Ant colonies

C. fellah queens were collected following a mating flight in Tel Aviv, Israel, in 2007. The colonies were reared at 27°C under a 12-h/12-h light/dark cycle, provided water and sugar solution ad libitum, and fed weekly with flies and an artificial ant food [34]. For 1 year prior to the experiment (C. fellah workers generally do not live longer than a year [20]), all newly eclosed workers were paint-marked weekly with a color code indicating their week of birth (Fig 1). C. fellah queens mate only once, so within colonies, all workers are full sisters [35].

One colony was analyzed in 2018, and 3 colonies were analyzed in 2020. Due to technological advances within the 2 years separating the analyses, the methods used were slightly different. The 2020 methods are reported below, and the differences are detailed in the Supporting information.

Behavioral tracking

Each colony was created by randomly subsampling approximately 100 workers (approximately 20% of the worker population) from lab stock colonies, along with the queen and approximately 20% of the brood. The colonies were housed in double-chamber setups, with a nest box (170 × 123 mm) kept in constant darkness connected via a plastic tube (internal diameter 19 mm) to a foraging arena (170 × 123 mm) under a 12-h/12-h light/dark cycle. Unique 1.4 mm² matrix barcodes from the ARTag library [36] were fixed to the thorax of each ant using SAUER skin adhesive (Fig 1). These tags weigh a fraction of what ants can carry, and their presence does not alter behavior [37].

Colonies were continuously video recorded for 7 days at 6 frames per second. The tracking system saves video files and the position and orientation of each tag in each frame. Full technical specifications and source code are available at: https://github.com/formicidae-tracker.

Tracking data processing

Workers that were detected in fewer frames than 2 standard deviations below the colony mean (5 or 6 workers per colony) were excluded from the analysis (“01-MergeData.R”). Each worker’s physical environment was quantified by discretizing the arena into an equilateral grid of hexagonal cells using the mean ant body length as the internal cell diameter and by counting the number of frames in which each worker was detected in each hexagon (“spatial fidelity.ipynb”) [38]. To quantify each worker’s social environment, we inferred pairwise social interactions from the tracking data as in [10]. Head and body regions were annotated for each ant, and interactions were defined as occurring when the head regions of 2 ants overlapped. The total number of pairwise interactions was used to create weighted social networks (Fig 2; “interaction network.ipynb”). This proximity-based definition of social interactions captures trophallaxis (approximately 16% of interactions), grooming (approximately 8% of interactions), and antennation events (approximately 35% of interactions), as well as occasions when 2 individuals move past one another or pause next to one another (approximately 42% of interactions). This high rate of “false positive” interactions has no effect on the data analysis because the frequency of pairwise false positives was well correlated with the frequency of real pairwise interactions, and the absolute numbers are not important because interaction counts are effectively normalized within colonies during analysis. To illustrate this point, we artificially increased the proportion of false positive interactions by comparing the pairwise interaction counts considering only the overlap of head regions (blue in Fig 1B) and the interaction counts when considering also the overlap of body regions (green in Fig 1B), which increases the average number of pairwise interactions from 96 to 260, and which massively increases the false positive rate (e.g., many instances of body-to-body overlap). The R-squared value between the pairwise interaction scores measured in these 2 ways was 0.812.

To quantitatively characterize individual position in the social network (“social maturity”), we used the soft community detection FacetNet (https://c4science.ch/source/facet_unil) [20,39,40]. This approach allows for overlapping social communities by outputting a continuous number in the range 0 to 1 denoting the extent to which a given node belongs to a given community (i.e., where a given worker is positioned between the nurse and forager communities). For this analysis, the number of communities was fixed at two based on previous analyses of C. fellah social networks [20]. In line with this classification, and consistent with previous results, there was a U-shaped distribution of social maturity scores, with most workers being deeply embedded in one or the other community (Fig D in S1 Text; “02-Main.R”).

