Introduction

Obesity, physical inactivity, high-carbohydrate diets, alcohol intake, smoking, pregnancy, and type 2 diabetes mellitus (T2DM) all increase triglyceride concentrations1. The magnitude of the increase varies substantially across individuals, which has been attributed in part to gene-environment interactions.

An alternative to gene-environment interaction is quantile-dependent expressivity, i.e., a dependence of genetic effects upon whether the phenotype (e.g., triglycerides) is high or low relative to its distribution in the population2. Specifically, different genetic effects can be obtained by selecting subjects for characteristics that distinguish high vs. low portions of the triglyceride distribution. We have shown that the effect size of a 31-SNP genetic risk score (GRSTG) increased significantly with increasing percentile of the triglycerides distribution2. Specifically, the effect of the GRSTG on triglyceride concentrations was 3.3-fold greater at the 90th percentile of the triglyceride distribution than at its 10th percentile. Within individuals, we have also shown that the genetic effect size for polymorphisms associated with ABCA1, APOA1, APOA2, APOA4, APOA5, APOB, APOC3, APOE, CETP, FABP2, FATP6, GALNT2, GCKR, HL, IL1b, LEPR, LOX-1, LPL, MC4R, MTP, NPY, SORT1, TNFA, TCF7L2, and TM6SF2 became significantly greater as the average triglyceride concentrations over all genotypes increased during postprandial lipemia3. Quantile-dependent expressivity has also been demonstrated for total cholesterol2, high- and low-density lipoprotein cholesterol2, body mass index2, and coffee consumption4.

Only about 11% of the triglyceride variance is currently explained by the 36 single nucleotide polymorphisms (SNP) showing genome-wide significance for fasting plasma triglyceride concentrations5,6. In contrast, heritability (h2) calculated from monozygotic twins raised together and apart suggest that additive genetic effects account for 54% to 65% of the triglyceride variance7. Verification of quantile-dependent expressivity was therefore sought using a more inclusive genetic measure in a larger population. To this end, we applied quantile regression8,9 to sibships and offspring-parent pairs from the Framingham Study10,11 to estimate heritability in the narrow sense (h212) at different quantile of the plasma triglyceride and very-low-density lipoprotein (VLDL) cholesterol distributions. Its importance is illustrated in fourteen published reports of drug or other treatment by genotypes interactions that were originally interpreted from a precision-medicine perspective13,14,15,16,17,18,19,20,21,22,23,24,25, and sixty-five other published examples originally attributed to biological interactions between genes and environment26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79,80, which might be more simply ascribed to quantile-dependent expressivity.

Methods

The Framingham Study data were obtained from the National Institutes of Health FRAMCOHORT, GEN3, FRAMOFFSPRING Research Materials from the NHLBI Biologic Specimen and Data Repository Information Coordinating Center. The Offspring Cohort consisted of 5,124 adult children of the original Framingham Study participants and their spouses who were first examined between 1971 and 1975, reexamined eight years later, and then every three to four years thereafter10. Children of the Offspring Cohort were recruited to form the Third Generation Cohort11. Subjects used in the current analyses were at least 16 years of age and were not taking medications to control lipid levels. Triglyceride concentrations were measured fluorometrically81 for all 9 exams of the Offspring Cohort, and exams 1 and 2 of the Third Generation Cohort. VLDL-cholesterol at exams 1–3 of the Offspring Cohort was determined by subtracting the bottom fraction cholesterol from total cholesterol81. Individual subject triglyceride values calculated as the average of the age- and sex-adjusted triglyceride concentrations over all available exams (i.e. the average of up to 9 exams for the Offspring Cohort, and up to 2 exams for Third Generation Cohort). VLDL-cholesterol was the average of up to three age- and sex-adjusted measurements.

Offspring-parent correlations (rOP) and regression slopes (βOP) were computed by assigning a weight of one-half to the offspring-father and one-half to the offspring-mother pair (if both parents were available), and assigning a weight of one to the offspring-parent pair if only one parent was available. Age and sex adjustment was performed separately in the Offspring and Third Generation Cohorts using standard least-squares regression with the following independent variables: female (0,1), age, age2, female x age, and female x age2. Offspring-midparental correlations (rOM) and regression slopes (βOM) were computed by comparing each child’s age- and sex-adjusted value to the average of the age and sex-adjusted parental values in those families having both parents. Full-sibling correlations (rFS) and regression slopes (βFS) were obtained by forming all ki(ki-1) sibpair combinations for the ki siblings in sibship i and assigning equal weight to each sibling82. The Lawrence Berkeley National Laboratory Human Subjects Committee approved use of the Framingham Cohort data for analysis.

Simultaneous quantile regression was performed using the sqreg command of Stata (version. 11, StataCorp, College Station, TX) and one thousand bootstrap samples were drawn to estimate the variance-covariance matrix for the 91 quantile regression coefficients between the 5th and 95th percentiles of the offsprings’ triglyceride distribution8,9. Post estimation procedures (test and lincom) were used to test linear combinations of the slopes with Σki -2 degrees of freedom for βOP and Σ (ki-1) degrees of freedom for βFS, where ki is the number of offspring in family i, and the summation is taken over all family sets. Quantile-specific expressivity was assessed by: 1) estimating quantile-specific βOP ± SE and βFS ± SE for the 5th, 6th… 95th percentiles of the sample distribution using simultaneous quantile regression; 2) plotting the quantile-specific βOP and βFS coefficients vs. the quantile of the offsprings’ trait distribution; and 3) testing whether the resulting graphs was constant, or changed as a linear, quadratic, or cubic function of the percentile of the trait distribution using orthogonal polynomials. Heritability in the narrow sense (h2) was estimated as h2 = 2βOP/(1 + rspouse) from the offspring-parent regression slope (βOP), as h2 = βOM from the offspring-midparental regression slope (βOM), and as h2 = [(1 + 8rspouseβFS)0.5−1]/(2rspouse) from full-sibs regression slopes (βFS), where rspouse is the spouse correlation12.

Published reports of gene-treatment and gene-environment interactions were identified through PubMed and the citations within each paper retrieved. In many cases27,28,29,30,32,33,35,36,40,43,48,49,62,64,65,67,70, mean triglyceride concentrations had to be estimated from published figures using the formatting palette of Microsoft Powerpoint to extract their quantitative information (version 12.3.6 for Macintosh computers, Microsoft corporation, Redmond WA). Vertical lines were drawn showing the vertical distances between each plotted point and the X-axis, and overall height of the Y-axis, from which triglyceride concentrations were derived3.

Results

There were 3325 Third Generation subjects who had one or more parents in the Offspring Cohort (1089 had one parent, 2236 had both parents). There were 1016 sibships with two or more full siblings in the Offspring Cohort (532 with two, 302 with three, 122 with four, and 60 with ≥five full sibs) and 1171 sibships with two or more full siblings in the Third Generation Cohort (576 with two, 333 with three, 155 with four, and 107 with ≥five full sibs). Unadjusted average triglyceride (SD) for subjects used in the analyses was 2.390 (1.934) mmol/L in the Offspring Cohort and 1.279 (0.914) mmol/L in the Third Generation Cohort. In addition, sibships from the Offspring Cohort had an unadjusted average VLDL-cholesterol concentration of 0.585 (0.294) mmol/L for exams 1–3.

Correlational analyses showed spouses were concordantly related for age- and sex-adjusted triglycerides (rspouse = 0.15), log triglycerides (rspouse = 0.31), and VLDL-cholesterol (rspouse = 0.09). Table 1 presents the traditional least squares regression slopes between offspring and parent (βOP) and offspring and midparent (βOM) and among full sibs (βFS). Triglyceride heritability (h2 ± SE) was significant as traditionally estimated from βOP (0.146 ± 0.013), βOM (0.131 ± 0.012), or βFS (0.456 ± 0.031). Heritability was even stronger for log triglycerides when estimated from βOP (0.360 ± 0.023), βOM (0.380 ± 0.023), or βFS (0.532 ± 0.030). Heritibility of VLDL-cholesterol was 0.343 ± 0.046 when estimated from full sibs (βOP unavailable because parents were not measured).

