Multiple solutions at the genomic level in response to selective breeding for high locomotor activity

Abstract

Replicate lines under uniform selection often evolve in different ways. Previously, analyses using whole-genome sequence data for individual mice (Mus musculus) from 4 replicate High Runner lines and 4 nonselected control lines demonstrated genomic regions that have responded consistently to selection for voluntary wheel-running behavior. Here, we ask whether the High Runner lines have evolved differently from each other, even though they reached selection limits at similar levels. We focus on 1 High Runner line (HR3) that became fixed for a mutation at a gene of major effect (Myh4^Minimsc) that, in the homozygous condition, causes a 50% reduction in hindlimb muscle mass and many pleiotropic effects. We excluded HR3 from SNP analyses and identified 19 regions not consistently identified in analyses with all 4 lines. Repeating analyses while dropping each of the other High Runner lines identified 12, 8, and 6 such regions. (Of these 45 regions, 37 were unique.) These results suggest that each High Runner line indeed responded to selection somewhat uniquely, but also that HR3 is the most distinct. We then applied 2 additional analytical approaches when dropping HR3 only (based on haplotypes and nonstatistical tests involving fixation patterns). All 3 approaches identified 7 new regions (as compared with analyses using all 4 High Runner lines) that include genes associated with activity levels, dopamine signaling, hippocampus morphology, heart size, and body size, all of which differ between High Runner and control lines. Our results illustrate how multiple solutions and “private” alleles can obscure general signatures of selection involving “public” alleles.

Introduction

By their very nature, complex traits can evolve in multiple ways. Thus, when a given form of directional selection is applied to replicate lines, adaptive responses are likely to be somewhat different (Mayr 1961; Cohan 1984a, 1984b; Tenaillon et al. 2012; Wone et al. 2019), a phenomenon often termed multiple solutions (e.g. see Bock 1959; Bennett 2003; Garland et al. 2011). These variable evolutionary pathways underscore the versatility of the genome and also provide opportunities for insight concerning the developmental and physiological mechanisms that underlie variation in complex traits.

The particular genomic and/or genetic features and processes that underlie a complex trait may affect the likelihood of multiple adaptive responses to a given type of selection. For example, duplications can create redundancy in genes, thus enabling altered function in 1 or both copies without detrimental effect on the organism. This has been seen in myosin MLC2 genes (Gerrits et al. 2012) and in hemoglobin (Natarajan et al. 2015; Storz 2016). Multiple solutions can also be modulated by highly impactful single-nucleotide polymorphisms (SNPs). A well-known example of this is malaria. Here, infection by a parasitic Plasmodium invokes typical immunological responses (Malaguarnera and Musumeci 2002), with a lethality rate of up to 30% in severe cases (i.e. multisyndromic and often manifesting as cerebral malaria, severe malaria anemia, and respiratory distress) (Karlsson et al. 2014). However, the sickle-cell mutation, which is an A-to-T substitution causing glutamate to be substituted with valine in the β-globin gene, is associated with substantial resistance to the disease in both heterozygotes and homozygotes, but with notable health detriments in homozygotes (Aluoch 1997; Griffiths et al. 2015). Despite the deleterious pleiotropic effects of the sickle-cell mutation, this allele has been favored by selection in populations where malaria is present (Karlsson et al. 2014), thus providing an alternative solution to the typical immunological responses.

One genomic feature that may promote multiple adaptive solutions is the presence of so-called genes of major effect (GOMEs), also referred to as major QTL, which are defined as genes whose allelic variants explain a large proportion of quantitative variation (Tanksley 1993). GOMEs may enhance the probability of divergent genomic pathways among replicate lines by affecting genetic variances and covariances (Agrawal et al. 2001; Garland 2003; Hannon et al. 2008; Stinchcombe et al. 2009). For example, Stinchcombe et al. (2009) demonstrated that the ERECTA allele in Arabidopsis thaliana had a small but clear impact on the G-matrix structure, although without a discernable impact on the response to selection. Epistatic genetic variance is also likely enhanced by the presence of GOMEs. Thus, if some populations have a given GOME and others do not, then they are likely to evolve genetically in somewhat different ways. (As noted in the Discussion, founder effects and random genetic drift can also increase the likelihood of multiple responses to selection.)

Replicated selection experiments offer excellent opportunities for discovering multiple adaptive responses to a defined and reproducible selective regime (Garland 2003; Garland and Rose 2009). Here, we test for multiple genomic responses to selection in the context of a replicated selection experiment that has a well-documented GOME that causes a phenotype termed mini-muscle (Garland et al. 2002: see below). Specifically, the High Runner (HR) mouse experiment includes 4 replicate lines of mice that have been bred (within family) for long daily distances of voluntary running on a wheel (HR3, HR6, HR7, and HR8) and 4 nonselected control lines (C1, C2, C4, and C5) (Swallow et al. 1998). A statistically significant response to selection could be detected by generation 6, and all lines reached selection limits around generations 17–27, running on average 2.5–3 times more than the control line (Careau et al. 2013). The 4 replicate HR lines vary in the extent to which daily running distance has evolved via increases in average speed vs duration of running, and a significant negative correlation between average running speed and duration of daily activity had evolved among the HR lines by generation 43 (Garland et al. 2011). For example, on average, mice from line HR3 (which became fixed for the allele underlying the mini-muscle phenotype) run faster but for fewer minutes per day than other HR lines, whereas the opposite is true for HR8 (see Fig. 3 in Garland et al. 2011).

Fig. 3.

a) Manhattan plot of the generation 61 individual mouse analyses, excluding HR3 (N = 5,931,993). Vertical, dashed lines included represent loci fixed for opposite alleles (155 total loci) in the C and HR lines (e.g. P = 2.43E−39). b) Manhattan plot of the generation 61 individual mouse analyses, excluding HR6 (N = 5,932,148). c) Manhattan plot of the generation 61 individual mouse analyses, excluding HR7 (N = 5,932,085). Vertical, dashed lines included represent loci fixed for opposite alleles (63 total loci) in the C and HR lines (e.g. P = 2.43E−39). d) Manhattan plot of the generation 61 individual mouse analyses, excluding HR8 (N = 5,931,663). Vertical, dashed lines included represent loci fixed for opposite alleles (485 total loci) in the C and HR lines (e.g. P = 2.43E−39). e) Manhattan plot of the generation 61 individual mouse analyses, all 8 lines (N = 5,932,148) (modified from Fig. 2 of Hillis et al. 2020).

