2.1 Ant Collection and Colony Maintenance

Eight Cataglyphis niger colonies (~100 workers each) were collected from the Tel Baruch sand dunes (32.1283 N, 34.7867 E) in January–September 2021 (Figure 1), transferred to the laboratory, and kept at ~25°C, 12:12 L:D in Plexiglas cages (50 × 20 × 5 cm). Around 2 weeks after collection, 50 worker ants of C. niger (no brood and no queen) per colony were separated and placed in an acclimation box. C. niger workers do not reproduce when reared in isolation (Aron, Mardulyn, and Leniaud 2016). The workers were randomly selected from diverse sizes and marked by gluing a paper sheet with a number (5 colors, numbered 0–9) to the thorax. Before the experiment started, the ants were starved for 10 days to increase foraging activity.

2.2 Spatial Exploration of the Maze

The maze presents a sequence of binary choices and repeating subunits with the same possible moves in each subunit, but only one route leading to the reward. The maze used is similar to those previously described (Bega et al. 2020; Saar et al. 2017, Figure 2A), but with the difference that the reward was placed in the inner chambers, requiring the ants to turn either left–right–left or right–left–right. Outer chambers were avoided to prevent the ants from “following walls,” which could make it easier for them to reach the outer chambers (e.g., right–right–right or left–left–left, Grüter et al. 2015). The acclimation box is set in front of the maze, and the cotton wool sealing the entry is removed. At the other side of the maze, an Eppendorf lid was placed with 0.5 g honey as a reward. The workers were allowed to search the maze for 50 or 30 min after the first worker reached the reward (whatever happens first). Subsequently, all workers that did not return to the acclimatation box by themselves were returned manually, and two more runs were performed after a 30-min interval. The transcriptional response of immediate early genes is around 20 min, and genes related to learning can take a few hours to produce transcripts (Bahrami and Drabløs 2016; O'Brien and Lis 1993; Walton et al. 1999). We recorded the identity of all workers that arrived at the reward, the time they entered the maze for the first time, and their time of arrival. Additionally, we video-recorded the whole run for in-depth analyses of the navigation behavior. After the third run, all workers were shock-frozen in liquid nitrogen and stored at −80°C.

2.3 Analysis of Navigation Behavior

In total, 51 workers reached the reward in at least two of the three runs. From this, a subset of 24 individuals was selected for further analysis and sequencing, based on highest variability across all workers in the proportion of time needed to discover the reward in the last run, compared to the first run (time improvement). The open software AnimalTA (v2.3.4, Chiara and Kim 2023) was used to manually track the route (1.88 fps) of the workers selected for sequencing through the maze. Their individual path was smoothed using the Savitzky–Golay filter in the program. For each worker and run, we noted the following behavioral variables (Tables S1 and S2): the distance and average speed to reach the chamber with the honey the first time in each run, the distance, and average speed in the complete run. In addition, the total time the honey was ingested and the time between the last food intake and sacrifice were recorded. Food intake was defined as the interaction of the mouthparts with the honey for more than three consecutive seconds, as a time below this threshold could be considered as tasting (antennation of the honey) but not necessarily feeding on the honey. For the following analysis, only individuals that fed on the reward were considered (n = 23) reducing the total colonies tested from eight to seven.

In terms of navigation in the maze, we tracked the number of wrong and correct turns made by each worker per run. Wrong or correct turns are defined here as forward movements (leftward through a door in Figure 2A) in the wrong direction to the reward's chamber or forward movements in the correct direction. The proportion of correct turns per run was calculated per individual until reaching the reward chamber and during the whole run (Corr_prop and Corr_compl, respectively, in Table S2). The latter includes information on the proportion of correct turns of unsuccessful runs (including decisions made between successful runs), because the ants gain experience by exploring the maze after finding the honey chamber in unsuccessful runs. As both variables were highly correlated (Pearson's correlation = 0.71, df = 21, τ = 4.61, p < 0.001), we kept the former as a proxy of efficiency when navigating the maze. Given that differences in time do not capture the nuances of the behavior (ants could stand still in the maze and groom themselves or enter back to the nest before reaching the reward), we decided to focus our analysis on the correctness of the turns as a proxy for learning the route to the reward, as it has commonly used (Grüter et al. 2015; Pasquier and Grüter 2016), rather than the time improvement.

