Genome-wide analyses of an avian herpesvirus identify 10 loci associated with tumorigenicity and vaccine escape

Abstract

Comparative genomic studies of Marek’s disease virus (MDV) conducted in the past two decades have proven powerful for identifying genomic loci driving historical shifts in virulence. However, past efforts were limited to comparing a few strains at a time, lacked access to certain regions of the MDV genome, and were restricted in their ability to characterize viral phenotypes. To address these limitations, we performed whole-genome sequencing on a collection of 65 MDV strains previously characterized using a standardized phenotyping assay. In addition to phylogenetic and recombination analyses, this approach enabled us to perform a genome-wide association study of virulence for this pathogen. In total, we identified 10 genome-wide significant loci associated with virulence, including a tandem repeat variant resulting from two alternative versions of a 132-bp repeating motif. Our findings support prior descriptions of virulence as a complex trait in MDV, and highlight the potential contribution of repeat-based and intergenic variants to the phenotypic diversity of herpesviruses.

Genome-wide analyses of Marek’s disease virus reveal tandem repeats, intergenic, and genic loci associated with virulence.

INTRODUCTION

Marek’s disease virus (MDV), also known as Mardivirus gallidalpha2 (and formerly as Gallid alphaherpesvirus 2), is an oncogenic large DNA alphaherpesvirus that infects poultry. Since it was first reported in 1907, circulating isolates of MDV have increased in virulence, with novel strains emerging that could overcome previously successful vaccines (1–4). Modern outbreaks of Marek’s disease (MD) are characterized by lymphomas, paralysis, weight loss, and immunosuppression, with up to 100% mortality for unvaccinated flocks. The Avian Diseases and Oncology Laboratory (ADOL) of the United States Department of Agriculture (USDA) introduced a standardized phenotyping assay to categorize MDV strains based on their ability to overcome the protection of commercial vaccines (5). This system includes four virulence categories, or “pathotypes”: mild (m), virulent (v), very virulent (vv), and very virulent plus (vv+). MDV strains of all four pathotypes can induce lymphoproliferative lesions in infected chickens; however, strains of the vv and vv + pathotypes can break through the immunity conferred by the HVT vaccine, derived from Herpesvirus of Turkeys (Mardivirus meleagridalpha1) (6). Strains of the vv + pathotype can also bypass the protection of the second-generation bivalent vaccine, which includes both HVT and strain SB-1 of the naturally apathogenic MDV serotype 2 (also known as Mardivirus gallidalpha3; formerly Gallid herpesvirus 3 (7). The third-generation live-attenuated MDV vaccine strain, CVI988/Rispens, has been the most effective commercial vaccine since its development in the 1970s. This is due to its unique ability to protect flocks against MDV strains belonging to all pathotypes, including vv+ (8). However, reports from Europe and Asia have since documented “hypervirulent” MDV strains capable of inducing tumors in CVI988/Rispens-vaccinated flocks, raising the possibility that circulating MDV strains may soon undergo another major increase in virulence (9–11). Consequently, there is ongoing interest in understanding the genomic changes driving historical increases in virulence and vaccine escape in MDV.

Several studies have attempted to identify MDV genomic loci contributing to virulence and vaccine escape through comparative genomic approaches (12–16). These efforts have faced two major obstacles. First, the relatively slow rate at which MDV genomes have been sequenced has limited the comparative power of most studies. Second, although the ADOL pathotyping assay is internationally recognized as the “gold standard” for phenotypic classification of MDV strains, it is highly technical, resource-intensive, and cannot be performed unmodified outside of the USDA (5, 10, 14–24). As a result, most published MDV genomes lack standardized phenotypic data, severely limiting cross-study comparisons (16, 23, 24). In an attempt to overcome these obstacles, Dunn et al. recently published pathotype data for an extensive collection of MDV isolates obtained from farms in the USA between 1962 and 2016 (12). These MDV strains, known as the “Witter collection”, were all characterized using the ADOL pathotyping assay, at the same location and under the same controlled conditions (5, 12). Although a substantial improvement over past efforts, this study nevertheless suffered from a number of limitations, including limited access to whole-genome sequencing (WGS) data, heavy reliance on reference-based genome mapping, and a focus on amplicon-based analysis of single-nucleotide polymorphisms (SNPs) in genic regions.

In the present study, we sought to address the limitations of the Dunn et al. study (12) by performing WGS based on de novo assembly for 65 MDV isolates belonging to the Witter collection, with the overall goal of conducting a viral genome-wide association study (GWAS) to identify genomic loci associated with virulence (25). Whole-genome data enabled us to account for a variety of genomic variants, including SNPs and insertions/deletions (INDELs) across coding, non-coding, and intergenic regions. We also took advantage of our recent characterizations of MDV tandem repeat loci using long reads to include a subset of tandem repeat variants in our GWAS analysis (26). In parallel, we explored the potential contribution of recombination events to MDV strain diversity, and identified seven non-recombinant regions in the MDV genome. To better account for the effects of linkage disequilibrium in our GWAS, we used these non-recombinant regions to define seven MDV chromosome-like regions, enabling us to apply a mixed linear model with a leave-one-chromosome-out (LOCO) approach (27). Overall, these analyses identified 10 genome-wide significant loci associated with virulence, with the strongest association corresponding to a tandem repeat variant. These data highlight the importance of assessing variation at a genome-wide scale and considering variants beyond SNPs, and suggest that recombination analyses can inform GWAS for large viruses.

