate genes detected were evenly distributed along the genome. This contrasts with the finding by Laayouni et al. (2007), who found a higher clustering of candidate genes within segregating inversions in D. subobscura populations under thermal selective regimes. Such clustering could be expected because of low recombination within and in the vicinity of inverted regions (Hoffmann and Rieseberg 2010), which could result in the identification of false positives due to hitchhiking (Schlötterer et al. 2015). Our data do not show such a prevalent clustering pattern. Importantly, our populations still presented polymorphisms for inversions by the start of the experiment with only one of the five chromosomes presenting an almost fixed inversion (average expected heterozygosity of 0.271). One simple explanation for the discrepancy between studies is that ours has tested a much higher number of genes than the Laayouni et al. microarray approach.

Common patterns were identified between the two geographical populations, with evidence for a consistent down-regulation of binding functions in both. It was anticipated that down-regulation would occur for non-essential genes (Mallard et al. 2018; Masoomi-Aladizgeh et al. 2022) and potentially involve a reduction in energy production (as described in Mallard et al. 2018). The consistent down-regulation of binding functions that we observed in both populations may be associated with this reduction in non-essential metabolic activities. This is also in agreement with our observations of specific genes showing common selective responses in both geographical populations. Of particular interest are genes coding for trichohyalin, larval serum proteins and fat body protein, which were highly down-regulated in the warming populations from both low- and high-latitude origin (Table S11). According to the Online GEne Essentiality (OGEE) database (Gurumayum et al. 2021), larval serum proteins1, gene CG31551 which codes for trichohyalin (The Uniprot Consortium 2023) and fat body protein 2 in D. melanogaster—are classified as non-essential or conditional (see also empirical studies by Meng et al. 2008; and Kasuya, Kaas, and Kitamoto 2009). The consistent down-regulation of larval serum proteins and fat body protein 2 suggests alterations in the fat body of the flies (Sato and Roberts 1983; Lehmann 1996; Jiang et al. 2005). This is in agreement with previous observations in Drosophila studies of thermal plasticity (Klepsatel et al. 2016) and adaptation (Mallard et al. 2018).

4.2 Selection Is Contingent on Previous History

A greater number of genes were identified as being candidates for selection in the high-latitude populations (4195 vs. 949). Importantly, there was no striking difference in coverage between latitudinal populations (see Figure S2). Also, differences between replicates (particularly in the candidate genes detected only in high-latitude populations) were, if anything, higher in the high-latitude populations (mean Jaccard index = 0.89 for low-latitude population and 0.90 in high-latitude populations). These observations suggest that the differences between low- and high-latitude populations were not due to reduced statistical power (due to increased replicate variability) or other sources of bias that would favour the detection of gene expression differences in the latter.

We also found clear functional differences between low- and high-latitude populations in genes that responded to selection by up-regulation. In fact, we found up-regulation of genes enriched for enzymatic activity (hydrolases such as peptidases and lipases) in low-latitude populations, whereas structural integrity was found to be more highly up-regulated in high-latitude populations. It seems plausible to suggest that these observed differences, along with the differing number of genes responding to selection, may be related to the historical adaptation of high-latitude populations to colder climates in nature. This would entail a subsequent need for greater adjustments in cellular metabolism and regulation to adapt to the warmer temperatures imposed in our experimental evolution regime. Consistent with this expectation, we do in fact observe higher evolutionary changes in the gene expression patterns of high-latitude populations. In particular, we found evidence for a convergence pattern for some of the candidate genes of the high-latitude populations, that evolved in those populations towards the expression patterns of their low-latitude counterparts. Our findings at the functional level also follow that expectation, with the high-latitude populations changing in a broader range of cellular functions than the low-latitude ones. Nevertheless, it is important to note that this potential association with natural environments must be interpreted with caution, given that our populations had already evolved for 70 generations under common, laboratory conditions when the new selective, warming regime was imposed (Santos et al. 2021, 2023). There is a conspicuous absence of studies addressing the interpopulation variation in transcriptome evolution. An exception to this is the geographical survey of D. subobscura European populations by Porcelli et al. (2016), which employed both common garden (i.e., laboratory) and in situ approaches to study gene expression. Their findings indicated that gene expression patterns reflect local adaptation along the European cline of the species, with a trade-off between metabolism and reproduction. The present study demonstrates that interpopulation variation can have a discernible impact on the evolution of gene expression in the context of climate change.

4.3 Candidate Genes for Thermal Selection

A survey was conducted on genes highlighted as relevant for heat-selective response in the thermal E&R literature. These genes were then compared with our own populations. Of the 35 selected genes from other studies, 44% and 17% were also highlighted in our high- and low-latitude populations respectively. In contrast, 16 genes (46%) of those 35 were not detected in our populations. This last finding is to be expected given the differences in thermal selective regimes and species used in the literature. Furthermore, of all compiled genes, only two presented a selective response in both high- and low-latitude populations. One of these genes, ornithine decarboxylase antizyme, shows an evolutionary up-regulation in both populations and is involved in positive regulation of protein catabolic process (Hayashi, Murakami, and Matsufuji 1996). This agrees with a plastic response of reduction in protein biosynthesis, growth and proliferation that occurs under heat stress (Klepsatel et al. 2016).

