Genetic basis of temporal variation in phenology in a woody perennial plant Authors: Yang Liu Published: 2026 ABSTRACT AbstractThe form of phenotypic plasticity in traits that are repeatedly and potentially reversibly expressed multiple times throughout an individual’s lifetime is labile plasticity. Yet, little is known about selection on, and genetic variation and inheritance in, such plasticity. Herein I embedded dynamic growth phenology traits into empirical understanding of the variability, inheritance, and predictability of their labile plasticity using a Populus common-garden experiment. I found no single nucleotide polymorphisms (SNPs) peak associated at P < 1e−4 with labile plasticity in phenological traits, overlapping in its two populations and no such plasticity-associated SNPs identified across study traits in single populations. Higher plasticity and more plasticity SNPs were identified in the population from climatically harsher conditions. While 6% of phenology peak-associated SNPs underwent adaptive selection, 95% of the identified plasticity SNPs were also peak associated with a given phenology, but none of them were found under selection. Plasticity SNP-containing genes highlighted environmental responsiveness functions for phenology plasticity. Projections for labile plasticity in phenology consistently showed that plasticity or trait peak-associated SNPs generated a higher predictive accuracy than non-associative counterparts. Altogether, the study elucidates the idiosyncratic characteristics of labile plasticity in growth phenology and deepens our understanding of its genetic modulation and evolution. Subject terms: Computational biology and bioinformatics, Evolution, Genetics, Plant sciences Introduction As our global climate continues to change with an estimated temperature increase of up to 5 °C by the end of this century1, there is an urgent need to increase understanding of how organisms will respond to rapid climate change2,3. The capacity to persist and thrive across varying environmental conditions requires adaptation, either through allele frequency change4 or phenotypic plasticity5. Phenotypic plasticity—the property of single genotypes to produce environment-induced phenotypic variation without mutation—is ubiquitous in natural populations, providing organisms with the ability to rapidly adjust morphology, physiology, metabolism, or genetic transcripts to reduce the probability of extinction in changing environments6,7. Decades of theoretical and empirical work on the quantitative genetics of plasticity have revealed that phenotypic plasticity is heritable8 and that heritable variation in plasticity can respond to natural selection9. Phenotypic plasticity evolves when genotypes of a population vary in their response to environmental variation10,11; such varying environments comprise a strong resultant selective pressure driving plasticity evolution (e.g., shade tolerance vs. avoidance12). Plasticity could facilitate adaptation to novel environments and population divergence in different environments13. Meanwhile, plasticity could impede adaptation by changing the distribution of phenotypes in the population, thereby shielding them from natural selection14. Whether plasticity plays a facilitative or hindering role in adaptive evolution depends on the properties of environmental change such as rate of change, variance, and autocorrelation15. If adaptation relies on change in the frequency of alleles that increase fitness through adaptive plastic responses, then genetic variation in plasticity would be critical for population persistence under changing environments11,16,17, providing opportunities for selection to act on heritable variation18,19. Change in phenology20—altered seasonal timing of recurring life-history events—is among one of the most dramatic biological responses to climate change21–24 with a potentially cascading influence on plant activities25,26 and ecosystem-level properties27,28. Plasticity in phenology represents an adaptation to variable environments29 and is expected to undergo natural selection30,31. While spatial variation in phenology reflects genetic differences and plastic responses to environmental heterogeneity32, climate change is causing shifts in the timing of phenology and the duration of the growing season over time33–35. Global warming has given rise to earlier bud flush in the spring, leading to an elevated possibility of frost damage and loss of synchronization with pollinator activities36,37. Of note, spring phenology has advanced by 2–7 days per degree of temperature increase in the temperate zone, depending on tree species and location38,39, which is on par with global shifts in phenology at rates of 4–6 days per decade40. Spring phenology change directly affects plant growth and survival41. Relatively, autumn phenological response (e.g., leaf senescence) to global warming is inconsistent, with advanced, delayed, and unchanged trends documented42. Temperate trees strictly regulate phenology by controlling the onset of bud dormancy (i.e., buds in deep sleep status) in the winter and the cessation of dormancy followed by the initiation of growth in the spring. Precisely timed phenology represents an adaptive strategy involving tradeoffs between maximization of the growing season and frost damage avoidance43. Photoperiod and temperature are two key agents in mediating the spatial synchrony of phenology in most temperate woody plants43,44. Compared to temperatures that fluctuate diurnally, seasonally, and annually, photoperiod provides a more consistent and reliable signal of seasonality, reflecting the consequence of life-history adaptations45. When photoperiodic cue falls below a certain threshold in late summer or early autumn, tree species such as Populus cease growth and form dormant buds46. Such short day-induced Populus growth cessation is mediated by a conserved signaling pathway of photoperiod-controlled flowering time as documented in Arabidopsis47. Temporal shifts in phenology in forest trees can be viewed as longitudinal within-individual phenotypic variation. This form of plasticity takes on a reversible individual expression of alternative phenotypes when exposed to changing environments through an individual’s lifetime, termed labile phenotypic plasticity48–50. Yet, there is a limited understanding of genetic variation and inheritance in labile plasticity of growth phenology and its ensuing influence on growth activity due to altered growing seasons. Nonetheless, prior knowledge suggests that most ecologically important traits have high polygenicity due to their genetic covariance with other traits51 and pleiotropy (i.e., a single gene affects multiple traits) and that organisms could alter allelic effect size to regulate multiple biological functions and trait variation (e.g., ref. 52). A powerful method to study phenotypic plasticity is the common-garden approach wherein multiple source populations are grown in a common environment and performance is compared among populations. Here, I utilized a Populus trichocarpa common-garden experiment data to investigate the genetic underpinnings of labile plasticity in spring and autumn phenology (bud break, bud set, and leaf drop) and growth activity (growth rate). A measure of labile plasticity for all four study traits was calculated as the difference in trait values over 2 years. I sought to reveal the genetic basis of labile plasticity and the biological functions of labile plasticity genes in the populations of the species. To that end, I interrogated three questions about the variability, inheritance, and predictability of labile plasticity. I hypothesized that: H1: Higher within-year trait variation in a population, higher plasticity variability of genotypes of the population. While a high population-level trait variation represents a high genetic variation among its genotypes, more divergent genotypes could produce an elevated discrepancy in their responses to changing environments. I therefore expect that the population of a higher trait variation within a year has a higher variability of labile plasticity between years. H2: Inheritance of temporal phenotypic variation is highly aligned with that of respective phenotypic traits. The omnigenic model indicates that core genes underpin the genetic basis of phenotypic traits53,54. Similarly, core genetic variants are significantly associated with phenotypic traits across environmental conditions. As such, alleles with the strongest signatures of association (i.e., peak associated at P < 1e−4 used in this study) for labile plasticity are expected to be also highly associated with its corresponding phenotypic trait and such alleles-containing genes have functional characteristics of environmental responsiveness. In addition, trait-associated alleles are dependent on population structure (e.g., ref. 55) and thus labile plasticity alleles would differ between populations. H3: Use