RESULTS

Whole-genome sequencing and multiple sequence alignment of 65 MDV strains

A total of 65 MDV strains from the Witter collection were used for Illumina-based whole-genome sequencing. The selected strains were collected between 1962 and 2016 from commercial flocks across 19 states and included virulent (v) (n = 19), very virulent (vv) (n = 23), and very virulent plus (vv+) (n = 23) strains (Fig. 1A, Table 1, and table S1). To minimize reference bias, viral consensus genomes for all 65 strains were generated using de novo assembly. Average genome-wide sequencing coverage ranged from 22-1405x (Fig. 1B, Table 1, and table S2). A trimmed and repeat-masked multiple sequence alignment of all 65 viral consensus genomes was then generated to identify nucleotide differences across strains (see Materials and Methods for details). Visual inspection of this alignment revealed 784 single-nucleotide polymorphisms (SNPs) (781 biallelic, 3 multiallelic, table S3) and 17 INDELs (listed in table S4). SNP variants were found to occur in the structural repeat regions (Internal Repeat Long = IRL, Internal Repeat Short = IRS) at nearly twice the rate of unique genomic regions (Unique Long = UL, Unique Short = US) relative to their size (∼0.01% in IRL/IRS vs ∼0.005% in UL/US, respectively) (Fig. 1C).

Fig. 1. Whole-genome sequencing and pairwise comparisons of 65 Witter MDV strains reveal genome-wide variation.

(A) A total of 65 MDV strains were sequenced using Illumina technology. These strains belong to the USDA Witter collection, a large collection of MDV isolates obtained from commercial farms in the United States between 1962 and 2016 and characterized using the ADOL pathotyping assay (see table S1 for further details) (5). The map indicates states from which strains were sourced (shaded in grey), as well as the number of strains isolated from each state (shaded circles) and their respective pathotypes (virulent = green, very virulent = orange, very virulent plus = red). The pie chart shows the number of strains of each pathotype. (B) Average genome-wide sequencing coverage for each strain. Strains are grouped based on pathotype (virulent = green, very virulent = orange, very virulent plus = red) and sorted by coverage, from lowest to highest. Horizontal dashed lines are shown for 100x, 200x, 300x, 400x, 500x, and 1,000x coverage thresholds. Strain names associated with strain IDs are provided in Table 1. (C) A trimmed (i.e., terminal repeats excluded) and tandem repeat-masked alignment of all 65 strains was used to identify MDV genomic variants. Vertical blue bars (y axis) represent the number of consensus genomes exhibiting a variant for each position in the MDV genome (x axis). A graphical representation of a trimmed MDV genome is provided along the x axis (UL = Unique Long, IRL = Internal Repeat Long, IRS = Internal Repeat Short, US = Unique Short).

Table 1. Sequencing statistics for 65 Witter collection MDV strains.

ID	MDV strain	GenBank accession	Pathotype*	Virulence rank*	Year	Location†	Coverage
1	617A	PV246980	v	24	1993	OH	22x
2	596A	PV246973	v	38	1991	WI	141x
3	709B/CVI988	PV247016	v	30	2010	PA	149x
4	Md3	PV247020	v	16	1977	MD	154x
5	295	PV246964	v	5	1980	CO	226x
6	Md8	PV247022	v	10	1977	MD	229x
7	MIS-X	PV247023	v	5	1980	GA	248x
8	JM/102W	PV247018	v	28	1962	MA	254x
9	571	PV246969	v	20	1989	CA	319x
10	747A	PV247012	v	29	2014	GA	408x
11	747C	PV247014	v	29	2014	GA	417x
12	232/1	PV246962	v	0	1978	MI	421x
13	CU-2	PV247015	v	0	1973	NY	433x
14	747B	PV247013	v	21	2014	GA	438x
15	GA/22	PV247017	v	25	1965	GA	455x
16	RPL39	PV247026	v	11	1969	GA	510x
17	RB1B	PV247025	v	41	1982	NY	589x
18	738	PV247009	v	26	2013	PA	591x
19	MSU-2	PV247024	v	18	1980	MI	635x
20	611	PV246977	vv	91	1992	PA	135x
21	643G	PV246981	vv	94	1994	NE	138x
22	568B	PV246968	vv	85	1988	NC	166x
23	643P	PV246982	vv	100	1994	NE	171x
24	723	PV247005	vv	100	2011	PA	174x
25	568A	PV246967	vv	66	1988	NC	206x
26	583A	PV246970	vv	55	1990	IA	258x
27	670	PV246990	vv	45	1997	ME	293x
28	549A(Del-S)A	PV246965	vv	73	1987	DE	298x
29	691	PV246996	vv	55	1999	GA	307x
30	656C	PV246988	vv	70	1995	VA	340x
31	549A(Del-S)B	PV246966	vv	73	1987	DE	399x
32	685	PV246993	vv	53	1997	GA	505x
33	610B	PV246976	vv	61	1992	MD	510x
34	653A	PV246987	vv	75	1995	DE	574x
35	718B	PV247000	vv	79	2011	PA	629x
36	612	PV246978	vv	65	1992	ME	631x
37	615K	PV246979	vv	88	1993	DE	632x
38	Md11	PV247019	vv	32	1977	MD	706x
39	287 L/1	PV246963	vv	40	1979	AL	710x
40	608	PV246974	vv	58	1992	AR	712x
41	Md5	PV247021	vv	83	1977	MD	793x
42	718A	PV246999	vv	76	2011	PA	833x
43	648B	PV246985	vv+	100	1994	OH	67x
44	648B2	PV246984	vv+	100	1994	OH	127x
45	722C	PV247003	vv+	100	2011	IA	135x
46	584B	PV246972	vv+	94	1990	NC	150x
47	730B	PV247007	vv+	100	2013	IA	185x
48	610A	PV246975	vv+	97	1992	MD	186x
49	676	PV246992	vv+	100	1997	PA	227x
50	701	PV246997	vv+	93	2007	PA	228x
51	652	PV246986	vv+	100	1995	NY	237x
52	709A	PV246998	vv+	91	2010	PA	452x
53	674	PV246991	vv+	93	1997	DE	499x
54	645	PV246983	vv+	100	1994	PA	559x
55	722B	PV247002	vv+	100	2011	IA	560x
56	686	PV246994	vv+	100	1999	IA	565x
57	660A	PV246989	vv+	97	1995	OH	589x
58	722A	PV247001	vv+	45	2011	IA	594x
59	690	PV246995	vv+	64	1999	GA	595x
60	730C	PV247008	vv+	94	2013	IA	610x
61	722D	PV247004	vv+	97	2011	IA	689x
62	739A	PV247010	vv+	88	2013	DE	790x
63	730A	PV247006	vv+	100	2013	IA	795x
64	739B	PV247011	vv+	88	2013	DE	797x
65	584A	PV246971	vv+	97	1990	NC	1405x

^*As previously report in Dunn et al. (12).

