Distinct constraints but shared dynamics shape the macroevolution of volatile organic compound chemodiversity

Abstract

Complex metabolic systems have evolved across the tree of life, producing remarkable chemodiversity, yet the structure and evolutionary modes underlying this diversity remain poorly understood. Here, we investigate the macroevolution of chemodiversity using a comprehensive dataset of gymnosperm foliage volatile organic compounds (VOCs) collected from the literature. We find that chemodiversity is dominated by a small fraction of “core” compounds. Although few in number, these core VOCs are widespread and abundant. To understand the full picture, we investigate the macroevolution of both high-dimensional global chemodiversity, and the dominant core chemodiversity. We find that both global and core chemodiversity have evolved through similar dynamics: punctuated evolution rather than gradual drift. We also identify a dichotomy in constraint, where global chemodiversity exhibits weak covariance—reflecting high evolutionary lability—and the core chemodiversity is characterized by high covariance and strong integration. We perform ancestral state reconstruction and find that this complex system is not a recent innovation; the major, core VOCs were already established in ancestral lineages prior to the Carboniferous period. These findings not only reveal the hierarchical structure of the gymnosperm VOC system, with its underlying constraints and dynamics, but also provide a framework for exploring the structure and evolutionary dynamics of chemodiversity.

Organisms produce volatile organic compounds for communication, defense and environmental responses, but how this chemodiversity evolved is unclear. This study shows that in gymnosperms, chemodiversity shows weak covariation, core compounds are highly integrated, and both evolve in a punctuated mode.

Introduction

Macroevolution shapes patterns of biodiversity, manifesting across multiple levels, from genetic diversity to ecosystem diversity. The structure of biodiversity and its underlying evolutionary trajectories are important parts of scientific inquiry, and have been actively studied across a wide range of biological systems^1–5, with most studies focusing on morphological traits. Chemodiversity, the diversity of metabolites, is also a vital layer of biodiversity. Despite its importance being repeatedly emphasized^6–8 and several scholars proposing quantitative models^9,10, the structure and underlying evolutionary mechanisms of chemodiversity remain largely unexplored.

Chemodiversity arises from metabolic networks and, like morphological trait diversity, is shaped by natural selection. While morphological diversity is often limited by intrinsic developmental constraints to relatively simple, low-dimensional structures, chemodiversity stems from various isozymes resulting from gene duplication and neofunctionalization¹¹, enzyme promiscuity⁶, and flexible metabolite modification⁸. This suggests that, theoretically, chemodiversity has a higher dimensionality and greater combinatorial potential than morphological traits. Metabolic systems are organized in a hierarchical manner with highly shared ‘core’ metabolic pathways and more flexible ‘peripheral’ pathways^6,12,13. This generates a complex structure where a few shared ‘core metabolites’ are widespread and abundant, while the core components and a vast array of other metabolites accumulate to form the total diversity. Given these differences, it is plausible that the highly abundant core metabolites evolve under different regimes compared to the broader global landscape. Consequently, decoding this potentially chaotic chemical space actually merits a dual examination: the overall landscape of global chemodiversity on the one hand, and the landscape of core metabolites on the other. So, what are the actual structures of global and core chemodiversity, and what are their evolutionary trajectories?

The diverse metabolites and corresponding metabolic pathways are highly intertwined, appear like a teeming chaos, and could collectively form a trait system of unfathomable complexity^2,14,15. Those diverse metabolites can be embedded into a multidimensional chemical trait space (hereafter, phylochemospace, adapted from phylomorphospace^16,17), which is capable of reflecting correlation, convergence, and the magnitude and direction of phenotypic change¹⁸. The structure of this phylochemospace is shaped by the development of these chemical traits and how they interact with, influence, and respond to their environment. In other words, genetic, metabolic, and ecological constraints among these chemicals limit the range of chemodiversity that a species can exhibit, resulting in so-called ecological ‘trade-offs’ that shape the phenotypic profiles of species¹⁹. Therefore, elucidating the phylochemospace can help us understand the internal constraints and structure of chemodiversity from an evolutionary perspective.

On the other hand, chemodiversity is essentially a combination of various individual compounds. Each metabolite is influenced by different selection regimes, resulting in varying evolutionary modes and tempos. Under completely selectively neutral conditions, metabolites will drift continuously according to an unbiased random walk, mathematically akin to Brownian motion characterized by an evolutionary rate (σ²)^4,20,21. In this scenario, the adaptive landscape of each metabolite is independent and flat, lacking significant correlation and optima^4,22. This theoretical assumption can be relaxed in two ways. On the one hand, once natural selective pressure is introduced, certain metabolites may confer a selective advantage to the species, guiding the evolutionary trajectory of the metabolites toward one or more adaptive optima ($\theta$)²³. On the other hand, with diversification and niche subdivision, the rate of metabolite evolution towards the optimum may also change, allowing for an alteration of the evolutionary rate (σ²)^21,23,24. As posited since Simpson’s time²⁵ and later supported by Gould and Eldredge²⁶, evolution is often far from a perfectly continuous process; it may also exhibit numerous large and often abrupt phenotypic “jumps”^25–27, which is called punctuated equilibrium. Studying the underlying dynamics could offer a comprehensive macroevolutionary framework to decode this complexity.

In this study, we investigated the macroevolution of chemodiversity by focusing on both the global structure and the core components of gymnosperms’ volatile organic compound (VOC) system²⁸. These compounds represent a diverse group of specialized metabolites that are small, hydrophobic molecules capable of volatilizing at ambient temperatures^7,8,29. Each VOC functions as an independent trait, while collectively, they form a complicated trait system. They are ecologically important, capable of signaling a plant’s unpalatability^30,31, attracting natural enemies of herbivores to alleviate selective pressure on the plant³², acting as vehicles for signal transduction^29,33, and facilitating allelopathy³⁴ (Fig. 1a). As an important part of the specialized metabolism system, the ecological importance and integrity of gymnosperm VOCs make them an ideal system for studying the evolution of chemodiversity.

