Advances in the discovery and analyses of human tandem repeats

Abstract

Long-read sequencing platforms provide unparalleled access to the structure and composition of all classes of tandemly repeated DNA from STRs to satellite arrays. This review summarizes our current understanding of their organization within the human genome, their importance with respect to disease, as well as the advances and challenges in understanding their genetic diversity and functional effects. Novel computational methods are being developed to visualize and associate these complex patterns of human variation with disease, expression, and epigenetic differences. We predict accurate characterization of this repeat-rich form of human variation will become increasingly relevant to both basic and clinical human genetics.

Introduction.

Tandem repeat DNA are hotspots of genetic variation and have a longstanding association with human disease. At least five categories of tandem repeats are recognized, defined primarily based on the length of the underlying repeating sequence motif and the total size of the resulting repeat structure. The shortest, termed short tandem repeats (STRs; aka microsatellites, Table 1), typically have a repeat motif length ≤6 bp and were used during early human genetic mapping studies because they exhibit a high degree of heterozygosity among human haplotypes [1]. The overall length of STR arrays rarely exceeds 1 kbp for any given locus, although there are disease-associated exceptions such as the FMR1 CGG repeat that leads to Fragile X syndrome in a hyperexpanded disease state with over 200 triplet repeats[2]. The motif length of variable number tandem repeats (VNTRs; aka minisatellites) are larger (≥7 bp) with the majority of VNTRs harboring motif length repeats between 10–60 units [3]. Larger tandem repeats, typically >1 kbp with respect to the tandem repeat motif, are most often annotated as tandem segmental duplications (SDs; >1 kbp and >90% sequence identity) and are not annotated as other identified classes of repeats such as LINEs or endogenous retroviruses (ERVs), to name a few. Such tandem SDs are distinguished from interspersed SDs, which refer to non-retrotransposed repetitive segments of DNA [4] separated from the nearest identity paralog by >1 Mbp. The structure of these larger tandem SDs can be complex in their organization, often composites of different sequences, and particular subsets were recently dubbed as composite elements during the analysis of the finished human genome [5,6]. Finally, in humans the most abundant centromeric satellite repeats consist of a tandemly repeated alpha-satellite monomer repeat of 170–171 bp. Alpha-satellite repeats are often organized into higher-order repeats (HORs) of specific sets of monomers that expand to reach multiple megabase pairs in length. A subset of these alpha-satellite HORs correspond to the site of kinetochore attachment. While alpha-satellites have a well-established function for mitotic and meiotic segregation of chromosomes, the role of other satellite classes is less well understood. In humans, these are most often associated with shorter motifs (Table 1) and have expanded into megabase-pair structures mapping most frequently within heterochromatic pericentromeric and acrocentric portions of human chromosomes [6,7]. Satellite DNA (HSAT2 &3), for example, typically consists of thousands to tens of thousands of tandem repeats of the pentamer “CATTC” and, in contrast to STRs, are isolated as a distinct fraction from centrifugation gradients of human DNA representing multiple Mbp of contiguous sequence.

Table 1:

Classes of tandem repeats in the human genome

Class		Motif length	Overall size*
STR	Short tandem repeat	1–6 bp	~200 bp
VNTR	Variable number tandem repeat	≥7 bp	~500 bp
SD	Tandem segmental duplication	>1 kbp	10–100 kbp
SAT	Alpha centromeric satellite DNA	~171 bp	500 kbp - multiple Mbp
	Human satellite 1 (HSat1A)	42 bp	890 kbp - multiple Mbp
	Human satellite 1 (HSat1B)	~2.4 kbp	~14 Mbp
	Human satellite 2 (HSat2)	5 bp	620 kbp or ~12.6 Mbp
	Human satellite 3 (HSat3)	5 bp	890 kbp - multiple Mbp
	Human satellite 4 (HSat4)	35 bp	10kbp - 100kbp
	Beta satellite	68 bp	Variable clusters
	Gamma satellite	220 bp	10–220 kbp clusters

^*Overall size is based on the average length observed in the T2T (telomere-to-telomere) genome assembly but some STR and VNTR alleles, especially pathogenic events, can expand to multiple kilobase pairs in length. Satellite classification is based on telomere-to-telomere assemblies [7,8].

When plans to sequence and assemble the human genome were first laid out, there was considerable enthusiasm that the “anatomy” of these regions would be fully revealed [9]. Unfortunately, these regions were often the last to be sequence resolved [3,10–12]. The sequence identity and length of the tandem repeats, in particular, prevented both reliable assembly and cloning of the largest and most mutable regions of the human genome [12–15]. Other factors like skewed GC composition contributed to amplification and cloning biases, such that both the regions and their variation could not be properly assessed. For almost two decades, these regions remained gaps in all human genomic analyses—often regarded as sequence non grata [16]. With the advent of long-read sequencing (LRS) of native DNA (i.e., no bacterial cloning), the telomere-to-telomere (T2T) completion of the first human genome [11], and sequencing of more than 100 human genomes [17,18], a more complete view of the genomic architecture of these regions has emerged. In this review, we apprise our understanding of human genomic organization, emerging properties regarding their variation, and outstanding challenges for visualizing, classifying, and genotyping tandem repeats. While many dozens of diseases have now been described in association tandem repeats as summarized in this review, we anticipate many more as human genetic variation in these dynamic regions becomes more routinely assessed, often for the first time, in the context of controls and patients with disease.

Distribution of tandem repeats in the human genome.