Table 1 Traditional and quantile regression analyses of triglycerides and very low density lipoprotein (VLDL)-cholesterol from the Offspring and Third Generation Framingham Cohorts.
Full size table

Plasma triglyceride concentrations

Figure 1 (upper panel) presents the offspring-parent regression slopes (βOP) for selected quantiles of the offspring’s plasma triglyceride distribution with their associated heritability estimates (h2 the narrow sense). Heritability became progressively greater with increasing quantiles of the offspring’s distribution, and differed significantly between the 10th and 90th percentiles (P =  2.2 × 10−11). These selected quantile-specific heritability estimates were included with those of other quantiles to create the quantile-specific heritability function in the lower panel, i.e., where h2 (Y-axis) is plotted as a function of the quantile of the offspring’s sample distribution (X-axis). Specifically, the Y-axis represents heritability at the 5th quantile, the 6th quantile,…, and the 95th quantiles of the offspring’s distribution. The shaded area presents the 95% confidence intervals for the individual slopes at each quantile. The figure shows that h2 increased from 0.033 ± 0.010 at their 10th percentile (P = 0.0009), 0.050 ± 0.016 at the 25th (P = 0.001), 0.111 ± 0.015 at the 50th (P = 1.3 × 10−13), 0.230 ± 0.030 at the 75th (P = 1.7 × 10−14), and 0.428 ± 0.059 at the 90th percentile of the offspring’ distribution (P = 6.4 × 10−13). If the heritability was the same for all offspring quantiles as traditionally assumed, then the upper panel would display parallel regression lines, and the lower graph would present a simple horizontal line. In fact, the graph shows that heritability became progressively stronger with increasing quantiles of its offsprings’ triglyceride distribution, such that on average each 1-percent increase in the offspring distribution was associated with a 0.0047 ± 0.0007 increase in heritability (P = 2.9 × 10−14). Moreover, the increase in quantile-specific h2 with increasing offspring’s triglyceride concentrations was significantly nonlinear, exhibiting both quadratic (P = 1.7 × 10−6) and cubic (P = 0.0007) effects. With respect to individual quantiles, heritability was statistically significant (P < 0.003) at every percentile between the 5th and the 95th percentiles of the offspring’ distribution, and was 13-fold greater at the 90th than at the 10th percentile.

Figure 1
figure 1

(upper panel) presents the offspring-parent regression slopes (βOP) for selected quantiles of the offsprings’ total triglyceride concentrations, with corresponding estimates of heritability (h2 = 2βOP/(1 + rspouse)). The slopes became progressively greater (i.e., steeper) with increasing quantiles of the triglyceride distribution. These quantile-specific regression slopes were included with those of other quantiles to create the quantile-specific heritability function in the lower panel. The statistical significance of the linear, quadratic and cubic trends and the 95% confidence intervals (shaded region) were determined by 1000 bootstrap samples. 1 mg/dL = 0.01129 mmol/L.

Full size image

Full-sib quantile regression

Figure 2 shows that the full-sib regression slope for triglyceride concentrations (βFS): 1) was 3.8-fold greater at the 90th (0.487 ± 0.081) than the 10th percentile (0.121 ± 0.010) of the sib distribution; and 2) increased 0.0042 ± 0.0007 (P = 1.8 × 10−9) for each percentile increase in the sibs’ distribution, and 3) exhibited significant nonlinearity (quadratic: P = 0.0007; cubic: P = 0.003). The full-sib slopes were statistically significant (P < 0.0001) at every percentile between the 5th and the 95th percentiles of the offspring’ distribution. Figure 2 also shows that siblings exhibited quantile-specific associations that were significantly greater at the 90th than 10th percentiles of the VLDL-cholesterol distribution (P = 0.04), and exhibited significant linear increases with each one-percent increment in their VLDL-cholesterol (0.0026 ± 0.0009, P = 0.003).

Figure 2
figure 2

Full-sib regression slopes (βFS) vs. quantiles of the sib’s triglyceride and VLDL-cholesterol distribution.

Full size image

Log-transformed triglyceride concentrations

Quantile-dependent effects persisted when triglyceride concentrations were log transformed for offspring-parent (Fig. 3, P = 4.9 × 10−10) and full-sib regression slopes (Table 1, P = 0.007).

Figure 3
figure 3

Offspring-parent regression slopes (βOP) vs. quantiles of the offsprings’ log-transformed triglyceride concentrations.

Full size image

Replication

Significant quantile-specific increases in βFS were detected separately for fasting triglycerides measured in 2792 sibling in 1016 sibships in the Framingham Offspring Cohort (linear: P = 7.7 × 10−5; quadratic: P = 0.003; cubic: P = 0.003), and in 3384 sibling in 1171 sibships in the Framingham Third Generation Cohort (linear: P = 0.001; quadratic: P = 0.84; cubic: P = 0.55).

Discussion

Genome-wide association studies have identified 36 single nucleotide polymorphisms (SNP) associated with triglyceride concentrations5,6. The most significant SNPs are associated with the glucokinase regulator (GCKR, P = 2 × 10−239), apolipoprotein A1 (APOA1, P = 7 × 10−224), and lipoprotein lipase genes (LPL, P = 2 × 10−199). Because only about 11% of the triglyceride variance is explained by these 36 loci5,6, the current paper investigated heritability in the narrow sense (h2) as a more comprehensive, albeit less specific, estimate of genetic transmission. It showed that h2 increased significantly with increasing percentiles of the triglyceride distribution. This result was replicated for βFS in the Framingham Offspring Cohort and Framingham Third Generation cohort separately. This confirms our previous analyses of fasting plasma triglyceride concentrations vs. GRSTG2, and postprandial triglyceride concentrations vs. individual SNPs3. The current analyses also demonstrated quantile-dependency for VLDL-cholesterol concentrations in sibs. Quantile dependence was also significant for log-transformed triglyceride concentrations. These analyses were based on simple robust estimates of heritability with nonparametric statistical significance determined from 1000 bootstrap samples.

Pharmacogenetics

Quantile-dependent expressivity predicts that genes affecting triglyceride concentrations should have a greater genetic effect prior to drug treatment when concentrations are high, than post-treatment when triglycerides are low. Moreover, smaller genotype differences when triglycerides concentrations are reduced pharmacologically might appear as gene-drug interactions in the absence of any true biological interactions. This prediction was assessed in 14 published reports purporting gene-drug or gene-treatment interactions on triglyceride response13,14,15,16,17,18,19,20,21,22,23,24,25.

For example, the histogram in Fig. 4A (insert) shows the reductions in fasting triglyceride levels reported by Lai et al. after three-week fenofibrate treatments15. The average decrease was significantly greater in APOA5 56 G carriers than non-carriers (35.8% vs. 27.9% decreases, P = 0.006). An accompanying editorial heralded its potential contribution to personalized medicine83. There is, however, an alternative interpretation of Lai et al.’s results from the perspective of quantile-dependent expressivity. Figure 4A shows that average triglyceride levels were higher before (1.58 ± 0.04 mmol/L) than after treatment (1.01 ± 0.02 mmol/L) and that triglyceride difference between genotypes were greater at the higher pre-treatment triglyceride levels (1.99–1.52 = 0.46 mmol/L difference, P = 0.01) than at the lower post-treatment triglyceride levels (1.06–1.00 = 0.06 mmol/L difference, P = 0.22), consistent with quantile-dependent expressivity. The smaller genetic effect size at the lower (post-treatment) than higher (pre-treatment) average triglyceride concentration requires that the trajectories of triglyceride reductions cannot move in parallel for different genotypes when triglycerides are decreased pharmacologically. Subtracting the pre-treatment from the post-treatment triglyceride levels will necessarily require a relatively greater triglyceride decrease for the genotype with the higher pre-treatment triglyceride level vis-à-vis the genotype with the lower pre-treatment level.

Figure 4
figure 4

Precision medicine perspective of different mean triglyceride reductions by genotypes following 160 mg/d fenofibrate or fenofibrate/statin combination therapy (histogram inserts of mean changes by genotype) vs. quantile-dependent expressivity interpretation (larger pre-treatment genetic effect size when average triglycerides concentrations were high vs. lower, requiring nonparallel triglycerides reductions by genotype), for: (A) Lai et al.’s 2007 report of 87 APOA5 56 G carriers vs. 703 non-carriers (genotype difference in mean triglyceride reduction P = 0.006)15; (B) Cardona’s et al.’s 2009 report of 14 APOA5 -1131C carriers vs. 22 non-carriers17; (C) Perez-Martinez et al.’s 2009 report of protected group (N = 236) consisting of the common allele homozygotes for GCKR rs780094C > T (CC), APOA5 −1131 T > C (TT), and APOA5 56 C > G (CC); an intermediate group (N = 490) consisting of homozygotes for GCKR rs780094C > T (CC) and carriers of the rare allele for either APOA5 −1131 T > C (CT or CC) or APOA5 56 C > G (CG or GG) or carriers of the rare allele for GCKR rs780094C > T (CT or TT) and homozygotes for both APOA5 −1131 T > C (TT) and APOA5 56 C > G (CC); and a risk group (N = 118) consisting of carriers of the rare allele for GCKR rs780094C > T (CT or TT) and carriers of the rare allele for either APOA5 −1131 T > C (CT or CC) or APOA5 56 C > G (CG or GG) with triglycerides > 1.69 mmol/L at baseline16; (D) Brautbar et al.’s 2011 report on 47 GG, 256 GA and 371 AA genotypes of rs3741298 in the APOA5-ZNF259 gene region who were also statin treated14; (E) Brautbar et al.’s 2011 report on 27 CC, 202 CG and 445 GG genotypes of rs964184 in the APOA5-ZNF259 gene region who were also statin treated14; and (F) Irvin et al.’s 2010 report on 81 APOE ε2-carrier, 454 ε3ε3, and 203 ε4-carriers18.