Fig. 1. Ecological functions and core metabolic pathways of VOCs.

a Functions of specialized metabolites, using gymnosperm volatile organic compounds as an example. The green triangle indicates promoting effects, while the red prohibition represents inhibitory effects. The green hexagons depict compounds released by plants. b The diversity and schematic diagram of core VOCs metabolism. The terpenoid biosynthesis is modular and can be divided into several stages. In stage one, terpenoids’ diverse specialized metabolic pathways begin with the primary metabolite isopentenyl pyrophosphate (IPP), which is the universal C₅ building block of terpenoids. The second stage involves the assembly of the structural core of each terpenoid class. In the last stage, terpenoid skeletons are further elaborated to yield a wide range of end products via redox, conjugation, and other reactions⁶⁵. Metabolism can be much more complex, with a greater diversity of metabolites and multiple pathways for synthesizing a single compound. This simplified diagram reflects only the most recorded pathways.

Here, we find that only a small subset of VOCs is widely distributed across species in gymnosperms, and these widely distributed, core VOCs are highly abundant. The intrinsic evolutionary dynamics of both global and core chemodiversity are characterized by a punctuation mode, suggesting that VOCs actively participate in ecological interactions with their surrounding environment. We also find that core VOCs exhibit strong phylogenetic covariance, whereas global chemodiversity displays a pronounced lack. Moreover, by tracing the phylogeny back to deep time, we discern that the major axes of gymnosperm VOC chemodiversity were already established in ancestral lineages prior to the Carboniferous period.

Results

Dataset compilation

We selected literature where VOCs were extracted from foliage via hydrodistillation, identified by gas chromatography-mass spectrometry (GC-MS), and quantified their relative abundance with gas chromatography-flame ionization detection (GC-FID). Following the above criteria, through the literature search (Supplementary Data File 1, 133 selected studies), we manually compiled 984 distinct VOCs from 234 gymnosperm species into a VOC matrix (Supplementary Data File 2). Since only one species from Sciadopityaceae and Cycadaceae has relevant literature, we retained these two families (species) to maintain phylogenetic integrity but excluded them in subsequent analyses.

The structure of the gymnosperm VOC System

The gymnosperm VOC landscape is characterized by high dimensionality, where the diversity of compounds outstrips the number of sampled taxa. Despite this potential complexity, the system is notably sparse³⁵ (Supplementary Data File 2). The abundant core VOCs are generally widely distributed across the phylogenetic tree, while minor compounds display a pattern of lineage-restricted distribution at both the species (Fig. 2a, b) and family levels (Fig. 2c, d).

Fig. 2. The distribution of chemodiversity at the species and family level.

a Bar plot illustrating the distribution frequency of VOCs across gymnosperm species, categorized from narrowly distributed to widely distributed. b Locally Estimated Scatterplot Smoothing (LOESS) regression showing the positive relationship between VOC distribution breadth (number of species producing each VOC) and their mean relative abundances across gymnosperm species. The solid line indicates the LOESS fitted regression curve, and the shaded region represents the 95% confidence interval of the fitted regression. The goodness-of-fit was assessed using the coefficient of determination. Statistical significance of the relationship was evaluated using a two-sided likelihood ratio test (LRT) comparing a null model (intercept-only model) with a linear model including VOC distribution breadth as the predictor. c,d The sharing pattern of VOCs amongst the six gymnosperm families (in various combinations or isolation) is shown in the UpSet plot, which is ordered based on the frequency and Venn diagram is colored to differentiate different families.

To visualize the high-dimensional VOC data, we constructed a phylochemospace from all 984 VOC by projecting the data onto a two-dimensional plane (Fig. 3a and S3, 4). The resulting projection (dimension-reduced phylochemospace) has irregular shapes but generally occupies a single contiguous region concentrated around a central area of high species density. There is substantial overlap in the global phylochemospace occupied by each family, with their centroids positioned relatively close to one another (Fig. 3b). This pattern suggests that different gymnosperm families have converged toward similar adaptive optima. This stands in contrast to the morphospaces of birds² or genital morphology in Protohermes³⁶, where the overlap between different taxonomic groups is comparatively weak. Given sampling noises arising from differences in growth phenology, environmental factors, and the different experimental conditions between research groups, the phylochemospace is a “perturbed” representation of the actual biological system. To evaluate/simulate the impact of these potential noises, we subjected every element of the matrix to perturbation. Specifically, we introduced random variations to the relative abundance of each compound within a base standard deviation plus a proportional rate, trying to mimic the natural fluctuations caused by environmental or sampling inconsistencies. We found that the deviation between the perturbed and original datasets is minimal, as evidenced by the high overlap of centroid distributions in our sensitivity analysis (Fig. 3c). The distribution of the perturbed data remained stable compared with the original. This indicates that the gymnosperm phylochemospace remains structurally robust to such perturbations.

Fig. 3. Phylochemospace of the high-dimensional VOC system.

a Contour plot projecting the high-dimensional VOC distribution onto a two-dimensional plane. Darker green indicates higher data density, while lighter green indicates lower density. The labels within boxes mark the centroid position of the phylochemospace for each family. b Centroid-ellipse network. Ellipses represent the 95% confidence ellipses for the VOC distribution of each family, while colored dots indicate the centroids. The shade of the connecting lines reflects the degree of similarity between VOC profiles (measured by Mahalanobis distance). c Comparison of the original phylochemospace with the range of centroid drift after perturbation. The gray area represents the projection of the original phylochemospace in 2D space (defined by the convex hull of all species within a family). The green dots constitute the perturbed centroids, representing the shifting positions of family averages under multiple rounds of simulated random noise. Compared to the total area occupied by each family, the perturbation of centroids is negligible.