Our ability to accurately resolve the sequence of tandem repeats has significantly improved through the combination of LRS from population genomes where parental data are available or where phasing information can be obtained from specialized next-generation sequencing (NGS) methods, such as Strand-seq and Hi-C [19,20]. As a result, the average full repeat length of VNTRs in the finished human genome (T2T) has more than doubled (585 bp) when compared to the average VNTR length reported for GRCh38 (215 bp). These larger tandem repeats have been placed into the context of chromosomal-level phased genome assemblies [21]. Concerted efforts from the Human Genome Structural Variation Consortium (HGSVC) [17], T2T Consortium [11], and Human Pangenome Reference Consortium (HPRC) [18] have produced valuable resources of long-read genome assemblies and structural variation from diverse human populations. These datasets have provided our first glimpse of the landscape of large tandem repeats in the human genome (Fig. 1).

Figure 1. Genome-wide distribution of tandem repeats in 47 HPRC human genomes.

The ideogram depicts the genome-wide distribution for VNTRs (n=166,918, purple) and STRs (n=131,679, green). We also show HOR regions of the genome enriched in satellite sequences (red) and tandem segmental duplications (SDs; blue) that map less than 1 Mbp apart. The average STR and VNTR lengths are 174 bp and 516 bp, while their average motif lengths are 3 bp and 49 bp, respectively. SDs and centromere satellite annotation are based only on the T2T reference genome.

Early studies of polymorphic tandem repeat markers identified an enrichment of variation in subtelomeric regions of the genome [22–24]. These studies were accurately quantified using long reads that found 55% of polymorphic VNTRs map to subtelomeric portions of human chromosome arms—the final 5 Mbp of sequence proximal to the telomere [25,26]. There is a modest yet significant association between the genome-wide distribution of VNTRs and sites of double-strand breaks (DSB, R2=0.23, P-value < 1E-22), suggesting a link between genomic regions prone to DSB and VNTR formation and we estimated that DSBs account for 23% of the variance in VNTR abundance across the genome. There are two distinct classes of VNTRs. The first is SINE-VNTR-Alu (SVA) retrotransposition-mediated events responsible for the expansion of VNTRs into GC-rich regions of the genome with a bias against genic regions. In contrast, non-SVA-associated VNTRs appear to accumulate preferentially near genes and are the most likely to be enriched in subtelomeric regions [3].

The recently published phased genome assemblies from 47 diverse humans confirm these original observations [18]. Using structural variants reported by the long-read base caller PBSV [27], we identify 145,966 polymorphic VNTRs with repeat motif ≥7 bp and an overall length <1 kbp. The subtelomeric bias is still observed though less pronounced as more genomes of greater diversity are included; 33.5% and 9.6% of the VNTRs and STRs are located in the subtelomere, respectively. The difference in subtelomeric enrichment is likely due to a combination of the increased population diversity in the HPRC genomes and any biases introduced by the different variant-calling methods employed in both studies. An additional 20,952 VNTRs are greater than 1 kbp in length and are potentially more likely to become unstable generationally. STR elements (n=131,679) are more uniformly distributed within the human genome and have an average size of 217 bp. Even larger tandem SDs and nearly complete satellite arrays have now been assembled, although these regions still pose challenges for the complete T2T assembly of diploid genomes [28]. SDs are shown to have higher mutation rates and interlocus gene conversion rates than unique DNA [14]. Figure 1 depicts our current understanding of the landscape of tandem repeats based on the analysis of 47 recently released HPRC genomes using Tandem Repeats Finder [29] and the GRCh38 UCSC table browser tracks for SDs and satellite sequences.

Advances in the discovery and alignment of tandem repeats.

Computational analysis of tandem repeat sequences is performed by de novo identification in reference genomes followed by targeted analysis of variation from sample sequence reads. While the sequence and assembly of large tandem repeat sequences have now been made possible by LRS, the characterization of their underlying variation has been challenging. Generalized algorithms for discovering variation from short reads, including single-nucleotide variants and small indels [30,31], as well as structural variation [32–35] have been shown to perform poorly in repetitive DNA [21,36–38]. Additionally, although variation in tandem repeats is accurately identified in current long-read data analyzed by context-agnostic LRS software [21,27], there are difficulties in comparing variant calls in repetitive DNA between methods and across individuals [27,39,40]. To address this, several sequence analysis tools have been developed to analyze specific classes of tandem repeat variation, with the computational approach governed by the relative length of repeat alleles to read lengths, prior knowledge of repeat domain structure, and read error profile.

Tools that discover variation in tandem repeats rely on de novo computational annotation of tandem repeat loci in reference genomes. RepeatMasker [41] performs a naive scan for mono, di, and trinucleotide repeats, while Tandem Repeats Finder (TRF) [29], TANTAN [42], and ULTRA [43] use more sophisticated algorithms based on k-mer repeats and hidden Markov models to find inexact repetitions of longer motif patterns. While multiple algorithms and theoretical studies have demonstrated more sensitive results for tandem repeat annotation than TRF [44–47], this method is used by RepeatMasker and in the University of California, Santa Cruz Genome Browser [48] annotations and, thus, forms the basis for most targeted repeat studies. Reliance on the human reference GRCh38 annotations misses some tandem repeat loci; an analysis of TRF annotation of 148 haplotype-resolved assemblies discovered an additional 5,294 annotations missing from the default annotations. Larger repeats including SD may be annotated using DupMasker [49], as well as SEDEF [50] and BISER [51] that are recently developed and are more computationally efficient. Finally, methods including SRF [52], RepeatNet [53], and Alpha-CENTAURI [54] have been written to annotate alpha-satellite repeats from sequencing reads because de novo assembly of centromeric DNA remains challenging.