Numerous differences among the HR lines have been identified at various points during the selection experiment, although these results have yet to be synthesized or approached from the perspective of a meta-analysis. These include pleiotropic effects attributable to the mini-muscle allele (Myh4^Minimsc) when in the homozygous state, in addition to differences that do not involve the mini-muscle phenotype. Those nonmini-muscle differences among the replicate HR lines have been documented for a variety of traits at the level of behavior, whole-animal performance, morphology, and physiology. For example, the HR lines have been shown to differ in both male-male (Klomberg et al. 2002) and maternal aggression (Gammie et al. 2003), as well as behavior in an open-field arena test (measuring aspects of exploration and risk-taking behavior) and in a plus maze (measuring aspects of anxiety) (Jónás et al. 2010). Among-line differences in performance and physiology have been documented for daily energy expenditure (Rezende et al. 2009), basal metabolic rate (Kane et al. 2008), endurance capacity during forced treadmill exercise (Meek et al. 2009), the ability to clear a parasitic nematode species (Nippostrongylus brasiliensis) from the small intestine (Malisch et al. 2009), and circulating corticosterone levels under both baseline conditions and after 40 min of restraint stress (Malisch et al. 2007), among other traits. Body mass differs among the HR lines (e.g. Klomberg et al. 2002; Hiramatsu et al. 2017), as do the masses of individual hindlimb muscles (controlling statistically for variation in body mass and even excluding those with the mini-muscle phenotype)(Houle-Leroy et al. 2003). Muscle fiber-type composition differs among lines and, at the level of muscle biochemistry, HR lines differ in the mass-specific activities of various metabolic enzymes (e.g. palmitoyl transferase, citrate synthase, cytochrome C oxidase) (Guderley et al. 2008). As these differences are in traits of functional relevance for endurance running, they suggest multiple solutions.

The mini-muscle phenotype noted above is caused by the recessive Myh4^Minimsc allele (a single base pair replacement) at the Minimsc locus in the eleventh intron of the Myosin heavy polypeptide 4 gene (chr11:67,244,850, GRCm38/mm10 assembly) (Kelly et al. 2013). The mini-muscle GOME was serendipitously discovered relatively early in the HR mouse selection experiment, based on systematic muscle dissections (Garland et al. 2002). The Myh4^Minimsc allele was uncommon in the base population (frequency ∼7%) and the phenotype has only been observed in 2 of the HR lines and in 1 control line (Garland et al. 2002; Syme et al. 2005). Of these lines, C5 apparently lost the allele to drift by generation 36 (Syme et al. 2005), HR3 becomes fixed for the Myh4^Minimsc allele by generation 36 (Garland et al. 2002; Syme et al. 2005), and HR6 has remained polymorphic for the allele through generation 98 (unpublished data; Cadney et al. 2021). Population genetic modeling indicated positive selection on the allele in the HR lines and either neutrality or negative selection in the C lines (Garland et al. 2002).

When present in the homozygous condition, the Myh4^Minimsc allele causes a 50% reduction of the mass of the triceps surae (calf) muscle, as well as total hindlimb muscle mass, earning it the name mini-muscle allele or phenotype (Garland et al. 2002; Hannon et al. 2008; Bilodeau et al. 2009; Kelly et al. 2013). The Myh4^Minimsc allele has a variety of pleiotropic effects when in the homozygous condition, such as increasing the mass of several organs, including the heart, spleen, liver, kidney, lung, stomach, and soleus muscle (Garland et al. 2002; Swallow et al. 2005; Syme et al. 2005; Guderley et al. 2006; Hannon et al. 2008; Kelly et al. 2017; and references therein), and altering the size and/or shape of various skeletal elements (Castro, Karakostis, et al. 2021; Castro, Rabitoy, et al. 2021). Possible effects in heterozygotes have not yet been studied. Perhaps most relevant for the concept of multiple responses to selection, mice from line HR3 (fixed for mini-muscle) and mini-muscle individuals in general tend to run faster but for fewer minutes per day as compared with the other HR lines (Kelly et al. 2006; Hannon et al. 2008; Dlugosz et al. 2009; Garland et al. 2011).

Loci with such far-reaching pleiotropic effects as mini-muscle have great potential to result in nonadditive epistatic effects with other genes, which may enhance their benefits or compensate their detriments (Pavlicev and Wagner 2012). Thus, we expected that the genomic basis of high voluntary wheel running in HR3—beyond the change in frequency of this 1 underlying allele—would differ from that of the other 3 HR lines. Although previous analyses involving all HR lines detected signatures of selection at various genomic regions (Hillis et al. 2020), we hypothesized that fixation for the Myh4^Minimsc allele in HR3 may mask additional signatures when this genetically divergent line is included in the analyses. To test this, we have repeated analyses using single-nucleotide polymorphism (SNP) data, dropping each of the HR lines. After confirming that dropping HR3 produced more novel selection signatures than when dropping any other HR line, we incorporated additional analyses used by Hillis et al. (2020) to highlight signatures of selection. Overall, our results illustrate how multiple solutions and “private” alleles (those unique to 1 or 2 lines) can obscure general signatures of selection involving “public” alleles (those present in all lines) (cf. Partridge and Gems 2002).

Materials and methods

HR mouse model

As described previously (Swallow et al. 1998; Careau et al. 2013; Hillis et al. 2020), 112 male and 112 female mice were obtained from Harlan Sprague Dawley (outbred Hsd: ICR strain) in 1993. Following 2 generations of random mating, 10 breeding pairs were randomly chosen to be founders for each of 8 closed lines (generation 0). Four of these lines were randomly designated as HR lines (lab designated HR3, HR6, HR7, and HR8), which would undergo selection based on voluntary wheel running. The remaining 4 lines would serve as unselected control (C) lines (lab designated C1, C2, C4, and C5) (Fig. 1). Each generation, all mice were given access to wheels at 6–8 weeks of age for 6 days. The highest-running (total revolutions on days 5 plus 6) male and female of each HR family were used to propagate the line (within-family selection, no sib-mating). This selection criterion was continued even after reaching selection limits at around generation 17–27 (Careau et al. 2013). The male and female from each C family were chosen randomly with respect to wheel running.

Fig. 1.

Schematic illustration of the HR mouse artificial selection experiment, begun in 1993 with a base population of 224 outbred mice.

Whole-genome sequencing

As described previously (Hillis et al. 2020), DNA was collected from 80 mice (10 from each line), from generation 61, via phenol-chloroform extraction and sequenced on an Illumina HiSeq 2500 1T platform. Libraries were constructed using Nextera kit and reads were trimmed and aligned to the GRCm38/mm10 mouse genome assembly as described in Didion et al. (2016). This generated an average read depth of 12× per mouse. SNPs were filtered to keep those with genotype quality (“GQ”) >5, read depth >3, MAF >0.0126 for all samples (as done by Hillis et al. 2020 to preserve all variable loci in data set), and Mapping Quality (“MQ”) >30. Of the 80 mice, 1 was excluded due to likely contamination, as in Xu and Garland (2017), leaving 79 for the following analyses. SNPs not found to be present in at least 2 of the 80 mice were also removed from analysis. This leaves 5,932,148 SNPs for analyses involving all 8 lines. The number decreased when dropping certain lines due to the remaining 7 lines being fixed for the same allele. Although Xu and Garland (2017) had identified these 80 mice from generation 61 as females, they were in fact all males with exception of 1 female from line C5.

Principal components analysis

Principal components analysis (PCA) was performed in R with the SNPRelate library (Zheng et al. 2012). Of the 5,932,148 variable SNPs across all lines (HR and C), we used 4,679,533 variable SNPs across the 9–10 mice within each of the HR lines.