The evolutionary mode and tempo of gymnosperm VOCs

Each component of the gymnosperm VOC system experiences distinct selective pressures, leading to divergent evolutionary trajectories. To decipher the underlying evolutionary mode of each VOC, we fitted nine candidate models of trait evolution to our VOC dataset to identify the best-fitting mode of evolution (results in Supplementary Data Files 4 and 5). The nine models are: Brownian motion (BM), Ornstein-Uhlenbeck (OU), early burst (EB), jump normal (JN), normal inverse Gaussian (NIG), variance gamma (VG), Brownian motion + jump normal (BMJN), Brownian motion + normal inverse Gaussian (BMNIG), and Brownian motion + variance gamma (BMVG). The first three models are pure gradual models, the next three are pure pulsed models, and the last three are pulsed models with gradual Brownian motion. We found that the predominant (97.2%) evolutionary mode across all 984 VOCs is pulsed evolution (Fig. 4a, c), indicating that global chemodiversity evolves in a punctuated mode rather than through strictly continuous, gradual drift.

Fig. 4. Evolutionary tempo and mode of global and core chemodiversity.

a Relative support for evolutionary models across all 984 VOCs. The bar plot displays the cumulative Akaike Information Criterion (AIC) weights for each model, summed across all datasets and normalized to a total of 1. The height of each bar represents the overall proportion of support for that specific model within the entire dataset. Higher bars indicate that the model provides the best fit for a larger number of compounds. b The degree of fit for each VOC to different models is represented by the heatmap. A better fit corresponds to a higher AIC.w value and, therefore, a deeper color. c Schematic plots illustrate the phenotypic evolution processes under continuous, punctuated, and mixed models. d Each phylorate plot illustrates the tempo or dynamics of one type of core gymnosperm VOC. The red-filled circles positioned on the branches indicate the locations of rate shifts. The branches of the phylogeny are colored based on the distribution of evolutionary rates along each branch, with warmer colors representing high evolutionary rates and cooler colors indicating low rates. Next to each VOC phylorate plot, a bar chart shows the evolutionary rates of that VOC across different families. The red dashed line indicates the average evolutionary rate across different clades. Overall, the evolutionary trajectory of VOCs demonstrated a distribution characterized by relatively lower evolutionary rates (cooler colors) in ancestral lineages, followed by a subsequent acceleration in evolutionary rates (warmer colors) as modern species emerged.

To elucidate the evolutionary dynamics of core chemodiversity, adhering to the criteria established by Barua & Mikheyev³⁷, we identify those core VOCs as $\alpha$-pinene, myrcene, terpinen-4-ol, $\alpha$-terpineol, caryophyllene, (+)-$\delta$-cadinene, camphene, $\beta$-pinene, and limonene, which have a mean relative abundance greater than 1% of the total VOCs and found in more than 50% of species, as highlighted with a green halo background in Fig. 1b. Model fitting for the nine core VOCs also supported a punctuated mode (Fig. 4b, c). Specifically, the evolutionary patterns of terpinen-4-ol, $\alpha$-terpineol, caryophyllene, camphene, and $\beta$-pinene are similar, all of which are best characterized by a JN model (i.e., large-scale shifts between adaptive zones) interspersed with varying degrees of Brownian motion. The evolutionary trajectory of (+)-$\delta$-cadinene exhibits small and large jumps (i.e., VG model), with gaps between jumps filled by Brownian motion (BMVG). Amongst those core VOCs, the evolutionary process of $\alpha$-pinene is particularly complex, as different models show similar weights, which reflects that $\alpha$-pinene has a complicated evolutionary landscape beyond the descriptive power of single models. Nevertheless, we still noticed that the BM model slightly outperforms the others, indicating that neutral drift is an important factor in the evolution of this $\alpha$-pinene. Overall, most core VOCs are well explained by models that include a pulsing component.

To directly visualize the tempo of this punctuated evolution across time, we mapped the evolutionary rates of specific core VOCs onto the gymnosperm phylogeny (Fig. 4d). The resulting phylorate plots confirm this punctuated dynamic, capturing the precise phylogenetic locations of abrupt, large-scale rate shifts (marked by red circles). Furthermore, mapping these rates reveals a clear temporal pattern: evolutionary trajectories typically exhibit relatively low rates of phenotypic change in deep ancestral lineages (depicted by cooler branch colors), followed by significant, subsequent accelerations (warmer branch colors) as modern species and distinct family clades emerged and diversified.

Global and core VOCs covary differently

To understand the structural integration and modularity of the chemical trait system, we conducted an evolutionary covariance analysis, which reflects the degree to which traits change together due to shared genetic, developmental, or functional constraints. By analyzing these patterns, we can discern whether chemodiversity evolves as a tightly integrated complex or as a collection of independent, labile modules. This is crucial for identifying the internal constraints shaping metabolic evolution. We observed distinct covariance patterns between the global and core chemodiversity. In the global analysis, pairwise correlations among the vast majority of compounds are strikingly weak (Fig. 5a), revealing a highly modular structure where the global chemodiversity is not tightly integrated but rather consists of numerous independent, freely evolving modules. In sharp contrast to this modular pattern of global chemodiversity, we observed strikingly high covariance between core VOC components (Fig. 5b), suggesting a tightly constrained evolutionary integration. For example, limonene exhibits a significant positive covariance with $\beta$-pinene, while it shows a notable negative covariance with (+)-$\delta$-cadinene.