We deduce a rate in navigation improvement (L_prop) as the change in information with successive runs in the same maze as the difference in proportion of correct movements between runs: $$$ {L}_{\mathrm{last}-\mathrm{first}}=\Delta \left(\frac{C}{C+W}\right) $$$ similar to previous calculations in a study using the same species (Bega et al. 2020). The first time a naive ant enters the maze, she will explore it without any prior knowledge of a reward, without inclination to reach any specific chamber, and is here defined as the training run. In contrast, the following times it enters the maze, we assume that the ant will be actively searching for the honey chamber, leading to an increase in the L rate, if the ant remembers the path or uses other cues to orient itself. We define the last run for each individual as the final time they reached the reward, which could be during the second or third test run, depending on the individual. The first run is defined as the first time the ant enters the maze and reaches the reward, which could be on the first or second run, depending on the individual. The rate of improvement in navigation is calculated between the first and last runs, based on the movements made from the start of the run until the honey chamber is reached. Similarly, a distance improvement (L_dist) is calculated for the traveled distance in the last and first run.

2.4 Statistical Analysis of the Navigation Behavior

To describe the navigation process and select the variables of interest for expression analysis, we tested for possible correlations between the measured behavioral variables described above, using the function cor (Pearson correlation, R stats v.4.3.1), and cor_pmat from the ggcorrplot package (v. 0.1.4.1). To test if ants that improve in the proportion of correct turns across runs and move at a higher speed, we used a linear model (glmmTMB, v. 1.1.8). We expected that ants navigating a third time might show greater improvement compared to those navigating only twice, so we aimed to control for it by including the number of times the ant ran through the maze (either twice or three times) as a random factor. While it is clear that Cataglyphis ants do not use pheromones to recruit their nestmates to food sources, it is unknown whether they can perceive the footprints of their nestmates who have previously traveled through the maze, as has been demonstrated in Lasius niger (Wüst and Menzel 2017). To rule out the possibility that this influences the correct turns to reach the reward chamber, we tested whether the number of ants that reached the reward before them is linked to their navigation accuracy. Therefore, we tested the effect of the number of ants visiting the reward chamber before our tracked worker reached it on the proportion of correct turns of the tracked worker. Moreover, to rule out that the encounters with other ants affected the targets' correctness, we estimated the proportion of ants in a 2 cm radius of the tracked worker at a 10-frame rate until reaching the reward and correlated it to the targets' correctness. The model's diagnostics were checked using DHARMa (0.4.6).

2.5 RNA Extraction, Sequencing, and Preprocessing of RNA-Seq Data

The central brain, tissue including the antennal lobes and the mushroom bodies, and the two optic lobes tissue from the 23 individuals were dissected. Each of the two tissues was placed separately into 1.5 mL tubes containing 50 μL of Trizol and stored at −80°C. Total RNA was extracted after manual tissue homogenization with plastic pistils and using the Direct-zol RNA miniprep Zymo following standard instructions. We obtained on average a total yield of 87 ng of mRNA from samples of the central brain and 59 ng from the optic lobes. Afterward, samples were sent to Novogene (Cambridge, UK) for RNA quality control, library construction, and sequencing at 150 bp paired-end reads on an Illumina NovaSeq6000. After quality control, we sequenced 18 samples for the central brain and 15 for the optic lobes that resulted in an average of 30 Mio. reads and 25 Mio. reads, respectively (Table S3). Reads were adapter- and quality-trimmed using TrimGalore (v. 0.6.7) and ran on the Galaxy platform usegalaxy.org (The Galaxy Community et al. 2022). Read quality was assessed using FastQC (v. 0.12.1) and summarized using MultiQC (v. 1.11.). Trimmed reads were mapped against the reference genome of C. niger (v. 3.1) (Cohen, Inbar, and Privman 2023) using STAR (genome-mode, version 2.7.10b) on Galaxy using default parameters and specifying --quantMode GeneCounts, for later use in DESEq2 analysis. Back mapping of the samples was an average of 87.15% (average uniquely mapped transcripts, central brain: 84.80%, optic lobes: 89.94%, Table S3). We mapped the counts to 14,889 genes and filtered out genes with counts below 10 reads in at least six samples. Therefore, we kept 9583 genes expressed in the central brain 9622 genes in the optic lobes for further DESEq2 analysis (Tables S4 and S5).