SNP analyses excluding individual HR lines

To assess the hypothesis that fixation of the mini-muscle allele would cause HR3 to differentiate from the other HR lines in genomic regions relevant to wheel-running behavior, a mixed-model ANOVA was used to calculate differentiation between C and HR lines while dropping each of the other HR lines. The mixed-model ANOVA used minimum variance quadratic unbiased estimation (mivque) method of variance estimation (Rao 1971; Xu and Garland 2017). In addition, P-values and the Aikake Information Criterion corrected (AICc) for small sample sizes were calculated for 4 models with different variance structures (equal within-line variance for HR and C lines, equal among-line variance, both variances equal, and both variances different) and we then used AICc scores to choose the best model (following Hillis et al. 2020). The results of each of these analyses were then compared, with the expectation that more selection signatures would be present after dropping HR3 as compared with dropping any other HR line. “Differentiated regions” were defined by the following 3-step process. First, we identified all SNPs differentiated with P-value ≤0.001. Second, we considered that any 2 such SNPs within 1mbp of each other were part of the same region. Finally, we considered any gap between SNPs (with P ≤ 0.001) larger than 1 mbp as delineating separate regions.

Power and type I error simulations

All else being equal, dropping one of the 8 lines from the analyses would be expected to reduce the power to detect differentiation between the HR and C lines, due to the loss of a denominator degree of freedom. To estimate this expected drop in power, we performed simulations. Data reflecting the alternative hypothesis were simulated by taking a region from chromosome 17 that had been shown to be differentiated by Hillis et al. (2020). Approximately 22,700 SNP loci in this region (chr17:17,846,983–23,586,163) were variable and were differentiated across the region for the 8-line analyses (mean P-value = 0.104, median = 0.137, lowest = 7.54E−05, highest = 0.952). To generate simulated data, a variable locus was randomly sampled from the region, then the alleles for each line were created by randomly sampling (with replacement) from the alleles at that locus for that line. This was done for each of the 8 lines and the whole process was repeated to produce 100,000 simulated loci. Membership of each line within the set of either HR or C lines was always retained. Simulated data were analyzed using the multimodel ANOVA method (Hillis et al. 2020), first with all 8 lines and then dropping each of the HR lines 1 at a time.

For calculating relative type I error rate, data reflecting the null hypothesis were generated with a method similar to that for the power analyses. Alleles were sampled (with replacement) from a single line in the previously indicated chromosome 17 region, but then assigned to any of the 8 lines at random. This process was repeated for all 8 lines in sequence. A total of 100,000 loci were thus created, and multimodel ANOVA was performed with all 8 lines as well as dropping each of the HR lines.

Haplotype and nonstatistical analyses excluding HR3 (mini-muscle)

Following Hillis et al. (2020), we performed 2 additional analyses to gauge differentiation between the 4 C and 3 HR lines (excluding HR3). First, we used the haplotype data that were used by Hillis et al. (2020) and applied the mixed-model ANOVA method used for the SNP analyses, dropping HR3. A critical threshold of P ≤ 0.00526 was used for these haplotype analyses, following Hillis et al. (2020). Next, loci that were fixed for a given allele (either reference or alternative) for all HR lines (excluding HR3) and simultaneously polymorphic for all C lines, were identified as “FixedHR/PolyC” (as in Hillis et al. 2020). Any loci or genomic region identified as differentiated in all 3 tests (SNP ANOVA, haplotype ANOVA, and FixedHR/PolyC) are referred to as “consistent” regions and regarded as having the strongest evidence of differentiation. Selection signatures implicated by these analyses were compared to those implicated by analyses including all 8 lines (as reported in Hillis et al. 2020).

Gene annotations and knockout phenotyping

Gene annotations were determined using the University of California, Santa Cruz Genome Browser for GRCm38/mm10 (http://genome.ucsc.edu/, accessed Oct 2021) (Kent et al. 2002) and the Rat Genome Database for the GRCm38/mm10 mouse genome browser (https://rgd.mcw.edu/, accessed May 2022) (Smith et al. 2019). Mouse Genome Informatics’ Batch Query database was used for the knockout phenotyping (http://www.informatics.jax.org/batch/, accessed Nov 2021) (Bult et al. 2019).

Results

Principal components analysis

PCA across all lines (79 individuals) produced 7 eigenvalues >1: PC1 = 10.6 (13.6% of variance), PC2 = 9.4 (12.0% of variance), PC3 = 8.8 (11.3% of variance), PC4 = 8.4 (10.7% of variance), PC5 = 7.9 (10.1% of variance), PC6 = 6.9 (8.9% of variance), and PC7 = 6.8 (8.7% of variance). The 3D scatterplot of eigenvectors for PC1, PC3, and PC5 demonstrates a clear differentiation between the HR and C lines and also that HR3 differs from other HR lines (Fig. 2a).

Fig. 2.

Principal components analysis of variable SNP loci. a) Scatterplot of principle components (PCs) using all 8 lines: PC1, PC3, and PC5 account for a combined total of 35.0% of the variance. b) Bivariate scatterplot of scores on PC1 vs PC3 (49.4% of variance), from an analysis of variable SNP loci for the 4 HR lines only. c) PC1 vs PC3 (46.5% of variance). d) PC2 vs PC3 (43.3% of variance). As can be seen in (b) and (d), line HR3 is very different from the other 3 HR lines for scores on PC2.

PCA of the 39 individuals in the HR lines included 4,679,533 variable SNP loci and produced 3 eigenvalues >1: PC1 = 10.0 (26.3% of variance), PC2 = 8.8 (23.1%), and PC3 = 7.7 (20.2%). Line HR3 was remarkably different from the other 3 HR lines for scores on PC2 (Fig. 2, b–d).

SNP analyses excluding individual HR lines

Each of the analyses dropping 1 of the HR lines produced some new peaks, as compared with the original analyses (Fig. 3). However, dropping HR3 produced generally lower P-values across the genome than dropping any of the other HR lines (paired t-test, t = −149.91, −126.2, and −163.56, when comparing results after HR3 to dropping lines HR6, HR7, and HR8, respectively). The overall reduction in P-values when dropping HR3 is due largely to the increase in SNPs with P < 0.001, which is 4 times greater than in the analyses including all 4 HR lines (Table 1). More specifically, this difference is attributable mainly to a large increase in loci with P-values in this range in 2 genomic regions (chr3:46,438,071–52,624,971 and chr10:101,652,005–106,038,129). Both regions contain some loci with P-values ≤1e−03 after dropping any of the other lines; however, dropping HR3 produces about 40,000 additional loci with low P-values in these 2 regions (Table 2). Dropping HR8 resulted in a notable increase in loci with P-values in the 1e−06 to 1e−08 range (Table 1), largely due to a single region containing 1,414 loci with uniquely low P-values (chr7:115,169,726–116,129,821) (Table 2). This region contains Sox6, a gene whose knockout phenotypes include abnormal skeletal muscle fiber-type ratio (van Rooij et al. 2009), and was also identified in the original 8-line analyses (Hillis et al. 2020).

Table 1.

Number of SNPs below different P-value thresholds in generation 61 individual mouse analyses (N = 5,923,148).