Full size image

Figures 4 and 5 display additional reports, initially interpreted from the perspective of personalized medicine, that are consistent with quantile-dependent expressivity, i.e., larger pre-treatment genetic effects when average triglycerides are high, followed by smaller post-treatment genetic effects when average triglyceride concentrations are low. Cardona et al. reported that the triglyceride reduction from fenofibrate treatment was over twice as great in TC/CC genotypes than TT homozygotes of the APOA5 -1131T polymorphism (2.34 vs. 1.15 mmol/L decreases, Fig. 4B histogram)17. The graph show that there was a greater triglyceride difference between carriers of the C-allele and TT homozygotes before treatment (5.80–3.74 = 2.06 mmol/L) when average triglycerides were high (4.54 mmol/L) than after treatment (3.46–2.60 = 0.86 mmol/L) when average triglycerides were lower (2.93 mmol/L). Perez-Martinez et al16. identified three genetic risk groups in hypertriglyceridemic subjects (pre-treatment triglycerides >1.69 mmol/L) derived from the GCKR-APOA5 loci: a protected group, an intermediate group, and a risk group. The histogram in Fig. 4C shows the decreases in plasma triglyceride concentration differed significantly between these groups after three-week fenofibrate treatment (P = 0.003): greatest in the risk group, intermediate in the intermediate group, and least in the protected group. However, the cross-sectional genotype differences were greater at baseline (risk: 3.08; intermediate: 2.71; protected: 2.40 mmol/L, P = 0.009) when the average triglyceride concentration over all genotypes was high (2.68 mmol/L), than after treatment (all genotypes approximately the same, P = 0.20) when average triglycerides were low (1.57 mmol/L, estimated from their Fig. 316).

Figure 5
figure 5

Precision medicine perspective of mean changes in triglyceride concentrations by genotypes (histogram inserts) vs. quantile-dependent expressivity perspective of larger genetic effect size when average triglycerides concentrations were high vs. low requiring nonparallel changes in triglycerides by genotype, for: (A) Brisson et al.’s 2015 report on 160 mg/d fenofibrate therapy in 44 carriers of LPL P207L mutation vs. 247 non-mutants19; (B) Pedro-Botet et al.’s report on 10 mg/day of atorvastatin’s effect in 10 male APOE ε2-carriers, 111 male ε3ε3, and 74 male ε4-carriers20; (C) Carmena et al.’s 2012 report on 80 mg/d lovastatin’s effect on 7 APOE ε2+, 58 ε3ε3, and 29 ε4+ familial hypercholesterolemia (FH) patients21; (D) Anagnostopoulou et al. 2007 report of 10–40 mg/d simvastatin in 160 carriers of the I-allele and 20 VV homozygotes of the CEPT I405V polymorphism22; (E) Balakrishnan et al.’s 2002 report of 93 patients who received pancreas transplants by APOE isoforms24; (F) Cabello et al.’s 2018 report on the effect of bexarotene treatment on carriers of the APOA5 -1131T > C or APOC3 388 T > C mutations vs. non-mutations25.

Full size image

Brautbar et al. reported that three SNPs in the ZNF259-APOA5 gene region on chromosome 11 showed substantially smaller genotype differences on fenofibrate/statin combination treatment when average triglyceride levels were low (1.64 mmol/L) compared to pretreatment differences when average levels were higher (3.15 mmol/L)14. Specifically, treatment reduced differences between GG, GA, and AA genotypes of rs3741298 from 3.77, 3.20, and 3.04 (P = 3.2 × 10−5) to 1.67, 1.65, and 1.64, respectively (P = 0.79, Fig. 4D), between the CC, CG and GG genotypes of rs964184 from 3.51, 3.41, 3.01 mmol/L (P = 2.3 × 10−7) to 1.77, 1.71, and 1.61 mmol/L, respectively (P = 0.18, Fig. 4E), and between GG, GA, and AA genotypes of rs10750097 from 3.37, 3.26, and 3.05 mmol/L (P = 0.002) to 1.62, 1.66, and 1.63 mmol/L, respectively (P = 0.86). Although the mean triglyceride reductions by genotype did not differ significantly by genotype in Brautbar’s paper (0.25 ≤ P ≤ 0.82), one by Aslibekyan et al. did report that the rs964184 polymorphism affected fenofibrate-induced triglyceride change significantly (P < 0.001)13.

Irvin et al.18 (Fig. 4F) reported that fenofibrate-induced changes in plasma triglyceride concentrations differed significantly (P = 0.05) between APOE ε3ε3 homozygotes (−29.0%), ε4-carriers (−26.4%) and ε2-carriers (−34.4%), an expected result from quantile-dependent expressivity given that the triglyceride difference between ε3ε3, ε4-carriers, and ε2-carriers were greater before treatment (1.48, 1.60, 1.75 mmol/L, respectively, P = 0.02) when average triglycerides were high (1.54 ± 0.03 mmol/L) than after treatment (0.97, 1.04, 1.10 mmol/L, P = 0.09) when average triglycerides were low (1.00 ± 0.02 mmol/L).

The histogram in Fig. 5A shows substantially greater triglyceride reductions due to fenofibrate treatment in 44 carriers of LPL P207L mutation than in 247 non-carriers who were hypertriglyceridemic (13.3 vs. 4.5 mmol/L average reductions). Brisson et al. attributed the difference to the mutation’s modulating effect on the fenofibrate response19. Alternatively, quantile-dependent expressivity would attribute the difference to the greater genetic effect size of the mutation (18.93–7.38 = 11.5 mmol/L difference) when average triglycerides were elevated before treatment (9.13 mmol/L) compared to post-treatment genetic effect size (5.60–2.88 = 2.72 mmol/L difference) when average triglycerides were much lower (3.29 mmol/L).

Several studies report statin-induced triglyceride reductions that were genotype specific. Pedro-Botet et al. reported that the average decrease in plasma triglyceride concentrations was significantly affected by APOE isoforms in a multicentric, double-blind clinical trial of 328 patients who received 10 mg/day of atorvastatin for one year20. The histogram in Fig. 5B shows that ε2-carriers had the greatest average decrease (0.64 mmol/L), ε4-carriers intermediate decrease (0.44 mmol/L), and ε3ε3 the smallest average decrease (0.34 mmol/L). However, the genetic effect size of the ε2-allele vis-à-vis ε3ε3 homozygotes was greater at baseline (0.46 ± 0.26 mmol/L) when average triglycerides were higher (1.94 mmol/L) than after one-year (0.16 ± 0.19 mmol/L) when average triglycerides concentrations were reduced (1.55 mmol/L), consistent with quantile-dependent expressivity.

These results agree with an earlier report by Carmena et al. of 94 patients with familial hypercholesterolemia (FH) who received 80 mg lovastatin for 80 days21. Figure 5C’s histogram shows that ε2-carriers showed the greatest average decrease (−0.69 mmol/L), ε4-carriers intermediate decrease (−0.52 mmol/L), and ε3ε3 the smallest average decrease (−0.41 mmol/L). Again, the genetic effect size of the ε2-allele vis-à-vis ε3ε3 homozygotes was greater at baseline (0.39 ± 0.29 mmol/L difference) when average triglycerides were higher (1.70 ± 0.02 mmol/L) than after one-year of lovastatin treatment (0.12 ± 0.15 mmol/L difference) when average triglycerides concentrations were reduced (1.23 ± 0.01 mmol/L), consistent with quantile-dependent expressivity.

Anagnostopoulou et al. reported significantly greater triglyceride reductions in carriers of the I-allele of the CEPT I405V polymorphism than VV homozygotes (0.70 vs. 0.47 mmol/L decreases, P = 0.04) from 10–40 mg/day simvastatin (Fig. 5D)22. Average triglyceride concentrations decreased from 2.01 to 1.34 mmol/L, and accordingly the genotype difference decreased from 0.27 mmol/L before treatment to 0.05 mmol/L after.

As a final example involving lipid-lowering drugs, Tuteja et al. reported that the C-allele of PDXDC1 rs3198697 accentuated the decrease in triglycerides from niacin-statin combination therapy (P = 0.02, not displayed)23. Once again, a larger pre-treatment genetic effect size per dose of the C-allele (0.14 mmol/L per copy) occurred when average triglycerides were high (1.88 mmol/L), and smaller post-treatment effect size occurred (β = −0.04 mmol/L per copy) when average triglycerides were low (1.35 mmol/L).

A non-pharmacological example is provided by Balakrishnan et al. who reported that triglyceride concentrations were significantly reduced, from 1.87 to 1.50 mmol/L, in 93 patients who received pancreas transplants (P = 0.002)24. The triglyceride difference between APOE ε4-carriers and ε3ε3 homozygotes went from being significant (0.45 mmol/L as estimated from their Fig. 2, P = 0.04) to nonsignificant (0.08 mmol/L) after the transplant, consistent with quantile-dependent expressivity (Fig. 5E).