The approaches to analyze tandem repeat variation using short reads fall into three groups: methods to genotype exact STR alleles fully contained by individual reads, targeted approaches to detect pathogenic repeat expansions, and methods that estimate VNTR length or composition by read depth. The RepeatSeq [55] and lobSTR [56] methods detect STR variation spanned by short reads (up to 100 bases) and provide exact counts of repeat motifs in each haplotype (Fig. 2a). The hipSTR method includes modeling of artifacts of polymerase chain reaction to improve genotype accuracy and enable identification of de novo repeat expansions. Furthermore, hipSTR identifies the non-reference repeat motifs common at highly variable loci.

Figure 2. Methods for discovery, genotyping and annotations.

Top, methods to genotype STR expansions using short reads. a, STR genotyping methods that require reads to fully span alleles: RepeatSeq, lobSTR, and hipSTR. These may provide exact repeat counts as well as phased variants. b, STR genotyping methods that can genotype expansions larger than the length of a read or paired-end insert: TREDPARSE, GangSTR, STRetch, STRling, and ExpansionHunter Denovo. The resulting calls are an estimate of motif repeat counts. c, ExpansionHunter genotypes long expansions that fit predefined patterns. Middle, methods to genotype VNTR expansions using short reads. d, adVNTR and adVNTR-NN use a hidden Markov model to estimate repeat unit copy number. e, Copy number defined by CNVnator correlates with ground-truth copy number with sufficient accuracy for association analysis. f, Genotyping VNTR length using repeat pangenome graphs can detect changes in motif composition. Bottom, methods to genotype VNTR alleles with long reads. g, Schematic of algorithmic approach used by TRviz and vamos to annotate motif copies in LRS and genome assemblies using wrap-around dynamic programming that aligns an optimal number of copies of a motif sequence to a genomic sequence by copying alignment scores from the last row to the first and allowing the trace back path index (red) to wrap around from the beginning to the end of the motif.

Pathogenic repeat expansions such as those at the HTT and C9orf72 loci may be many times greater than the length of a short read and cannot be measured in short-read data using methods requiring reads to span the repeat allele. Nevertheless, they may be detected indirectly in short-read data by excess read alignments and paired-end discordance at known tandem repeat loci. TREDPARSE [57] and GangSTR [58] incorporate information from paired-end reads and likelihood models to genotype repeat expansions, including triplet repeat expansions with nearly 1,000 motif copies (FIg. 2b). exSTRa[59] uses similar read information but is designed to detect outliers rather than specific alleles. ExpansionHunter [60,61] genotypes pathogenic repeat expansions with known motif patterns, such as the HTT repeat expansion (CAG)*CAACAG(CCG)* that may have a variable number of CAG and CCG trinucleotide repeats (Fig. 2c). Because reads often misalign at tandem repeat loci, the GangSTR method and STRetch [62] include reads mapped to off-target or STR decoys to improve the estimated tandem repeat length. To allow the discovery of novel repeat expansions, STRling [63] and ExpansionHunter Denovo [64] can similarly estimate repeat expansions without specific tandem repeat intervals being specified but often fail to report both the true length or underlying sequence architecture.

The motif diversity and length of VNTR loci make them recalcitrant to variation analysis with software designed for STRs [65]. For relatively short VNTRs (up to 21 motif repeats), adVNTR and adVNTR-NN [65,66] use a hidden Markov model to estimate motif counts for targeted VNTR loci, with adVNTR-NN using a neural network to improve read recruitment (Fig. 2d). For relatively large VNTR loci, >100 bases with repeat motifs at least 10 bases, read depth measured by CNVnator [67] calibrated by flanking sequences provides a relatively accurate measurement of tandem repeat length [68] (Fig. 2e). Read alignment to a repeat pangenome graph constructed from VNTR sequences using danbing-tk can identify VNTR expansions [69] as well as expansions of individual motifs (Fig. 2f) [70]. Finally, there are bespoke methods for genotyping specific disease or biologically relevant loci, such as LPA [71] and telomeres [72].

Exact measurement of tandem repeat variation using short reads is thus limited to short loci, and remaining loci only have estimations of repeat variation. These challenges are solved using LRS because the long reads or their assemblies span the full length of the STR and VNTR loci enabling precise discovery of tandem repeat variation from alignments [21,73]. The initial methods for tandem repeat analysis using long reads focused on overcoming the high error rates of single-molecule sequencing to count repeat copy number from Pacific Biosciences (PacBio) continuous long reads [74], or from Oxford Nanopore Technologies (ONT) current signal data [75,76]. However, with the accuracy provided by PacBio HiFi (high-fidelity) sequencing [27] and pore improvements paired with consensus sequencing from ONT [77], low-level methods are not necessary. Instead, a greater challenge is in organizing highly polymorphic VNTR sequences at the population level. The SV-Merger [78] and Jasmine [40] methods use graph clustering to combine variant calls across individuals; however, they can underestimate diversity and do not distinguish tandem repeat motif composition. Two similar methods, TRviz [79] and vamos [80], annotate the motif composition of tandem repeats from long-read sequences. Both methods use an algorithmic approach called wrap-around dynamic programming (Fig. 2g) that in contrast to standard local alignment [81] aligns a single query sequence (motif) to a target (genomic VNTR sequence) and finds the optimal matching between repeated motifs to LRS or genome assemblies [82,83]. A summary of software and their targeted classes of variation is given in Table 2.