of more genetic variants yields a higher predictive accuracy for labile plasticity. The tenet of Fisher’s infinitesimal model claims that the phenotype is determined by an infinite number of alleles, each with a small effect size56. If phenotypic plasticity is deemed a type of trait, it should be better predicted using more genetic variants. Considering the complexities of nonlinear relationships between genetic variants and phenotypes in a multi-dimensional trait space, I used machine learning models to test this hypothesis. Machine learning modeling and its regularization techniques can mitigate issues of multicollinearity and overfitting due to a plethora of explanatory variables. Such models can automatically learn data patterns, optimize model parameters, and have great efficiency and power in capturing complex nonlinear patterns57,58. They are thus suited to making predictions with genomic data. In addition to within-population cross-validation, I examined how well the model trained on one population over 2 years can be transferred to another population or the same population using independent data. Results Magnitude of within-year phenotypic variation and between-year plasticity variability In total, 362 individuals were assigned to two genetic clusters, termed North and South populations thereafter (Figs. 1A and S3), which had 129 and 233 individuals, respectively. This population stratification, based on a k-means algorithm, was congruent with that previously inferred from a Bayesian or Maximum Likelihood algorithm59,60. The two populations originated from habitats with significantly different climates in terms of annual mean temperature and precipitation (t-test P < 0.0001; Figs. 1A and S2B). Their dominant climate types were mediterranean continental and subarctic interspersed with tundra for the North, and oceanic and humid continental for the South (Fig. S2A). This suggests that the climate of the North is likely more heterogenous than the South with regard to the intensity of climatic variation. For trait measurements in the year 2010 relative to 2009, the common garden had fewer frost-free days by 73 days and a higher mean temperature by 1.88 °C prior to the growing season commencement, leading to earlier bud break by 37 days in 2010 (Fig. 1B). In addition, the common garden seemed significantly more arid than both populations’ places of origin (P < 0.05)61, although a drip irrigation system was applied in 2009. Timing, density distribution, and plasticity pattern of population-level phenological events. A There were 129 and 233 individuals from 15 and 12 drainages sampled that were classified into two populations, North and South, respectively, based on the outcome of population structure analysis (Fig. S3). Each filled circle is a drainage. The blue and red color code is used for North and South populations, respectively, throughout the study. The background color of the map represents the annual mean temperature over the period 1970-2000, extracted from WorldClim v. 2.1. Note that annual mean temperature was identified as a vital evolutionary force for the species’ population divergence (ref. Liu et al.61). B Plot of phenological events and mean temperature of the common garden over the study period. The blue and red colors indicate the year 2009 and 2010, respectively. The dashed lines at the bottom indicate the number of frost-free days available for growth based on minimum temperature ≥0 °C (Fig. S2C). The green and yellow bands denote the period of bud break and leaf drop, respectively. The horizontal solid lines above the curves indicate these periods in each of the 2 years. C Variation and distribution of phenotypic traits by population-and-year combinations. D Phenotypic plasticity for study traits over 2 years. Crossing of lines indicates that the performance of a genotype relies on the environment, thus reflecting phenotypic plasticity. Population-level average and individual plasticity are shown in dark and light colors. The common garden experiment revealed that natural selection drove the phenotypic divergence of canopy phenology (i.e., bud break, bud set, and leaf drop) and stem growth activity (i.e., growth rate) based on QST » FST, as evidence of local adaptation61. All phenology and growth traits significantly varied between years (Year P < 0.05; Table 1), indicating significant provenance-specific labile plasticity of these traits. Moreover, there were significant G × E interactions across traits (Prov × Year P < 0.05; Table 1). This indicates a significant genetic variation in labile plasticity of the traits, satisfying a necessary but not sufficient condition for plasticity to evolve. Only bud break and leaf drop were significantly different between provenances (Prov P < 0.05; Table 1); except bud break, all the other traits had a significant difference between populations (P < 0.0001), based on a two-way ANOVA for testing population, year, and their interaction effects. In addition, differences in leaf drop and growth rate were significant between genotypes (Genotype P < 0.05; Table 1). Compared to the South, the North had an earlier bud set and leaf drop and a lower growth rate (Fig. 1C). Furthermore, both populations had a significant shift in trait means between years (ΔNorth −0.05 to 0.04 and ΔSouth −0.02 to 0.03 for all standardized traits; paired, two-tailed t-test P < 0.05), with smaller values in the North in single years (one-tailed t-test P < 0.0001) for all traits except for bud break. With an exception to bud break in 2009, all traits in the North had a higher variation in single years than in the South (ΔσNorth 0.72-1.61 and ΔσSouth 0.38-1.27; one-tailed t-test P < 0.0001; Fig. 1C). Consistently, the magnitude of genotype-specific labile plasticity in bud set and growth rate was significantly higher in the North than in the South (absolute Δtrait-North 0.59–0.88 and Δtrait-South 0.33–0.59; one-tailed t-test P < 0.0001; Fig. 1D). Table. 1Variance partitioning of growth phenology and activity traits using REML-linear mixed models for the comparison of trait responses across 125 provenances over 2 years, three environmental covariates, and G × E interaction effectsBud breakBud setLeaf dropGrowth ratedfdf.resF/χ2Pdf.resF/χ2Pdf.resF/χ2Pdf.resF/χ2PProvenance (Prov)1243691.550.0033731.190.2143632.17<10−43500.710.987Year12395.030.02524413.240.00124111.620.00223410.310.003Prov × Year (G × E)1242411.82<10−42443.22<10−42412.12<10−42351.520.003PC1forcing14411.660.199445490.78<10−4470476.88<10−446374.40<10−4PC1photoperiod14654.130.1704710.050.8254503.110.2384241.190.550PC1chilling146216.17<10−44671.770.18446937.55<10−44535.190.045Genotype (G-matrix)10.570.8980.400.89812.660.00122.97<10−4Model fit statisticsR2m0.5720.9390.9540.700R2c0.5970.9420.9660.811R20.7060.9630.9820.899AIC4427.393957.073181.82422.83F-values for fixed effects and χ2-values for random effects are reported. Significant P-values after the sequential Bonferroni correction (α = 0.05) are shown in bold. Degrees of freedom for residuals (df.res) were calculated using Satterthwaite’s approximation. Marginal R2 (R2m) is a proportion of the variance explained by the fixed effects; conditional R2 (R2c) is a proportion of the variance explained by both fixed and random effects; R2 was calculated as the correlation between model fitted and observed values. Variance partitioning of growth phenology and activity traits using REML-linear mixed models for the comparison of trait responses across 125 provenances over 2 years, three environmental covariates, and G × E interaction effects F-values for fixed effects and χ2-values for random effects are reported. Significant P-values after the sequential Bonferroni correction (α = 0.05) are shown in bold. Degrees of freedom for residuals (df.res) were calculated using Satterthwaite’s approximation. Marginal R2 (R2m) is a proportion of the variance explained by the fixed effects; conditional R2 (R2c) is a proportion of the variance explained by both fixed and random effects; R2 was calculated as the correlation between model fitted and observed values. Labile plasticity SNPs were highly population- and trait-dependent I identified only one intergenic SNP significantly associated with leaf drop labile plasticity in the North (Chr04:22968318_T/C P < 1e−6) after Bonferroni corrections or a genomic inflation control. I thus used a suggestive threshold at α = 1e−4 to identify potentially associated SNPs, termed peak-associated SNPs (paSNPs hereafter). I interpreted signals above the threshold, while those below it were treated as exploratory signals. While an average of 155 SNPs were peak associated with labile plasticity (i.e., plasticity SNPs) at P < 1e−4 in the North across traits, only about 22 SNPs had a peak-association in the South (Fig. 2A). The North had more plasticity paSNPs than the South across traits (right-tailed t-test P < 0.05; Fig. 2A). Consistently, SNP heritability was higher in the North (h2: 0.92 ± 