^†For further details, see table S1.

Very virulent plus MDV strains isolated in the United States share a common ancestor

A maximum-likelihood (ML) phylogenetic tree based on the trimmed and masked alignment of all 65 Witter MDV genomes revealed five distinct phylogenetic clades (Fig. 2). Clades 1 and 2 formed one monophyletic group that accounted for all 23 vv + strains, suggesting a common ancestry for strains of the vv + pathotype. Clade 1 exclusively contained vv + and vv strains, while Clade 2 included one v strain and more vv strains than vv + strains. Clade 2 also contained more deeply branched subclades. Clades 3 and 4 both consisted of vv and v strains. Clade 5 was exclusively composed of v strains. The five clades together accounted for 62 of the 65 MDV strains included in the study, with Md8 and CU-2 forming independent branches, and JM/102W being used as the root. MDV strains isolated from the same farm (table S1) were found to typically fall within the same clade, especially if they were isolated the same year. Neither the geographical region of origin (e.g., Northeast vs. Midwest) nor the year of isolation for strains that were not isolated from the same farm were found to be predictive of clade membership (Table 1, table S1, and Fig. 2). An extended ML tree based on a multiple sequence alignment of all 65 Witter MDV genomes and 32 previously published global MDV genomes (fig. S1) showed Eurasian strains clustering with Clade 5 Witter strains, while North American strains clustered with Witter strains from Clades 1 and 2.

Fig. 2. Phylogenetic analysis of 65 Witter MDV strains reveals five distinct clades.

The trimmed and masked alignment of all 65 Witter MDV strains was used to generate a maximum-likelihood (ML) tree using the K3Pu + F + R3 substitution model. Bootstrap values ≥70% are labeled. The root of the tree corresponds to the oldest isolate, strain JM/102W, which was isolated in 1962. Strains are shaded based on pathotype (virulent = green, very virulent = orange, very virulent plus = red). Strain pathotype, virulence rank, state of origin, and year of isolation are indicated in order after each strain name. Strains were found to cluster into five phylogenetic clades with distinct phenotypic characteristics (Clades 1–5). Brackets span all of the strains belonging to a specific clade. The root branch giving rise to each of the five clades is indicated using a black circle. Recombinant strains detected in later analyses are starred (see Fig. 4 for details).

Assessing variation patterns in three genomic regions harboring tandem repeats

In parallel to our phylogenetic analyses of non-repetitive genomic regions, we assessed variation in a subset of MDV tandem repeat regions where the presence of alternative repeating motifs facilitates genotyping using short-read sequencing. These three regions included the 132-bp repeats overlapping the genes MDV006.5 and MDV075.2, the Meq-proline rich repeat (PRR) region, and the UL36-PRR region. We recently characterized the repetitive patterns at each locus using high-fidelity long reads (26). Briefly, the MDV006.5/MDV075.2 genes, which are located in the TRL and IRL regions, partially overlap a set of 132-bp repeats. Two alternative versions of the 132-bp repeating motif have been reported to date (132A, 132B), distinguished only by a synonymous C > T transition in position 67 (Fig. 3 and table S5 and S6). We found that all 65 strains showed two copies of the 132-bp repeating motif based on de novo assembly of short reads. The 132B motif (T in position 67) was only observed in vv + (96%) and vv (43%) strains.

Fig. 3. Repeat patterns in the Meq-PRR and 132 bp repeats overlapping the MDV006.5/MDV075.2 genes reflect phylogenetic relationships and phenotypic differences across strains.

The MDV006.5/MDV075.2 and Meq-PRR tandem repeat regions were analyzed to assess their repetitive patterns and reveal potential genotype-phenotype associations at these loci. The MDV006.5/MDV075.2 genes each overlap a set of 132-bp nucleotide repeats. Two versions of the 132-bp motif have been reported to date, which are distinguished by a synonymous C > T transition in position 67 of the repeating unit. The proline-rich region of Meq (Meq-PRR), the major MDV oncoprotein, has three lengths of repeating units at the amino-acid level: 27-AA, 19-AA, and 14-AA. Of these, the 27-AA has the most variable motifs. Repeat patterns at these loci were found to differ across pathotypes and across the clades/subclades identified in Fig. 2. Strains are shown grouped into their respective clades, following the same order (top to bottom) as the ML tree in Fig. 2. Dendrograms based on the ML tree phylogeny are used to show relationships between strains within each clade. Strain names are shaded based on pathotype (virulent = green, very virulent = orange, very virulent plus = red).

The Meq-PRR, which also exists in the TRL and IRL regions, has been found to exhibit highly complex tandem repeat patterns at the amino acid level, involving multiple alternative versions of three distinct repeating motifs, which are 27-AA, 19-AA and 14-AA in length, respectively (Fig. 3 and tables S5 and S6) (26). In the Meq-PRR, the most common genotype among vv + strains (57%) and among vv strains (70%) consisted of two copies of the 27F repeating motif interspaced by a 14A motif (Fig. 3). Notably, all of the vv + and vv strains included in the study contained at least one copy of the 27F motif. Conversely, the 27F motif was only found in 21% of v strains, where it was always accompanied by a 27A repeating motif. The most common genotype among v strains consisted of one 27A motif and one 27E motif interspaced by a 14A motif (Fig. 3). The 27E and the 27A-P3 motifs were exclusively found in v strains. Likewise, genotypes showing higher copy numbers and containing the 19-AA repeating unit were only found in v strains (3 out of 19 strains). As part of these analyses, we also found two alternative versions of the 27-AA repeating motif that had not been described previously, hereafter referred to as 27H and 27I, respectively.