Fig. 5. Correlation structure and multivariate patterns of global and core chemodiversity.

a Correlation pattern of global chemodiversity. The isosceles triangular matrix visualizes correlations. Left- and right-angle sides represent all 984 VOCs. Color ranges from white (low or no correlation) to dark green (high correlation). b Phylogenetic covariance between nine core VOCs. The diagonal panels show kernel density estimates of VOC posterior distributions. The lower triangular panels display pairwise correlation plots with linear regression fits and 95% confidence intervals. A green background represents a significant positive correlation, while a khaki background indicates a significant negative correlation. A blank background signifies no significant correlation between specific pairs of VOCs. Asterisks denote statistical significance, with * indicating p values between 0.05 and 0.01, ** indicating p values between 0.01 and 0.001, and *** indicating p values less than 0.001. Statistical significance of correlations was assessed using two-sided Bayesian posterior inference based on the posterior distributions of correlation coefficients. c Phylogenetic PCA of nine core VOCs. The first two principal component axes capture VOC relationships across the nine types of core VOCs. In this phylochemospace, gymnosperm VOCs cluster along four major axes: $\alpha$-pinene, $\alpha$-terpineol, $\beta$-pinene, and caryophyllene, representing four core VOC formulations that converge in gymnosperms.

To visualize these internal constraints of core chemodiversity, we projected the phylochemospace of nine core VOCs, which exhibits an effectively low dimensionality characterized by four core axes (Fig. 5c). The first two principal axes capture 84.07% of the variation in the phylochemospace. The first axis (PC axis 1) captures a trade-off: caryophyllene, camphene, and (+)-$\delta$-cadinene align in one direction (also supported by positive covariance in Fig. 5b), while limonene and $\beta$-pinene align in the opposite direction. Crucially, while components within each group are positively correlated, the groups themselves are negatively correlated, suggesting a fundamental metabolic or ecological trade-off along this primary axis. Meanwhile, $\alpha$-pinene and $\alpha$-terpineol define the independent variation along PC axis 2. We also notice that $\alpha$-pinene lacks significant covariance with other compounds (Fig. 5b), suggesting it may be expressed constitutively or regulated by distinct mechanisms decoupled from the main terpenoid network.

Estimating the ancestral state of the gymnosperm VOC system

Because the core compounds represent converged, shared evolutionary themes with deep phylogenetic time, we applied ancestral state estimation to infer their evolutionary past and ancestral states of the gymnosperm VOC system (Fig. 6a, b). Based on the PCA of core chemodiversity, we categorized the gymnosperm VOC systems into five major composition patterns defined by the dominance of specific core metabolites: (1) $\alpha$-pinene-dominant, (2) $\beta$-pinene-dominant, (3) $\alpha$-terpineol-dominant, (4) caryophyllene-dominant, and (5) an ‘other’ pattern, which encompasses various alternative core strategies with lower individual frequencies. We fitted these ‘dominant patterns’ into different continuous-time, discrete k-state Markov processes, namely, the equal-rates (ER) model, the symmetric transition model (SYM), and the all-rates-different (ARD) model (Fig. S6). By comparing the different models’ likelihood ratios and AIC values, we found the all-rates-different model had the statistically significant highest fit (Supplementary Data File 6).

Fig. 6. Ancestral state estimation of core VOCs evolution in gymnosperms.

a The best-fitting discrete 5-state Markov model depicts the transition rates between dual VOC characters. Rates are individual changes per million years. The arrows’ colors represent the transition speed (cooler being lower and warmer being faster) b Pie charts at nodes represent the marginal empirical Bayesian posterior probabilities of estimated ancestral states of five dominating VOC types.

If a gymnosperm employed caryophyllene, $\beta$-pinene, and ‘other’ as its primary VOC composition, its ancestor would possess considerable evolutionary potential toward $\alpha$-pinene. However, if a gymnosperm primarily features $\alpha$-terpineol in its VOC composition, its ancestor likely lost the direct evolutionary potential towards $\alpha$-pinene (Fig. 6a). Once a lineage’s ancestor establishes $\alpha$-pinene as the primary composition, its descendants tend to conserve this trait, resulting in a sharp decline in evolutionary potential towards other major composition patterns. We also observed very high transition rates between the states ‘other’ and ‘$\alpha$-terpineol’, indicating that these two states are likely randomly distributed concerning one another on the tree. Suppose $\alpha$-terpineol dominates a species’s VOC formulation. In that case, it is just as likely that a very close relative will have an ‘other’ formulation as share the $\alpha$-terpineol formulation, and vice versa. We reconstructed the gymnosperm VOC system’s ancestral state based on the Markov transition matrix underlying this schematic (Fig. 6b). Among existing species, $\alpha$-pinene is the predominant composition pattern in all major families we analyzed except for Ephedraceae. Nevertheless, the probabilities of the five major VOC composition patterns were equal in this study’s common ancestor of all gymnosperms.

Discussion

Metabolic allocation and chemical convergence of the VOC system

In Fig. 2a, b, we found a positive correlation between distribution breadth and abundance, which is consistent with natural selection: when a compound confers a selective advantage, it tends to be (i) retained across descendant lineages (increasing its phylogenetic breadth) and (ii) upregulated (enhancing its abundance) to amplify fitness. Our quantitative models mathematically demonstrate the strength of these two aspects of chemodiversity under natural selection (Fig. S1). This suggests a metabolic allocation strategy where resources are concentrated into core compounds, which subsequently act as substrates for low-cost diversification. In contrast, lineage-specific/restricted, low-abundance derivatives serve as a reservoir for innovation. These compounds represent a low-risk exploration of chemical space that, while currently low in abundance, may provide critical adaptive advantages under changing ecological pressures^11,13. Because disparate gymnosperm families invest their resources into similar core VOCs, their overall chemical phenotypes have converged heavily toward shared adaptive optima. This is visually evident in the global phylochemospace (Fig. 3), where the chemical distributions of different taxonomic families exhibit massive overlap. In morphological systems, species typically undergo adaptive radiation by diverging into distinct, non-overlapping physical phenotypes to partition niches and avoid direct resource competition (e.g., evolving entirely different beak shapes to exploit different food sources)^2,38. In contrast, in chemodiversity, “niche partitioning” does not require abandoning the underlying structural architecture. Gymnosperms can share the exact same core chemical “optimum” for general defense or signaling³⁹, achieving specificity merely by tweaking the low-abundance innovation reservoir at the metabolic tips^11,13. Thus, while extreme morphological traits always diverge in morphospace, chemodiversity could remain structurally convergent.