0.01) than in the South (h2: 0.38 ± 0.04) across the plasticities of study traits (Wilcoxon test P < 0.05; Fig. 2B). Across populations, there were few plasticity SNPs that overlapped between traits; only 5 paSNPs were identified in bud set and growth rate, and none of them were identified in more than two traits (Fig. 2C). More than 95% of plasticity SNPs were also found to be peak associated with respective phenotypic traits in single or both populations combined (Fig. 2D). There was no pronounced shift in allele frequencies towards rare or common alleles for plasticity paSNPs (Fig. 2E). Integration of multi-trait GWAS for study traits and their labile plasticity. A The Manhattan plots depict -log10(P-values) from a genome-wide association scan for labile plasticity against SNP positions on each of the 19 chromosomes. Results of the North and South populations are shown above and below the horizontal axis, respectively. The horizontal blue or red lines indicate a suggestive, peak association threshold at α = 1e−4 for the North and South populations, respectively. A summary of plasticity SNPs in North and South is given in Venn diagrams. Trait paSNPs of the North and South populations are highlighted in blue and red, respectively. B SNP heritability for study traits in single traits across years. C The overlap between trait-based plasticity SNPs of either population. D Components of plasticity peak-associated SNPs in trait peak-associated counterparts identified by single and/or two populations combined approaches. E Allele frequency spectrum for plasticity paSNPs in single populations. No significantly trait-associated SNPs overlapped between study traits or across populations SNP heritability had differentiation between traits over years (Fig. S5). There was a higher heritability in the North (point estimate h2: 0.99) than in the South (h2: 0.61 ± 0.15) across traits in 2010 (Fig. S5). The two autumn phenology traits (i.e., bud set and leaf drop) had a higher heritability in the North (h2: 0.99) than in the South (h2: 0.64 ± 0.12) across years (Fig. S5). Bud break h2 in 2009 was small with a high variation in both populations (Fig. S5), indicating strong environmental influences on the trait in the year. A total of nine, two, 71 and, 12 SNPs were significantly associated (Bonferroni corrected) with bud break in the North, bud set for both populations combined, leaf drop for either population or both populations combined, and growth rate in the North, respectively (Fig. S6). None of these significantly associated SNPs overlapped between traits or single populations. Similarly to the abundance pattern of plasticity paSNPs in populations, the number of trait paSNPs in the North was five to 16 times as many as in the South, depending on traits (Fig. S7A). None of trait paSNPs identified by single populations overlapped, whereas trait paSNPs had an overlapping between a single population and both populations combined (Fig. S7A). In the four study traits, both populations combined generated an average of 148 distinct paSNPs, which were not identified in single populations (Fig. S7A). In addition, there were paSNPs identified in more than two traits (Fig. S7B), including six SNPs peak associated with autumn phenology and growth activity (Fig. S7B). Adaptive selection and biological functions for labile plasticity and traits associated genes While the hapFLK approach identified 114 SNPs under selection (Fig. S4), a total of 713 putative plasticity SNPs were identified in at least one study trait in either population (Fig. 2). However, none of these plasticity SNPs overlapped with those under selection (Fig. 3A top panel). By contrast, about 6% of paSNPs with autumn phenology and growth activity traits underwent adaptive selection (Fig. 3A bottom panel). Moreover, there was a significant negative relationship between selective strength and plasticity peak-association signal (slope P < 0.05; Fig. 3A); however, there were trait peak-associated SNPs under weak and strong adaptive selection (positive quadratic term P < 0.05; Fig. 3A). Co-localization of paSNPs and their related genes governing growth phenology and activity with adaptive selection on a chromosome 15 region. A Association strength relationships between trait or plasticity paSNPs and SNPs under selection. Associative strength denotes the signal of SNP-plasticity (top) or SNP-trait (bottom) associations using -log10(P-value) from GEMMA. The vertical dashed line indicates a cutoff for suggestive selection from hapFLK at P < 1e−4. B GWAS for growth traits highlighting SNPs exhibiting significant signals for all phenology traits within a 100 kilobase region on chromosome 15 and selection scan for this genomic region. The last panel displayed a genomic scan for adaptive selection within the region using a hapFLK (positive) or FLK (negative axis) approach, and orange circles labeled trait paSNPs. The dashed horizontal lines indicate the P-value threshold. A total of 12 trait paSNPs were identified in the region which are within four genes including their regulatory regions marked in light green rectangles with gene IDs given in the top panel. The relative length of the coding sequences of the four genes was demarcated in green boxes on the top of the first panel. Information about these SNPs and genes is given in Table 3. Refer to Figs. S4 and S6 for the full results of selection scan and trait-based GWAS. C The allele frequency of six trait peak-associated SNPs for the North and South populations. They were also identified under selection based on hapFLK and FLK. D The allele frequency of 12 labile plasticity peak associated SNPs in either North or South population identified on this chromosome. Notably, 12 of the 19 paSNPs with autumn phenological traits were within a genomic region of 100 kb of chromosome 15, spanning four genes and their regulatory regions (Fig. 3B). Moreover, 13 of them were identified under adaptive selection using both hapFLK and FLK methods (Fig. 3B). I particularly highlighted the trait paSNPs identified within one annotated (Potri.015G142000 with 2 paSNPs) and one unannotated (Potri.015G141800 with 7 paSNPs) genes were subject to adaptive selection (Fig. 3B). Within these two genes and their regulatory regions, a total of 47 (10) and 100 (25) SNPs were identified with population divergence based on hapFLK (or FLK), respectively (Fig. 3B). By examining the pattern of genetic variation at each of these paSNP loci by gene in each population, I found a large shift in the allele frequency of three paSNPs within the gene, Potri.015G141800, with adaptive haplotypes or deletions rising to a high frequency of 0.8 and above in the South (Panels 1, 3, and 4 in Fig. 3C). For the rest of paSNPs from the other three genes, there was a high-to-modest change in allele frequency ranging between 0.82 and 0.62 in the North (Panels 2, 5, and 6 in Fig. 3C). On the chromosome, 12 plasticity SNPs within five genes were identified (Table S2). Except that two of the SNPs related to four genes showed a marginal difference in allele frequency between populations (Δp < 0.1; Panels 4 and 6 in column 2 of Fig. 3D), the other 10 plasticity SNPs had allele frequency differences of 0.5 between populations (Fig. 3D). Across chromosomes, there were 69 phenological labile plasticity gene candidates with known biological functions, in which 61 were identified from the North and 25 were from leaf drop trait (Table 2 and S2). The functions of these plasticity genes were involved in flowering, reproduction, senescence, biotic (e.g., NBS-LRR), abiotic (e.g., oxidation/reduction reaction, dehydration, and phytochrome), and hormones (e.g., ethylene) responsiveness (Table 2). Nonetheless, a functional enrichment for these genes indicates they had no propensity for a specific ontology domain (all FDR > 0.05) but might be inclined to be cellular components (Fisher’s exact test P < 0.05). Table. 