Hypertriglyceridemia is the most common reason for discontinuing bexarotene, a drug used for treating cutaneous T-cell lymphomas84. Cabello et al. proposed that carriers of the APOA5 -1131T > C or APOC3 c.40 C > G mutations were the best candidates for bexarotene treatment because of their smaller triglyceride response25. Figure 5F presents the triglyceride differences between genotypes before and after oral bexarotene therapy while receiving prophylactic hypolipidemic therapy and 50 μg/d of levothyroxine sodium. From the perspective of personalized medicine, carriers of either minor allele experienced smaller triglyceride increases than non-carriers (1.25 vs. 2.39 mmol/L), whereas quantile dependent-expressivity would ascribe some of the effect to the smaller genetic difference between carriers and non-carriers before treatment (effect size: + 0.12 mmol/L) when average triglyceride concentrations were lower (1.59 mmol/L) than after treatment (effect size: −1.01 mmol/L, P = 0.02) when average triglyceride concentrations were higher (3.53 mmol/L).

To summarize, whereas other papers advocate individualized drug prescriptions using genetic markers to target patients (e.g., the histograms of Figs. 4 and 5), quantile-dependent expressivity postulates that these genetic markers follow different trajectories due to smaller genetic effects at lower triglyceride concentrations. It is unnecessary to hypothesize pharmacologic interactions of these genetic markers with treatment, rather APOA5, GCKR, APOA1, and APOE are simply among the brightest genetic signals tracking the reduced heritability.

Implications regarding gene-environment interactions

Environmental factors that distinguish higher vs. lower triglycerides (e.g., obesity, physical inactivity, smoking, alcohol, high-carbohydrate diets, T2DM) are predicted to produce different genetic estimates under quantile-dependent expressivity. Traditionally, these differences have been attributed to gene-environment interactions, where: 1) the effect of the genotype on the phenotype differs by environment26,27,28,29,30,31,32,33,34,35,36,37,41,42,43,45,46,47,48,49,50,51,52,53,54,55,58,59,60,61,62,63,65,66,67,74,75,76,77,78,79,80, or equivalently: 2) the effect of the environment on the phenotype differs by genotype36,38,39,40,44,56,57,59,64,68,69,70,71,72. In almost every case, these were explicitly interpreted as arising from a biological interaction between gene product and treatment. Not one of the cited reports considered the differences in average triglyceride levels between environmental conditions as an explanation of their observed results. As a causal model, quantile-dependent expressivity may arise from concentration-dependent effects of the mutations affecting triglyceride concentrations, e.g., impaired catabolism due to slower lipoprotein lipase activity having a greater effect when triglycerides were high than when low. Biologically, and from the perspective of chemical reactions, this makes more sense than the traditional fixed effect size. In our heritability analyses, it is possible that shared environmental effects contributed to offspring parent-regression slopes and that these too were quantile dependent. In contrast, our original GRSTG2 and the published examples to follow26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79, show quantile-dependent expressivity for genetic variants that are independent of shared environmental effects. The examples to follow represent interactions that are consistent with quantile-dependent expressivity because they show larger genetic effect sizes at higher average triglyceride concentrations.

Body mass index and waist circumference

Meta-analyses suggest that plasma triglyceride concentrations decrease 0.015 mmol/L per kg of weight loss85. BMI and waist circumference are associated with higher triglyceride concentrations due, at least in part, to the release of free fatty acids from visceral depots causing greater hepatic synthesis of VLDL86. Most reports of gene-weight interactions appear to be at least partially attributable to quantile-dependent expressivity, including six studies based on genetic risk scores (GRSTG)26,27,28,29,30,31. Cole et al. reported that the effect of their GRSTG on triglyceride concentrations was significantly larger in obese subjects (effect size ± SE: 0.480 ± 0.053 mmol/L for BMI ≥ 35 kg/m2) then in lean subjects (0.261 ± 0.034 mmol/L for BMI ≤ 23 kg/m2), and intermediate in subjects with an intermediate BMI (0.354 ± 0.029 mmol/L)26. Ali et al. reported that each unit increment in their GRSTG was associated with a 2.4% triglyceride-increase in overweight/obese subjects, and a 1.5% triglyceride-increase in normal weight subjects27. Quantile-dependent expressivity would attribute these differences to the higher average triglyceride concentrations of the obese subjects reported by Cole et al. (>0.50 mmol/L higher) and the overweight/obese subjects reported by Ali et al. (estimated >0.40 mmol/L) than their lean comparison group.

Klimentidis et al.28. report that increasing tertiles of waist-to-hip ratio were associated with progressive increases in the GRSTG effect size (estimated β1st tertile = 0.16, β2nd = 0.18, and β3rd = 0.22, PInteraction = 3.9 × 10−8). This, however, was in the context of highly significant increases in average triglyceride concentrations with increasing waist circumference (P = 1.3 × 10−56). Zubair et al. reported that the triglyceride difference between a high and low GRSTG score was greater in overweight/obese women than leaner women (0.49 vs. 0.29 mmol/L, Pinteraction = 0.03) and greater in broad-waisted than slim-waisted women (0.54 vs. 0.27 mmol/L, Pinteraction = 0.02)29. Again, these differences are consistent with quantile-dependent expressivity given that average triglycerides concentrations were greater in overweight/obese than leaner women (1.38 vs. 1.22 mmol/L) and greater in broad-waisted than slim-waisted women (1.44 vs. 1.19 mmol/L, calculated from their published data).

Justesen et al. reported significant interactions between adiposity and their 39-SNP GRSTG in two Danish cohorts: the Inter99 cohort (N = 5961 subjects) and the Health2006 cohort (N = 2565 subjects)30. BMI was divided into normal weight, overweight, and obese. Waist circumferences were divided into normal, centrally overweight, and centrally obese. Both cohorts showed triglycerides that were significantly affected by BMI x GRSTG interactions (Inter99: P = 0.002; Health 2006: P = 0.02; combined; P = 9.8 × 10−5) and waist circumference x GRSTG interactions (Inter99: P = 0.0001; Health 2006: P = 0.05; combined; P = 2.0 × 10−5), with a larger genetic effect among individuals who were obese. However, average triglyceride levels for normal weight, overweight, and obese increased from 0.92, to 1.23 to 1.55 mmol/L in the Inter99 cohort, respectively, and from 0.94 to 1.23 to 1.54 mmol/L in the Health2006 cohort. Similarly, average triglyceride levels for normal, centrally overweight, and centrally obese subjects increased from 0.96 to 1.26 to 1.49 mmol/L in the Inter99 cohort and from 0.95 to 1.18 to 1.41 mmol/L in the Health2006 cohort. From the perspective of quantile-dependent expressivity, greater adiposity was an indicator of higher average triglyceride concentrations and its larger genetic effect.

Ahmad et al. reported that each unit increase in their 40-SNP GRSTG produced a significantly stronger effect on triglycerides in overweight and obese (1.013% triglyceride increase) than healthy weight women (1.011%, Pinteraction = 0.004), and a significantly stronger effect in centrally overweight and obese (1.012%) than centrally healthy weight women (1.010%, Pinteraction = 0.005)31. These results are consistent with quantile-dependent expressivity and the higher triglyceride concentrations of the overweight and obese vs. healthy weight women (1.8 vs. 1.3 mmol/L, P < 0.0001), and the centrally overweight and obese vs. centrally normal weight women (1.7 vs. 1.2 mmol/L, P < 0.0001).

Gene-weight interactions have also been reported for individual SNPs, including those associated with APOA526,32,33,34, LPL26,35,36,37,38,39,40,41,42,43, GCKR26,44, insulin receptor substrate-1 (IRS-1)45,46, methylenetetrahydrofolate reductase (MTHFR)47,48, proprotein convertase subtilisin/kexin type 9 (PCSK9)48, APOB49, APOE50, peroxisome proliferator-activated receptor γ 2 (PPARγ2)51, and patatin-like phospholipase domain-containing protein 3 gene (PNPLA3)52.

The APOA5 gene is the strongest genetic determinant of plasma triglyceride concentrations87. Four studies report interactions between BMI and APOA5 polymorphisms that are consistent with quantile-dependent expressivity. Wu et al. reported that the effect size for the Gly185Cys polymorphism at APOA5 rs3741297 was accentuated in Filipinos with a higher waist circumference32. Specifically, it increased from 0.13, 0.06, 0.30, to 0.96 mmol/L from the first to the fourth quartiles of waist circumference in mothers (Pinteraction = 0.01), and from 0.19, 0.03, 0.13, to 0.58 mmol/L from the first to the fourth quartiles in offspring (Pinteraction = 0.007). Average triglycerides levels also increased from the first to the fourth quartiles of waist circumference, i.e., 0.96, 1.24, 1.38, 1.55 mmol/L in mothers, and 0.86, 0.90, 1.03, and 1.23 mmol/L in offspring. A second study, by Kim et al.33, reported that the triglyceride difference between C-carriers and TT homozygotes of the APOA5 -1131T > C polymorphism was greater in overweight vs. normal weight Koreans at baseline (0.31 vs. 0.10 mmol/L) and their 3-year follow-up (0.55 vs. 0.19 mmol/L), corresponding to the higher average triglycerides in overweight than normal weight Koreans at baseline (1.47 vs. 1.02 mmol/L) and follow-up (1.60 vs. 1.06 mmol/L, estimated from their published graphs). Hsu et al.’s reported that the effect of the C-allele of rs662799 on plasma triglyceride concentrations was greater in obese than lean patients (0.473 vs. 0.142 mmol/L per C-allele) in accordance with their higher average triglyceride concentrations (1.51 ± 0.07 vs. 0.90 ± 0.02 mmol/L)34. The fourth study, by Cole et al.26., reported a significantly greater effect size for APOA5 rs964184 in obese than lean subjects (β = 0.159 ± 0.03 vs. 0.140 ± 0.03 mmol/L per G allele, Pinteraction = 0.009) whose average triglycerides differed by >0.5 mmol/L.