2.6 Gene Expression Analysis

To model gene expression on brain transcriptional activity, we selected variables that could capture information on the navigation, such as the proportion of correct turns (measured as the average of proportion of correct turns in the test runs), and the correctness improvement, indicating changes in efficiency of the navigation after subsequent runs. We expected transcriptional activity in the brain to correlate to behaviors exhibited in the last performed run; therefore, we modeled the distance traveled in the last run before sampling. To avoid the results being heavily influenced by a few samples, we filtered potential outlier genes based on the maximum Cook's distance per gene and sample, using a Cook's cutoff of 6.55 for the central brain and 6.70 for the optic lobes samples (calculated by default following DESeq2 v. 1.40.2, Love, Huber, and Anders 2014). Therefore, 1051 genes expressed in the central brain and 882 in the optic lobes were filtered out and we compared full models that included either the last distance traveled, the correctness in the last runs and the correctness improvement as explanatory variables to reduced models without them using a likelihood ratio test (LRT) and FDR-corrected p-value < 0.05 as a significance threshold. A principal component analysis was performed with the samples using plotPCA. The functional annotations were obtained for C. niger (v. 3.1) (Cohen, Inbar, and Privman 2023) to perform Gene ontology (GO) enrichment analysis (topGO v. 2.52.0, Alexa and Rahnenfuhrer 2010), from the differentially expressed genes using weight01 algorithm and calculating the Fisher test. For visualization purposes, we summarize and remove redundant GO terms using Revigo (v.1.8.1, Supek et al. 2011), to report GO terms of the central brain expression data. Revigo finds a representative subset of the terms using a clustering algorithm that relies on semantic similarities.

3.1 Navigation Behavior

Based on the time improvement in reaching the food reward between the first and the last run (divided by the total run time), we selected 23 ant workers from seven colonies for further analysis, representing the highest variability within the workers (exhibiting improvement, no change, or worsening in time). For these workers, we recorded the proportion of correct turns on the way to the reward chamber in their first, second, and third runs, if available (Figure 2B). Notice that ants can make more than three correct turns to reach the reward if the ant backtracks but continues to move forward in the correct direction. The time improvement did not correlate significantly to the correctness improvement ($$$ {\mathrm{cor}}_{\mathrm{Time}\ \mathrm{improvement}-{L}_{\mathrm{prop}}} $$$ = −0.28, p = 0.10, Figure 2C), nor to the distance improvement ($$$ {\mathrm{cor}}_{\mathrm{Time}\ \mathrm{improvement}-{L}_{\mathrm{dist}}} $$$ = −0.28, p = 0.10, Figure 2C); therefore, we focus on the correctness and distance improvement further on.