Gymnosperm VOCs evolved with intermittent pulses

Our model-fitting analyses demonstrated that 97.2% of global VOCs and most core VOCs are best explained by models incorporating evolutionary jumps (Fig. 4), which establish that gymnosperm chemodiversity evolved through intermittent pulses. The concept of pulsed evolution is increasingly recognized, transitioning from heterodoxy to a well-accepted idea supported by empirical research across various taxa and traits. Evidence ranging from skink morphological novelties⁴ to vertebrate body size evolution²⁷, and from early echinoderm anatomy³ to plant reproductive innovations⁴⁰, suggests that many phenotypic traits have evolved through pulses rather than through purely incremental change. In the context of gymnosperm VOCs, among nine VOCs studied, most exhibited pulsed components, which indicates that the macroevolutionary dynamics of chemodiversity within the gymnosperm VOC system share a similar tempo and mode of evolution with those of morphological traits. However, it is important to note that model fitting results indicate that many VOCs possess gaps between evolutionary jumps filled by Brownian motion (BMVG, BMNIG, and BMJN). We consider the evolutionary process of gymnosperm VOCs more akin to the punctuated gradualism model⁴¹.

Core compounds such as myrcene, terpinen-4-ol, $\alpha$-terpineol, caryophyllene, camphene, and $\beta$-pinene evolved through large-scale shifts. This suggests that those metabolites have often escaped random diffusion (Brownian motion) to conduct rapid explorations of trait space, ultimately landing in new adaptive zones (also reiterated by red dots in phylorate plots), and therefore confer the selective advantage of their producers. For example, myrcene, a host-derived precursor involved in bark-beetle pheromone communication⁴², has been linked to mass mortality in gymnosperms⁴³. Indeed, myrcene exhibits allelopathic effects that inhibit the growth of neighboring plants⁴⁴, conferring higher fitness and therefore explaining the high evolutionary rate of this compound during early evolutionary stages. However, in later stages, the attraction of pine beetles resulted in decreased fitness for gymnosperms, leading to notable declines in the evolutionary rate of myrcene. Terpinen-4-ol has been implicated in suppressing the growth of other plants through allelopathy⁴⁵, which corresponds with numerous significant increases in evolutionary rates observed in the phylorate plot. This indicates that during the evolution of gymnosperms, terpinen-4-ol conferred significant selective advantages at various times. On the one hand, the evolutionary rate of $\alpha$-terpineol remains high in certain lineages, likely due to its allelopathic effects⁴⁶ and its ability to attract natural enemies of Fabaceae plants⁴⁷, thus providing ecological advantages by reducing competition. On the other hand, significant declines in its evolutionary rate in some lineages may be due to similar reasons underlying the trends observed with terpinen-4-ol. Caryophyllene has demonstrated direct inhibitory effects on bacterial growth⁴⁸, potentially explaining its high evolutionary rates across most lineages. Interestingly, $\beta$-pinene exhibits unique evolutionary rates within the Pinaceae and Araucariaceae families, which may not be adequately explained by stress responses and natural selection, suggesting a fixation process resulting from genetic drift, and many other factors may also contribute.

Overall, the global and core chemodiversity exhibit pulses and significant rate variations across multiple lineages, indicating their active involvement in chemical ecology (ecological interaction mediated by chemical substances). This is attributed not only to their volatile nature, which facilitates long-distance dispersal, but also to their lipophilic properties, allowing them to freely traverse cellular membranes and influence signaling pathways in other organisms⁴⁹. In general, we can assert from an evolutionary perspective that VOCs should no longer be regarded as mere waste products of metabolism³³ that undergo random diffusion. Instead, we found many scenarios of pulsed evolution where selection may be particularly effective.

It is important to acknowledge an inherent limitation of historical chemical occurrence data: the lack of a reported detection does not necessarily indicate the true biological absence of a compound. Intuitively, this may lead to an overestimation of pulsed models, as “absence-to-presence” trait evolution can be interpreted as discrete jump events. To account for this, we conducted sensitivity analyses by simulating trace abundances for undetected compounds. Our results confirm that the evolutionary inferences are highly robust to these perturbations (Figure S9). Because trace amounts exert minimal leverage on the overall variance compared to the highly expressed compounds, the punctuated dynamics observed likely reflect true adaptive shifts rather than analytical artifacts. Interestingly, the perturbation increased the support for the VG model. While JN captures isolated major shifts against a static background, the VG model accommodates an infinite number of small jumps alongside major leaps, capturing the trace-level perturbations without losing the large shift signal.

Global Modularity and core integration of the VOC system

Evolutionary covariance reflects the degree to which traits change together due to shared genetic, developmental, or functional constraints. Because the global chemodiversity involves many lineage-specific evolutions, most VOCs rarely show high co-occurrence across the phylogeny, statistically precluding the formation of strong, system-wide correlations. Besides, many metabolites are often synthesized by specialized enzymes acting at the terminal branches of metabolic pathways. By diversifying at these endpoints rather than at central metabolic junctions, these compounds avoid determining the flux of the entire pathway, thereby allowing them to vary without dragging other traits along with them. This pattern mirrors the covariance pattern of the snake venom system³⁷, where no significant covariance in the expression of venom genes was detected, and the covariance matrix was nearly algebraically singular. This is because the venom system comprises several genetic subunits that can vary independently of one another to fill specific ecological niches rapidly. This type of developmental modularity is highly advantageous in specialized plant metabolism, as it reduces pleiotropic constraints and allows plants to rapidly assemble novel chemical cocktails in response to dynamic, fluctuating herbivore pressures without compromising fundamental physiological functions⁵⁰.