P threshold	8-Line	Regions	No HR3	Regions	No HR6	Regions	No HR7	Regions	No HR8	Regions
0.05	248,549	455	315,791	474	283,789	467	275,305	488	296,948	433
0.01	87,600	129	132,865	196	100,845	185	88,311	193	97,760	173
0.001	16,844	44	68,372	75	16,561	70	23,107	63	21,256	69
0.0001	5,667	19	15,852	35	6,237	23	6,694	19	8,308	18
0.00001	3,522	6	4,862	19	3,924	12	1,364	11	2,806	9
0.000001	984	3	482	8	272	6	167	7	2,176	5
1.00E-07	2	1	215	6	7	4	138	5	1,148	4
1.00E-08	2	1	207	5	2	1	126	5	608	4
1.00E-09	0	0	155a	4	0	0	63a	1	485a	3

Number of SNPs below different P-value thresholds in generation 61 individual mouse analyses (N = 5,923,148). Results for the 8-line comparison are from Hillis et al. (2020). This includes 155 from the analyses that excluded line HR3 that were fixed for opposite alleles in the C and HR lines producing low P-values. See Fig. 3 for the Manhattan plot of these regions (Fig. 3e for 8-line Manhattan plot).

^aThese SNPs have very low P-values (2E−39) due to fixation for opposite alleles in the C vs HR lines.

Table 2.

Genomic regions with different response to selection among the 4 replicate HR lines.

Chr	minPOS	maxPOS	Width (bp)	loci_noHR3	loci_noHR6	loci_noHR7	loci_noHR8	Significant SNPs in 8-line analysis
1	155,052,375	157,767,127	2,714,753	1,063	NA	NA	NA
1	163,002,979	163,450,173	447,195	NA	18	NA	NA
1	189,994,733	190,372,872	378,140	NA	NA	NA	277
2	42,643,461	43,450,389	806,929	744	NA	NA	NA
2	141,166,416	145,485,766	4,319,351	NA	30	NA	NA
3	35,723,757	36,887,906	1,164,150	631	NA	NA	NA
3	46,438,071	52,624,971	6,186,901	18,517	1,167	1,169	2,126	x
3	75,213,797	76,058,839	845,043	NA	NA	NA	324
3	148,414,742	148,630,252	215,511	240	NA	NA	NA
4	87,042,703	87,114,728	72,026	64	NA	NA	NA
4	96,750,866	97,151,452	400,587	307	NA	NA	NA
4	98,501,069	99,358,924	857,856	126	NA	NA	NA
5	61,441,085	63,038,738	1,597,654	NA	58	NA	NA
5	103,680,064	104,223,799	543,736	NA	38	NA	NA
5	118,866,355	119,908,192	1,041,838	3	2	816	3	x
6	23,817,774	25,279,190	1,461,417	NA	NA	1,807	NA
6	144,501,313	144,545,689	44,377	32	NA	NA	NA
7	115,169,726	116,129,821	960,096	NA	NA	NA	1,414
7	117,620,491	119,188,853	1,568,363	372	NA	NA	NA
9	38,651,820	39,097,109	445,290	NA	260	NA	NA
9	47,437,476	48,268,374	830,899	109	NA	NA	NA
10	101,652,005	106,038,129	4,386,125	23,176	6	5	23	x
11	64,166,197	64,725,826	559,630	NA	NA	167	NA
11	73,267,237	74,873,424	1,606,188	551	NA	NA	NA
11	79,985,635	80,596,899	611,265	1	1	383	1	x
11	82,262,541	82,854,023	591,483	NA	NA	198	NA
11	114,193,898	114,629,516	435,619	NA	NA	241	NA	x
12	71,753,929	72,115,917	361,989	NA	NA	24	NA
13	48,002,204	48,725,034	722,831	222	NA	NA	NA
13	57,093,620	57,687,257	593,638	NA	387	NA	NA
13	115,026,145	115,159,660	133,516	NA	58	NA	NA
14	51,204,847	54,600,493	3,395,647	11	10	195	6	x
14	111,028,017	111,103,065	75,049	NA	855	NA	NA
15	38,434,855	40,095,009	1,660,155	NA	NA	1,115	NA
15	61,388,625	61,598,676	210,052	175	NA	NA	NA
15	69,108,096	71,591,605	2,483,510	40	496	36	2,414	x
16	31,440,034	33,322,583	1,882,550	144	772	103	1,999	x
16	87,078,841	88,667,433	1,588,593	5,078	NA	NA	NA
17	44,043,299	44,712,180	668,882	3,480	NA	NA	NA
17	67,989,641	68,777,695	788,055	NA	NA	NA	212
17	87,221,754	87,282,285	60,532	168	NA	NA	NA
17	88,660,339	88,885,305	224,967	NA	NA	632	1
18	59,052,362	60,490,520	1,438,159	31	NA	NA	NA
19	18,635,879	18,895,311	259,433	NA	NA	480	NA
19	53,052,787	53,102,861	50,075	NA	NA	11	NA

Genomic regions (N = 45; and counts of SNP loci within those regions) that became significant at the P ≤ 0.001 level (1) for at least 10 SNPs after dropping the indicated line but no other HR line (e.g. chr1:155,052,375–157,767,127) and (2) for substantially more SNPs than any other dropped line (e.g. chr3:46,438,071–52,624,971).

Although Table 1 seems to generally show that dropping HR3 produces more differentiated regions (N = 75) than dropping any other HR line (N = 63–70), some of these regions will contain only 1 or a few SNPs, which may be a result of sampling error and thus a Type I error (see section on Type I error, below). Therefore, Table 2 and Supplementary Table 1 concentrate on those regions with at least 10 SNPs with P ≤ 0.001.

Supplementary Table 1 contains all regions with at least 10 SNPs with P ≤ 0.001 for any analyses where an HR line was dropped. Dropping HR3 from the analyses resulted in 34 such regions, which is more than those identified after dropping any of the other HR lines (noHR6 = 23 regions, noHR7 = 27 regions, and noHR8 = 19) and also more than the 21 regions that were produced when analyzing all 8 lines.

The 45 regions listed in Table 2 are a subset of those shown in Supplementary Table 1, excluding regions where similar numbers of SNPs with P ≤ 0.001 were produced when dropping any HR line. The regions in Table 2 are highlighted because they are where the HR lines responded differently from each other to the selection protocol. This leaves regions with (1) at least 10 SNPs with P ≤ 0.001 after dropping only 1 specific HR line (e.g. chr1:155,052,375–157,767,127, see Fig. 4 for illustration) or (2) a substantial increase in significant loci when dropping a specific line (e.g. chr3:46,438,071–52,624,971). Of the 45 regions listed in Table 2, dropping the line fixed for Myh4^Minimsc (HR3) produced more of these unique regions than dropping any other line (19 regions for HR3, 8 regions for HR6, 12 regions for HR7, and 6 regions for HR8).

Fig. 4.

Allele frequencies for individuals SNPs within example regions from Table 2. a) Chr1:155,052,375–157,767,127 illustrates a region that was detected as differentiated only after dropping line HR3, which has allele frequencies similar to the 4 control lines. b) Chr1:163,002,979–163,450,173 illustrates a region that was detected as differentiated only after dropping line HR6. c) Chr1: 189,994,733–190,372,872 illustrates a region that was detected as differentiated only after dropping line HR8.

Although none of the SNPs were fixed for opposite alleles between all 4 HR and 4 C lines (Hillis et al. 2020), dropping individual lines did produce some loci where the remaining C and HR lines were fixed for opposite alleles (Table 3). When dropping HR3, 155 SNPs are fixed for opposite alleles between the C and HR lines, clustered in 4 regions. Dropping any other HR line produces 0–3 such regions (Table 3).

Table 3.

Fixation for opposite alleles after dropping a single line.