A significant increase in correct turns by 20% in the subsequent runs showed an improvement in navigation in the maze (glmmTMB, Z-value = 3.15, p = 0.001). Workers that exhibited an increase in correct turns (positive values of L_prop) over time also walked over shorter distances to reach the reward in subsequent runs (negative values of L_dist, $$$ {\mathrm{cor}}_{L_{\mathrm{prop}}-{L}_{\mathrm{dist}}} $$$ = −0.62, p < 0.014, Figure 2C). Thus, workers that made more correct navigational decisions took shorter routes. Yet, workers who made more correct turns navigated the maze slower in the final run ($$$ {\mathrm{cor}}_{\mathrm{Correctness}\ \mathrm{last}\ \mathrm{runs}-\mathrm{Speed}\ \mathrm{last}\ \mathrm{run}} $$$ = −0.33, p = 0.02, Figure 2C). Ants that improved in their navigation moved shorter total distances in the last run ($$$ {\mathrm{cor}}_{L_{\mathrm{prop}}-\mathrm{Last}\ \mathrm{distance}} $$$ = −0.57, p = 0.003, Figure 2C), but did so at lower speed ($$$ {\mathrm{cor}}_{L_{\mathrm{prop}}-\mathrm{Speed}\ \mathrm{last}\ \mathrm{run}} $$$ = −0.41, p = 0.043, Figure 2C). Conversely, the longer the distance traveled by workers in the last run, the higher their speed ($$$ {\mathrm{cor}}_{\mathrm{Last}\ \mathrm{distance}-\mathrm{Speed}\ \mathrm{last}\ \mathrm{run}} $$$ = 0.56, p = 0.022, Figure 2C). Finally, we did not find a correlation between the average time drinking the reward and the proportion of correct turns in the test runs ($$$ {\mathrm{cor}}_{\mathrm{Correctness}\ \mathrm{last}\ \mathrm{runs}-\mathrm{Mean}\ \mathrm{time}\ \mathrm{drinking}} $$$ = 0.26, p = 0.29, Figure 2C), nor to the improvement ($$$ {\mathrm{cor}}_{\mathrm{Correctness}\ \mathrm{last}\ \mathrm{runs}-{L}_{\mathrm{prop}}} $$$ = 0.15, p = 0.34, Figure 2C).

Interestingly, of the 23 tracked workers, seven reached the reward chamber first compared to the other seven chambers, which is more than chance would suggest (X² = 6.76, p < 0.01). We found weak evidence for a link, between the correctness in the first run and the correctness in subsequent runs (cor_{Correctness training vs. Correctness last runs} = 0.37, p = 0.08). Thus, naïve workers that reached the reward chamber with a high accuracy during the first run tended to do so also in subsequent runs. We wanted to rule out that workers were following CHC footprints of nestmates when first exploring the maze, as shown for other insects (Rottler, Schulz, and Ayasse 2013; Saleh et al. 2007; Wüst and Menzel 2017). We analyzed whether the number of workers that visited the reward chamber, before the tracked worker reached it, influenced her correctness. We found no such effect of the number of ants previously entering the chamber on the first run correctness (n = 23, glmmTMB, Z-value = −0.14, p = 0.89, Figure A1 in Appendix). We also investigated the possibility that naïve ants have been influenced by the decisions of other ants. We showed that the correctness of the first run did not depend on the proportion of encounters with other nestmates (n = 46, glmmTMB, Z-value = 0.81, p = 0.42), nor an interaction between the encounters and the type of run (first or last run × Encounters, glmmTMB, Z-value = −0.217, p = 0.83).

3.2 Gene Expression in the Central Brain

The two principal components of the PCA of the central brain of 18 individuals (both with λ > 1) explained 51% of the total variation in gene expression (Figure 3A). No grouping pattern was noticeable based on colony identity (Figure 3B). No differentially expressed genes were found in response to the proportion of correct turns made during the test run(s) and correctness improvement in the central brain using DESeq2. However, we identified 478 DEGs, 271 up and 207 downregulated genes specifically (Figure 4A and Table S6), whose expression was better explained in a model with the distance moved in the maze during the last run. Among the 10 most significant DEGs, whose expression increased with increasing walking distance, we found genes involved in oxidative processes, such as cytochrome P450, peroxidase, and xanthine dehydrogenase. Other upregulated genes play a role in muscle activity, for example, matrix metalloproteinase-14 and frizzled-4, and the gene Bestrophin-2, that encodes a calcium-activated chloride channel (George et al. 2023). The GO term analysis showed enrichment in upregulated genes for 145 GO terms (74 represented by more than one annotated gene, Table S7 and Figure 4B). Central brain transcriptomes of ants moving over larger distances in the last run were enriched for processes related to energy intake and consumption, such as the “insulin receptor signaling pathway,” “cholesterol efflux,” “cholesterol homeostasis,” and “trehalose metabolism.” We also found enrichment for muscle activation processes such as “motor neuron axon guidance,” “response to muscle stretch,” “actin cytoskeleton reorganization,” “regulation of smooth muscle cell migration,” “regulation of actomyosin contractile ring contraction,” “muscle attachment,” and the “regulation of cardiac muscle contraction.”