Table 2:

Methods for tandem repeat variant annotation and discovery

Reference annotation
Class of repeat	Method				Reference
STR, VNTR	Repeat Masker				[41]
STR, VNTR	Tandem Repeats Finder				[29]
STR, VNTR	TANTAN				[42]
STR, VNTR	ULTRA				[43]
Satellite	SRF				[52]
Satellite	Repeatnet				[53]
Satellite	Alpha-CENTAURI				[54]
Segmental Duplication	DupMasker				[49]
Segmental Duplication	SEDEF				[50]
Segmental Duplication	BISER				[51]
Variant discovery, short-reads
Class of variant	Method	Size range*	Precision	Untargeted**
STR	RepeatSeq	< RL	exact		[55]
STR	LobSTR	< RL	exact		[56]
STR	hipSTR	< RL	exact		[95]
STR	TREDPARSE	< 500 bp	exact, estimate		[57]
STR	ExpansionHunter	< 1kb	exact, estimate		[60]
STR	exSTRa	Any	estimate		[59]
STR	ExpansionHunter-denovo	Any	estimate	yes	[64]
STR	STRetch	Any	estimate		[62]
STR	STRling	Any	exact, estimate	yes	[63]
STR,VNTR	GangSTR	Any	exact, estimate		[58]
VNTR	adVNTR-NN	Any	estimate		[66]
VNTR	CNVnator	Any	estimate		[67]
VNTR	danbing-tk	Any	estimate		[69]
Variant discovery, long-reads
VNTR	TRViz	< RL	exact		[79]
VNTR/STR	vamos	< RL	exact		[80]

^*Size range <RL are limited to the sample read length.

^**Untargeted methods do not require a list of input tandem repeat loci to analyze.

The PacBio and ONT technologies each currently offer distinct advantages in the analysis of tandem repeats. ONT supports an ultra-long sequencing protocol with N50 read lengths (minimum length contained by 50% of data) over 100 kbp [84]. In contrast, PacBio HiFi reads are consistently more accurate than ONT (r10.4) with error rates less than ~1 in 1000 base pairs. Furthermore, more computational support is available for phased genome assembly using PacBio HiFi reads [85]. This combination of HiFi and ONT reads can produce telomere-to-telomere assemblies that resolve the majority of centromeres [86]. This is particularly useful for fully resolving the longest tandem repeat structures and was critical for the contiguous sequence assembly of the first human centromeric satellites [7]. While both platforms can now be used to reliably detect CpG methylation [87], it is possible to determine longer range phasing of methylated sites using ONT reads. For practical considerations, ONT sequencing is capable of higher-throughput sequencing of up to 48 multiplex samples at once using the PromethION, offers portable sequencing using the MinION, and requires less up-front capital investment than the PacBio Revio. However, both technologies are rapidly advancing and future specifications may differ.

Expansions of tandem repeat sequences are often associated with changes in methylation [88–90], yet mapping methylated bases with the standard approach of bisulfite sequencing with short-read technologies is challenging because of the sequence divergence and alignment ambiguity. A promising application of LRS is the ability to detect methylated bases directly from native DNA. The Remora base caller for ONT sequences can detect 5-methylcytosine (5mC) bases from current (squiggle) data. These data have been used to characterize methylation patterns in satellite repeats of a complete human genome [91]. When combined with phasing, this can produce haplotype-resolved methylation maps that have a genome-wide correlation of 0.949 with whole-genome bisulfite sequencing[92]. Recently, PacBio released the Jasmine software to detect 5mC from HiFi sequencing data. Both approaches rely on detecting changes in kinetics unique to methylated bases from low-level sequencing data. These properties have been used to map protein-DNA interactions by treating DNA with a methyltransferase such that interacting bases are protected from methylation and interactions are read as the inverse of the methylation signal [93,94].

Advances in visualization methods.

Sequenced-based resolution of the tandem repeats has revealed complex higher-order structures even for alleles with the same overall repeat length necessitating the need to develop more sophisticated tools to investigate their underlying architecture. Several groups have begun to tackle the challenge of resolving tandem repeat architecture by creating programs to extract sequence information, often specifically for STRs, VNTRs, and centromere satellites. Some of the early strategies for demonstrating variability in VNTR sequence and structure involve color-coding internal repeats, as is the case for PRDM9 [96,97]. Dot plot graphs are often used to demonstrate commonality of repeat structure within a genomic region and variability of length; however, these matrices do little to demonstrate the diversity of sequence information (Fig. 3a). In the past, custom scripts were developed to align VNTRs and reveal their structure, such as in CACNA1C [98] and WDR7 [99,100]. Other programs are now emerging that specifically address the need to visualize VNTRs, producing, for example, “waterfall” plots [3] (Fig. 3b) that show changes in cassette architecture based on tracking of k-mers as a function of VNTR length. The TRviz program takes VNTR sequences that have been extracted, identifies the most common k-mers (repeat motifs), next assigns an ASCII character to the most common character, and then converts the character to a color to produce an alignment based on internal sequence homology, what we refer to as a “Seattle plot” (Fig. 3c) [79]. A similar program, vamos, adjusts parameters for generating efficient motif sets at VNTRs and has the benefit of collapsing repeats into fewer colors to help visualize patterns when merging rare or private motifs [80]. Other tools like TRVZ have been made available to visualize VNTRs from long-read sequences based on the knowledge of individual repeat motifs [101] in addition to the tandem repeat annotation library (TRAL), which has been used to generate what are termed “Mola plots” for several VNTRs in SLC6A3 [102].