2Information about putative labile plasticity genes with annotated biological functionsGene ID (v.3.0)Functional annotationIdentified popltn● Bud breakPotri.001G060200Peptidyl-prolyl cis/trans isomeraseNorthPotri.002G035000MYB family transcription factor family proteinNorthPotri.002G159900CRYPTIC PRECOCIOUS family protein (related to flowering regulation)NorthPotri.002G160800Cysteine synthaseNorthPotri.003G192200Cysteine proteinase inhibitorNorthPotri.004G193700Kinesin motor family proteinNorthPotri.011G120300Laccase (related to oxidation reactions)NorthPotri.011G132400Basic helix-loop-helix (HLH) family proteinNorthPotri.011G132600HexosyltransferaseNorthPotri.012G119000Zinc finger family proteinNorthPotri.015G083800Leucine-rich repeat (LRR) family proteinNorthPotri.016G057300GlycosyltransferaseNorthPotri.017G039200AminopeptidaseNorthPotri.018G084500Casein kinase family proteinNorthPotri.018G084700Methyltransferases superfamily proteinNorth● Bud setPotri.001G083200EndoglucanaseSouthPotri.002G200300Cellulose synthase family proteinNorthPotri.004G092900Histone deacetylaseNorthPotri.005G026400RecQ helicase SIMNorthPotri.005G047700Glyoxal oxidase-related family proteinNorthPotri.005G241300DEA(D/H)-box RNA helicase family proteinNorthPotri.008G107900Calmodulin-binding family proteinNorthPotri.009G060800Cellulose synthaseSouthPotri.009G102700LaccaseNorthPotri.011G064500Proline-rich family proteinSouthPotri.013G013500Epoxide hydrolase family proteinNorthPotri.017G016300DUF679 domain membrane protein 7 (related to reproductive development)NorthPotri.019G073200Short-chain dehydrogenase/reductase family proteinNorth● Leaf dropPotri.001G008800histone mono-ubiquitination 2NorthPotri.001G085500Kinase family proteinNorthPotri.001G048800Uncharacterized protein (fragment)SouthPotri.001G215500Chloroplast outer membrane family proteinSouthPotri.001G259100Aldehyde dehydrogenaseNorthPotri.001G357900Damaged DNA- binding protein 1 B (related to DNA repair)NorthPotri.002G004600RNA recognition motif-containing family proteinNorthPotri.002G090700Transcription factor TGA7 family proteinSouthPotri.003G197700SNF2 domain-containing family proteinNorthPotri.004G138600VHS domain-containing family proteinNorthPotri.004G166600Pentatricopeptide repeat (PPR) superfamily proteinNorthPotri.006G149000CRS1/YhbY (CRM) domain-containing proteinNorthPotri.006G194900Kinesin motor family proteinNorthPotri.008G176900Leucine-rich repeat (LRR) family proteinNorthPotri.010G011700Senescence-associated family proteinNorthPotri.010G012900Cyclophilin-like peptidyl-prolyl cis/trans isomerase family proteinNorthPotri.010G220200CarboxypeptidaseNorthPotri.011G024800Pre-mRNA cleavage complex-related family proteinNorthPotri.012G048900LaccaseSouthPotri.012G051300T-complex protein 1 subunit deltaNorthPotri.013G102400GDSL-like Lipase/Acylhydrolase superfamily protein (hydrolytic enzyme)NorthPotri.014G142900Vesicle-mediated transportNorthPotri.014G143100Protein kinase superfamily proteinNorthPotri.014G143600Glycosyl hydrolase family 38 proteinNorthPotri.018G076400Heavy metal transport/detoxification superfamily proteinNorth● Growth ratePotri.001G065600CTD phosphatase-like protein 3NorthPotri.001G261200CarboxypeptidaseNorthPotri.001G410600Ethylene-responsive family proteinSouthPotri.003G097600bZIP transcription factor 3 family proteinNorthPotri.003G155000A subunit ribosomal protein S33NorthPotri.005G094500Cytochrome P450 family proteinNorthPotri.005G114300FAT domain-containing family proteinNorthPotri.005G211500Putative senescence-associated protein SAG102NorthPotri.T155200ER lumen protein retaining receptorNorthPotri.006G253900Phospholipase DNorthPotri.010G180000HXXXD-type acyl-transferase family protein (related to hypersensitivity)NorthPotri.013G022900Hexosyltransferase (Fragment)NorthPotri.014G094100GDSL-like Lipase/Acylhydrolase superfamily proteinNorthPotri.015G088700Ribosomal RNA processing Brix domain proteinNorthPotri.015G088900DNAse I-like superfamily proteinNorthPotri.016G053600Protein detoxificationNorthAll these genes were identified based on SNPs above a peak association threshold at α = 1e−4. Information about putative labile plasticity genes with annotated biological functions All these genes were identified based on SNPs above a peak association threshold at α = 1e−4. Through the 94 significantly trait-associated SNPs (Fig. S6), I identified 15 genes with known biological functions (Table 3). Three of six trait-associated genes on chromosome 15 (Table 3) were functionally annotated—copper-binding family protein, rhomboid-like family protein, and pectinesterase—with putative functions in plant growth and development, hormone response, and response to biotic and abiotic stresses. Furthermore, 595 paSNPs with at least one study trait were identified and an enrichment analysis for these SNPs-related genes at P < 0.05, albeit not FDR < 0.05, showed their involvement in alcohol/carbohydrate metabolic or catabolic processes. Table. 3Information about genes containing significantly associated SNPs with study traitsSNP ID (v.2.0)aLocation rel. to geneGene ID (v.3.0)Functional annotationChr01:5201679_C/TwithinPotri.001G030600Glycosyltransferase family proteinChr04:13632425_A/GupstreamPotri.004G129800Pyruvate Dehydrogenase E1 beta family proteinChr05:1998693_T/GwithinPotri.005G030600Uncharacterized protein (Fragment)Chr05:14486371_C/GupstreamPotri.005G162900Serine/threonine-protein phosphatase 2 A activatorChr05:19656168_C/AwithinPotri.005G183700PHD finger family proteinChr08:4929891_T/CwithinPotri.008G080300Enhancer of AG-4 1 (HUA1) family protein (related to flowering)Chr09:12778305_T/CwithinPotri.009G168200phosphate transporterChr09:12778468_T/CwithinPotri.009G168200phosphate transporterChr09:12778174_T/CwithinPotri.009G168200phosphate transporterChr14:11516440_A/TwithinPotri.014G156400LysM domain-containing receptor-like kinase 4 family proteinChr14:16431671_T/GdownstreamPotri.014G191400Zinc finger family proteinChr14:8338150_A/GwithinPotri.014G116900Golgi apparatus membrane protein TVP23Chr15:376392_G/CdownstreamPotri.015G003900Copper-binding family proteinChr15:378603_A/GwithinPotri.015G004000-Chr15:379601_G/AwithinPotri.015G004000-Chr15:382584_G/CwithinPotri.015G004100-Chr15:13322026_T/AwithinPotri.015G142000Rhomboid-like proteinChr15:13324388_G/AwithinPotri.015G142000Rhomboid-like proteinChr15:13336485_T/AwithinPotri.015G141800-Chr15:13336494_G/CwithinPotri.015G141800-Chr15:13336519_T/AwithinPotri.015G141800-Chr15:13336546_A/GwithinPotri.015G141800-Chr15:13337837_G/AwithinPotri.015G141800-Chr15:13337893_A/TwithinPotri.015G141800-Chr15:14220826_G/CupstreamPotri.015G127800PectinesteraseChr16:5469074_G/AwithinPotri.016G075300Casparian strip membrane protein 3Chr16:5469257_G/AdownstreamPotri.016G075300Casparian strip membrane protein 3Chr16:5470483_G/TdownstreamPotri.016G075300Casparian strip membrane protein 3Chr16:5470578_A/TwithinPotri.016G075300Casparian strip membrane protein 3Chr16:5470639_C/TdownstreamPotri.016G075300Casparian strip membrane protein 3Chr16:5470570_G/AdownstreamPotri.016G075300Casparian strip membrane protein 3Chr16:7123500_A/TwithinPotri.016G090400Phototropic-responsive NPH3 family proteinSNP IDs in bold are for SNPs peak associated with three of the four study traits at α = 1e−4. Other than those on the highlighted chromosome 15 in Fig. 3, only significantly associated SNPs with known linked gene’s functions were displayed.aSNP ID consists of information about chromosome number, position, and reference/alternative alleles. All these SNPs were significantly associated with at least one study trait after Bonferroni corrections. Information about genes containing significantly associated SNPs with study traits SNP IDs in bold are for SNPs peak associated with three of the four study traits at α = 1e−4. Other than those on the highlighted chromosome 15 in Fig. 3, only significantly associated SNPs with known linked gene’s functions were displayed. aSNP ID consists of information about chromosome number, position, and reference/alternative alleles. All these SNPs were significantly associated with at least one study trait after Bonferroni corrections. Plasticity SNPs had a minor-to-moderate change in effect size over time and best predicted plasticity Based on the distribution of allelic effect sizes for each trait by year (Fig. S8), all SNPs excluding the paSNPs identified were discretized into major/minor effect size sets with high/low β values, respectively. While these major/minor effect size SNPs were split at the threshold of µ ± 0.5 σ (mean ± half standard deviation), the number of SNPs within each set differed (Fig. 4A), indicating that the number of high/low effect size SNPs varies with traits and populations. SNPs with absolute, high effect size over the 2 years in the North and South populations were most abundant in bud break, with 16-37% more in number than their second highest trait (Fig. S9). More than 90% of plasticity SNPs had a within 30-fold change in β ratio [i.e., log10(β-ratio) = ±1.5] over 2 years for each trait across populations (Fig. 4B), which corresponded to about 50% quantile of the β ratio distribution, shown in Fig. S11 (range circled with a red dashed line). On average, 97% of plasticity SNPs had no more than 30-fold change in β ratio across traits and populations (Fig. 4B). Moreover, 66 ± 16% of plasticity SNPs had a consistent positive or negative effect size over 2 years (Fig. 4B). Similarly, about 96% of trait paSNPs had a fold change of less than 30 across traits (Fig. S12); 35 ± 18% of trait paSNPs had flipped effect size sign over time across traits (Fig. S12). Overall, both plasticity and trait paSNPs, labeled by dots at the y = 2.5 or 0.5 position in Fig. S10, respectively, had no bias in the distribution of effect sizes (Fig. 4C and Fig. S10). Effect size of phenotypic traits for plasticity-associated or major/minor SNPs sets. A Alluvial plot showing relative abundance in the number of SNPs of effect-size categories (major [+/-] or minor [+/-]) for each trait by population. Cutoff