Dropped line	Chr	First SNP	Last SNP	Size (bp)	Total fixed	Identified in 8-line analyses	Nearby genes
C1	16	4,429,565	5,003,974	574,410	43	No	Srl, Tfap4, Glis2, Pam16, Coro7, Vasn, Dnaja3, Nmral1, Hmox2, Cdip1, Ubald1, Mgrn1, Nudt16l1, Anks3, Sept12
C2	5	105,140,011	105,818,176	678,166	5	Yes	Gbp10, Gbp6, Gbp11, Lrrc8b, Lrrc8c, Lrrc8d
C2	11	78,446,157	79,935,912	1,489,756	997	Yes	Slc46a1, Sarm1, Vtn, Sebox, Tmem199, Poldip2, Tnfaip1, Ift20, Tmem97, Nlk, Fam58b, Lyrm9, Nos2, Lgals9, Ksr1, Wsb1, Nf1, Omg, Evi2, Evi2b, Evi2a, Rab11fip4, Utp
C4	5	109,133,489	109,133,489	1	1	Yes	Vomeronasal region
HR3	3	51,572,179	51,606,225	34,047	149	Yes	Setd7
HR3	5	108,438,229	108,438,230	2	2	Yes	Atp5k
HR3	5	133,019,521	133,451,500	431,980	3	No	None
HR3	14	39,477,931	39,477,931	1	1	Yes	Nrg3
HR7	5	119,606,810	119,896,957	290,148	63	Yes	Tbx3, Tbx5
HR8	7	115,139,726	115,344,641	204,916	472	No	Sox6 a
HR8	9	42,193,286	42,434,034	240,749	11	Yes	Sc5d, Tecta, Tbcel
HR8	16	14,215,165	14,473,951	258,787	2	Yes	Myh11, Fopnl, Abcc1

Frequency of fixation for opposite alleles in C and HR lines after dropping different lines. We do see unique regions when dropping other lines. Although more regions are identified when dropping HR3 than any other line, this is not by a large margin. However, as 8-line analyses (Hillis et al. 2020) identify 3 of these regions, at least some of these fixed regions are likely biologically significant across 3, if not all 4, HR lines.

^a Sox6 was mentioned by Hillis et al. (2020) as being a suggestive gene that the mixed-model analyses struggled to identify due to low within-line variance

Power and type I error simulations

Dropping any one of the 8 lines generally reduced the number of P-values lower than 0.05 and lower than other relevant significance thresholds, although the difference was sometimes negligible (Supplementary Table 2). Overall, these comparisons suggest that, as would be expected, the statistical power to detect differentiation between the HR and C lines is reduced when an HR line is excluded from the analyses.

The relative change in P-value appears to increase as the P-value decreases. For example, those loci whose 8-line analyses produced a P-value in the 0.05 < P < 0.01 range, had an average increase in P-value by about 4.5% when a line was dropped, whereas loci whose 8-line analyses produce a P-value <1.00E−05 had an average increase of about 36.6% in P-values when a line was dropped.

Table 4 illustrates that type I error rates for α = 0.05 are deflated in the 8-line analyses, as was noted previously (Hillis et al. 2020), and a similar deflation occurs for α = 0.01. Dropping an HR line from the analyses increases the type I error rate for both α = 0.05 and α = 0.01 (Table 4). For α = 0.001, the type I error rates were inflated for both the 8-line analyses (0.00319) and when dropping a line (range = 0.00276–0.00286), and even more so for α = 0.0001 (range = 0.00060–0.00078). For some of the P thresholds (e.g. P ≤ 0.001), the increase was quite large relative to the type I error rate for the 8-line analyses (Table 4).

Table 4.

Type I error rates and estimated true positives.

P threshold	Expected	8-Line type I errora	8-Line true positivesb	noHR3 type I error	noHR3 true positives	noHR6 type I error	noHR6 true positives	noHR7 type I error	noHR7 true positives	noHR8 type I error	noHR8 true positives
0.05	5,000	677	3,512.9	2,009	3,314.4	2,014	2,769.9	2,021	2,619.9	1,949	3,056.7
0.01	1,000	443	1,033.7	654	1,585.7	647	1,053.0	640	848.7	624	1,024.0
0.001	100	319	−35.1c	278	874.6	276	3.2	284	105.5	286	72.3
0.0001	10	67	28.5	64	203.2	60	45.1	61	51.8	78	62.1
0.00001	1	3	56.4	7	75	8	58.1	7	16.0	12	35.3
0.000001	0.1	0	16.6	1	7.1	0	4.6	1	1.8	1	35.7
1E-07	0.01	0	0	0	3.6	0	0.1	0	2.3	0	19.4
1E-08	0.001	0	0	0	3.5	0	0	0	2.1	0	10.2
1E-09	0.0001	0	0	0	2.6	0	0	0	1.1	0	8.2

^aType I error columns include the number of false positives for a given P threshold after sampling alleles from a given line and randomly shuffling lines between the linetypes (see Materials and Methods).

^bEstimated true positive in the real data calculated by converting the number of loci found differentiated at a given P threshold into a ratio out of 100,000 loci and subtracting from these loci the number of estimated type I errors.

^cNegative values (i.e. fewer total positives than predicted false positives).

To compare type I error rate to the P-values for the real data, total P-values below each of these thresholds (found in Table 1) were scaled to be out of 100,000 to match the simulation. When the estimated type I error rate (Table 4) is subtracted from the frequency of calculated total positives (scaled from Table 1), many signatures of selection remain, particularly when dropping HR3 (Table 4).

Haplotype and nonstatistical analyses excluding HR3 (mini-muscle)

Hillis et al. (2020) had identified 13 “consistent” regions (i.e. differentiated in SNP ANOVA, haplotype ANOVA, and FixedHR/PolyC) when performing analyses using all 8 lines. All 13 of those regions are listed in Table 5, including 1 region (chr16:40,742,298–41,357,426) that was inadvertently not identified as consistent by Hillis et al. (2020). When dropping HR3 from the analyses, 17 regions were identified as consistent, 7 of which were not identified in the 8-line analyses by any of the 3 analytical methods (Table 5). These 7 regions included genes associated with systems known to be different in the HR lines as compared with the C lines, including skeletal, heart, and neuronal development (see Discussion). For completeness, Table 5 also lists 15 additional regions identified by at least 2 of the 3 analytical approaches when dropping HR3.

Table 5.

Regions identified by analyses excluding HR3.