Figure 3. Methods to visualize tandem repeats and their variation.

a, A 63 bp VNTR in MIR4435-2 host gene (MIR4435-2HG) is used as an example (top). A dot plot alignment of one allele each from two individuals is shown (bottom). b, A Waterfall plot where repeat motifs are assigned various colors based on sequence identity (top) and then sorted by total length (bottom). This strategy is also often used to demonstrate individual reads separated by allele for one individual. c, A Seattle plot, which takes color-coded alleles and organizes them by internal sequence similarity to highlight clusters of related alleles. Insertions and deletions (representing whitespace gaps) are observed in this context. d, A StainedGlass plot (reproduced from [104]) demonstrating sequence homology for a tandem repeat in heatmap form, often used for centromeric repeats. Heat map defines % identity of the higher-order repeat (red~99% identical versus blue ~70% identity) e, A Pangenome graph highlighting different common paths that a repeat can take in the context of the genome including other structural variants, like two that are shown here.

Some tools, such as StainedGlass and its derivative ModDotPlot, have been designed to interactively handle the massive amount of sequence associated with centromeric and pericentromeric satellite DNA [103]. StainedGlass, for example, generates colored heatmaps based on sequence identity across megabase pairs of DNA and can uncover long-range higher-order structures and regions of recent homogenization identifying distinct evolutionary layers in the formation of these regions (Fig. 3d) [104]. Given that higher sequence identity often reflects more recent evolutionary events, the tool provides a snapshot of evolutionary change and recent gene conversion across otherwise highly identical repeats. Other tools, such as DupMasker [49] and RepeatMasker [41], have proven very useful for defining the composition of tandem SDs by identifying the evolutionary structure of the underlying cassette that is tandemly repeated [5,11,103].

These programs notwithstanding, providing a robust and streamlined approach to visualize tandem repeats remains a challenge to the genetics community. All current programs, for example, prioritize stand-alone visualization of tandem repeats without consideration of how they vary among different individuals. Such realizations may be uncovered through integrating repeat structure and variation via pangenome graphs [69,105] (Fig. 3e) although these regions have proven particularly challenging. Most of the popular tools, unfortunately, do not adequately resolve these details at present, at least for the largest and most variable tandem repeats in our genome. Part of the difficulty in visualizing high-copy tandem repeats arises from the diversity and complexity of the underlying sequence that is present. VNTRs differ in repeat length but also internal sequence differences that often cannot be evenly divided into individual repeat units. In some VNTRs, the repeat motif length is highly variable, or harbor repeats-within-repeats (such as in a VNTR near PROP1 [106]), which presents obstacles to identifying the various internal repeat sequence motifs. Centromere satellites similarly contain new higher-order repeat units that are derivative from a larger higher-order repeat unit sequence already present in the surrounding sequence making alignments challenging. These issues are further compounded by assembly errors and sequence errors, especially at homopolymer runs creating artifacts of variation. Moreover, VNTR start and end positions are often not precisely recorded, so portions of a repeat motif will appear at the end of a VNTR. For example, the WDR7 repeat ends in 46 bp of the 69 bp motif [99]. Convention dictates that the repeat motif starts with the first nucleotide in the VNTR, but the true origination of the repeat may differ, and the repeat may originate and expand from a 3’ to 5’ (rather than 5’ to 3’) direction. Altogether, programs for visualizing VNTRs need to accommodate several pieces of information to produce an output that is accurate yet also clearly informative of the internal sequence.

Nevertheless, with appropriate methodology often geared to a particular class or even an individual locus, tandem repeat visualization methods have become powerful tools to quickly identify and demonstrate differences in repeat length and internal sequence composition. For instance, color-coding is effective at highlighting where a shift in repeat motif occurs in ATXN10 [107], or interruptions in the CAG repeat in HTT that correlate with age of onset of Huntington’s Disease. In the case of FMR1, the identification of alleles that have lost these interruptions has become a predictor of those that will ultimately become permutation alleles [108]. In a more complex example, a VNTR in ART1 ranges from 9 to 500 copies (0.5 to 31.5 kbp) in HGSVC samples. Only after breaking the repeat into color-coded repeat units do patterns emerge for samples, consisting of higher order ~3 kbp (47 repeat motif copies) duplication events[106]. Overall, a universal tool for handling each repeat is not currently available, but conceivably could consist of a combination of tools that provide a comprehensive overview of a repeat, beginning with a high level view of the repeat structure (e.g. via StainedGlass), breaking the repeat apart into individual repeat units and clustering individuals based on length or sequence similarity (e.g. via vamos or TRviz), and then providing a combined output via a pangenome graph. As increasing numbers of long-read genomes become available, having the appropriate programs to visualize tandem repeats will be critical in deducing patterns variation and mechanisms of expansion events as well as their evolutionary origin.

Tandem repeats and disease.