The LPL enzyme hydrolyzes triglycerides, and it participates in hepatic triglyceride-rich lipoprotein (TRL) clearance via the LDL receptor-related protein1. Multiple studies suggest that purported interactions between LPL polymorphisms and BMI on triglycerides are consistent with quantile-dependent expressivity, in that greater adiposity is associated with higher average triglyceride concentrations. Fisher et al. first reported a significant interaction between LPL S291 and BMI on triglycerides (Pinteraction = 0.02)35. Compared to those with a BMI < 25, their Fig. 2 showed heavier men had a greater triglyceride difference between genotypes (heavier vs. leaner: 0.42 vs. −0.17 mmol/L difference) corresponding to their higher average triglyceride (1.94 vs. 1.54 mmol/L) in the Northwich Park Heart Study II project35. The European Atherosclerosis Research Studies reported that S291-carriers had greater increases in plasma triglycerides with increasing BMI than non-carriers (P < 0.01)36. Correspondingly, the genotype differences and average triglyceride concentrations were −0.08 and 0.89 mmol/L in the lowest BMI tertile, respectively, 0.18 and 1.00 mmol/L in the intermediate BMI tertile, respectively, and 0.18 and 1.13 mmol/L in the highest BMI tertile, respectively. Mailly et al. reported a marginally greater triglyceride difference for carriers vs. non-carriers of the N9 mutation in overweight men (0.53 ± 0.27 mmol/L difference) with higher average triglycerides (1.86 ± 0.05 mmol/L) than in leaner men (0.02 ± 0.26 mmol/L difference) with lower average triglycerides (1.51 ± 0.05 mmol/L)37, as did Gerdes et al. for the highest BMI tertile (0.25 mmol/L genotype difference) with higher average triglycerides (1.12 mmol/L) vis-à-vis leaner men (0.10 mmol/L genotype difference) with lower average triglycerides (0.93 mmol/L)36, although neither reached statistical significance.

Figure 6A presents Jemaa et al.’s findings for a 10 week restricted calorie diet by the LPL HindIII polymorphism38. From a precision medicine perspective, the histogram (insert) shows plasma triglyceride concentration decreased significantly more in H2H2 homozygotes than H1-carriers (0.27 vs. 0.04 mmol/L decreases, P = 0.03). Consistent with quantile-dependent expressivity, the difference between genotypes was greater at baseline than after weight loss (0.32 ± 0.13 vs. 0.09 ± 0.11 mmol/L) in accordance with the higher average triglycerides at baseline (1.23 ± 0.07 vs. 1.08 ± 0.05 mmol/L). Again, the smaller genetic effect size at the lower (post-treatment) than higher (pre-treatment) triglyceride concentrations require that the effects of the genotypes do not move in parallel when triglycerides are decreased by weight loss. Subtracting the pre-treatment from the post-treatment triglyceride levels will necessarily create a relatively greater triglyceride decrease for the genotype with the higher pre-treatment triglyceride level vis-à-vis the genotype with the lower pre-treatment level.

Figure 6
figure 6

Precision medicine perspective of different mean changes in triglyceride concentrations by genotypes (histogram inserts) vs. quantile-dependent expressivity perspective of larger genetic effect size when average triglycerides concentrations were high vs. low requiring nonparallel changes in triglycerides by genotype, for: (A) Jenaa et al.’s 1997 report on the triglyceride response to 10-week weight loss diet in 58 H2H2 homozygotes and 57 H1-carriers of the of the LPL Hind III polymorphism (P = 0.03)38, (B) Yamasaki et al.’s 2015 report on the effect of a 3 month weight loss intervention in 87 TT, 163 TC, and 43 CC genotypes of the APOA5 -1131T > CT polymorphism39, (C) Pollin et al.’s 2011 reported on the effect of a one-year lifestyle intervention in 919 subjects by the GCKR rs1260326 P446L polymorphism44; (D) Ruaño et al.’s report on the effect of 6-month exercise training in 53 homozygotes and 22 A-allele carriers of the APOA1 −75 G/A polymorphism57; (E) Lin et al.’s report on the effect of going from a 54% carbohydrate/31% fat diet to a 70% carbohydrate/15% fat diet on 36 TT and 20 C-carriers of the APOA5 -1131T > C polymorphism68; (F) Humphries et al.’s 1996 report on the triglyceride response to a high saturated fat (26% SFA, 10% MUFA, 2% PUFA) and high polyunsaturated fat diets (9% SFA, 6% MUFA, 23% PUFA) in 45 H + and 10 H- genotypes of the LPL Hindlll gene loci69.

Full size image

Figure 6B presents Yamasaki et al.’s results for a 3-month lifestyle weight loss intervention that reduced dietary calorie intake from 2066 ± 27 to 1691 ± 22 kcal/d, increased energy expenditure from 2100 ± 34 to 2266 ± 36 kcal/d, and reduced BMI from 25.65 ± 0.18 to 24.75 ± 0.17 kg/m239. The intervention produced twice as much triglyceride decrease in CC than CT or TT genotypes of the APOA5 -1135T > C polymorphism. Quantile-dependent expressivity would attribute the difference to the smaller genotype differences after weight loss than before (CC vs. TT difference ± SE: 0.19 ± 0.12 after vs. 0.43 ± 0.19 mmol/L before) due to the lower average triglyceride concentrations after weight loss (1.27 ± 0.04 vs. 1.55 ± 0.05 mmol/L).

Data extracted from Vohl et al.’s Fig. 1 showed that triglycerides increased significantly with increasing adiposity in LPL HindIII H + homozygotes but not H- carriers as measured by BMI (slope ± SE: 0.137 ± 0.052 vs. 0.020 ± 0.041 mmol/L per kg/m2), and visceral adipose tissue area (0.009 ± 0.003 vs. 0.001 ± 0.004 mmol/L per cm2)40. The difference between genotypes increased with increasing BMI (0.12 ± 0.07 mmol/L per kg/m2, P = 0.08), and visceral area (0.008 ± 0.005 mmol/L per cm2, P = 0.10), and average triglycerides concentrations also increased with increasing BMI (0.069 ± 0.033 mmol/L per kg/m2, P = 0.04) and visceral area (0.006 ± 0.003 mmol/L per cm2, P = 0.02). In another study, Senti et al.’s data showed that women with higher waist-to-hip ratios had somewhat greater triglyceride difference between H- carriers and H+ homozygotes (0.255 mmol/L difference) than women with lower waist-to-hip ratios (0.194 mmol/L difference), which probably corresponds to their difference in average triglycerides concentrations (1.26 vs. 0.90 mmol/L, respectively, P < 0.001)41.

Huang et al. reported that the triglyceride difference between SS and SX/XX genotypes of the LPL S447X polymorphism was greater in centrally obese than nonobese twins (0.24 vs. 0.06 mmol/L differences, P = 0.16), corresponding to the greater average triglycerides in centrally obese than nonobese twins (1.39 vs. 0.99 mmol/L)42. Garenc et al.’s data showed that the triglyceride difference between homozygotes for the S477-allele and X477-carriers was greater in obese men (0.80 mmol/L difference, P = 0.002) and women (0.53 mmol/L difference, P = 0.01) than normal weight men (0.09 mmol/L difference) and women (0.15 mmol/L difference), which corresponds with the higher average triglycerides of the obese men (2.02 mmol/L) and women (1.25 mmol/L) compared to the normal weight men (1.15 mmol/L) and women (1.10 mmol/L)43. Cole et al.’s reported significantly greater effect size for LPL rs12678919 in obese than lean subjects (β = −0.148 ± 0.03 vs. −0.050 ± 0.03 mmol/L per G allele, Pinteraction = 0.0007) whose average triglycerides differed by > 0.5 mmol/L26.

The GCKR Pro446Leu polymorphism (rs1260326) affects triglyceride concentrations by increasing hepatic glucokinase activity44. Pollin et al. reported that there was a significant interaction (P = 0.04) between the triglyceride response to lifestyle intervention and rs1260326, with the 446L (T)-allele showing enhanced triglyceride reduction (P = 0.04, Fig. 6C)44. Consistent with quantile-dependent expressivity, triglycerides increased more per T-allele at baseline when triglycerides averaged 1.84 mmol/L (β = 0.141 mmol/L, P = 1.8 × 10−9) than after weight loss intervention when estimated triglycerides averaged 1.57 mmol/L (β = 0.084 mmol/L). This agrees with Cole et al.’s report of a significantly greater effect size in obese than lean subjects (β = 0.093 ± 0.03 vs. 0.067 ± 0.02 mmol/L per T allele, Pinteraction = 0.03) whose average triglycerides differed by > 0.5 mmol/L26.