Chromosome	Low BP	High BP	SNPs	Haplotype	FHRPC	Identified in 8-line
1	156,378,655	156,798,297	x	x
2	42,722,345	43,214,647	x	x	x	LM
3	35,754,848	36,551,567	x		x	LM
3	46,440,016	51,798,758	x	x	x	LM, Hap
3	148,414,742	148,638,738	x	x
4	57,735,866	58,314,736	x		x	LM, FHRPC
4a	96,840,333	98,580,378	x	x	x
4	155,480,343	155,654,426	x	x		LM
5	104,331,989	106,773,650	x		x	FHRPC
5	108,000,623	109,929,498	x	x	x	ALL
6	37,440,411	39,293,531		x	x	FHRPC
6	41,584,862	42,841,617	x	x	x	ALL
6	43,004,011	43,431,434	x	x		Hap
6	143,921,896	144,545,685	x		x
7	29,603,841	30,672,131		x	x	ALL
7a	119,020,988	119,188,853	x	x	x
8	121,821,514	126,241,684		x	x
9	41,240,184	42,314,823	x	x	x	ALL
9a	48,043,076	50,172,437	x	x	x
10	75,061,742	75,698,409		x	x	LM
10	101,672,276	105,807,985	x	x	x	LM, Hap
11a	74,340,021	74,838,159	x	x	x
11	78,166,291	80,218,837		x	x	ALL
11	112,075,283	114,258,210	x	x	x	ALL
12	88,904,818	92,138,006	x		x	LM, FHRPC
13a	48,002,204	48,380,363	x	x	x
13	56,280,106	56,618,730	x	x
14	52,072,148	53,779,979		x		ALL
14	97,210,445	99,269,397		x	x	ALL
15	18,196,530	20,632,943	x	x	x	ALL
15	61,361,016	61,653,459		x	x	FHRPC
15	70,665,271	71,547,239	x	x	x	ALL
16	31,686,517	31,779,261		x	x	ALL
16	40,742,298	41,357,426		x		ALLb
16	87,080,074	88,667,433	x	x
17	17,661,786	23,599,776	x	x	x	ALL
17a	44,063,088	44,685,529	x	x	x
18a	56,147,892	60,494,440	x	x	x

This table contains regions identified by at least 2 of the 3 analytical methods used. The column on the far right indicates which of these regions were identified by which method for the Hillis et al. (2020) 8-line analyses (LM = local maximum, implicating most significant SNPs of a suggestive region).

^aThese were regions implicated by all 3 analytical methods (mixed-model analyses of individual SNPs and haplotypes, as well as FixedHR/PolyC) after dropping HR3, but not by any method with all 8 lines.

^bThis region was accidentally not implicated as a consistent region by Hillis et al. (2020) but meets the criteria.

Discussion

In the present study, we took advantage of the serendipitous discovery of a gene of major effect, named mini-muscle, which is part of the adaptive response to selection for high voluntary wheel running (see Introduction). Given its major effect on muscle mass and fiber-type composition, the observation that mini-muscle mice (and line HR3 in general) tend to run faster but for fewer minutes per day, as well as its pervasive pleiotropic effects on other behaviors, physiological traits, and organ sizes, we hypothesized that line HR3, which became fixed for the Myh4^Minimsc allele, would show evidence of multiple solutions at the genomic level, as compared with the other 3 HR lines. Our results provide substantial support for this hypothesis, and encourage the application of similar analytical approaches to other replicated selection experiments.

SNP analyses excluding individual HR lines

Much of the increase in significant SNPs that we see when dropping HR3 can be attributed to 2 regions, chr3:46,438,071–52,624,971 and chr10:101,652,005–106,038,129 (Table 2), which had been identified by the 8-line analyses (Hillis et al. 2020). Because these regions were also detected with the 8-line analyses, it stands to reason that they responded to selection in all 4 HR lines. However, the wider areas implicated by the other 3 HR lines may correspond to stronger selection and hence a faster response to selection, as compared with HR3, thus not allowing sufficient time for recombination to break the haplotype in the other 3 HR lines (Smith and Haigh 1974; Kaplan et al. 1989; Kim and Stephan 2002).

Multiple solutions among the HR lines

A new genomic region that emerges as statistically significant only after dropping one of the HR lines (i.e. 4 C lines vs 3 HR lines) implies: (1) the region is likely relevant to wheel-running behavior (though not as strongly supported as genomic regions identified with all 4 HR lines) and (2) the HR line that was dropped does not show the same response to selection as the other 3 HR lines (Fig. 5). Therefore, each of the 37 new regions listed in Table 2 may be thought of as relevant to voluntary wheel running in 3 of the 4 HR lines, thus providing evidence of “multiple solutions” at the genomic level.

Fig. 5.

Illustration of different analysis strategies for detection of “private” alleles. The 4 possibilities shown include all 4 HR lines vs all 4 C lines [as was done by Hillis et al. (2020)], 3 HR vs 4 C (as is done in the present study), as well as 1 HR vs 4 C and 1 HR vs 3 HR (both of which are expected to have increased type I error rate as compared to the previous 2 analyses).

Possible explanations for different responses to selection among the HR lines include:

• Founder effects. Different starting allele frequencies (i.e. founder effects, sensu Mayr 1942) could alter the response to selection (James 1970; Simões et al. 2008). For example, if certain biologically significant loci were already fixed or close to fixed in a given line, then that line would be forced to respond to selection via changes at other loci. The Myh4^Minimsc allele was present in the base population at a frequency of ∼7%, and so may have been absent in some lines (Garland et al. 2002), although the probability is low even for lines that were not observed to have the phenotype (∼0.07, based on calculation of posterior probabilities). Indeed, the phenotype was only ever observed in 1 C line and in 2 HR lines (e.g. see Fig. 1 in Garland et al. 2002).

• Random genetic drift. Following the founding of a small population, if the effective population size remains low, then drift may eliminate an allele despite some positive selection (or fix an allele despite some negative selection). This would be especially likely to occur for an allele that was present at a low frequency when the experiment began, such as Myh4^Minimsc. Thus, drift can exacerbate founder effects and constrain the genetic options available to a given population.

• Epistatic effects. If an allele with large epistatic effects (nonadditive interactions with alleles at other loci) increases in frequency within a given line, then substantial changes in allele frequencies at the epistatically related loci would be expected. For example, if allele A at the A locus positively affects wheel running, then alleles at other loci that increase wheel running only when allele A is present will be favored by selection only when allele A is present. If allele A were present in only some populations under uniform selection, then the likelihood of multiple adaptive response would be increased.

• Selection limits and constraints. Suppose that mice are subject to a constraint on wheel running caused by joint pain: they stop running when the pain becomes intolerable. In this scenario, joint pain is sufficient to limit wheel running and it serves as a “weak link” in the physiological and neurobiological systems that are required for high levels of wheel running. Then suppose 10 alleles located at 10 independent biallelic loci, with entirely additive effects, are capable of increasing wheel running. Suppose further that only 5 such alleles are needed to achieve the amount of wheel running that causes intolerable joint pain. In such a scenario, fixation of the favorable allele at any 5 of the loci will coincide with a selection limit, but these alleles may be different among replicate lines.

Signatures of selection after dropping 1 line at a time

Although no loci were fixed for opposite alleles between the HR and C linetypes in Hillis et al. (2020) when considering all 8 lines, dropping any 1 HR or C line from the SNP analyses usually produced loci fixed for opposite alleles between the different linetypes (Table 3). These SNPs unsurprisingly tend to be clustered into specific regions (separated by at least 1 mbp), some of which have been detected either in the present study or by Hillis et al. (2020). Most of the regions listed in Table 3 were also identified by the 8-line SNP analyses, which may suggest that the dropped line is not drastically different from the others within its linetype. However, 3 new regions emerge.

The first new region is seen after dropping C1 (chr16:4,429,565–5,003,974) and contains various genes whose knockouts have been associated with heart morphology (Yoshida et al. 2005; Hayashi et al. 2006; Cota et al. 2006; Dickinson et al. 2016). Since all of the HR lines fixed for the same allele, this would not be an example of different responses to selection, but an example of variation among control lines disrupting our ability to detect selection signatures in the HR lines.