One of the primary motivations for systematic discovery and characterization of these loci has been their association with genomic instability and disease. Tandem repeats have been estimated to mutate at orders of magnitude higher than most other unique regions of the genome [109] and variation at these loci has been implicated in Mendelian, as well as seemingly non-Mendelian, and more complex genetic disease. Pathogenic mutations result typically when repeat lengths either go below or exceed some threshold often resulting in changes of expression of some nearby genes wherein the repeats are embedded. Such is the case with reductions of the D4Z4 repeat associated with facioscapulohumeral muscular dystrophy type I [110] where contraction below 11 copies leads to a permissive chromatin state and the expression of DUX4 in muscle tissue and disease. Similarly, reductions of the tandem SD encoding the two exons of the Kringle IV domain of lipoprotein A gene lead to higher levels of lipoprotein (a) in the circulating blood and is one of most significant genetic risk factors for coronary heart disease and stroke especially among individuals of African ancestry [111].

More often, however, an increase in repeat length is associated with pathogenicity. This is the case with many triplet repeat disorders where expanded repeats lead to disruption of the normal function of the protein or hypermethylation of the promoter and silencing [88–90]. In the case of Huntington’s disease, expansion of a protein-coding CAG repeat within HTT beyond 41 units results in the formation of an abnormal protein that accumulates in the brain, causing neurodegeneration [112]. In the case of myotonic dystrophy, expansion of the CTG nucleotide repeat beyond the normal range (5–37 repeats) in the 3′ untranslated region (UTR) of the dystrophia myotonica protein kinase (DMPK) gene leads to mRNA instability and decreased expression (Table 3) [113,114]. Several genes, such as XYLT1 (Baratela-Scott syndrome) and FMR1 (Fragile X syndrome), possess CGG repeats within the 5′ UTR of the gene. In both cases expansion of the CGG repeat beyond the premutation size range (100–200 repeats) results in hypermethylation of the nearby promoter region followed by silencing of gene expression and disease [88,115]. Importantly, the discovery of expanding triplet repeats and their transmission within families provided the molecular basis for genetic anticipation for several genetic disorders, including increased penetrance, severity, as well as earlier age of onset in subsequent generations [116].

Table 3:

Tandem repeats associated with disease.