Additionally, we uncovered pathways related to the neurological activity of the brain such as “neuron projection development” with 26 DEGs out of 461 annotated genes, for example, Serine/threonine-protein kinase MARK1, Discoidin domain-containing receptor A, Histone lysine acetyltransferase CREBBP, and Protein spaetzle encoding genes. The GO terms “synaptic target inhibition and recognition,” represented by the Extracellular sulfatase SULF-1 homolog and Protein toll encoding genes, and “neurotransmitter receptor transport, postsynaptic endosome to lysosome” represented only by the gene AP-3 complex subunit delta-1 were enriched. Other enriched processes as the “ganglion mother cell fate determination,” “positive regulation of axon regeneration,” “axon ensheathment in central nervous system,” and “ommatidial rotation” are worth mentioning. Furthermore, we found enrichment in processes regulating the transcription such as “positive regulation of CREB transcription factor activity” represented by the Histone lysine acetyltransferase CREBBP and the CREB-regulated transcription coactivator 3 encoding genes.

Downregulated genes in workers moving over larger distances included genes encoding the Transcription factor IIIB 90 kDa subunit, Surfeit locus protein 1, and the Post-GPI attachment to proteins factor 3. GOterm of downregulated genes (99 GO terms, 35 represented by more than one gene) exhibit enrichment in “signal peptide processing,” “mitochondrial respiratory chain complex III assembly,” “triglyceride homeostasis,” “anaphase-promoting complex-dependent catabolic process,” and “eye pigment granule organization” processes, among others (Table S8 and Figure 4C).

3.3 Gene Expression in the Optic Lobes

The first two axes of the PCA of the optic lobes of 15 individuals with Λ > 1 explained 73% of the total gene expression variation (Figure 5A). No grouping pattern was apparent based on colony identity (Figure 5B). We identified 81 DEGs (9 up and 72 downregulated genes, Table S9), whose expression was better explained in a model with the average proportion of correct turns made during the test(s) runs (Figure 6A). The upregulated genes include the Protein O-mannosyl-transferase TMTC2, Glutamate receptor ionotropic, Peroxynitrite isomerase THAP4, Ovarian-specific serine/threonine-protein kinase Lok, and the Carboxylesterase 5A. No functional annotation was found for four of the upregulated DEGs (gene_5625, gene_12173, gene_14297, and gene_14765). GO term analysis showed enrichment based only on a single annotated gene for the “response to glycoside,” “positive regulation of anoikis,” “cellular response to bisphenol A,” and “nitrate metabolic process” (Table S10).

Among the most significant downregulated genes, we found the RWD domain-containing protein 2A, Gonadotropin-releasing hormone II receptor, mRNA decay activator protein ZFP36L2-A, Spondin-1, Elongation of very long chain fatty acids protein 2, T-cell acute lymphocytic leukemia protein 1, and the Cyclic AMP-responsive element-binding protein 3-like protein 1 (Creb3l1). GO term enrichment for more than one annotated gene in downregulated DEGs occurred for the “response to sucrose,” “purine nucleotide biosynthetic process,” among other 41 enriched GOterms (Table S11 and Figure 6B). No differentially expressed genes were found in response to improvement and the distance walked in the last run in the optic lobe samples using DESeq2.