The second and third regions were identified by dropping HR3 and HR8, respectively (Table 3). These regions might implicate different responses to selection among the HR lines. One of these regions contains the Sox6 gene described above and by Hillis et al. (2020) for its effect in regulating muscle fiber type, hematopoiesis, bone growth and heart function (Smits et al. 2001; van Rooij et al. 2009). While 3 of the HR lines were fixed for the reference allele, HR8 became fixed for the alternate allele. The region identified when dropping HR3 (chr5:133,019,521–133,451,500) does not contain any annotated sequences. Some possible explanations include: a relevant gene being present but simply not yet annotated; this region serving an unknown regulatory role for other genes; or this region having undergone this fixation pattern purely by drift (i.e. it does not influence running behavior). One potential gene regulated by this region would be Auts2 (approximately 480 kbp downstream of the region), which has been implicated in neurodevelopment (Oksenberg and Ahituv 2013). Auts2 is also thought to be associated with the Runx1 pathway: an intriguing association when Runx2 is found in a separate region identified when dropping HR3 (Table 6).

Table 6.

Genes identified by all analyses, after excluding HR3

Chromosome	Low BP	High BP	Genes within region
4	96,840,333	98,580,378	Nfia, Tm2d1, Patj
7	119,020,988	119,188,853	Gpr139, Gprc5b (downstream)
9	48,043,076	50,172,437	Nxpe2, Nxpe4, Nxpe1-ps, Rexo2, Rbm7, Nnmt, Zbtb16, Htr3a, Htr3b, Usp28, Cldn25, Zw10, Tmprss5, Drd2, Ankk1, Ttc12, Ncam1
11	74,340,021	74,838,159	Olfr411, Olfr412, Rap1gap2, Ccdc92b, C1uh, Pafah1b1, Mettl16, Mnt
13	48,002,204	48,380,363	Id4
17	44,063,088	44,685,529	Enpp4, Enpp5, Clic5, Runx2, Runx2os3
18	56,147,892	60,494,440	Gramd3, Aldh7a1, Mir1258, Phax, Tex43, Lmnb1, Marchf3, C330018D20Rik, Megf10, Prrc1, Ctxn3, Ccdc192, Slc12a2, Fbn2, Slc27a6, Isoc1, Adamts19, Minar2, Chsy3, Mir6355, Iigp1, Smim3

These regions were implicated by all 3 analyses (individual SNP, haplotype, and fixation in HR/polymorphic in C) only after dropping line 3 from the analyses. Brief descriptions of these genes are presented in Supplementary Table 3.

Variation in olfactory response to selection

Olfaction is known to play an important role in some motivated behaviors. Our previous 8-line analyses showed that several vomeronasal genes have responded consistently to selection (Hillis et al. 2020; Nguyen et al. 2020). The vomeronasal organ is part of the overall olfactory system and functions primarily to detect nonvolatile organic compounds. Table 2 includes regions with genes that have an olfactory, but not vomeronasal, function. These regions were identified when dropping lines HR3, HR6, or HR7, with a different region appearing important after dropping each of the lines (HR3 = chr11:73,267,237–74,873,424; HR6 = chr9:38,651,820–39,097,109; HR7 = chr14:51,204,847–54,600,493). We interpret this as evidence for multiple solutions occurring in a given physiological system, at 2 different levels. In other words, olfaction seems to be important in the evolution of the HR phenotype (Dewan et al. 2019), and this may occur by either vomeronasal or nonvomeronasal pathways (or both). Although multiple vomeronasal genes in multiple regions on multiple chromosomes were identified in the previous 8-line analyses, here we did not find evidence of differences among the HR lines for these genes. However, we did find that multiple nonvomeronasal olfactory genes seem to have been important in the response to selection, and with different genes being important in different HR lines.

Power and type I error simulations

An increase in type I error rate when dealing with low sample size is not a new observation for some types of genetic data (Baldi and Long 2001). In any case, the inflated type I error rate may draw into question some of our “significant” results for the 7-line analyses in Table 2. To gauge the magnitude of this problem, we subtracted the expected false positives (Table 4) from our total positives (scaled from Table 1). As shown in Table 4, dropping HR3 produces many more P-values of 0.001 or lower than expected under the null hypothesis, and more than when dropping any of the other lines. This observation increases our confidence that the genomic response to selection by line HR3 truly differs from that of the other 3 HR lines.

Chromosomal regions identified when excluding HR3

Despite the many expected similarities between the 8-line analyses and analyses dropping HR3, the present study identifies 7 genomic regions implicated by all 3 tests (SNPs, haplotype, and FixedHR/PolyC) that were not identified by any tests when analyzed with all 8 lines (Table 6). These regions contained nearly 61 genes; however, 3 in particular caught our attention, Ncam1, Drd2, and Minar2.

Ncam1 codes for a cell adhesion protein whose knockouts are associated with altered hippocampus, cerebellum, and olfactory bulb development (Tomasiewicz et al. 1993; Holst et al. 1998), as well as shortened circadian period (Shen et al. 2001). Differential circadian rhythms have been found by Koteja et al. (2003), who showed that HR mice have a shorter free-running period (tau) under both constant dark and constant light. In addition, human GWAS have implicated Ncam1 in playing a role in heel bone mineral density (Kim 2018; Morris et al. 2019) and addictive behaviors, specifically smoking (Kichaev et al. 2019; Karlsson Linnér et al. 2019; Liu et al. 2019). Several bone differences between HR and C lines have been documented (particularly in limb bone size and shape). This includes a number of differences between mini- and normal-muscled mice (Kelly et al. 2006; Middleton et al. 2008, 2010; Wallace et al. 2010, 2012; Castro, Rabitoy, et al. 2021). Moreover, HR mice show withdrawal symptoms when wheel access is removed (Kolb et al. 2013).

Hippocampal function in the HR lines has been explored through indirect methods (Rhodes et al. 2003; Johnson et al. 2003; Bronikowski et al. 2004). For example, Bronikowski et al. (2004) found some genes related to transcription and translation that had increased expression in the hippocampus in HR vs C lines, whereas some associated with neuronal signaling and immune function had decreased expression in HR mice. The HR lines also had increased brain-derived neurotrophic factor in the hippocampus after having access to wheels for 7 days (Johnson et al. 2003). As for response to wheel running, the C lines showed a positive correlation between wheel running and neurogenesis in the dentate gyrus of the hippocampus, whereas the HR mice did not, with all HR mice having a high level of neurogenesis (Rhodes et al. 2003). Moreover, wheel access improved learning in the Morris water maze for C mice but not for HR mice. With body mass as a covariate, Schmill (2021) found that the total volume of the hippocampus is larger in HR than in C mice, both for animals housed with wheels for 10 weeks and those housed without wheels.

Drd2 is a dopamine receptor that has been associated with a wide variety of disorders, addictions, and compulsive behaviors (Blum et al. 1995; Hung Choy Wong et al. 2000; Noble 2003; Bronikowski et al. 2004; Munafò et al. 2004; Foll et al. 2009). Drd2 has also been tied to wheel running in mice based on differential expression in high- and low-running lines (C57L/J and C3H/HeJ, respectively) (Dawes et al. 2014). In addition, Drd2 knockouts have altered wheel-running behavior (Roberts et al. 2017).