Behavioral annotation

We quantified individual performance of 6 of the most important and commonly performed behaviors. Foraging frequency was automatically quantified as the number of frames in which the individual was detected in the foraging arena divided by the total number of frames in which the individual was detected (“02-Main.R”). The other 5 behaviors were quantified manually, selecting 1 frame every 2 hours (a total of approximately 80 frames) and manually annotating the identities of all workers that were:

1. Tending the queen: Positioned near, and orientated toward the queen.
2. Guarding the nest: Positioned near, and orientated toward the nest entrance, and stationary.
3. Nursing the brood: Stood on the brood pile, or with antennae or mouthparts in contact with brood.
4. Cleaning: Stood on the trash pile, or with antennae or mouthparts in contact with the trash pile, or carrying dead ants/ debris.
5. Engaging in trophallaxis: Engaged in visible fluid sharing with another individual.

These manually annotated behaviors were not normalized by detection frequency because when an individual’s identity was not detected in the selected frame, we followed the individual through the videos until the identity was resolved. Hence, foraging was bound by 0 and 1, while all other behaviors were bound by 0 and 80, though with different distributions. To give equal weight to all behaviors during dimensionality reduction, each behavior was normalized between 0 and 1 within colonies (“02-Main.R”). PCA of the normalized behavioral data yielded similarly V-shaped plots across the 4 colonies independently, and when data were pooled across colonies (Fig E in S1 Text; “02-Main.R”). For the correlational analyses, including differential expression analysis, PC1 of behavioral space was calculated separately for workers from each colony. For the machine learning–based analysis PC1 of behavioral space was calculated using individuals from all colonies together. The performance of the 6 annotated behaviors mapped onto the social networks in a consistent manner (Fig F in S1 Text; “02-Main.R”).

RNA extraction, library preparation, and sequencing

Immediately following the tracking experiment, all workers were flash frozen and stored individually in 1.5 ml Eppendorf tubes at −80°C. Brains were later dissected out in 1× PBS and homogenized in 1 ml of TRIzol reagent with ceramic beads in a PRECELLYS Evolution SUPER Homogenizer. Homogenized samples were incubated for 5 min at room temperature (RT) before adding Chloroform (200 μl), vortexing, and incubating for a further 5 min at RT. Samples were centrifuged (25 min at 12,000 rpm and 4°C) and the upper aqueous layer (approximately 500 μl) transferred to a new tube with Isopropanol (650 μl) and Glycogen blue (1 μl, RNAse-free, Invitrogen, 15 mg/ml, #AM9516). Samples were vortexed and incubated overnight at −20°C. Samples were then centrifuged (30 min at full speed at 4°C), the supernatant was discarded, and EtOH (1 ml at 80%) added. Samples were vortexed and centrifuged again (5 min at full speed at 4°C). The supernatant was discarded, and EtOH (1 ml at 70%) added. Samples were vortexed and centrifuged final time (5 min at 12,000 rpm at 4°C). All supernatant was removed, and the pellet was allowed to dry (10 to 15 min) at RT. The pellets were resuspended in nuclease-free water. The KAPA Stranded mRNASeq Library Preparation Kit (#KK8421) was used for library preparation, and samples were sequenced (150 bp, paired-end) using a full S4 FlowCell (4 lanes) on the Novaseq 6000 at the Genomic Technology Facility of the University of Lausanne, yielding 42 ±7 million (mean ± SD) reads per sample.

Gene expression analysis

The transcriptomic reads were mapped to the C. fellah reference genome (BioProject: PRJNA901066) with STAR v2.7.8a and counted with FeatureCounts using default parameters at all steps (mapping at the gene, not the transcript level) [41,42] (“Mapping.sh”; “Counting.sh”). After mapping and counting, we obtained 38.7 ± 8.3 million (mean ± SD) reads per individual.