Clausen et al. reported that the IRS-1 G972R mutation and obesity interacted to significantly increase plasma triglyceride concentrations (Pinteraction = 0.04)45. The difference between the R-carriers and GG homozygotes was seven-fold greater in obese than lean subjects (0.70 ± 0.42 vs. 0.10 ± 0.09 mmol/L), which could be due to the higher average triglyceride concentrations of the obese subjects (1.46 ± 0.10 vs. 0.91 ± 0.02 mmol/L). Baroni et al. also reported that R-carriers showed a greater difference from GG homozygotes in obese subjects (2.11–1.73 = 0.38 mmol/L difference) than lean subjects (1.72–1.78 = −0.06 mmol/L difference) corresponding to the higher average triglyceride concentrations of the obese vs. lean subjects (1.80 ± 0.10 vs. 1.73 ± 0.09 mmol/L)46.

Zhi et al. reported that differences between the CC, CT, and TT genotypes of the MTHFR C677T polymorphism were significantly greater (P = 0.02) in women with BMI > 24 kg/m2 (0.90, 0.99, 1.09 mmol/L, respectively) who had higher average triglycerides (1.00 mmol/L), than in leaner women (0.65, 0.71, 0.63 mmol/L, respectively) who had low triglycerides (0.67 mmol/L)47. Our analyses of Yin et al.’s data (their Fig. 3) showed significant differences between obese and lean Chinese for the effects of the MTHFR C677T genotype (0.387 vs. 0.029 mmol/L per dose of the T allele, Pinteraction = 0.006) and PCSK9 E670G (AG-AA difference: 1.64 vs. −0.17 mmol/L, Pinteraction < 0.0001) consistent with the higher average triglycerides of the overweight/obese Chinese (1.74 vs. vs. 1.21 mmol/L)48.

Other gene-environment interactions involving body weight also appear attributable to quantile-dependent effects. With respect to the APOB XbaI polymorphism, Turner et al.’s data showed the effect size per dose of the X+ allele became progressively greater when going from the lowest (−0.012 mmol/L), to the intermediate (0.035 mmol/L) to highest BMI tertile (0.053 mmol/L, Pinteraction = 0.015)49. The lowest BMI tertiles had low average triglycerides (estimated as 0.84 mmol/L), the intermediate BMI tertiles had intermediate average triglycerides (0.89 mmol/L) and the highest BMI tertile had the highest average triglycerides (1.01 mmol/L).

Jemaa et al. reported that triglyceride concentrations were significantly lower for APOE ε3ε3 homozygotes than ε2- or ε4-carriers only in Tunisians who had BMI ≥ 30 kg/m2 (and presumably higher triglycerides)50.

Becer et al. reported that triglyceride concentrations were more strongly related to the dose of the A-alleles of the PPARγ2 Pro12Ala polymorphism in obese (0.115 mmol/L per A allele, P = 0.05) than nonobese subjects (0.054 mmol/L per A allele, P = 0.52) which probably relates to the higher average triglyceride concentrations of the obese subjects (1.71 ± 0.03 vs. 1.16 ± 0.03 mmol/L)51.

Finally, Stojkovic et al. reported a significantly stronger trend (P = 0.01) from the GG, CG, to CC genotypes of the PNPLA3 rs738409 polymorphism in overweight (i.e., 1.39, 1.50, 1.57 mmol/L for BMI > 25) than normal weight subjects (from 1.26, 1.20, 1.20 mmol/L for BMI ≤ 25, Pinteraction = 0.003) consistent with the higher triglycerides in the overweight than normal weight subjects (1.54 vs. 1.20 mmol/L)52.

Physical activity

Aerobic physical activity decreases triglyceride concentrations by facilitating triglyceride hydrolysis and use by skeletal muscles88. Meta-analyses suggest that triglyceride concentrations average 0.11 mmol/L less for those who walked ≥6000 vs. <2000 steps/day, and 0.23 mmol/L less for those who exercised at 50% of VO2max for three 30-minute sessions per week compared to less active subjects89. Our analyses of Senti et al.’s data53 showed that the each dose of the H+ allele of the LPL HindIII polymorphism was associated with a triglyceride increase of 0.148 mmol/L in the least active men (expending ≤291 kcal/d), 0.135 mmol/L in men expending 292–525 kcal/d, and 0.105 mmol/L in the most active men (>525 kcal/d) in an apparent gene-environment interaction. However, average triglyceride concentrations decreased with increasing physical activity: from 1.432, 1.250, to 1.106 mmol/L, respectively, suggesting an effect size for the H+ allele consistent with quantile-dependent expressivity.

Pisciotta et al. reported that the −265 T/C polymorphism of the APOA2 gene had a greater effect on triglyceride concentrations in sedentary men (TT, TC, CC: 2.12, 1.64, 1.24 mmol/L) than active men who cycled 120–150 km/wk (1.35, 1.33, 1.09 mmol/L, respectively), consistent with the higher triglyceride-cholesterol concentrations of the sedentary men (1.74 ± 0.11 vs. 1.30 ± 0.08 mmol/L)54.

Tanisawa et al. reported that triglyceride concentrations increased with increasing tertiles of their GRSTG (1st: 0.93 ± 0.06, 2nd: 1.41 ± 0.13, 3rd: 1.46 ± 0.14 mmol/L) in Japanese men with low cardiorespiratory fitness, but not in those with higher fitness (1st: 0.92 ± 0.07, 2nd: 0.77 ± 0.06, 3rd: 1.05 ± 0.08 mmol/L, Pinteraction = 0.03) as predicted by quantile-dependent expressivity given the higher average triglycerides of the low vs. high fitness groups (1.30 ± 0.07 vs. 0.93 ± 0.04 mmol/L)55.

A small training study by Hagberg et al. reported larger triglyceride decreases in the +/− and +/+ than −/− genotypes of the LPL Pvull polymorphism (−0.68 ± 0.28 vs. −0.35 ± 0.18 mmol/L)56. However, training reduced average triglyceride concentrations from 2.04 to 1.41 mmol/L, and quantile-dependent expressivity would therefore predict the larger genotype difference at baseline than follow-up (+/− and +/+ vs. −/− difference: 0.25 vs. −0.08 mmol/L, respectively), producing nonparallel triglyceride decreases by genotype.

Ruaño et al. reported that 6 months of supervised aerobic exercise training produced significantly greater percent reductions in triglyceride concentration in A-carriers of the APOA1 -75G > A polymorphism than in GG homozygotes (P = 0.05)57. Figure 6D shows that average triglyceride concentrations were lower after training than before (1.30 vs. 1.49 mmol/L) corresponding to smaller genotypic differences after training than before (0.38 vs. 0.72 mmol/L).

Smoking

Smokers are insulin resistant and exhibit impaired lipid metabolism, including impaired triglyceride clearance after a mixed meal90. Meta-analyses suggest that triglyceride concentrations of smokers average 9.1% higher than nonsmokers, and show a dose-dependent relationship from light (10.7%), moderate (11.5%) to heavy smokers (18%)91. Quantile-dependent expressivity would predict greater genetic effects on triglycerides in smokers than nonsmokers because of the smokers’ higher triglyceride concentrations. Czerwinski et al. in fact reported that the heritability of plasma triglyceride concentrations was higher in smokers (h2 = 0.70, average triglycerides 1.68 ± 0.06) than nonsmokers (h2 = 0.42, average triglycerides 1.58 ± 0.03)58. With respect to individual loci, smoking is reported to modify the effects on triglycerides of the upstream stimulatory factor 1 (USF1) gene polymorphism rs251683959, C242T polymorphism of the cytochrome b-245 alpha chain (CYBA) gene60, –482 C > T in the insulin-responsive element of APOC361, LPL HindIII53,62, and LPL rs26363.

There are several reports of LPL polymorphisms affecting the triglyceride response to smoking. Peacock et al. found larger differences between H+H+ homozygotes and H- carriers of the LPL– HindIII polymorphism in smokers than nonsmokers (sexes combined: 0.23 vs. 0.01 mmol/L difference, P < 0.02) consistent with the smokers’ higher average triglyceride concentrations (1.03 vs. 0.92 mmol/L)62. Senti et al. reported a significant difference between H+ H+ homozygotes and H- carriers of the LPL HindIII polymorphism in sedentary smokers (0.53 ± 0.26 mmol/L, P = 0.04) but not nonsmokers (0.05 ± 0.14 mmol/L), again, consistent with the higher average triglycerides of the sedentary smokers (1.54 ± 0.12 vs. 1.39 ± 0.07 mmol/L)53. Pyun et al.’s data showed a greater difference between CC homozygotes and T-carriers for LPL rs271 in smokers (2.20–1.96 = 0.24 mmol/L) than nonsmokers (1.66–1.63 = 0.03 mmol/L, Pinteraction = 0.009) consistent with the smokers’ higher average triglycerides (2.11 vs. 1.65 mmol/L)63.