for major or minor is based on an absolute effect-size value of 3e−6 (i.e., μ ± 0.5σ). B Negative logarithm of allelic effect size ratio for plasticity SNPs from GEMMA between years in each population by trait. C Effect size means with 50% and 95% CI bands for SNP sets. Refer to Fig. S10 for the effect size distribution of all SNPs and the effect size values of plasticity SNPs and trait paSNPs for each of the study traits by population. Further, the set of labile plasticity SNPs (Set 1) consistently had the best prediction for plasticity in all study traits of either population with the interquartile range (IQR) of the point R2 values of 0.33–0.61 (Fig. 5). Meanwhile, the set of trait paSNPs for single or both populations combined (Set 2 and 3) also predicted plasticity better (IQR R2: 0.13–0.41) than non-associative SNPs sets and randomly selected subsets (Set 4 and 5; IQR R2: −0.05 to 0.13; Fig. 5). Using any paSNPs set (i.e., Set 1, 2, or 3) across the trait and population combinations, bagging and adaptive boosting achieved a better prediction performance (IQR R2: 0.27–0.57) than random forest and gradient boosting (IQR R2: 0.20–0.44; Fig. 5). Overall, model performance for within-population evaluations was better than cross-population or independent data evaluations (Fig. 5) due in part to distinct peak-associated SNPs identified in different populations. In addition, a regularization hyperparameter, α (RidgeCV), in linear models considerably affected labile plasticity predictive accuracy, particularly for paSNPs sets (Set 1–3; Fig. S13). Results of performance evaluation for the five machine-learning models used for predicting labile plasticity in phenology and growth activity by population. The model selection results for all five SNP sets and two randomly selected SNP subsets were based on five predictive machine-learning models using a 10-fold CV on the training data (80%). Points are average values with error bars denoting s.d. over the CV folds. The asterisks (*) showcase the estimates of model performance using the held-out test data (20%) for the same population over the same period as the training data. Model performance estimates shown in green crosses (×) are based on cross-population prediction; the blue plus signs (+) are for evaluations using independent data. No independent data were available for growth rate. For visualization convenience, all R2 and/or RMSE estimates (s.d.) exceeding the x-axis range were masked, although R2 > 1 or negative R² values were treated appropriately in machine-learning models. Discussion In cold regions, woody plants alternate phases of intensive growth activity and reproduction with dormancy to survive under unfavorable conditions. Synchronizing phenological processes represents a regulation of growth phase transitions and is paramount to preventing sensitive growing tissues from frost damage. While numerous studies have investigated phenotypic plasticity over space, scant studies have revealed temporal variability and genetic variation in labile plasticity and elucidated how natural selection affects the evolution of phenotypic plasticity62–64. This study unified genomic association and allelic effect size to elucidate the genetic basis of labile plasticity in response to environmental change in a P. trichocarpa common-garden experiment over 2 years. The study substantiates the first two hypotheses raised: (H1) populations of higher within-year trait variation (e.g., North relative to South in this study) have a higher magnitude of labile plasticity in growth phenology and activity between years and (H2) plasticity-associated SNPs are also highly associated with its corresponding phenotypic trait and such SNPs-related genes have biological functions related to environmental responsiveness. However, the study did not support the hypothesis (H3) that inclusion of more alleles improves predictive accuracy than just using peak-associated SNPs. Instead, the study revealed that plasticity or trait peak-associated SNPs had a higher predictive accuracy than non-associative counterparts across phenology and growth activity traits, in stark contrast with the best prediction outcomes for complex traits often from the infinitesimal model. The outcome of a high predictability from peak-associative SNPs could be ascribed to the attribute of distinct, strong genetic determination of labile plasticity. The North population responded to temporal environmental variation more strongly than the South, as manifested by a higher trait variance and stronger labile plasticity for the North. The North is at a higher latitude than the South and in general, higher latitude regions get warmer faster than lower latitudes as climate changes1. This finding indicates that temperature rise may lead to greater magnitudes of phenological variation within and between years in the North than in the South, possibly being an important strategy to mitigate adverse influences on high-latitude populations due to climate change. As exposure to an accumulative amount of cold temperature in the winter (chilling) is needed for establishing competency to respond to temperature rise in the spring (i.e., accumulation of thermal time or growing degree days; forcing), global warming may prolong the time required for winter chilling and spring forcing, leading to a decline in bud break advancement65,66. However, if chilling requirements can be met during periods of cold autumn, then a warm winter may advance the period of awakening and stimulate premature bud flush. At higher latitudes or in cooler regions, populations may be more responsive to warming possibly because a slight warming still allows chilling requirements to be met in cooler climates, where even a decrease in chilling days due to warming winter temperatures would be sufficient to break winter dormancy67, leading to an extended growing season. Nonetheless, if the growing season increases, the risk of growing season frost damage may also augment due to earlier budburst in the season extending the period in which plants are susceptible to frost68–70. Altogether, warming likely has a greater influence on the phenology of the North relative to the South population. Prior to bud dormancy, trees have to accumulate sufficient carbohydrates (e.g., sugar and starch) to withstand winter coldness and reactivate growth in the spring71,72. Previous studies indicate that a deferred bud break in many trees is associated with low carbohydrate availability73–75. Increased carbon assimilation may directly result in advanced autumn phenology76 and then foster earlier spring phenology77,78. Consistently, this study highlighted the importance of carbohydrate and energy metabolism (e.g., carbon fixation, photosynthesis, and oxidative phosphorylation) to growth phenology plasticity (e.g., genes with relevant functions in Table 2). On the other hand, phytohormones (e.g., auxins, abscisic acid, and ethylene) have been reported to play an important role in seasonal growth of perennial plants79. This study identified an ethylene responsive gene with relevance to growth phenology labile plasticity. Furthermore, environmental cues are thought to be important for initiating phenological processes. The study identified plasticity genes involved in environmental responses, such as oxidation/reduction reaction, dehydration, and phytochrome. Besides, the biotic responsive gene, NBS-LRR, was found to be associated with phenology plasticity. A flowering regulator and a circadian clock control gene related to senescence were identified with relevance to phenological labile plasticity. Nonetheless, the study did not pinpoint the relevance of short-day induced growth cessation related genes to growth phenology plasticity. Notably, FLOWERING LOCUS T (FT) paralogs regulate different phenology traits during the annual growth cycle80; particularly, FT2 is associated with the timing of bud set81,82 and its paralog FT1 is involved in the release of winter dormancy80. Adaptive plasticity is a product of genetic evolution83. The finding of a high overlapping between plasticity and trait paSNPs agrees with positive genetic correlations between phenotypic traits and plasticity documented in the literature (e.g., ref. 6). This implies that plasticity in a trait might be a by-product of selection on the trait but does not necessarily exclude the possibility of selection, drift or other evolutionary mechanisms acting exclusively on its plasticity. The finding of labile plasticity SNPs lacking a strong signal of positive selection suggests that relaxed selection could constrain plasticity evolution84,85. Meanwhile, increase in plasticity may impose more constraints on adaptive evolution by masking genetic variation from natural selection86,87. The question of whether plasticity has evolved could be interpreted as the difference between the average fitness across environments. Natural selection operates on phenotypes that are expressed