Before running the differential expression analysis, we filtered out genes with <100 reads across all samples (i.e., 1,667 out of 14,664 genes). We used DESeq2 [43] to identify genes that were differentially expressed by each of age, social maturity, and PC1 of the behavioral, microbiota, and physical environment data, as well as for each aforementioned variable when controlling for each other variable, for each of the 4 colonies separately (“03-GeneExpression.R”). In DESeq2, significant differential expression was assessed with a Wald test, and the Benjamini and Hochberg method was used to obtain multiple testing adjusted p-values. Genes were considered as differentially expressed when the adjusted p-value was <0.05.

Microbiota

Workers were surface sterilized by dipping in 95% ethanol, soaking for 1 minute in 5% bleach, then rinsing with sterilized water. The abdomens were then removed and crushed in PowerBead tubes, and DNA was extracted using the DNeasy PowerSoil kit following the manufacturer’s protocol (https://www.qiagen.com/de/resources/download.aspx?id=91cf8513-a8ec-4f45-921e-8938c3a5490c&lang=en). For each batch of DNA extractions, we also performed blank DNA extractions in which no tissue was added, to control for possible contaminants in the reagents. A mock community composed of 16S rRNA gene plasmids [10] was processed and sequenced along the experimental samples to check for biases introduced during PCR and sequencing. To characterize the microbiota, we amplified the V4 hypervariable region of the 16S rRNA gene following the Illumina 16S metagenomic sequencing preparation guide with minor modifications (https://support.illumina.com/documents/documentation/chemistry_documentation/16s/16s-metagenomic-library-prep-guide-15044223-b.pdf) using primers 515F-Nex (TCGTCGGCAGCGTCAGATGTGTATAAGAGACACCGCGGTAA) and 806R-Nex (GTCTCGTGGGCTCGGAGATGTGTATAAGAGACAGGGACTACHVGGGTWTCTAAT), which contain the adapter sequences for Nextera XT indexes and the primers for the V4 region of the 16S rRNA gene [44]. Briefly, we performed PCR amplifications in a total volume of 25 μl, using 2.5 μl template DNA, 12.5 μl of Invitrogen Platinum SuperFi DNA Polymerase Master Mix, 5 μl MilliQ water, and 2.5 μl of each primer (5 μM). PCR conditions were as follows: 98°C for 30 s followed by 25 cycles of 98°C for 10 s, 55°C for 20 s, and 72°C for 20 s, and by a final extension step at 72°C for 5 min. Amplifications were verified by 2% agarose gel electrophoresis. The PCR products were then purified with Clean NGS purification beads (CleanNA) in a 1:0.8 ratio of PCR product to beads and eluted in 27.5 μl Tris (10 mM, pH 8.5). A second PCR step was performed to append dual-indexes to each sample using the Nextera XT index kit (Illumina). Second-step PCR amplifications were performed in a total volume of 25 μl using 2.5 μl of the products from the first PCR, 12.5 μl of Invitrogen Platinum SuperFi DNA Polymerase Master Mix, 5 μl MilliQ water, and 2.5 μl of each Nextera XT index primer. Thermal cycle conditions were as follows: a first denaturation step at 95°C for 3 min followed by 8 cycles at 95°C for 30 s, 55°C for 30 s, 72°C for 30 s, and a final extension step at 72°C for 5 min. We again purified the PCR products using Clean NGS purification beads in a 1:1.12 ratio of PCR product to beads and eluted them in 27.5 μl Tris (10 mM, pH 8.5). The amplicon concentrations were quantified by PicoGreen, and amplicons were then pooled in equimolar concentration with the exception of the negative control and blank DNA extractions, which were diluted 10×. These controls are intended to show any background noise and contamination, so the same sequencing depth is unnecessary. We verified that the final pool was of the right size using a Fragment Analyzer (Advanced Analytical) and performed sequencing on an Illumina MiSeq sequencer at the Genomic Technology Facility of the University of Lausanne, producing 2 × 250 bp reads.