Niemiec et al. reported that the USF1 rs2516839 polymorphism modified the triglyceride response to smoking, however, triglyceride differences between the CC, CT and TT genotypes were greater in smokers (2.27 ± 0.26, 1.80 ± 0.09, 1.53 ± 0.10 mmol/L, respectively) in accordance with their higher average triglycerides (1.79 ± 0.07 mmol/L) than in nonsmokers (1.49 ± 0.11, 1.46 ± 0.06, 1.57 ± 0.08, respectively) in accordance with their lower concentrations (1.51 ± 0.05 mmol/L)59. Ge et al. reported that the difference between the CC homozygotes and carriers of the T-allele of CYBA C242T polymorphism was significant in smokers (0.17 mmol/L, P = 0.01) but not nonsmokers (0.04 mmol/L, P = 0.76), which quantile-dependent expressivity would partially attribute to the smokers higher average triglyceride concentrations (1.33 vs. 1.21 mmol/L)60. Waterworth et al. reported that the smoking-triglyceride relationship was modified by both APOC3 –482 C > T (Pinteraction = 0.009) and 3238 C > G polymorphisms (Pinteraction = 0.04)61. Specifically, smokers’ had higher average triglyceride concentrations than nonsmokers (1.74 vs. 1.59 mmol/L), and as predicted, a greater effect per dose of the –482T-allele (0.135 vs. −0.009 mmol/L) and per dose of the 3238G-allele (0.380 vs. 0.113 mmol/L, calculated from their table two) than nonsmokers61.

Smokers did not have higher triglycerides than nonsmokers in the 41,000 subjects of the Population Architecture Using Genomics and Epidemiology (PAGE) study (mean ± SE: 1.476 ± 0.010 vs. 1.486 ± 0.005 mmol/L)92. Consistent with quantile-dependent expressivity, their meta-analysis did not show any significant SNP by smoking interactions.

Diet

Each 1% isoenergetic replacement of carbohydrates with fat is expected to decrease plasma triglyceride concentrations by an average of 0.021 mmol/L if saturated, 0.019 if monounsaturated, and 0.026 mmol/L if polyunsaturated93. Adherence to a Mediterranean diet decreases plasma triglyceride concentrations by an average of 0.069 mmol/L94. Quantile-dependent expressivity would predict larger genetic effects on low-fat high-carbohydrate diets than high-fat low-carbohydrate diets, and larger genetic effects on Western than Mediterranean diets, in accordance with the expected higher triglycerides of the former.

Gomez-Delgado et al. reported that decreases in plasma triglyceride due to adopting a Mediterranean diet were significantly greater in 203 GG homozygotes of the tumor necrosis factor alpha gene (TNFA, rs1800629) than in 48 carriers of the A-allele, i.e. approximately 0.31 vs. 0.12 mmol/L, respectively (P = 0.005)64. However, plasma triglyceride concentrations averaged approximately 1.80 mmol/L at baseline and 1.52 mmol/L on the diet, and correspondingly, the differences between the GG and GA/AA genotypes were 0.38 vs. 0.19 mmol/L, respectively. A quantile-dependent interpretation of these results is that the Mediterranean diet decreased plasma triglyceride concentrations, which in turn produced a smaller difference between genotypes.

Pyun et al.’s data showed a greater triglyceride difference between CC homozygotes and T-carriers for LPL rs263 with increasing energy intake: 0.005 mmol/L difference for ≤1500 kcal/d, 0.14 mmol/L difference for 1501–2000 kcal/d, 0.13 mmol/L for 2001–2500 kcal/d, and 0.20 mmol/L for > 2500 kcal/d, Pinteraction = 0.02) corresponding to the increasing average triglycerides concentrations with energy intake (1.73, 1.78, 1.78, 1.84 mmol/L, respectively)63.

Garcia-Rios et al. reported significant interactions between plasma concentrations of n-6 polyunsaturated fatty acids and LPL rs238 (Pinteraction = 0.05) and LPL rs1059611 (Pinteraction = 0.04)65. Below median n-6 PUFA concentrations, the rs1059611 triglyceride difference between AA homozygotes and carriers of the G allele was 0.33 mmol/L and the average triglyceride concentration across genotypes was 2.14 mmol/L. Above the median, the genotype difference was smaller (−0.09 mmol/L) in accordance with lower average triglyceride concentrations (1.37 mmol/L), consistent with quantile-dependent expressivity. Nearly identical results were reported for rs238, which was in strong linkage disequilibrium with rs1059611.

Garcia-Rios et al. also reported a significant interaction between plasma saturated fatty acids concentrations and the circadian clock gene Period 2 (PER2) rs2304672 on plasma triglyceride concentrations (Pinteraction = 0.004)66. Above the median plasma SFA concentration of 30.9 mmol/L, plasma triglyceride concentrations differed significantly between the carriers of the G allele and CC homozygotes (2.61–1.98 =  0.63 mmol/L, P = 0.001) but not below the median (1.43–1.51 = −0.08 mmol/L), consistent with the higher average triglyceride concentrations in those with the high plasma SFA concentrations (2.06 ± 0.06 vs. 1.50 ± 0.06 mmol/L).

Samoan triglyceride concentrations were elevated (average 1.18 mmol/L) if they consumed a modern dietary pattern, intermediate for a transitional dietary pattern (average 0.96 mmol/L), and low for a neo-traditional diet (average 0.83 mmol/L)67. Correspondingly, the difference between CC homozygotes and carriers of the T allele of the insulin induced gene 2 (INSIG2) rs9308762 differed significantly on the modern diet (0.88 mmol/L), showed intermediate difference for transitional diet (0.33 mmol/L), and showed no significant difference (−0.10 mmol/L) on the neo-traditional diet (Pinteraction  = 0.04)67.

Figure 6E presents Lin et al.’s report of a two-fold greater triglyceride increase in C-carriers of the APOA5 -1131T > C polymorphism vs. TT homozygotes in going from a 54% carbohydrate/31% fat diet to a 70% carbohydrate/15% fat diet68. Consistent with quantile-dependent expressivity, the genotype difference went from 0.13 ± 0.10 to 0.22 ± 0.10 mmol/L while average triglycerides increased from 0.83 ± 0.08 to 0.94 ± 0.05 mmol/L.

Figure 6F displays the significantly greater triglyceride decreases in LPL Hindlll H- carriers than H+ homozygotes when switching from a high saturated fat to a high polyunsaturated fat diet (0.35 vs. 0.10 mmol/L decreases, P = 0.05) reported by Humphries et al.69 However, the high polyunsaturated fat diet produced smaller differences between H- and H+ genotypes than the high saturated fat diet (0.31 vs. 0.56 mmol/L) in accordance with its lower average triglyceride concentrations (2.15 vs. 2.29 mmol/L).

Figure 7A presents Carvalho-Wells et al.’s finding that switching from a low-fat diet to a high-fat diet containing 3.45 g/d DHA produced significantly greater triglyceride reductions in APOE ε3ε4 heterozygotes (−0.48 ± 0.11 mmol/L) than ε3ε3 homozygotes (−0.22 ± 0.06 mmol/L, Pinteraction = 0.03). Average triglyceride concentrations were higher on the low-fat than high-fat diet (1.43 vs. 1.08 mmol/L), and the difference between genotypes was correspondingly greater on the low-fat than the high fat diet (0.33 vs. 0.06 mmol/l difference)70.

Figure 7
figure 7

Precision medicine perspective of different mean changes in triglyceride concentrations by genotypes (histogram inserts) vs. quantile-dependent expressivity perspective of larger genetic effect size when average triglycerides concentrations were high vs. low requiring nonparallel changes in triglycerides by genotype, for: (A) Carvalho-Wells et al.’s 2012 report on the triglyceride response to switching from a low-fat (24% fat, 59% carbohydrate) to high-fat diet (38% fat, 45% carbohydrate with 3.45 g DHA/d) in 44 APOE ε3ε3 homozygotes vs. 44 ε3ε4 heterozygotes (Pinteraction = 0.03)70; (B) Kang et al.’s 2014 report of switching from their usual to a refined rice diet in 43 TT homozygotes and 50 C carriers of the APOA5 -1131 T > C polymorphism71. (C) Vallée Marcotte et al.’s 2016 report on starting omega-3 (n-3) fatty acid supplementation in 142 AA homozygotes and 66 AC heterozygotes of the neurexophilin-1 (NXPH1) rs7806226 polymorphism72; (D) Vallée Marcotte et al.’s 2016 report on starting omega-3 (n-3) fatty acid supplementation in 155 TT homozygotes and 53 CT heterozygotes of the V-MYB avian myeloblastosis viral oncogene homolog (MYB) rs1115479472.