and cannot identify whether an organism can adapt to more environments unless individuals encounter multiple environments in a lifetime. Experiencing one environment in a life span does not allow for selection on plasticity, as natural selection disfavors plastic phenotypes in each short-term environment. Only in the long-term does adaptive plasticity maximize fitness. When selection disfavors plasticity, plasticity alleles are transmitted from one generation to another, and their fixation or loss will hinge on their fitness across the environments they have experienced5,16. As such, natural selection might differentiate between plastic and non-plastic alleles. On the other hand, environmental heterogeneity is a primary driver for the evolution of phenotypic plasticity88 and greater heterogeneity selects for higher plasticity89. This notion is in line with the finding of higher labile plasticity in the North inhabiting a region with more heterogenous landscapes and higher climatic variation intensities than the South. Furthermore, phenotypic plasticity evolves likely in relation to the variability and predictability of the environment, with a greater plasticity for populations inhabiting more predictably variable environments16 with reliable environmental cues90. It would be interesting to examine the relationships between climate predictability and plasticity for study populations in the future. Phenotypic plasticity in traits could evolve via alleles of major or antagonistic (i.e., flipped sign over time) effects. In contrast to no plasticity SNPs identified under selection in this study, selection scan identified about 6% of the trait paSNPs that underwent adaptive selection, indicating adaptive relevance in phenology and evidencing the genetic underpinnings of the evolution of phenological traits. As a result, some phenology-associated SNPs had a large shift in allele frequency with adaptive haplotypes rising to higher than 80% in the North or South population. Meanwhile, some plasticity paSNPs had an allele frequency differentiation of >0.5 between populations, indicating that population demography or other evolutionary mechanisms might have driven shifts in allele frequencies for SNPs related to phenological labile plasticity. Different paSNPs were identified from the two study populations. Assuming that an ancestral population adapted to its home environment is displaced to a foreign environment, forming two divergent populations, natural selection would increase the frequency of certain alleles until the new population can gain the highest fitness in the new environment91. Consequently, such a process of polygenic adaptation leads to subtle allele frequency shifts at many loci91–93. This indicates that changes in polygenetic traits are likely mediated by different loci in different populations94; similarly, plasticity-associated SNPs could vary with populations. The characteristics of plasticity SNPs seldomly overlapped between traits might be due to genetic tradeoffs or conditional neutrality driving the evolution of plasticity. An allele (or gene) of antagonistic pleiotropy could engender variable fitness under different environmental conditions95, indicating a possible constraint to evolutionary responses. In phenotypic selection, the optimal phenotype under one environmental condition is inferior to the optimum in a different condition, making it difficult for selection to maximize fitness across environments95. Alternatively, a trait value is favored only in one environment but unrelated to fitness in another, that is, conditional neutrality96. There is growing evidence showing that conditional neutrality is more common than antagonistic pleiotropy52,96–98. The study found that phenology labile plasticity SNPs had major or antagonistic effects, providing genetic potential to alter phenotypic traits under changing environmental conditions. Meanwhile, few labile plasticity SNPs had a substantially variable effect size over time (e.g., Δ >30×). No polygenic SNPs identified for plasticity in growth phenology and activity might represent a bet-hedge strategy at the population or trait level. Plasticity or trait paSNPs yielded the highest prediction for plasticity, in stark contrast with complex traits having genetic architecture determined by many loci distributed across the genome, each with a small effect size—a concept rooted in Fisher’s infinitesimal model56. In addition, as opposed to plasticity SNPs, trait paSNPs with major and/or antagonistic effects were highly superposed across growth phenology and activity traits in both populations. Theory points to the contribution of many small effect loci and a few large effect loci to phenotypic traits and prompts that large-effect loci should be favored, especially when a population is far away from its optimum99. Analysis of empirical QTL data also revealed a general architecture of a few large effect size loci with many small-to-intermediate effect loci100,101. Nonetheless, SNPs of absolute large effects are rare because they mostly have deleterious effects and are thus selected against102. Relatively, labile plasticity in bud break had more high effect size SNPs by >16% than in the other study traits at a β cutoff of 3e−6 (i.e., μ ± 0.5σ) across populations. This underpins the genetic discrepancy between spring and autumn phenology. The genetic basis of plasticity could be reflected by overdominance, stating that plasticity is negatively correlated with the number of heterozygous loci103, that is, the more heterozygous a genotype is, the more plastic it is. In addition, the finding of only a few trait paSNPs overlapped between canopy greening and stem growth rate traits agrees with a loose coupling of leaf and stem phenology due to differential environmental cues stimulating their resumption104,105. Further, previous ecological studies have demonstrated that spring phenology (e.g., bud break) had a more consistent trend of shifting towards earlier spring onset than autumn phenology (e.g., bud set and leaf drop), as summarized in the Introduction. Possibly because advanced and deferred autumn phenologies represent two ecological strategies for forest trees to store resources. While early leaf drop allows for transferring nutrients to the next year in advance to reduce energy waste on leaves in the fall or winter, late leaf drop extends photosynthesis, leading to an elevated accumulation of carbon reserves available for the next season106. Moreover, the timing of warm temperatures could have an opposing effect on autumn phenologies. For example, warming at night leads to a delay in leaf senescence, whereas daytime warming advances it107. This does not mean that there is a higher genetic basis for spring phenology than autumn phenology. This study demonstrated that there was a highly variable h2 for bud break but similar high h2 and H2 for bud set and leaf drop over 2 years, given environmental confounders in 2009 due to irrigation. Consistently, h2 for labile plasticity in bud set and leaf drop was higher and more stable than that in bud break. However, plasticity in these spring and autumn phenologies was predicted with relatively high accuracy by paSNPs. This indicates that both phenologies had high genetic control. Nonetheless, spring phenology had stronger G × E interactions with less genetic constraints by population structure, leading to a higher cross-population predictability and genetic (co)variance variability with environmental change. By contrast, autumn phenology had a reduced magnitude of plasticity, weaker G × E interactions, and more pronounced population structure, which could explain a high variability and inconsistent predictability with climate-based models in the literature. Collectively, the study revealed that labile plasticity SNPs had a minor-to-moderate change in effect sizes over time for growth phenology and activity and can predict plasticity better than plasticity or trait non-associative counterparts. Moreover, growth phenology plasticity genes had biological functions in flowering regulation, senescence, (a) biotic and hormone responsiveness. The work has implications for strengthening our understanding of the genetic basis of labile plasticity. Nonetheless, it remains challenging to study plasticity in long-lived trees, owing to age or developmental effects during years of common-garden establishment and temporally lagged impacts. As recently indicated, the effect of the environment on phenology may be temporally lagged108,109 and, in turn, lagged environmental effects are amenable to shifts in phenology110. In addition, photosynthetic carbon assimilation in prior seasons is related to advanced spring phenology of the current year111. Future research should tease apart confounding factors affecting growth phenology. Methods Study system Branch cuttings from naturally growing trees of Populus trichocarpa Torr. & A. Gray located in 125 provenances belonging to 27 