We obtained a total of 11,552,825 raw sequences, from the 289 abdominal samples, 4 negative PCR controls, 4 mock community samples and 16 blank DNA extractions. Raw sequencing data quality filtered with Trimmomatic [45] using LEADING:3, TRAILING:3, SLIDINGWINDOW:4:15, and MINLEN:180. The quality-filtered data were analyzed with the Divisive Amplicon Denoising Algorithm 2 (DADA2) pipeline (“dada2” package version 1.20.0 in R) [46]. All functions were run using the recommended parameters (https://benjjneb.github.io/dada2/tutorial.html) except that we set randomize = TRUE and nbases = 3e8 at the learnErrors step, and pool = TRUE during the sample inference step. The SILVA database v.138 was used for taxonomy assignments of the identified ASVs. We removed any ASV classified as mitochondria, chloroplast, or Eukaryota (“phyloseq” package v1.36.0 [47], “subset taxa” function). We then used both the “prevalence” and “frequency” methods (method = “either”) in the R package “decontam” v. 1.12.0 [48] to identify and remove contaminants introduced during wet lab procedures, using the negative PCR controls and the blank samples as reference, which allowed us to identify 14 such ASVs (Renibacterium sp., Ralstonia sp., Microbacterium sp., Leifsonia sp., Cutibacterium sp., Enhydrobacter sp., Gordonia sp., 2 Sphingomonas ASVs, 3 Methylobacterium-Methylorubrum ASVs, a Comamonadaceae, and a Finegoldia). The final data set consisted of 10,459,506 reads belonging to 79 ASVs (“Microbiota ABC.csv”; Fig G in S1 Text).

In the analyses, we excluded Blochmannia, which is an obligate intracellular endosymbiont of Camponotus ants providing essential amino acids and with a likely role in nitrogen recycling [49,50]. This bacterium was present in all individuals (“02-Main.R”).

Multiplex network visualization and analysis

Euclidean distance was used to calculate pairwise in behavior, brain gene expression, gut microbiota profile (using relative abundances of different ASVs), and physical environment. To generate the layout for each layer in the multiplex network, we used the R package “multinet” [51]. Intralayer edges connect nodes whose similarity/interaction frequency was above the upper quartile, while interlayer edges connect like nodes (i.e., node i in layer a with node i in layer b). We used default values for all parameters except for gravity (set to 1) and iterations (set to 1,000). The networks were plotted using python package “mnet.” To compare the correlation between layers, each layer was reduced in complexity to be univariate. Social maturity was used for the social layer and age is de facto univariate. For the other 4 layers, we used PCAs (“02-Main.R”; “04-InterlayerCorr.R”). We extracted PC1 of the microbiota, behavior, and physical environmental data. We extracted PC2 of the gene expression data because all biological variables were best correlated with this principal component, while PC1 was best correlated with extraction batches from the molecular lab (Fig H in S1 Text).

Prediction with machine learning

To investigate how accurately gene expression data could be used to predict social maturity, behavior, and age, we first used DESeq2 to run a differential expression analysis on the variable of interest, using all samples and controlling for colony identity (model: colony + social maturity; model: colony + PC1 of behavioral data; model: colony + age; model: colony + foraging score, etc.; “05-ML.R”). Genes with fewer than 100 counts across all individuals were removed prior to the differential expression analysis, and genes were considered as differentially expressed when Benjamini–Hochberg adjusted p-values were <0.05. The support vector machine model (R package “e1071” with type = “eps-regression” and esp = “linear”) was trained using only genes that were differentially expressed by the variable of interest. We used half of all workers for training and the other half for testing and repeated this procedure for 100 iterations. At each iteration, the R² between the predicted and the observed values was calculated and the distribution of these scores over the iterations was used to evaluate prediction accuracy. We used 2-sample t tests to compare the distributions of R² values across the hundred iterations between variables.