Overall, we found that the repetitive patterns in both the MDV006.5/MDV075.2 repeats and the Meq-PRR varied across the phylogenetic clades and subclades identified as part of our earlier analyses of non-repetitive genomic regions (Fig. 2). Clade 1 strains all contained the 132B motif in the MDV006.5/MDV075.2 repeats, and 86% showed a Meq-PRR genotype with two copies of the 27F repeat (Fig. 3). Clade 2 strains were nearly evenly split into two distinct genotypic groups. Just over half of the strains in this clade had a Meq-PRR genotype with two 27F repeats and a MDV006.5/MDV075.2 repeats genotype with two 132A repeats. The other half had a Meq-PRR genotype with a 27H repeat and a 27F repeat, and a MDV006.5/MDV075.2 repeat genotype consisting of one 132A and one 132B repeat. One exception to this pattern was v strain 738, which did not contain any copies of the 132B repeat (Fig. 3). Strains in Clades 3, 4, and 5, as well as the strains forming the root or singular branches (i.e., JM/102W, Md8, CU-2) all lacked any copies of the 132B repeat, and at most showed one copy of the 27F repeat in the Meq-PRR. Clade 5 showed the highest genotypic diversity in the Meq-PRR, with strains in this clade encompassing five distinct Meq-PRR genotypes (Fig. 3).

In addition to our analyses of the MDV006.5/MDV075.2 repeats and the Meq-PRR, we also assessed the repetitive patterns of the UL36-PRR in the 46 strains where this locus was resolved (see Materials and Methods, fig. S2, and table S6). The UL36-PRR also exhibits complex tandem repeat patterns at the amino-acid level, involving multiple alternative versions of two repeating motifs, which are 6-AA and 10-AA in length, respectively (fig. S2) (26). Of these, 41 strains showed a repeat pattern that we recently proposed as being typical of strains isolated in North America (26). The defining sequence features of this pattern include a distinct lack of 6D and 6E repeats and the presence of a 6C repeat in the last segment of 6-AA repeats (fig. S2). Strains 571 and MIS-X showed a repeat pattern proposed as being typical of Eurasian strains, where the 6D repeat is present in the second segment of 6-AA repeats and the last segment only contains 6B repeats. Strain 709B/CVI988 showed a truncated version of the typical repetitive pattern described for this vaccine-related strain, which is largely defined by the presence of 6F repeats and the absence of 6B repeats. The last two strains, 747B and 747C, both showed an atypical repetitive pattern that did not strictly match any of the previously described patterns (26). UL36-PRR patterns were not found to differ in a substantial manner across phylogenetic clades nor across strains with different levels of virulence.

Eight MDV strains show mosaic structures resulting from past recombination events

In parallel to our assessments of genome-wide variation, we explored the contribution of homologous recombination to the genomic diversity of MDV strains in the United States. Using 3SEQ, we first performed non-parametric tests for sequence mosaicism on all possible combinations of sequence triplets, with two sequences randomly assigned as parents and one sequence assigned as offspring (28). These analyses revealed 14 breakpoint-free regions (BFRs), with 43 strains identified as potentially exhibiting mosaic structures (see Materials and Methods for details, Fig. 4A and table S3). Maximum likelihood trees were then constructed based on each BFR to enable assessments of phylogenetic incongruence. Adjacent BFRs that did not exhibit evidence of phylogenetic incongruence in at least 70% of replicate trees (i.e., bootstrap value ≥70) were concatenated to form putative non-recombinant regions (NRRs) (29). In total, we identified seven NRRs in the MDV genome (Fig. 4A and table S3). Maximum likelihood trees were then constructed from each NRR to infer parent-offspring relationships (Fig. 4B). Based on these analyses, a total of 27 strains were found to have been involved in past recombination events either as parents or offspring, with a total of eight strains identified as recombinants: 610B, 656C, 615K, 653A, 232/1, RPL39, GA/22, and 571 (Figs. 2 and 4B, starred). Strains 610B, 656C, 615K, and 653A, which all shared the vv pathotype and belonged to Clade 2, were found to have resulted from recombination events between Clade 2 and Clade 3 strains (Fig. 4, NRR1 vs. NRR2). Strains 232/1 and 571, which both shared the v pathotype and belonged to Clade 5, were found to have resulted from recombination events exclusively involving Clade 5 strains (Fig. 4, NRR2–4). Strains RPL39 and GA/22, which also shared the v pathotype and belonged to the same subclade of Clade 5, were found to have resulted from multiple recombination events between Clade 2 and Clade 5 strains (Fig. 4, NRR5–7). Clade 1 and Clade 4 strains showed no evidence of recombination and no evidence of having contributed to any of the identified recombinants.

Fig. 4. Identification of mosaic structures in MDV genomes based on phylogenetic incongruence reveals eight recombinant Witter strains.

(A) To identify positions in the MDV genome potentially associated with recombination breakpoints, strains were randomly assigned into triplets and tested for the presence of mosaic structures using 3SEQ (69). A trimmed MDV genome diagram is shown at the top (grey and orange bars, UL = Unique Long, IRL = Internal Repeat Long, IRS = Internal Repeat Short, US = Unique Short). Vertical bars (dark blue) shown below the genome diagram represent the number of consensus genomes supporting a breakpoint at any given position along the genome (x axis). Regions completely lacking breakpoint support were designated as breakpoint-free regions (BFRs), and then adjacent BFRs that lacked phylogenetic incongruence were concatenated to form non-recombinant regions (NRRs). In total, this resulted in seven NRRs across the MDV genome (green bars below the genome diagram, labeled 1–7). Alternating shades of green are used to distinguish adjacent NRRs. (B) For each NRR, we show an ML tree generated using the GTR + Γ model from each NRR, with bootstrap values ≥70% labeled. Number labels correspond to NRRs 1–7 as shown in (A). For each tree, strains giving rise to phylogenetic incongruence signals between the trees of two adjacent NRRs are labeled in color. Strains belonging to the same parent/offspring triplet are shown using the same color family (e.g., red, dark red, pink), with shading used to track offspring strains across trees. Bold and starred names indicate the eight recombinant MDV strains detected in these analyses (also starred in Fig. 2 and fig. S3).