drainages (i.e., topographic units separated by watershed barriers; Figs. 6A and 1A), extending 11° in latitude (49.1–59.6°N) across two-thirds of the species’ distribution range (44–60°N, −121 to −138°W), were collected by British Columbia Ministry of Forests, Lands and Natural Resource Operations112. Cuttings were rooted and outplanted at Surrey, British Columbia (49.19°N, −122.85°W, 134 masl) in 2000. In spring 2008, propagules were collected from the Surrey site and used to establish a common garden at the University of British Columbia Research Totem Field (49.26°N, −123.25°W, 82 masl). In the common garden, biological replicates were assigned using a randomized block design for the clones with 4–20 clonal replicates of each genotype. All individuals used in this study were presented by trees of similar age and conditions. Field assessment of phenotypic traits in the common garden was conducted in two consecutive years 2009–2010 for 362 P. trichocarpa genotypes of 1500–2000 individuals (Text S1). The average growing season was 80 days longer in the year 2010 than 2009 (Fig. 1B), and the mean temperature before growing season commencement was 1.88 °C higher in 2010 compared to 2009 (Fig. 1B). In addition, the common-garden study had phenological traits for the year 2008 or 2011, depending on specific traits; I used these data as independent data for machine learning modeling evaluations (see “Machine-learning models” section below). Flowchart of a common-garden experiment for phenotyping and genotyping and data analysis pipeline. A A common-garden experiment was performed at the southernmost site relative to all sampling locations. Four phenotypic traits (i.e., bud break, bud set, leaf drop, and growth rate) were measured in 2 years 2009 and 2010. B A multi-trait GWAS was performed for phenotypic plasticity in the four traits using GEMMA and BGLR. Phenotypic plasticity was calculated as a trait value difference between the 2 years. For study traits, multi-trait GWAS was conducted for all traits in each of the years using BGLR or single traits over years using GEMMA as well as BGLR (as a comparison only). The β values are SNP effect weights based on multi-trait GWAS. Low (high) β denotes SNP effect size value within (outside) μ ± 0.5σ (mean ± half standard deviation), respectively. All the above GWAS were performed by populations. For each population, all the SNPs were classified into five subsets based on peak associations or effect sizes. C Each SNPs subset was used for machine learning. Input data comprised instances (individuals) with phenotypic plasticity values (a target for prediction) and categorical or numerical features (predictors). Example features, as shown in the tabular data, include population, provenance (categorical), environmental covariates (i.e., forcing, photoperiod, and chilling), and allele frequency at each locus (numerical). The instances were split into training (80%) and testing (20%) data sets. Training (categorical and numeric) and target data are shown in green, blue, and red lines, respectively. The training data were used to construct the models, whereas the testing data were used to evaluate the performance of the final models. Phenotyping for growth characteristics The project team measured two characteristics of tree growth dynamics: (1) growth phenology including the onset and cessation of a growth period based on canopy greening and yellowing in the spring and autumn (i.e., bud break, bud set, and leaf drop), respectively and (2) growth activity within a growth period based on stem growth (i.e., growth rate). The measurement of the traits was described previously61. Succinctly, seasonal canopy events were recorded directly from observations of trees and calculated additional traits based on phenological date information. The day of the year (DOY; d) of phenological events was recorded for each tree, including bud break (i.e., the timing of bud unfolding and the start of leaf emergence), final bud set, and leaf drop. Phenology events were marked using visual observations of the terminal bud on the tree. Active growth period was calculated based on days from bud break to final bud set. Active growth rate (cm d−1) was determined from yearly height gain divided by the growth period. All phenotypic traits were converted to Z-scores (i.e., Z-standardization) before use. The labile plasticity of each trait for an individual was calculated as the mean difference in trait values between the two measurement years. Genotyping DNA was extracted from P. trichocarpa mature leaves, and three different sample sets of the individuals were used to call single nucleotide polymorphisms (SNPs). Genotypes were sequenced at an expected coverage ranging from 15× to 30× using the Illumina HiSeq2000 platform. Short reads from the sequencing libraries were independently aligned to the P. trichocarpa v.3.0 genome using Burrows-Wheeler Aligner (BWA v.0.7.17) with default parameters. Mate pair metadata and marked duplicate molecules were corrected using the FixMateInformation and MarkDuplicates methods in the Picard package. Reads present in areas surrounding InDels were re-aligned using the IndelRealigner method from Genome Analysis Toolkit (GATK v.4.2.5.0). Subsequently, SNPs and small indels were independently called using the UnifiedGenotyper method from GATK. All SNPs were then filtered to exclude variants within 3 bp of any identified variants, having a mapping quality less than 5, and a variant quality less than 30. Each SNP was annotated using SNPeff113 with P. trichocarpa genome v.3.0. Phenotypic variance partitioning REML-linear mixed models were used to partition phenotypic variation into provenance (Prov), measurement year, and Prov × Year (G × E) interactions. A significant Prov, Year, and G × E interaction effect indicates provenance-level phenotypic differentiation, labile plasticity, and genetic variation in labile plasticity, respectively. Moreover, the first principal components (PC1) for forcing, photoperiod, and chilling were included as envionmental covariates (see details in Text S2). Selection of these environmental variables for the models was determined by a multicollinearity metric—Variance Inflation Factor (VIF) < 5 for a suite of environmental variables. All the aforementioned variable terms (i.e., Prov, Year, Prov × Year, and three PC1’s) were fixed effects. The random effect included was genotype ~N(0, σ2G), in which the G-matrix accounts for population structure. This genetic random effect controlled for potential confounding effects on associations between phenotypes and independent variables due to close genetic relatedness. The structure of each model was as follows: The model was fitted with the “relmatGlmer” function in the R package lme4qtl v.0.2.2114. The significance of both random and fixed effects was assessed using likelihood ratio tests. Model assumptions including normality, homoscedasticity, and linearity of residuals were checked using diagnostic plots in the R package performance v.0.15.2115. Population genetic structure and genomic scan for adaptive selection By pruning SNPs with LD < 0.2 to avoid the strong influence of SNP clusters on population genetic structure analysis, I selected 40,582 SNPs from 1,614,064 biallelic SNPs filtered by missingness <10% and MAF > 5%. Allele counts for each SNP were centered but not scaled. I applied a k-means clustering algorithm to detect the number of groups by maximizing the genetic variation between groups116. The k-means clustering was run on all principal components (PCs) of genetic data (362 PCs = the number of variables). The optimal number of groups in the k-means clustering algorithm is the one with the lowest Bayesian information criterion (Fig. S3A). To describe these genetic groups, I performed a discriminant analysis of principal components (DAPC)117,118. To avoid overfitting, only the first PCs that cumulatively explained 90% of the variance (p = 302 PCs) were retained in DAPC. As results showed the optimal number of groups was 2 (k = 2), only one discriminating function was rendered [i.e., min (p, k - 1)]. The discriminant function was constructed as a linear combination of the principal components to maximize the between-group variance and minimize the within-group variance. All these analyses were performed using “glPca”, “find.clusters”, and “dapc” functions in the R package adegenet v.2.1.1116. Subsequently, I carried out a genomic scan to calculate genetic differentiation between the two genetic clusters using the program hapflk v1.4119. This outlier approach detects adaptive selection based on population’s kinship (F) matrix Lewontin-Krakauer tests119,120. The analysis was implemented using the following parameter values: ten haplotype clusters (−K = 10) and 25 EM runs to fit the LD model (--nfit = 25). P-values were obtained by fitting a χ2 distribution to the empirical ditribution using a custom script available at