BONABEAU APPS

System-level reorganization of gene co-expression

An interactive guide to Page & Bonabeau, “System-level reorganization of gene networks underlies the nurse-to-forager transition in honeybees,” PNAS 123(43), October 2026. All networks below are recomputed from the authors’ published expression data.

Worker bees spend their first weeks nursing brood, then leave the hive to forage. The paper asks whether that switch shows up as a few genes turning up or down, or as the whole correlation structure of the transcriptome rearranging itself. Pick a day and watch the network split.

correlates with vitellogenin anti-correlates positive edge negative edge
–significant gene–gene correlations
–of them negative
–mean correlation between the nurse and forager modules

What was measured

Newly emerged workers were marked and placed in nursery colonies. At days 1, 3, 6, 10 and 15, sixteen bees per age were sampled and their abdominal fat body, the tissue that makes the yolk protein vitellogenin, was assayed for 90 genes plus the vitellogenin protein titer in hemolymph. The genes were chosen in advance for roles in vitellogenin and juvenile hormone signaling, insulin and nutrient sensing, brood-food production, biogenic amine signaling, immunity, and epigenetic regulation.

Because each day is a different set of bees, nothing here is a trajectory of one individual. Each network is a snapshot of how genes covary across sixteen bees of the same age. An edge is drawn only when the Spearman correlation between two genes survives a Benjamini–Hochberg correction at q < 0.05.

Vitellogenin protein per bee by age. The titer rises and plateaus by day 6, but its spread keeps widening: the quartile coefficient of dispersion goes from 0.24 at day 1 to 0.99 at day 15. A Gaussian mixture test found no evidence for two discrete groups, so the spread is a continuum rather than a split.

That widening spread matters for everything that follows. By days 10 to 15, some bees in a cohort are starting to forage while others are still nursing. The cohort is becoming a mixture, and a correlation network computed across a mixture can pick up structure that no single bee contains.

Three scales of description

The paper borrows a habit from statistical physics: describe the same system at the level of the whole (macrostate), its parts (mesostate), and its components (microstate), and ask whether a change seen at one level is visible at the others.

Macro: the whole network

Global numbers that don’t care which gene is which. How many correlations are negative? Does the network break into modules more than chance would give? Is it internally consistent?

See the measures

Meso: modules

Groups of genes that move together. Which genes cluster, how stable the clusters are across resampling, and how membership shifts from one age to the next.

See the modules

Micro: genes and bees

Where individual bees sit relative to the modules, and whether an independent gene-knockdown experiment points in the same direction.

See the bees

Macro: the signs flip, the strength doesn’t

The simplest thing to count is the fraction of significant correlations that are negative. At day 1 it is 0.6%. At day 10 it is 5%. That is an eightfold rise, while the average strength of correlations (ignoring sign) barely moves. The network is not loosening or tightening; some of its ties are turning antagonistic.

Share of BH-significant edges that are negative, recomputed here from the raw data. The paper reports a Cochran–Armitage trend test over the underlying counts of Z = +5.52, p = 3 × 10-8.
Distribution of significant correlation values on the day selected at the top of the page. Before day 10 the histogram is one-sided; afterwards a second lobe appears near −0.7.

Modularity and balance

Signed modularity asks whether the network can be partitioned into groups with positive ties inside and negative ties between, more cleanly than in a shuffled version of the same data. On days 1, 3 and 6 the answer is no: observed modularity sits at or below the permutation null. On day 10 (Q± = 0.24, z = 4.4) and day 15 (Q± = 0.21, z = 2.3) it is clearly above it.

Two more global measures tell the same story from different angles. The smallest eigenvalue of the signed Laplacian, which is zero for a perfectly “balanced” two-camp network and grows with internal contradiction, stays near zero at every age (z between −15 and −19 against the null). And an exhaustive census of triangles finds a balance ratio of exactly 1 at every age: never a triangle with an odd number of negative edges. Early on that is trivially true because almost nothing is negative. By day 10 it is true because negative edges run strictly between two coherent camps.

Values reported in the paper’s Fig. 1: signed modularity Q± against the permutation-null mean (gray), and the second spectral gap λ3 − λ2, which dips at day 6 and peaks at day 15. The dip marks a reorganization; the later rise means extra genes sit outside the two-camp split.

Meso: two modules that pull in opposite directions

Consensus clustering (three algorithms × 300 bee-resampled datasets) partitions the day 15 network into four modules. Two of them carry the signal. The authors call them H, for genes that rise with vitellogenin, and L, for genes that fall with it. Hover a gene to see where it sits in the network at the top of the page.

H: the nurse program

13 genes, mean correlation with vitellogenin +0.66. Yolk-protein machinery (vg, both hexamerins, transferrin), the brood-food protein MRJP-3, the nurse-biased insulin ligand ilp1, and the juvenile-hormone esterase JHE.