Ten genomic variants are associated at genome-wide significance with virulence

To identify genomic loci associated with virulence, we applied a mixed-linear model of association (MLMA) to 800 genomic variants, which included 781 biallelic SNPs and 17 INDELs identified through inspection of the trimmed and masked alignment (tables S3 and S4), as well as 2 tandem repeat (TR) variants associated with the MDV006.5/MDV075.2 repeats and the Meq-PRR, respectively (Fig. 3, fig. S3A, and table S5). For the MDV006.5/MDV075.2 repeats, a biallelic genomic variant was defined based on the presence or absence of the 132B repeat (Fig. 3, fig. S3B, and table S5). Likewise, a biallelic genomic variant was defined for the Meq-PRR by binning variants based on whether they contained the 27F or the 27E repeat, respectively (Fig. 3, fig. S3B, and table S5). Initial assessments using a standard MLMA with pathotype as the phenotypic component resulted in a QQ plot showing middle deflations, indicative of a loss in statistical power (fig. S3C). To address the limitations of using a categorical outcome to define virulence, we tested using virulence rank instead of pathotype as the phenotypic component (fig. S4). Moreover, to better account for confounding effects due to population structure, we tested using a mixed linear model with a “leave-one-chromosome-out” (MLMA-LOCO) approach instead of a standard MLMA (27, 30). To enable a MLMA-LOCO approach in the context of a herpesvirus, we used the seven NRRs identified as part of our recombination analyses to define seven MDV chromosome-like regions (Figs. 4 and 5A and table S3). Both of these adjustments independently resulted in QQ plots showing a closer fit to the expected distribution; however, a combined implementation resulted in the closest fit (λ = 0.886) (fig. S3C). In total, 10 genomic variants (9 SNPs, 1 TR) were associated at genome-wide significance with virulence (P < 6.25 × 10⁻⁵) (Fig. 5A and Table 2). The strongest associations were observed for the 132-bp repeat variant that overlaps MDV006.5/MDV075.2 and for a non-synonymous SNP variant in MDV076/Meq (Fig. 5B and Table 2). Other implicated genes included MDV049/UL36 (2 SNPs), MDV054/UL41 (1 SNP) and MDV056/UL43 (1 SNP). One additional non-synonymous SNP variant in MDV076/Meq was also found to be significant. Additionally, 3 SNP variants at genome-wide significance were found to occur in intergenic regions. We found no INDEL variants associated at genome-wide significance with virulence. A screening of the 10 identified variant sites in previously published MDV genomes revealed low allelic diversity among Eurasian strains, with most strains from this region exclusively showing alleles associated with lower virulence levels in Witter strains (table S7). Non-Witter North American strains showed equivalent allelic diversity to Witter strains, while the previously sequenced vv + Witter strain 648A showed alleles associated with higher virulence levels across all 10 loci (table S7).

Fig. 5. Manhattan plot of MDV virulence GWAS shows 10 genomic variants at genome-wide significance.

(A) A Manhattan plot was generated based on the MLMA-LOCO analysis, with virulence rank used as the phenotypic component (figs. S3 and S4) (27, 30). The position in the MDV genome (x axis, UL = Unique Long, IRL = Internal Repeat Long, IRS = Internal Repeat Short, US = Unique Short) and the observed -log₁₀(P value) of all tested genomic variants (y axis) are shown. Shaded green rectangles above the x axis indicate chromosome-like regions assigned as chromosomes in the MLMA-LOCO analysis (1–7, from left to right), with different shades of green used to distinguish adjacent regions. Solid blue line indicates the P value threshold for genome-wide significance (P < 6.25 × 10⁻⁵). ORFs associated with each significant locus are labeled, with shaded ovals used to indicate multiple variants of the same label. A map of all ORFs in the MDV genome is shown below the x axis, with ORFs associated with genome-wide significant variants shown in color. (B) Protein domain diagrams for each ORF associated with variants at genome-wide significance, sorted by strength of significance (top to bottom, descending). Functionally relevant domains are colored and labeled. For each variant, the residue number and resulting amino-acid change (synonymous = blue/blue, non-synonymous = blue/red) is shown. Nucleotide triplets for each variant codon are shown below in parenthesis, with the variant nucleotide underlined (DUB = Deubiquitinase, XPG = Xeroderma Pigmentosum, BR = Basic Region, ZIP = Leucine Zipper, PRR = Proline-Rich Repeat region).

Table 2. Genome-wide significant loci identified by MLMA-LOCO.

Position in alignment	ORF	Type	Impact	P value
114,491.5	MDV075.2*	Repeat	Synonymous	2.91 × 10−7
120,510	MDV076/Meq	SNP	Non-synonymous	7.06 × 10−7
87,276	MDV056/UL43	SNP	Non-synonymous	2.62 × 10−6
83,750	MDV054/UL41	SNP	Non-synonymous	5.28 × 10−6
68,961	MDV049/UL36	SNP	Synonymous	3.24 × 10−5
73,765	MDV049/UL36	SNP	Non-synonymous	3.24 × 10−5
119,510	Intergenic (MDV076/Meq†)	SNP	-	1.55 × 10−5
119,914	Intergenic (MDV076/Meq†)	SNP	-	1.55 × 10−5
125,245	Intergenic (MDV080†)	SNP	-	1.55 × 10−5
120,281	MDV076/Meq	SNP	Non-synonymous	2.89 × 10−5

^*In the trimmed alignment, the MDV006.5 copy (located in TRL) is not present.

^†Nearest ORF.