Full size image

Figure 7B presents Kang et al.’s report of significantly greater triglyceride increases from a refined rice diet in carriers of C-allele than TT homozygotes of the APOA5 -1131 T > C polymorphism (0.53 vs. −0.01 mmol/L, P = 0.02)71. Again, the difference between genotypes was greater after the diet than before (0.92 ± 0.04 vs. 0.38 ± 0.03 mmol/L difference) when average triglycerides were higher (2.03 ± 0.02 vs. 1.75 ± 0.01 mmol/L).

Finally, Fig. 7C,D present Vallée Marcotte et al.’s report of a significantly different triglyceride responses to 5 g/day of fish oil by the neurexophilin-1 (NXPH1) rs7806226 polymorphism (Pinteraction = 0.004) and V-MYB avian myeloblastosis viral oncogene homolog (MYB) rs11154794 polymorphism (P = 0.02)72. The histograms present the greater triglyceide reductions in homozygotes. The difference between genotypes was greater at baseline when average triglycerides concentrations were higher (1.21 ± 0.04 mmol/L) than after treatment when average concentrations were lower (1.02 ± 0.04 mmol/L) for both rs7806226 (AA-AC: 0.187 ± 0.07 vs. 0.018 ± 0.08 mmol/L) and rs11154794 (TT-TC: 0.087 ± 0.111 vs. 0.014 ± 0.075 mmol/L), consistent with quantile-dependent expressivity.

Alcohol

Triglycerides increase an average of 0.11 mmol/L per 23 g/d of alcohol consumed, equivalent to 1 large beer89. Although De Vries et al. study of 394,584 subjects revealed no gene-alcohol interactions attaining genomewide significance73, there are several reports of larger genetic effects in drinkers than nondrinkers in accordance with their higher average triglyceride concentrations and quantile-dependent expressivity. These include those by Ruixing et al. for APOC3 3238 C > G (rs5128) genotypes in drinkers (CC/CG/GG: 0.97, 0.95, 1.28 mmol/L, P < 0.001) vs. nondrinkers (CC/CG/GG: 0.91, 1.01, 0.93 mmol/L, P = 0.002)74, by Yin et al. for GG vs. A-carriers of the APOA5 457 G > A (rs3135507) polymorphism in drinkers (1.01 vs. 0.95) vs. nondrinkers (0.97 vs. 0.99 mmol/L, PInteraction < 0.001)75, by Pyun et al.’s between LPL rs263 GG homozygotes and A-carriers in drinkers (1.97–1.78 = 0.19 mmol/L) vs. nondrinkers (1.65–1.64 = 0.01 mmol/L, Pinteraction = 0.009)63, and by Zhou et al. for the cholesteryl ester transfer protein (CETP) TaqIB polymorphism (rs708272) in drinkers (B1B1/B1B2/B2B2: 1.42, 1.01, 0.88 mmol/L, P = 0.02) than nondrinkers (0.94, 1.17, 0.99 mmol/L, respectively, P = 0.18)76. In each case, average triglyceride concentrations were greater in drinkers than nondrinker (Ruixing and Yin et al.: 1.09 ± 0.03 vs. 0.97 ± 0.03 mmol/L, P = 0.0174,75; Pyun et al.: estimated 1.90 vs. 1.65 mmol/L63; Zhou et al.: 1.17 ± 0.09 vs. 1.04 ± 0.06 mmol/L76).

Tan et al. deduced a significant interaction between alcohol intake and the aldehyde dehydrogenase 2 gene (ALDH2) in their effect on triglyceride concentrations (P = 3.3 × 10−5)77. Specifically, the triglyceride difference between GG homozygotes and A carriers increased from nondrinkers (−0.09 mmol/L), to drinkers consuming 1–10 g/d (0.15 mmol/L), 10–30 g/d (0.26 mmol/L), to ≥30 g/d (0.51 mmol/L). However, average triglyceride concentrations also increased from nondrinkers (1.21 mmol/L), 1–10 g/d (1.21 mmol/L), 10–30 g/d (1.42 mmol/L), to ≥30 g/d (1.48 mmol/L) in support of a quantile-dependent expressivity.

Insulin resistance

VLDL overproduction due to diminished degradation of newly synthesized apo B, increased free fatty acid flux to the liver, and increased de novo hepatic lipogenesis all contribute to hypertriglyceridemia in T2DM86. Klimentidis et al.28 reported that the effect of GRSTG on triglyceride concentrations increased progressively with increasing tertiles of fasting insulin (estimated β = 0.15, β  = 0.21, β  = 0.23, P = 2.7 × 10−11) and HOMA-IR (estimated β = 0.14, β = 0.21, β  = 0.24, P = 2.5 × 10−11), in the context of a highly significant triglyceride increases with both (fasting insulin: P = 2.4 × 10−100; HOMA-IR: P = 9.1 × 10−133). Justesen et al. also reported that higher HOMA-IR was associated with a greater affect of GRSTG on triglycerides concentration (Pinteraction = 0.0009), presumably in association with rising average triglyceride concentrations, although the triglyceride-HOMA-IR relationship was not reported30.

Inamdar et al. reported that T2DM patients, who had higher average triglycerides than non-T2DM patients (1.90 vs. 1.27 mmol/L), showed greater carrier-noncarrier triglyceride differences for APOE ε2 (0.56 vs. 0.31) and ε4 (−0.45 vs. −0.12 mmol/L)78. Data presented by Vohl et al. showed that the fasting triglyceride difference between LPL-HindIII H + H + homozygotes and H-carriers were greater for fasting insulin concentrations ≥71.5 than  < 71.5 pmol/L (0.65 vs. −0.12 mmol/L), consistent with the higher triglyceride concentrations of the former (2.01 vs. 1.18 mmol/L)40.

Pregnancy

There is a two-fold increase in circulating triglyceride levels during the third trimester due to enhanced VLDL-production and LPL supression86. Ma et al. reported that the effect of LPL deficiency had a much greater effect during pregnancy, when triglycerides are normally two- to three-fold higher, than when not pregnant, i.e., the LPL deficient women’s triglyceride were 20.2–22.5 mmol/L when pregnant vs. 3.4 mmol/L when not79.

Twin studies

Higher average triglyceride concentrations in MZ vs. DZ twins (1.33 vs. 1.07 mmol/L) could have contributed to the higher triglyceride correlations (rMZ = 0.527 vs. rDZ = 0.349) reported by Jermendy et al., affecting their estimation of genetic and environmental influences80.

Limitations

An important limitation of the analysis of the Framingham data its reliance on the simple formula h2 = 2βOP/(1 + rspouse) and h2 =  [(1 + 8rspouseβFS)0.5 − 1]/(2rspouse) to estimate heritability12. These formula are unlikely to embody the true complexity of triglyceride inheritance. With respect to the published examples cited, we wish to emphasize that consistency with quantile-dependent expressivity does not disprove gene-environment interactions, rather, it provides an alternative interpretation. The examples presented are those originally interpreted from the perspective of precision medicine and biological interactions that might be more easily explained by quantile-dependent expressivity. It is not our contention that all triglyceride gene-environment interactions are explained by quantile-dependent expressivity. For example, Wojczynski et al.’s report of the significant effect (P < 0.0001) of the APOB rs676210 variant on the triglyceride response to fenofibrate would not be attributable to the quantile-dependent expressivity of Fig. 1 due to their being larger genotype differences post-treatment when triglycerides were low than pretreatment when triglycerides were high95. Some gene-environmental interactions may arise because triglycerides and environmental factors may be coregulated by shared genes or genes in strong linkage equilibrium. For that reason, the examples presented in Figs. 4–7 may be particularly informative for testing whether the genetic effect size is affected by average triglyceride concentrations because they represent intervention affecting triglyceride concentrations directly. Among the various genetic variants discovered to date, the proportion of the total triglyceride heritability explained by any specific SNP is too small to noticeably affect h25,6. Thus quantile-dependence of triglyceride heritability estimated from parent and offspring phenotypes does not necessarily describe the interactions between any particular genetic variant and its environment. Many published reports do not provide the information required to evaluate their consistency with quantile-dependent expressivity, namely unadjusted triglyceride concentrations by genotype and condition.

In conclusion, assuming Falconer and Mackay’s formula apply12, these analyses suggest that triglyceride heritability is strongly dependent upon whether an individual is high or low relative to the triglyceride distribution in the population. Alternatively, quantile-dependent shared environmental effects could also give rise to the increase in βOP and βFS with increasing average triglyceride concentrations, however our previous findings showing increasing genetic effect size for GRSTG2 and during post-prandial triglyceride increases3, and the studies cited herein13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69,70,71,72,73,74,75,76,77,78,79 support a genetic interpretation. Quantile-dependent expressivity potentially provides a common principle underlying a plethora of published gene-drug and gene-environment interactions. Specifically, rather than attributing these interactions on the basis of triglyceride metabolism, gene functionality, and the specific metabolic effect of adiposity, physical activity, insulin resistance, diet, smoking, alcohol, and pregnancy, quantile-dependent expressivity postulates that the impaired functionalities of these genetic variants are simply triglyceride concentration dependent.