This notable pattern of strong core covariance is likely attributed to both metabolic and ecological constraints of the gymnosperm VOC system (i.e., shared biosynthetic pathways and ecological coupling). From a biosynthetic pathway perspective, the core VOCs all belong to the terpenoid family, many of which share common precursor compounds, and differentiation among different VOCs (metabolic branching points) occurs only at the end of a long metabolic pathway. On the one hand, any increase or decrease in enzyme activity prior to the metabolic branching point will simultaneously raise or lower the yields of both VOCs at the end of that metabolic pathway, thus explaining the source of significant positive covariance in the covariance matrix. On the other hand, if the enzymes responsible for synthesizing the two VOCs after the branching point are reciprocally regulated, this could lead to changes in the yields of the two VOCs in opposite directions, resulting in significant negative covariances. Furthermore, from a chemical ecology perspective, selection for synergistic or antagonistic functional traits may reinforce these correlations. Ultimately, such strong integration constrains the evolution of core chemodiversity to changes along limited, conserved evolutionary directions^51,52.

In line with the perspectives of Guillerme et al.⁵³, we propose that several core VOCs constitute a major axis (also referred to as elaboration), serving as a line of least resistance that provides a pool of variation upon which selection can act⁴. In metabolic terms, these core VOCs offer parent structures from which new compounds can be easily derived. Conversely, minor VOCs that evolve along numerous minor axes (also known as innovation) represent a high level of evolutionary dynamism and flexibility. Thus, the gymnosperm VOC system balances stability and plasticity, where major axes or core components (elaboration) reflect metabolic constraints, limiting free response to environmental changes, whereas widespread minor axes (innovation) facilitate adaptive shifts in expansive directions different from the core, thereby endowing gymnosperms with a multi-layered strategy for chemical adaptation.

Deep time origin of core compounds

The ancestral state reconstruction (Fig. 6) shows that the ‘heritage’ of gymnosperms was complete, indicating that the origins of these VOCs and their corresponding metabolic pathways are quite ancient, predating the Carboniferous period. Substances such as $\alpha$- and $\beta$-pinene, camphene, myrcene, limonene, camphor, and $\alpha$-terpineol can even be found in mosses and ferns (more primitive groups of embryophytes)^54,55. This suggests that to explore the origins of volatile compounds, we may need to look into clades that are even more primitive than gymnosperms. However, although the original formulation is quite complete and uniform, with natural selection, the chemodiversity in the gymnosperm VOC system opted for $\alpha$-pinene as a major axis, while a smaller number chose other VOC composition patterns. We also observed that lineage-specific specialization occurred relatively recently, around the past 100 million years ago, which coincides with the Cretaceous period—the era when angiosperms emerged and the competition between gymnosperms and angiosperms started.

In conclusion, the gymnosperm VOC system offers a promising platform for understanding the structure of chemodiversity and the underlying evolutionary processes that shape it. We show that a small proportion of compounds are widely shared across taxa, and these widespread compounds are abundant; we refer to these as core compounds. We have demonstrated that global and core chemodiversity share common evolutionary dynamics. Both global and core have been subject to strong natural selection throughout evolutionary history. In addition, they evolved in a punctuated mode. However, they diverge fundamentally in their structural constraints. We identify a dichotomy in evolutionary covariance: global chemodiversity exhibits weak covariance patterns, reflecting high evolutionary lability and freedom, whereas core chemodiversity is characterized by high covariance, indicating a strong degree of evolutionary integration. The diversity and complexity of the gymnosperm VOC system are not a product of later evolutionary stages but a characteristic that dates back to the ancestors of gymnosperms, as the core VOCs evolved from a profound depth of phylogeny and were already well established by the Carboniferous period.

The evolution of chemodiversity has gained increasing attention in the fertile ground of the phytochemistry renaissance. On the theoretical front, we take the gymnosperm VOCs as a model system to probe the macroevolution of chemical traits and provide a framework for exploring the tremendous chemodiversity of living organisms.

Methods

Phylogeny and VOC data matrix compilation

We used a stratigraphically calibrated plastid phylogeny of gymnosperms constructed in a previous study²⁸. For each species in the phylogeny, we cross-referenced the nomenclature with the International Plant Names Index (IPNI; https://www.ipni.org) to verify taxonomic accuracy. We then collected data regarding the following criteria: we searched Web of Science and Google Scholar, using the keywords: “essential oil” OR “volatile organic compounds” AND “species name” to collect relevant literature. Considering that different plant tissues, different extraction and analytical methods can introduce inconsistencies in the compound identification results reported in the literature, we therefore prioritized studies employing hydrodistillation for VOC extraction specifically from gymnosperm leaves, with compound identification conducted through gas chromatography-mass spectrometry (GC-MS) for qualitative characterization and quantification achieved via gas chromatography-flame ionization detection (GC-FID) to determine relative abundances. We excluded species for which no literature was found. After identifying the relevant literature (Supplementary Data File 1), we extracted the following information regarding the essential oils or volatile organic compounds: (i) compound name, (ii) CAS ID, and (iii) corresponding relative content. Articles that only provide absolute content without relative content were discarded. The data collection was focused on foliage whenever possible to account for metabolic differences in various plant parts. Ultimately, we compiled the collected data into a data matrix (Supplementary Data File 2) according to the CAS ID (consolidating different names for substances with the same CAS ID) and the corresponding species. The data matrix, which contains all compounds regardless of their prevalence, is referred to as the global VOC chemodiversity. It has dimensions of 234 × 984, where the (i,j)-th element represents the relative abundance of the j-th compound in the i-th species. To screen the core part of gymnosperm VOC chemodiversity, we have also restricted our dataset to include volatile chemical compounds that (1) are robust to perturbations, (2) are present in at least 50% of gymnosperm species, and (3) exclude VOCs with relatively low content (using 1% as a threshold) (Supplementary Data File 3). The phylogeny was then pruned to 234 species that matched our dataset using drop.tip function in the R package ape⁵⁶. By using this comprehensive dataset of previously detected gymnosperm VOCs from the literature, we conducted the following analysis.