Test statistics for the MLMA-LOCO analysis revealed a moderate phylogenetic signal (Blomberg’s K = 1.79, Mantel r = 0.329). A visualization of allele distributions for all 10 significant variants against whole-genome phylogenies showed phylogenetic-phenotypic overlap (fig. S5). To evaluate the effectiveness of our MLMA-LOCO approach in correcting for population structure, we performed complementary phylogeny-aware association tests using treeWAS (31). In total, treeWAS analyses identified 30 terminal associations (i.e., convergent evolutionary events), 6 simultaneous associations (i.e., coincident evolutionary events), and no subsequent associations (i.e., ancient evolutionary events, population structure artifacts) (see Materials and Methods, fig. S6). Test statistics for the treeWAS analysis revealed moderate inflation for all three tests (λ = 1.19–1.58). Terminal associations included all 10 genome-wide significant variants identified in the initial MLMA-LOCO analysis. The repeat variant associated with the MDV006.5/MDV075.2 repeats was also found to exhibit a simultaneous association with virulence (fig. S6). To further deconvolute population structure, we generated chromosome-decoupled phylogenies for all 10 genome-wide significant variants identified by MLMA-LOCO and TreeWAS (see Materials and Methods and figs. S3 and S7). Visualizations of allele distributions against chromosome-decoupled phylogenies showed reduced phylogenetic-phenotypic overlap relative to whole-genome phylogenies for strains 617A, 691, 643G, and 643P (fig. S7).

DISCUSSION

We report the results of the first large-scale, genome-wide comparative study in MDV based on whole-genome sequencing of the 65-strain USDA Witter collection. This extensive library of MDV isolates were sourced from commercial poultry farms in the United States between 1962 and 2016, and they have all been characterized using the current “gold standard” phenotyping assay (5, 12). In addition to providing evidence suggesting that highly virulent MDV strains in the United States share a common ancestor, we identified eight strains as recombinants and dissected variation patterns in three tandem repeat loci in relation to evolutionary history and virulence. In total, we identified 10 virulence-associated variants at genome-wide significance. Importantly, the strongest association was observed for a tandem repeat variant, where the variant reflects two alternative versions of a 132-bp repeating motif that overlaps the MDV006.5/MDV075.2 genes.

Our analyses of Witter strain phylogenies suggest that vv + strains in the United States likely emerged from a common ancestor. While this should narrow the genomic window for identifying causal variants, we found no variants that were completely exclusive to vv+, vv, or v strains (tables S3 and S4). The absence of pathotype-specific mutations largely stemmed from vv MDV strains having genotypic overlap with either vv + or v strains. MDV strain virulence exists on a continuum, with prior studies measuring replication rates, clinical neurological responses (i.e., neuropathotyping), and lymphoid organ atrophy all showing that the lowest-ranking vv + strain is very close to the highest-ranking vv strain (5, 32, 33). Our GWAS used virulence rank as the phenotypic metric, which was based here and in prior work (12) on the HVT protective index, to enable comparisons with low-virulence strains whose pathotyping predates the bivalent vaccine’s inclusion into this assay. The use of a virulence rank metric that includes the bivalent vaccine might improve the separation of vv + from vv strains, but it would lose statistical power due to an overall smaller dataset. Despite these limitations, we found a clear genotypic and phylogenomic separation between vv + and v strains. Specifically, 9 of the 10 variants at genome-wide significance identified by our GWAS were exclusive to either the vv + or the v pathotype. For the remaining variant, only vv + strain 690 broke this pattern; however, this strain is unique in having once been assigned as a vv pathotype (5, 12). Altogether, these data support MDV virulence as a complex trait, while also suggesting that genomic comparisons across phenotypic extremes (i.e., vv + vs. v) may yield the greatest insights into the molecular basis of MDV virulence.

To extend our analyses beyond non-repetitive regions, here we identified a subset of MDV tandem repeat loci where the diversity of alternative motifs facilitates genotyping using short reads (fig. S3). Expanding on past observations by Spatz and Silva, we found that the 132B repeat associated with the MDV006.5/MDV075.2 genes was almost always present in vv + strains, but completely absent from v strains (Fig. 3) (26, 34). The significance of this finding was confirmed in our GWAS, where this variant showed the strongest genome-wide association with virulence (Fig. 5). For the Meq-PRR, we found that the 27F and 27E repeats never appeared together, with the 27F motif typically found in higher-virulence strains and the 27E motif being exclusively found in v strains (Fig. 3). Nevertheless, when testing variants containing the 27F repeat against variants containing the 27E repeat as part of our GWAS, we found no significant genome-wide association with virulence. As such, it is likely that the highly polymorphic nature of the Meq-PRR necessitates larger datasets in order to test individual genotypes for genome-wide significance.

In the present study, eight MDV strains were identified as recombinants (Fig. 4). While homologous recombination is known to occur in MDV, available data suggest that it occurs at lower frequencies than in highly recombining herpesviruses like HSV-1 (35–39). Our findings likewise suggest that homologous recombination has only played a minor role in the evolution of MDV in the USA, and that it is unlikely to be a main driver of genomic diversity for MDV. However an important byproduct of these analyses was the identification of seven non-recombinant regions (NRRs) in the MDV genome. NRRs, or haplotype blocks, are segments of DNA exhibiting high linkage disequilibrium (LD) (40). Haplotype blocks have been previously described for other herpesviruses including HSV-1, VZV and HCMV (41–43). However, these regions have not previously been used to inform genome-wide associations in viral genomes. Here, we used these NRRs to define seven MDV chromosome-like regions and conduct mixed linear model associations (MLMA) with a “leave-one-chromosome-out” (LOCO) approach (27, 30). MLMA-LOCO approaches have higher statistical power compared to traditional MLMAs due to their ability to prevent proximal contamination; that is, the inclusion of test variants in the genetic relationship matrix (GRM) (44, 45). In testing various adjustments to calibrate our GWAS, we found that the implementation of the LOCO framework substantially improved model fit, suggesting that it substantially reduced confounding due to population structure. Crucially, beyond improving viral GWAS, our MLMA-LOCO approach could be extendable to other organisms without canonical chromosomes.