L: the forager program

14 genes, mean correlation with vitellogenin −0.56. The TOR/Rheb growth axis, the insulin receptor InR1, the forager marker Malvolio, the tyramine receptor TYR1, and two antimicrobial peptides. The most stable cluster: these genes co-cluster in 84% of resampling runs.

The other 63 genes form two further modules with little net relation to vitellogenin: one holding the small-RNA and DNA-methylation machinery together with the ecdysone receptor cassette, the other a FOXO/lipid-mobilization/dopamine group whose signed correlations with vitellogenin cancel out.

Mean significant correlation within H, within L, and between them, recomputed from the data. Within-module ties stay around +0.77 at every age. The between-module tie swings from +0.73 at day 6 to −0.70 at day 10.
Adjusted mutual information between module partitions of the 27 H and L genes at each pair of ages (paper Fig. 3B). Partitions at days 1, 3 and 6 barely resemble each other; days 10 and 15 are identical. Membership settles after the day 6 to 10 transition.

Micro: where each bee sits

Scoring every bee on its average standardized expression of the H genes and of the L genes puts each individual on a plane. Young bees cluster near the center. By day 15 they have moved to the edges: high on one program and low on the other. The population-level antagonism between H and L is, at least in part, the signature of a cohort splitting into nurse-like and forager-like individuals, not two programs fighting inside one bee.

Each point is one bee; color is its vitellogenin titer. Module scores use log(1 + expression), z-scored across all 80 bees, averaged over the module’s genes.

A check against an independent experiment

Wang et al. (2020) silenced the tyramine receptor TYR1 with RNAi and measured 12 genes that also appear in this dataset. Knockdown bees had a less forager-like profile. The authors fit a one-dimensional linear discriminant on rank-normalized knockdown data to separate knockdown from control, then projected the age-series bees onto that same axis without refitting.

Age-series bees slide from the knockdown-like end toward the control-like end as they age (Spearman ρ = −0.63, p = 3 × 10-10, reproduced below). By day 15 their mean is indistinguishable from controls and significantly different from knockdown bees. The axis was defined by a different experiment on different bees, so the agreement is not circular.

Discriminant score per sample. Dashed lines are the control and knockdown means (−0.78 and +1.10). The within-knockdown gap is optimistic by construction; the age-series contrast is out of sample with respect to the axis.

What to take away, and what not to

Reorganization, not rescaling

Correlation strength holds steady across ages while signs redistribute. The transition is a change in who is coupled to whom, which a list of differentially expressed genes would not reveal.

A sharp step between days 6 and 10

Every macro measure, the module partitions, and the between-module correlation all change most between those two samples. Day 6 looks like a hinge: the spectral gap dips, then two antagonistic camps appear.

Two attractor-like states

The day 15 network is nearly structurally balanced with two internally coherent, mutually negative modules, plus genes that belong to neither. Individual bees occupy one state or the other.

Direction is not established

The authors are explicit that “regulatory” and “dynamics” are descriptive. Network change may drive behavioral maturation, or foraging may reshape fat-body expression, or both. Sixteen bees per age also limits every estimate, which is why nulls and bootstraps appear throughout.

Terms used above

Vitellogenin

A yolk precursor protein made in the fat body. In workers it is high in nurses, supports brood-food production and immunity, and falls when a bee begins foraging. It interacts with juvenile hormone in a mutually repressive loop.

Spearman correlation and BH correction

Spearman measures monotone association between two genes’ ranks across the 16 bees of one age, which is more robust than Pearson at this sample size. With 4,095 gene pairs tested per day, the Benjamini–Hochberg procedure controls the expected proportion of false edges; an edge is kept when its adjusted q-value is below 0.05.

Signed modularity

A score for how well a network splits into groups with positive edges inside and negative edges between, compared with what a random network of the same degree would give. Here it is judged against a null built by shuffling each gene’s values across bees, which preserves every gene’s distribution but destroys the correlations.

Signed Laplacian and structural balance

A signed network is balanced if its nodes can be split into two camps with all positive edges within camps and all negative edges across. The smallest eigenvalue of the normalized signed Laplacian is zero exactly when that is possible and grows with “frustration.” The gaps between successive eigenvalues say how clean the two-camp split is and whether extra structure exists beyond it.

Consensus clustering and AMI

Rather than trust one clustering of one noisy dataset, the authors resample bees 300 times, cluster each resample with three methods, and record how often each pair of genes lands together. Adjusted mutual information compares two partitions on a scale where 0 is chance agreement and 1 is identity, corrected for differing cluster counts.

Linear discriminant analysis

Finds the single direction in gene space that best separates two labeled groups, by maximizing between-group variance relative to within-group variance. Projecting new samples onto that direction gives each a score of resemblance to either group.