Our findings indicate that genes with a robust selective response in one population may show only a slight (or even absence of) response in another population. Consequently, inferences on targets of thermal selection based on studies in a single population can be misleading, potentially resulting in inaccurate assumptions regarding adaptation to temperature. Standardising experimental protocols by pursuing responses to ecologically relevant conditions (e.g., fluctuating temperatures and progressive warming) could help to alleviate this trend.

4.4 Interplay Between Plasticity and Selection Associated With Historical Origin

Given the dynamic nature of the warming regime, we anticipated a greater number of plastic genes in the warming populations. This expectation was only met for the high-latitude populations, which exhibited a doubling of the number of plastic genes expressed after evolution compared with their controls (from 1567 genes—corresponding to 13%—to 3096%–26%). In contrast, for the low-latitude populations, the number was virtually identical between warming and control populations. An interesting aspect is that the ancestral populations presented different starting points for thermal plasticity, with low-latitude populations having more plastic genes (2427 corresponding to 20% of genes) than the high-latitude ones (1567 genes corresponding to 13%). Interestingly, these differences in thermal plasticity translate into a higher degree of historical differentiation in the warming environment than in the ancestral one. This increased differentiation in the ancestral populations when exposed to the new (warming) environment may have favoured the evolutionary differences between populations with contrasting histories observed in this environment.

During thermal evolution, a greater number of genes changed their degree of plasticity in the high-latitude populations (325 of all genes) than in the low-latitude ones (198). In the low-latitude populations, the plasticity evolved primarily through up-regulation, whereas in the high-latitude populations, it evolved through both up- and down-regulation. It is noteworthy that the up-regulated genes are not identical in the high- and low-latitude populations, with only one gene in common. The differential evolution of plasticity between populations reinforces the relevance of their historical genetic background. These results highlight how thermal responses may vary between different populations and reinforce the importance of conducting studies in several populations to make correct assessments on adaptation to temperature changes. The magnitude of plasticity changes observed in our populations is comparable to that reported in another study addressing thermal plasticity of gene expression in D. simulans (Mallard, Nolte, and Schlötterer 2020). In that study, 2.9% (325/11,200) of genes evolved a change in plasticity relative to the control population, after > 60 generations in the hot environment. Interestingly, we did not require as many generations to obtain a rather similar response (1.6% and 2.6% in low- and high-latitude populations). It remains to be seen whether Mallard, Nolte, and Schlötterer (2020) would have obtained similar results had they conducted the experiment over a shorter time span. Nevertheless, the selective regimes applied differ (milder in the Mallard et al. study), likely leading to different selective pressures and thus different rates of response. Surprisingly, in a thermal experimental evolution study using a fluctuating regime (daily variation between 13°C and 28°C), Manenti, Loeschcke, and Sørensen (2018) did not find indications of evolution of plasticity in the transcriptome of D. simulans populations after 20 generations. These populations also did not show differences in several phenotypic traits across environments (Manenti et al. 2015). Overall, the existence of several sources of differences between studies, namely the species used, the thermal regimes applied and the different time spans of thermal evolution, prevents wider comparisons and generalisations across studies.

It is important to note that the probability of plastic genes being recruited by selection was found to be higher than that of non-plastic genes. This pattern was observed to be more pronounced in low-latitude populations (34% of genes responding to selection were initially plastic in low-latitude populations) than in high-latitude populations (17% in high-latitude populations). Furthermore, we expected that populations with a stronger selective response would show lower ancestral adaptive plasticity. In fact, this pattern was observed in our populations, with the high-latitude populations exhibiting a higher selective response and lower ancestral adaptive plasticity (17%) relative to the low-latitude populations (80%; see Figure 5). It is crucial to note that this analysis only includes the subset of genes that both responded to selection and exhibited plasticity in the control group. We are assuming the levels of adaptive plasticity that we observed can be extrapolated to the entire transcriptome. If this assumption is met, it can be concluded that the observed patterns are consistent with a buffering effect of adaptive plasticity in the evolutionary response (Kelly 2019). Conversely, a lower initial adaptive plasticity implies that a population will be further away from the ‘optimum’ (Ghalambor et al. 2007, 2015; Morris and Rogers 2013; Gibert, Debat, and Ghalambor 2019). This can explain the higher evolutionary response observed in our high-latitude populations. Altogether, these findings demonstrate the influence of the interaction between plasticity and selection (Ghalambor et al. 2007) in the context of adaptation to warming environments (Fox et al. 2019; Gibert, Debat, and Ghalambor 2019; Kelly 2019). Furthermore, it emphasises the necessity of considering interpopulation variation, which enables the assessment of the relevance of different genetic backgrounds and enhances the capacity to identify the relative influence of plasticity and selection in distinct populations.