Gene	Repeat	Normal length	Pathogenic Length	Context	Cyto	Inheritance	Disease	Phenotype OMIM #	Clin testing for expansion?	Ref
ABCA7	25 bp	12–427	>200	intron	19p13.3	AD	Susceptibility to Alzheimer’s disease	608907	n	[132]
AFF2	CCG	4–39	>200–900	5’UTR	Xq28	XLR	Intellectual developmental disorder, X-linked 109	309548	n	[133]
AR	CAG	9–36	> 37–68	exon	Xq12	XLR	Spinal and bulbar muscular atrophy of Kennedy (SBMA)	313200	y	[134]
ARX	GCN	12–16	20–23	exon	Xp21.3	XLR	Developmental and epileptic encephalopathy 1 (EIEE1)	308350	n	[135]
ATN1	CAG	3–35	>47–93	exon	12p13.31	AD	Dentatorubral-pallidoluysian atrophy (DRPLA)	125370	y	[136]
ATXN1	CAG	6–38	>38–88	exon	6p22.3	AD	Spinocerebellar ataxia 1 (SCA1)	164400	y	[137]
ATXN10	ATTCT/ATTGT	10–32	>280–4500	intron	22q13	AD	Spinocerebellar ataxia 10 (SCA10)	611150	n	[138]
ATXN2	CAG	13–31	>31–500	exon	12q24.12	AD	Spinocerebellar ataxia 2 (SCA2)	183090	y	[139]
ATXN3	CAG	12–44	>54–87	exon	14q32.12	AD	Spinocerebellar ataxia 3 (SCA3); Machado-Joseph disease	109150	y	[140]
ATXN7	CAG	4–33	>36–460	exon	3p14.1	AD	Spinocerebellar ataxia 7 (SCA7)	164500	y	[141]
ATXN8 / ATXN8OS	CAG/CTG	15–50	>74	3’UTR	13q21	AD	Spinocerebellar ataxia 8 (SCA8)	608768	y	[142]
BEAN1/TK2	TAAAA*	variable	≥ 110–760	intron	16q22	AD	Spinocerebellar ataxia 31 (SCA31)	117210	n	[143]
C9orf72	GGGGCC	3–25	>30	5’UTR/intron	9p21.2	AD	Frontotemporal dementia and/or amyotrophic lateral sclerosis 1	105550	y	[144]
CACNA1A	CAG	4–18	19–33	exon	19p13.13	AD	Spinocerebellar ataxia 6 (SCA6)	183086	y	[145]
CACNA1C	30 bp	variable	na****	intron	12p13.33	AD	Schizophrenia and bipolar disorders	620029	n	[98]
CNBP	CCTG/CAGG	11–30	>50–11000	intron	3q21.3	AD	Myotonic dystrophy type 2 (DM2)	602668	y	[146]
CSTB	CCCCGCCCCGCG	2–3	≥ 30–75	promoter/5’UTR	21q22.3	AR	Unverricht-Lundborg syndrome (EPM1)	254800	y	[147]
DAB1	ATTTT**	7–400	>31–75	intron	1p32	AD	Spinocerebellar ataxia 37 (SCA37)	615945	n	[148]
DMPK	CTG	5 – 37	>50 – 2,000	3’UTR	19q13.32	AD	Steinert myotonic dystrophy syndrome (DM1)	160900	y	[149]
FGF14	GAA	50	>250	intron	13q33.1	AD	Late-Onset Spinocerebellar Ataxia 27B	620174	y	[123]
FMR1	CGG	5–50	>200	5’UTR	Xq27.3	XLD	Fragile X syndrome (FXS)	300624	y	[150]
FMR1	CGG	5–50	55–200	5’UTR	Xq27.3	XLR	Premature ovarian failure 1 (POF1)	311360	y	[151]
FMR1	CGG	5–50	55–200	5’UTR	Xq27.3	XLR	Fragile X-associated tremor/ataxia syndrome (FXTAS)	300623	y	[152]
FOXL2	GCN	14	19–24	exon	3q22.3	AD	Blepharophimosis, epicanthus inversus, and ptosis, type 1 (BPES)	110100	n	[153]
FXN	GAA	5–34	>66–1300	intron	9121.11	AR	Friedreich ataxia 1 (FRDA)	229300	y	[154]
GIPC1	CGG	12–32	≥ 97–120	5’UTR	19p13.12	AD	Oculopharyngodistal myopathy 2 (OPDM2)	618940	n	[155]
GLS	GCA	8–16	≥ 680 – 1400	5’UTR	2q32.2	AR	Global developmental delay, progressive ataxia, and elevated glutamine	618412	n	[156]
HOXA13	GCG	12–18	18–30	exon	7p15.2	AD	HFGSHand-foot-uterus syndrome (HFU)	140000	n	[157]
HOXD13	GCG	15	24	exon	2q31.1	AD	Synpolydactyly (SPD1)	186000	n	[158]
HTT	CAG	6–35	>36	exon	4p16.3	AD	Huntington disease (HD)	143100	y	[159]
JPH3	CAG	6–28	>40–58	exon	16q24.2	AD	Huntington disease-like 2 (HDL2)	606438	n	[160]
LRP12	CGG	13–45	90–130	5’UTR	8q22.3	AD	Oculopharyngodistal myopathy 1 (OPDM1)	164310	n	[161]
MARCHF6	ATTTT**	10–30	≥ 660–2800	intron	5p15.2	AD	Epilepsy, myoclonic, familial adult, 3 (FAME3)	613608	n	[120]
NOP56	GGCCTG	5–14	≥ 650–2500	intron	20p13	AD	Spinocerebellar ataxia 36 (SCA36)	614153	n	[162]
NOTCH2NLC	CGG	7–60	≥ 61–500	5’UTR/exon1	1q21.2	AD	Neuronal intranuclear inclusion disease (NIID)	603472	n	[163]
NUTM2B-AS1	CGG/CCG	3–16	40–60	noncoding RNA	10q22.3	AD	Oculopharyngeal myopathy with leukoencephalopathy 1 (OPML1)	618637	n	[121]
PABPN1	GCG	6–10	>11–17	exon	14q11.2	AD	Oculopharyngeal muscular dystrophy (OPMD)	164300	n	[164]
PHOX2B	GCN	20	25–29	exon	4p13	AD	Central hypoventilation syndrome 1 (CCHS)	209880	n	[165]
PLIN4	99 bp	27–31	40	exon	19p13.3	AD	Myopathy with rimmed ubiquitin-positive autophagic vacuolation (MPUPAV)	613247+	n	[166]
PPP2R2B	CAG	4–32	≥ 43–78	5’UTR	5q31	AD	Spinocerebellar ataxia 12 (SCA12)	604326	y	[167]
PRDM12	GCG	12	18–19	exon	9q34.12	AR	Neuropathy, hereditary sensory and autonomic, type VIII (HSAN8)	616488	n	[168]
PRNP	24 bp	1	6	exon	20p13	AD	Creutzfeldt-Jakob Disease	123400	n	[169]
RAPGEF2	ATTTT**	na	na	intron	4q32.1	AD	Epilepsy, myoclonic, familial adult, 7 (FAME7)	618075	n	[118]
RFC1	AAAAG / AAAGG / AAGAG / AGAGG***	variable	≥ 400 – 2000	intron	4p14	AR	Cerebellar ataxia with neuropathy and bilateral vestibular areflexia syndrome (CANVAS)	614575	y	[170]
RUNX2	GCN	17	27	exon	6p21.1	AD	Cleidocranial dysplasia 1	119600	n	[171]
SAMD12	ATTTT**	7-exp	≥ 440–3680	intron	8q24	AD	Epilepsy, myoclonic, familial adult, 1 (FAME1)	601068	n	[172]
SOX3	GCN	15	26	exon	Xq27.1	XLR	Intellectual developmental disorder, X-linked, with isolated growth hormone deficiency	300123	n	[173]
STARD7	ATTTT**	9–20	≥ 661–735	intron	2q11.2	AD	Epilepsy, myoclonic, familial adult, 2 (FAME2)	607876	n	[119]
TAF1	CCTCT	none	30–55	intron	Xq13.1	XLR	X-linked Dystonia-Parkinsonism (XDP)	314250	n	[174]
TBP	CAG (or CAG/CAA)	25–40	>42–66	exon	6q27	AD	Spinocerebellar ataxia 17 (SCA17)	607136	y	[175]
TCF4	CTG	5–31	> 50	intron	18q21.2	AD	Fuchs endothelial corneal dystrophy 3 (FECD3)	613267	y	[176]
TNRC6A	ATTTT**	na	na	intron	16p12.1	AD	Epilepsy, myoclonic, familial adult, 6 (FAME6)	618074	n	[118]
WDR7	69 bp	variable	na *****	intron	18q21.31	AD	Amyotrophic lateral sclerosis (ALS)	606640	n	[99]
XYLT1	GGC	9–20	> 100	promoter	16p12.3	AR	Desbuquois dysplasia 2; Baratela-Scott Syndrome (BSS)	615777	n	[177]
YEATS2	ATTTT**	7–400	na	intron	3q27.1	AD	Epilepsy, myoclonic, familial adult, 4 (FAME4)	615127	n	[178]
ZIC2	GCG	15	25	exon	13q32.3	AD	Holoprosencephaly 5 (HPE5)	609637	n	[179]