https://forge-dga.jouy.inra.fr/documents/588 (last access October 2022) and visualized in Fig. S4. Genotype-phenotype associations Prior to the analysis, I partitioned individuals (i.e., genotypes) into 80% (training) and 20% (testing) subsets for each of the North and South populations. The split was performed at the genotype level to avoid any information leakage. This split strategy was applied to both phenotypic and genotypic data accordingly and was used throughout the study. Using the training data, association tests were performed based on a multivariate linear mixed model and a Bayesian multivariate model implemented in Genome-wide Efficient Mixed Model Association (GEMMA)121 and Bayesian Generalized Linear Regression (BGLR)122, respectively (Fig. 6B). As GEMMA tests for association while controlling population structure through genetic relationship matrix, I ran a GEMMA to identify SNPs associated with a given trait over years across populations (Fig. S6), whereas I implemented a BGLR for each population to estimate SNP effect size. Genetic relationship matrices were estimated by GEMMA (e.g., Fig. S3B) and used as a covariate in the models to control population structure. Specifically, I first used each phenotypic trait measured over 2 years, a genetic relationship matrix, and 1.6 M SNPs for the marker association test with GEMMA, and significant associations were identified after Bonferroni corrections at α = 0.05. Besides, P-values were adjusted for inflation using a genomic controlled statistical method123. In addition, I estimated SNP heritability based on the proportion of variance in traits explained by genotypes using REML algorithms in GEMMA. Then, I ran a BGLR to estimate each SNP effect on each phenotypic trait by population while considering the correlation structure of all four traits measured in the same year (“multi-trait model”) or of a given trait measured over years (“single trait over time model”). The regression coefficients of the SNPs (β’s) indicate the magnitude and sign of their contribution to phenotypic variance. Mathematical specifications about both models were elaborated in Text S3. I used a β ratio indicating a fold-change in β over time for each trait. I chose a linear regression method suitable to tackle high-dimensional, collinear data—ridge regression—to demonstrate the distribution of the ratio for each trait, visualized in Fig. S11, using the R package glmnet v.4.1-7124,125. Functional enrichment for labile plasticity and trait associated genes Information about gene annotation and gene ontology (GO) terms was retrieved from PlantGDB (https://www.plantgdb.org/) and agriGO v2.0126, respectively. Moreover, gene functions were manually checked at https://phytozome-next.jgi.doe.gov/. SNPs were assigned to corresponding genes given that a SNP is within a gene or its regulatory regions near upstream or downstream within a range of 5 K bp, whereby gene functions and signaling pathways for given SNPs were enriched. GO enrichment analysis was performed for plasticity- or trait-associated genes using Fisher’s exact test with Benjamini-Yekutieli multiple test adjustment at FDR < 0.05127, implemented in agriGO v2.0126. In addition, it is worth noting that due to no more than 50% (i.e., 21 K) of the poplar genes with GO terms, this annotation incompleteness would limit the enrichment for some functional categories. Classification of alleles based on peak-associations and effect sizes β I generated three sets of SNPs based on peak-associations with plasticity or traits across populations or by population using GEMMA at P < 1e−4 (Set 1–3) and discretized the non-peak associated SNP data into two sets in terms of β values based on multi-trait models from BGLR (Set 4–5; Fig. 6C). In addition, I randomly selected 300 SNPs (approx. average number of Set 1–3) from Set 4 and Set 5, constituting sub-Set 4–5, respectively. The following provided the details of generating the Sets. Set 1 (plasticity peak-assoc.): labile plasticity peak-associated SNPs (plasticity SNPs); Set 2 (across-popltn trait peak-assoc.): trait peak-associated SNPs (paSNPs) across populations with the Set 1 SNPs, if any, excluded; Set 3 (popltn-spec. trait peak-assoc.): trait paSNPs by population with the Set 1 SNPs, if any, excluded; Set 4 (high β’s): SNPs of high β values using cutoff |β| > 3e-6 based on |Z-score| > 0.5σ across traits in the distribution of β, but not in the Set 1–3; and Set 5 (low β’s): SNPs of low β values using cutoff |β| ≤ 3e−6, but not in Set 1–3. Machine-learning models I built machine-learning models to predict labile plasticity using multiple bootstrapping, boosting, and linear regression algorithms (Table S1), implemented in the scikit-learn package128 in Python. The goals of the analysis were twofold: comparing optimal model performance using different subsets of SNPs classified and examining model performance across populations and environments. As most machine leaning regression models utilize complex and opaque algorithms, which renders it difficult to ascertain the effects of predictor variables, a linear regression model was used to estimate the relative importance of SNPs to predicted target. In addition to genomic data (i.e., allele frequency), each set included the first principal component (PC1) for each of the three envionmental variables—forcing, photoperiod, and chilling. For predictive models, I first preprocessed the data by performing feature selection if the number of SNPs exceeded 30 K using the Lasso algorithm, imputing missing predictor data, encoding categories, and scaling numerical variables (Fig. S1; also see programming scripts). For model selection and validation, I used each data subset partitioned as previously described. I run models with hyperparameter tuning within a cross-validation (CV). Model specific parameters were tuned by fitting models over a grid of parameter values using a 10-fold CV scheme and selecting the combination of values with the smallest predictive error. A final model for each data set was built with the best hyperparameters identified. Model performance was assessed by comparing the coefficient of determination (R2) and root mean squared error (RMSE). Standard deviation of R² and RMSE was generated across CV-folds. As prediction accuracy depends on the similarity between training and testing datasets, models trained on one population may transfer poorly to another if they differ in genetic architecture or environmental conditions129. Models trained from one environmental trial may have poor transferability to multi-environmental trials if environmental effects are strong, resulting in altered genetic correlations or new constraints. I additionally used the model trained on the North or South training data to test the model performance using the entire South or North dataset, respectively. In addition, I used phenology data collected in other years (i.e., 2008 or 2011) as testing data to examine population-based model performance across environments. Supplementary information Supplementary File Supplementary Software. Acknowledgements The work was supported in part by the Leverhulme Trust (RPG-2017-196) and the Australian Research Council Centre of Excellence for Plant Success in Nature and Agriculture (CE200100015). Data used in this work were generated through funds of the Genome Canada Large-Scale Applied Research Program (POPCAN; project 168BIO) to the leading PIs of the project team (C.J. Douglas, Y.A. El-Kassaby, R.D. Guy, S.D. Mansfield, and Q.C.B. Cronk). I am grateful to Cambridge High-Performance Computing Facility (CSD3) for implementing genomic data preprocessing and computational simulations and to Cambridge Centre for Mathematical Sciences for providing physical workspace for manuscript drafting. I also thank you to the Editor and two anonymous reviewers for their helpful suggestions. Finally, I would like to dedicate this manuscript to the late Professor C.J. Douglas for his significant contributions to understanding poplar biology and his lead role in the POPCAN project. Author contributions Y.L. conceived the study, composed programming scripts, performed all data analyses, interpreted the results, drafted and revised the manuscript. Data availability The whole-genome shotgun sequencing data and alignment information have been deposited on SRA under the accession PRJNA276056. Code availability The custom machine-learning code is included as one Supplementary Item. Competing interests The author declares no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Supplementary information The online version contains supplementary material available at 10.1038/s41540-026-00691-6. References Associated Data Supplementary Materials Supplementary File Supplementary Software. Data Availability Statement The whole-genome shotgun sequencing data and alignment information have been deposited on SRA under the accession PRJNA276056. The custom machine-learning code is included as one Supplementary Item.