The 10 genomic variants that we found to be associated at genome-wide significance with virulence involved several MDV genes previously highlighted by Dunn et al. (12). However, our list expands on these earlier findings by incorporating three intergenic SNP variants found upstream of MDV076/Meq (n = 2) and downstream of MDV080 (n = 1), in addition to the tandem repeat variant associated with MDV006.5/MDV075.2. Kim et al. showed that the incorporation of several variants identified by Dunn et al. into a vv + backbone reduced virulence only to a vv level (46). Our GWAS findings propose MDV006.5/MDV075.2 and MDV076/Meq as the two MDV genes with the strongest genome-wide association to virulence. A critical next step to elucidating the contribution of these loci to MDV virulence will be to engineer them into a recombinant MDV strain as above (46), and test for a reduction in pathotype level. MDV076/Meq, which encodes the major MDV oncoprotein, remains the single most studied MDV locus in relation to virulence (18, 47–50). In contrast, the MDV006.5/MDV075.2 genes have only recently been confirmed to be translated into a protein product, of unknown function(s) (51). Previous studies of this region were primarily focused on assessing the role of repeat expansions in attenuation, and exclusively used MDV strains lacking the 132B alternative motif (52, 53). As such, the contribution of the 132B repeat to MDV virulence presents a key opportunity for further study.

Another remaining open question relates to the generalizability of our findings in MDV strains from other geographical regions. An earlier study by Trimpert et al. suggested that North American and Eurasian MDV strains may have taken independent paths to virulence, and several studies have since further supported this notion (23, 24, 54). Although our phylogenomic data suggest a more nuanced evolutionary history than the currently accepted North American/Eurasian clade divide, the near-exclusive presence of low-virulence Witter alleles across all 10 variant loci in Eurasian strains (table S7) is nevertheless supportive of divergent molecular trajectories toward high virulence. Further studies focusing on direct experimental comparisons between Eurasian and North American MDV strains under standardized conditions are needed to fully clarify the extent of this divergence, especially given its potential implications for ongoing vaccination strategies.

One major advantage of our study was having access to comprehensive characterizations of MDV tandem repeat loci. Prior to this work, MDV tandem repeats were masked either through automated detection tools like Tandem Repeats Finder or by simply removing regions with alignment gaps using tools like Gblocks (23, 24, 55, 56). However, most MDV tandem repeat loci have now been shown to contain highly complex and polymorphic repeats that cannot be reliably resolved by current automated detection tools (26, 57). These repeats often lead to misalignments that can extend into non-repetitive genomic regions (58). As such, these approaches are likely to have led to historical inaccuracies. To help us avoid these issues, here we opted to rely exclusively on manual approaches to carefully curate and mask repeats prior to phylogenetic and recombination analyses. Likewise, we relied on manual approaches to assess variation patterns in a subset of MDV tandem repeat loci and enable their inclusion in our GWAS.

Despite these advantages, the current study still suffers from a number of limitations. First, the use of short reads for these de novo genome assemblies meant that we limited our analyses of tandem repeat regions to those with sufficient uniqueness and length to be accessible via this technology (58). Future studies using long-read high-fidelity sequencing (e.g., PacBio) would enable a complete analysis of all MDV repeat regions (26). Second, strains in the Witter collection exhibit extensive phylogenetic-phenotypic overlap (fig. S5 and table S1), resulting in relatively few phylogenetic contrasts and potentially limiting our ability to distinguish between true associations and population structure artifacts. While we implemented several approaches to account for phylogenetic confounding in our analysis, completely correcting for both evolutionary history and population structure in highly clonal microbial populations remains an ongoing challenge (59, 60). Finally, the MDV strains sequenced here were purified during their isolation from commercially farmed birds, to remove any coincident pathogens (e.g., chicken anemia virus) or live-attenuated vaccine virus (5). While this step may introduce culture-associated selection bias, it provides added consistency for downstream genetic and phenotypic assays. This enables an accurate link between the observed pathotype in experimental infections and the sequenced genomes here.

MATERIALS AND METHODS

Viral culture and DNA isolation

A pathotype and amplicon-based sequence analysis of the USDA “Witter” collection of MDV strains have been previously described (5, 12). Duck embryo fibroblasts (DEFs) were used for MDV culture (5). DEF media was a 1:1 mixture of Leibovitz’s L-15 and McCoy’s 5A media, supplemented with fetal bovine serum (FBS; 4% at plating; 1% for maintenance), 200 U/ ml penicillin, 20 μg/ml streptomycin, and 2 μg/ml amphotericin B. Viral DNA was isolated from a 5–6 day DEF infections at passages 6–7, using the Gentra Puregene DNA isolation kit (Qiagen).

DNA library preparation and sequencing

Libraries were prepared for Illumina high-throughput whole-genome sequencing (WGS) as previously described (16). DNA was quantified by Qubit (Invitrogen, CA), and viral genome copy number within each sample was estimated using qPCR for the viral pp38 gene (61). Total DNA was acoustically sheared using a Covaris M220 (60 s duration, peak power 50, 10% duty cycle, 4°C). Libraries were then prepared using the Illumina TruSeq Nano DNA prep kit according to the manufacturer’s recommended protocol, which includes 8 cycles of PCR. Additional quality controls on the libraries included Qubit (Invitrogen, CA), Bioanalyzer (Agilent), and library specific qPCR (KAPA Biosystems). Libraries were then multiplexed and run according to manufacturer’s recommendation on either an Illumina MiSeq or a HiSeq, for 600 cycles (i.e., 300 bp paired-end reads).

Processing of sequencing reads and genome assembly

MDV-specific reads were identified using Kraken2 v2.1.6 (default settings were used for all software unless otherwise specified) and extracted into a separate file using a custom Python script (62). The extracted reads were then subjected to the quality control and preprocessing step (Step 1) of our published viral genome assembly (VirGA) workflow, which performs trimming of adapters and low-quality bases (63). Quality-controlled reads were then used for de novo assembly using metaSPAdes v3.14.0 (64). The resulting contigs served as input for Steps 3 and 4 of VirGA, which include genome linearization, annotation and post-assembly quality assessments. For the reference-guided contig-ordering step, we used a trimmed version (TRL and TRS regions removed) of the published genome for strain RB-1B (GenBank accession: EF523390).