When line HR8 was compared to stock ICR mice, significant differences in expression of Drd1a and Drd2 receptors (downregulated in HR8) were found in the dorsal striatum (Mathes et al. 2010). In addition, HR and C mice in the wheel-running response to cocaine (Rhodes et al. 2001). Though surprisingly, Drd2 receptor antagonist does not appear to cause a different response in the HR lines than control (Rhodes and Garland 2003); however, this study did not separate HR3 or other mini-muscle mice in the analyses.

Minar2 is a NOTCH2-associated receptor whose knockouts have been associated with altered bone structure (Dickinson et al. 2016), impaired coordination and gait (Ho et al. 2020), decreased body mass and length (Dickinson et al. 2016), and loss of dopaminergic neurons (Ho et al. 2020). Mice from the HR lines are generally smaller than the C lines, and differ in bone properties (see above), dopaminergic function (see above), and some aspects of gait during treadmill running (e.g. see Swallow et al. 1999; Rhodes et al. 2001; Girard et al. 2007; Garland et al. 2011; Claghorn et al. 2017).

Limitations of the present study and concluding remarks

Given the complexity of voluntary wheel-running behavior, identical evolutionary pathways in the 4 replicate HR lines would be highly unlikely. The fixation of the Myh4^Minimsc allele in just line HR3 is a clear example an alternative “solution” to selection that favors high-activity levels. Here, we show that the other 3 HR lines also show evidence of somewhat unique responses to selection (Table 2). However, HR3 seemingly stands out from the rest of the HR lines. As explained in the Introduction, a plausible explanation for this is that the Myh4^Minimsc allele has such large direct and pleiotropic effects (particularly in systems relevant for wheel running) that much of the rest of the genome has had to evolve differently in response. We would also note that HR3 has higher heterozygosity than any other line (including C lines) (Hillis et al. 2020).

Although the mixed-model method using multiple models and mivque variance estimation seems to be a relatively powerful method of analyzing these data (Xu and Garland 2017; Hillis et al. 2020), dropping lines negatively impacts power and inflates type I error rates (Table 4). More powerful analytical methods may need to be developed to better identify signatures of selection. One possibility may be to incorporate inferences similar to those described by Baldi and Long (2001) to offset the low sample size. Genomic data from generations closer to when the selection limit was reached may also reduce the type I errors produced by drift, allowing for better detection of true positive results. In addition, the present study does not perform any functional analyses of the suggested genes to establish a causal relationship between the gene and wheel-running behavior or other phenotypes suggested by KO studies (see above). Further studies are needed to establish these functional connections within the HR mice or at least to demonstrate that KO mice for these genes differ from wildtype in wheel running when measured under conditions similar to those used in the HR selection experiment.

A noteworthy question that the present study does not address is: why did HR3 become fixed for the Myh4^Minimsc allele while HR6 has remained heterozygous despite continued selection? Possible explanations for this include heterozygote advantage or epistatic interactions with loci unique to HR6. These ideas could be tested by genomic analyses of current or historical (e.g. see Kelly et al. 2013) samples and associating genotype with wheel-running and other relevant phenotypes. In addition, the differences between mini-muscle and normal-muscled individuals for some muscle properties are greater in HR3 than in HR6 (Guderley et al. 2006), suggesting that selection favoring this phenotype may have been stronger in HR3.

Despite the limitations discussed above, the present study was able to identify 7 new genomic regions of differentiation in 3 of the lines bred for high voluntary wheel running, as compared with the 4 nonselected control lines. These regions contain genes that are both intuitive for voluntary-exercise behavior and correlate to known phenotypic differences between the HR and control lines. These regions also highlight some of the genomic differences between HR3 and the other HR lines, enabling us to begin to address multiple solutions in response to uniform selection.

Selection experiments involving replicate lines have demonstrated both similar and varying responses to selection (Garland and Rose 2009). Supporting the latter possibility, Ernst Mayr (1961, p. 1505) once wrote that “Breeders and students of natural selection have discovered again and again that independent parallel lines exposed to the same selection pressures will respond at different rates and with different effects, none of them predictable.” On the other hand, replicates involving asexually reproducing bacteria typically tend to implicate the same genes or pathways, although not necessarily the same SNPs (Long et al. 2015). For example, Tenaillon et al. (2012) demonstrated that evolving 115 populations of Escherichia coli for survival at increased temperatures resulted in replicates consistently implicating a limited number of genes. However, despite regular patterns in mutated genes, favorable mutations in the rho gene deterred the mutations that would normally have been favorable in the rpoBC gene, implicating a potential alternative solution.

Evolution of replicate Drosophila lines commonly results in similar responses to selection (Long et al. 2015). An example of this would be selection on Drosophila melanogaster wing venation (Cohan 1984a). Conversely, Cohan and Hoffmann (1986) identified different responses to selection for alcohol tolerance in D. melanogaster. The alcohol tolerance experiment began with different populations of flies taken from different geographic areas and so differences in starting genetic background is a potential explanation for these different responses. However, even with different populations, Cohan and Hoffmann (1986) concluded that genetic drift was no less a driving force in differential response to selection than genetic background. Furthermore, Cohan et al. (1989) later showed that models assuming large epistatic interactions were less consistent with response to selection than models assuming pure additivity. Epistatic interactions have commonly been found to influence outbred populations, potentially because recombination allows beneficial mutations to be found in a variety of alleles and genetic backgrounds (Long et al. 2015).

Given the large size of the commercial breeding colony from which our base population of 224 mice derived and with 2 generations of random mating in our lab before being divided into 8 closed lines (Swallow et al. 1998; Carter et al. 1999; Girard et al. 2002), the replicate HR and C lines should have started with largely homogenous genetic backgrounds. However, even if most lines had the mini-muscle allele, only 3 of 8 ever had mice with the mini-muscle phenotype due to the low allele frequency and recessive nature (Garland et al. 2002). Potentially, only those HR lines that happened to express the mini-muscle phenotype before it was lost to drift had the opportunity for the mini-muscle allele to be favored by selection, thus altering the genetic background through various pleiotropic and epistatic effects. However, HR3 is not the only line to differ in its response to selection. As shown here, each of the HR lines reveal new potential selection signatures when dropped from the analyses (Tables 2 and 3), implicating variation in their response to the selection criterion. We encourage workers to focus more on the utility of replicate lines for the study of multiple solutions at all levels of biological organization (see also Garland 2003; Garland and Rose 2009).

Data Availability

Original data were made available by Hillis et al. (2020) and can be found at https://doi.org/10.25386/genetics.12436649. Supplementary File 1 contains brief descriptions of supplemental tables. Supplementary Table 1 contains all regions with at least 10 SNPs with P ≤ 0.001 for any analyses where an HR line was dropped. Supplementary Table 2 contains results of power analyses performed by sampling from a locus in a differentiated region and sampling alleles from each line for that locus with mixed-model analyses used to produce a test statistic for each of 100,000 repetitions of this sampling method (see Materials and Methods). Supplementary Table 3 contains a list of annotated genes in the new genomic regions identified only after dropping line HR3, with content from Entrez database related to current understanding of the genes’ function.

Supplemental material is available at GENETICS online.