Data transformation and construction of phylochemospace

The dataset exhibits two key characteristics: it is compositional and high-dimensional. Given that the compositional data reside in Aitchison space, we applied the centered log-ratio (CLR) transformation to map them into Euclidean space. Subsequently, as the number of dimensions exceeded the number of samples, we estimated the covariance matrix of the global VOC chemodiversity dataset within a penalized likelihood (PL) framework⁵⁷. The penalized covariance matrix was then used to construct the phylochemospace. Next, species were partitioned by taxonomic units, and kernel density estimates (i.e., contour plots) were generated. Centroids for each family were computed based on the first two principal components (PCs). For the same dataset, we also calculated the variation radii for each family, with radii derived from the Mahalanobis distance and the chi-squared distribution. For a 95% confidence level (level = 0.95) and degrees of freedom df=2 (two-dimensional space). The radius formula stems from the ellipse boundary: $r=\sqrt{\max \left(\lambda \right)}\cdot \sqrt{{\chi }_{0.95\text{,}2}^2}$, where $\sqrt{\max \left(\lambda \right)}$ scales the standard deviation along the major axis, and ${\chi }_{0.95\text{,}2}^2$ is the chi-squared quantile with df =2, ensuring the ellipse encompasses 95% of the data points under a normality assumption. The phylochemospace construction of core VOCs is based on a phylogenetic generalized linear mixed model (GLMM). The phylogenetic GLMM was constructed using the MCMCglmm R package⁵⁸.

Matrix perturbation analysis

Since VOCs can vary due to differences in organs, geographic origins, and growth stages, introducing “noise” into the data matrix, we therefore treated the observed VOC matrix as a perturbed sample of the underlying true biological feature matrix. To assess robustness, we applied Gaussian perturbations to the original global VOC chemodiversity matrix (with a 10% base standard deviation and 20% proportional rate), followed by CLR transformation and penalized PCA. Centroids were then recalculated for each family (restricted to those with more than four species) along the first two PC axes. This process yielded a number of centroids per family equal to the number of perturbations. With axes fixed, these centroids were projected onto a two-dimensional plane, and convex hulls were delineated. By comparing the convex hull areas to the original data distribution, we evaluated the perturbation resistance of each family.

Trait evolution models

We fitted both global and core VOCs to various models of trait evolution: Brownian motion (BM), Ornstein-Uhlenbeck (OU), early burst (EB), Jump Normal (JN), Normal Inverse Gaussian (NIG), variance gamma (VG), BM + JN, BM + NIG, and BM + VG by using the pulsR package²⁷. The first three models represent incremental processes, while the next three models represent pulsed evolution. The final three models are a combination of incremental and pulsed evolution. Brownian motion (BM) is a stochastic, directionless process in which variation accumulates through time, with a rate parameter denoted σ². The Ornstein-Uhlenbeck (OU) model is an extension of the BM model, with an additional parameter ($\alpha$) that describes the tendency to evolve towards a central value $\left(\theta \right)$, where $\theta$ corresponds to the position of the adaptive optimum, and $\alpha$ to the strength of natural selection. Under the early burst (EB) model, the rate parameter σ² starts with the initial value ${\sigma }_0^2$ and then decreases monotonically with an exponential decay parameter $a$. The current Lévy processes are characterized by a Brownian motion with rate σ² and a jump measure $v\left(\text{d}x\right)$. In JN model, the waiting time between jumps is exponentially distributed, whereas jump sizes are normally distributed. The characteristic function is $\log {\phi }_{\text{JN}}\left(\xi \right)=t\lambda \left({\text{e}}^{-\frac{{\tau }^2{\xi }^2}{2}}-1\right)$, where $\xi$ is a dummy variable, $t$ is a given branch length of phylogeny, $\lambda$ is the jump rate, and $\tau$ is the standard deviation of jump sizes. In this model, a lineage stays in an adaptive zone for a duration that follows an exponentially distributed amount of time before jumping to a new one. The NIG process simulates an infinitely active process (constantly jumping) with a jump measure characterized by a Bessel function, and its characteristic function is given by $\log {\phi }_{\text{NIG}}\left(\xi \right)=t\delta \left({\alpha }^2-\sqrt{{\alpha }^2-{\xi }^2}\right)$. Here, $\delta$ determines the rate of jumps, while $\alpha$ determines the scale of the jump distribution. The VG process allows for both frequent small jumps and occasional large shifts. It has a characteristic function: $\log {\phi }_{\text{VG}}\left(\xi \right)=-\frac{t}{v}\log \left(1+\frac{{\xi }^2{\tau }^{2}v}{2}\right)$, where v is roughly inversely proportional to the jump rate, and τ determines the scale of jumps. Weighted Akaike information criterion (AIC.w) values were used to evaluate the goodness of model fit. Given the inherent limitation of chemical occurrence data (where a lack of detection does not necessarily indicate true absence), we conducted a sensitivity analysis to ensure our model fitting results are robust to scenarios where zero values actually represent trace amounts. We randomly selected 30% of the zero values for each compound and replaced them with minor random perturbations sampled from a uniform distribution $\text{U}$(0,0.1), representing trace, undetected compounds. We chose the 0 − 0.1% interval because in most GC-MS profiling, compound abundances above 0.1% are generally reported reliably, whereas the detection of trace constituents below 0.1% is highly variable and depends on instrument sensitivity, methodology, and operator thresholds. We then refitted all nine models to this perturbed global dataset to evaluate the stability of model selection.