Multiple-sequence alignment and phylogenetic analyses

Trimmed viral consensus genomes were aligned using MAFFT v7.505, resulting in an initial alignment 157,179 bp in length (65). A total of 10 tandem repeat regions were annotated and manually masked in Geneious Prime v2025.1.3 based on our recent characterizations using long reads (26). An as yet uncharacterized repetitive locus located near the TRS/US junction was also masked based on preliminary descriptions from prior studies (66, 67). The final trimmed and masked alignment was 149,135 bp in length. An initial screening of genomic variants was performed using a custom R script. The alignment file was then imported into Geneious Prime to manually verify genomic variants (see tables S3 and S4 for list). SNPs and INDELs associated with homopolymers or with stretches of Ns were excluded from all downstream analyses. Maximum-likelihood trees for whole-genome and chromosome-decoupled phylogenies were generated from multi-genome alignment files using IQTREE v3.0.1 (68). The Bayesian Information Criterion (BIC) in IQ-TREE was used to determine the most suitable substitution model. Bootstrap values were computed using 1,000 replicates.

Repeat analyses

For each consensus genome, tandem repeat sequences associated with the 132-bp repeats overlapping MDV006.5/MDV075.2, the proline-rich repeat (PRR) region of Meq (Meq-PRR), and the UL36-PRR were manually extracted and aligned using the “Geneious Alignment” tool in Geneious Prime. The resulting alignment was visually inspected, annotated, and manually re-aligned to optimize gap openings. Strains exhibiting stretches of Ns within repeating motifs or completely lacking repeats were excluded. Based on these criteria, a total of 19 strains were excluded from analyses of UL36-PRR diversity.

Recombination analyses

Pre-screening of genomic mosaic signals potentially indicative of recombination was performed using 3SEQ v1.7, as previously described (29, 69). Briefly, 3SEQ performed a hypergeometric random walk (HGRW) test on triplets to determine sequence positions corresponding to potential recombination breakpoints (assuming only two breakpoints are possible for any triplet). Continuous stretches of positions that were not associated with breakpoints in any of the query sequences were designated as breakpoint-free regions (BFRs). BFR boundaries are listed in table S3. Following a first round of 3SEQ, BFRs longer than 20-kb were extracted and subjected to a second round of 3SEQ to identify potential breakpoints occurring within them. Maximum-likelihood (ML) trees were then constructed for each BFR >500-bp using RAxML v8.2.8 with 1,000 replicates to identify phylogenetic incongruence (PI) signals (70). Briefly, ML trees belonging to adjacent BFRs were manually inspected to identify genomes that cluster with one group of genomes (parents) in one of the trees and then cluster with another group of parents in the adjacent tree. Situations where this relationship existed with sufficient bootstrap support (≥70%) were considered to be statistically supported PI signals. Adjacent BFRs that did not exhibit PI signals were concatenated to form non-recombinant regions (NRRs). ML trees based on each NRR were constructed using RAxML with 1,000 bootstrap replicates. Sequences that always clustered with another group of sequences were designated as offspring, while sequences that only clustered with offspring sequences in one of the ML trees were designated as parents.

GWAS analyses

A custom R script was used to obtain formatted variant tables from the trimmed and masked multiple sequence alignment. The resulting files were then used as input for PLINK v1.9.0-b.7.7 to generate a binary biallelic genotype table (71). Multiallelic loci were flagged and excluded (n = 3). The tandem repeat variants associated with the MDV006.5/MDV075.2 genes and the Meq-PRR were transformed into biallelic variants by defining one repeat-based genotype for each allele (e.g., Allele 1 = CAGCAGCAG, Allele 2 = CAGCTGCAG) and representing each allele with a distinct arbitrary nucleotide (e.g., Allele 1 = “A”, Allele 2 = “T”). A genetic relationship matrix (GRM) was obtained from the resulting BED file using GCTA v1.94.1 with the --autosome flag (30). Association mapping was then performed using GCTA MLMA (−-mlma) or MLMA-LOCO (−-mlma-loco), with a GRM cutoff of 2 (27). When using pathotype as the phenotypic component, the v pathotype was given a numeric value of 0, the vv pathotype a value of 0.5 and the vv + pathotype a value of 1. When using virulence rank as the phenotypic component, the corresponding numeric value between 0–100 was used. MDV chromosome-like regions for MLMA-LOCO analyses were defined based on NRRs identified as part of recombination analyses. Chromosome-like region boundaries are listed in table S3. Chromosome-like regions 1–6 encompassed all of the nucleotides in their respective NRRs, in addition to all prior nucleotides starting immediately after the end of the preceding NRR (or from the start of the genome for chromosome-like region 1). Chromosome-like region 7 also included all nucleotides following NRR7 up to and including the end of the genome. QQ plots and Manhattan plots were generated using the ggplot2 package in RStudio. Phylogeny-aware association tests were performed using the R package treeWAS (31). Briefly, treeWAS applies three complementary association tests while controlling for clonal population structure through phylogeny-aware null simulation. Terminal associations (Score 1) detect widespread patterns across phylogenetic tips representing independent evolutionary changes. Simultaneous associations (Score 2) identify coincident genotype-phenotype evolution on the same branches. Subsequent associations (Score 3) detect phylogenetically-driven patterns indicating potential confounding. Input files included the ML phylogenetic tree generated with IQTREE based on the trimmed and masked multiple sequence alignment of all 65 consensus genomes. Virulence rank was used as the phenotypic component (fig. S4). A total of 8,000 null simulations were used for significance testing. Statistical significance was assessed using Bonferroni correction for multiple testing (800 variants × 3 tests, α = 0.05), yielding a threshold of P < 2.08 × 10⁻⁵. Associations were interpreted as genuine convergent evolution if terminal or simultaneous associations were detected without corresponding subsequent associations.

Supplementary Materials

The PDF file includes:

Figs. S1 to S7

Legends for tables S1 to S7

Other Supplementary Material for this manuscript includes the following:

Tables S1 to S7