Evolutionary dynamics

We used Bayesian analysis of macroevolutionary mixtures (BAMM) to model the complex dynamics of the evolution of core VOCs on phylogeny⁵⁹. This method is oriented towards detecting and quantifying heterogeneity in evolutionary rates. Specifically, each single rate shift event obeys $\text{Bern}\left(p\right)$ distribution with p = b_k/T, where b_k is the length of branch $k$, and $T$ is the sum of all branch lengths. Therefore, the distribution of all rate changes in a phylogenetic tree follows a binomial distribution, $\text{Bin}\left(n,p\right)$. Under this condition, considering that we sample nearly infinitely and the occurrence of multiple events on a single branch is extremely rare, the distribution of events or rate shifts is added to or removed from the phylogeny according to a compound Poisson process. The occurrence of an event v_e defines a subtree τ_e descended from v_e. The rate of phenotypic evolution is defined as $\beta ,$ and $\mathit{\omega }$ is a vector of phenotypic data. To compute the likelihood of a phylogeny with character state data at the tips, the joint likelihood of the data was given by L (τ,ω) = L (τ)L(ω). In this framework, the algorithm treated node values as nuisance parameters and analyzed marginal distributions of evolutionary rates after integrating over node values using reverse jump Markov Chain Monte Carlo (rjMCMC). In practice, we ran BAMM on a normalized data matrix of core VOCs. We set the parameters as follows: modelType = trait, expectedNumberOfShifts = 1.0, numberOfGenerations = 20,000,000, mcmcWriteFreq = 1,000, eventDataWriteFreq = 1,000, and acceptanceResetFreq = 1,000, while the remaining parameters were set to their default values. The visualization is facilitated by the R package BAMMtools⁶⁰, and the MCMC chain convergence was determined by visual inspection.

Phylogenetic covariance analysis

To assess the correlation pattern of global VOCs, we extracted the covariance matrix from the penalized PCA result. The phylogenetic covariance matrix for core VOCs is constructed based on a phylogenetic generalized linear mixed model (GLMM), which was constructed using the MCMCglmm R package⁵⁸. This methodology is specifically designed to account for phylogenetic non-independence while modeling complex trait covariation. Bayesian MCMC sampling provides reliable parameter estimates and credible intervals for complex covariance structures, and it also accommodates non-Gaussian distributions. This model is composed of two major parts: (i) a link function and (ii) a standard linear mixed model for latent variables: $\mathbf{l}=\mathbf{X}\mathit{\beta }+\mathbf{Zu}+\mathit{\epsilon }$, where $\mathbf{l}$ is a vector of latent variables whose actual value is unknown, $\mathbf{X}$ is a matrix relating fixed predictors with parameter vector $\mathit{\beta }$, $\mathbf{Z}$ is a matrix of random predictors with parameter vector $\mathbf{u}$, and $\mathit{\epsilon }$ is a vector of residuals. For the prior settings, a multivariate normal prior distribution was used by default for fixed effects, while the mean values of random effects and the residual variance structure were specified as 0.002. We assumed Gaussian distributions with identity link functions. The MCMC was run for a total of 300,000 iterations, with a burn-in phase of 60,000 and a thinning interval of 100. Diagnostics for MCMC convergence were conducted by obtaining plots for the MCMC and assessing autocorrelation.

Ancestral state estimation

The estimation of ancestral character states for core VOCs uses an extended Mk model family⁶¹. This model family describes a continuous Markov chain with k possible states and is fully parameterized by a $k\times k$ matrix, Q. In this Q-matrix, all non-diagonal elements (q_i,j for any I ≠ j) represent the instantaneous transition rates between states $i$ and $j$, while the diagonal elements are set to the negative sums of the off-diagonal elements in their respective rows, ensuring that each row of the Q matrix sums to zero. Using the fitMk function in the R package phytools^62,63, we fitted our curated dataset and pruned tree into the equal rates (ER) model, the symmetric transition (SYM) model, and the all-rates-different (ARD) model, all of which belong to the Mk model family. The likelihood ratio test and Akaike Information Criterion (AIC) indicated the ARD model was the best-fitting model. We subsequently conducted the ARD model-based marginal reconstruction ancestral state estimation using the ancr function in the R package phytools^62,63.

Author contributions

Z.X. conceptualized the study. J.Z. collected the metabolite data. Z.X. designed and performed comparative analyses. Z.X. wrote the original draft with input from L.W. All authors contributed to the review and editing of the manuscript. X.L. supervised the data collection. L.W. supervised the evolutionary analysis.

Peer review

Peer review information

Nature Communications thanks the anonymous reviewers for their contribution to the peer review of this work. A peer review file is available.

Funding

This work is supported by the Early Career Scheme from the Research Grants Council of the Hong Kong Special Administrative Region, China (Project Reference Number: 26104824), and CORE. CORE is a joint research center for ocean research between Laoshan Laboratory and HKUST.

Data availability

All datasets analyzed in this study are also publicly available in the GitHub repository: https://github.com/Planula-Ziyue/Gymnosperm_Chemodiversity⁶⁴. The repository contains the literature list, global and core VOC trait matrices, and the phylogenetic tree used for phylogenetic comparative analyses. These resources provide the underlying data used for all analyses and figures reported in this study.

Code availability

Code necessary to rerun the analysis are provided in the GitHub repository⁶⁴: https://github.com/Planula-Ziyue/Gymnosperm_Chemodiversity

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41467-026-77351-0.