^*repeat motif is TGGAA/TAGAA when pathogenic

^**repeat motif is ATTTC when pathogenic

^***repeat motif is AAGGG when pathogenic

^****specific 30-mer sequences are associated with disease risk

^*****longer repeat associated with ALS

⁺MIM for PLIN4, as the phenotype is not in OMIM

Over the last decade other more heterogeneous complex disorders often neurological in nature with diverse genetic etiologies have been shown to result from the instability of tandem repeats. A hexanucleotide repeat expansion of the GGGGCC motif in C9orf72 to thousands of base pairs in length is the cause of chromosome 9p21-linked amyotrophic lateral sclerosis and frontotemporal dementia (ALS-FTD) [89,90]. Interestingly, even carriers of two intermediate-size alleles (<20 repeats) may be at risk for ALS and associated disorders [117]. Expansions of the TTTCA and TTTTA repeats in a variety of genes (e.g., SAMD12, TNRC6A, RAPGEF2) are now thought to underlie benign adult familial myoclonic epilepsy [118]. Later, the same pentanucleotide TTTCA expansion beyond 10 kbp in length was observed in association with familial autosomal myoclonic epilepsy, albeit mapping to the intron of different genes (MARCH6 and STARD7) [119,120]. The finding of the same pathogenic pentanucleotide expansion in genes that share no other property other than being highly expressed in the brain has led to speculation that it is the transcribed repeat itself instead of the specific function of the gene that underlies the epilepsy pathology. In a separate instance, a VNTR in an intron of CACNA1C was found to be associated with schizophrenia by motif composition rather than length [98]. In this light, understanding not only the repeat length but also the composition of the repeat appears critical to the pathogenic state [121] making both detection and deciphering the sequence of the expanded repeat structures critical to understanding disease risk (see above).

LRS has been used to discover novel repeat sequences in individuals with familial adult myoclonic epilepsy with negative clinical testing [122] and to identify new genes and expand the phenotype of genes associated with spinocerebellar ataxia [123,124]. For example, in a cohort with late-onset spinocerebellar ataxia 27B, Pellerin and colleagues used LRS to clarify a heterozygous trinucleotide GAA repeat expansion in the gene, FGF14, initially identified with short-read sequencing (SRS) [123]. LRS has also been used to identify disease-causing expansions within duplicated genes–so-called repeats within repeats, as in neuronal intranuclear inclusion disease (NIID, OMIM: 603472) and the expanding CGG repeat mapped to the 5′ UTR of NOTCH2NLC or NBPF19–a human-specific duplicated gene [118,125]. Expansion of this repeat beyond 100 repeat units has been associated not only with NIID but also with oculopharyngodistal myopathy with neurological manifestations (OMIM: 619473).

Until recently, clinical testing for tandem repeat disorders was limited to PCR-based approaches or Southern blot. Clinical testing labs have begun to evaluate limited loci using short-read-based approaches, likely due to the integration of ExpansionHunter [60] into the DRAGEN pipeline [126]. LRS has also been used to evaluate select repeats. In 2021, for example, Invitae began evaluating PCR products from FMR1 alleles with 55–90 triplet repeats on the PacBio platform [127]. While it is exciting to see both SRS and LRS approaches beginning to be used in the clinical space, there remains considerable need to broaden the number of tandem repeats that can be clinically evaluated (Table 3), which may be met by the introduction of LRS into the clinical environment [128]. Routine LRS of amplicons [129], CRISPR-CAS targets [122], and ultimately whole genomes will make routine clinical testing of these disorders possible [130] and perhaps drive novel therapeutic approaches [131].

Future prospects.

LRS has created a renaissance of interest in investigating the disease significance and biological mechanisms of expanding/contracting repeats and their epigenetic consequences. Tools are being developed to search for associations with gene expression [65,68,69,180] using GTEx genomes and trait association in the UK Biobank [181]. Intriguing candidate loci and genome-wide patterns of tandem repeat variation are being identified in families with autism and developmental delay[182,183] that will benefit from more extensive LRS characterization. Reports of enrichment of tandem repeat expansions in various cancer types [184] will justify LRS of tumor-matched patient samples. Still, the discovery of new disease or functional roles of tandem repeat variation is limited because almost all disease cohorts of sufficient sample size have been sequenced by short-read NGS platforms. New computational tools, such as PanGenie [73], are leveraging the sequence resolution and linkage disequilibrium provided by phased pangenomes to more accurately genotype and perform genome-wide associations in preexisting NGS datasets [17].

Current population-level analyses of tandem repeats using LRS with digital readouts of motif counts and tandem repeat composition are from a relatively modest number of human diversity samples [17,18]. Improvement in accuracy from the ONT PromethION and increases in throughput and reduction in costs such as the PacBio recent release of the Revio system will make it possible to begin to more cost-effectively consider larger disease and population cohorts. The low-coverage LRS of 3,622 Icelandic individuals has already led to the association of two new loci with height and atrial fibrillation [78]. A smaller pilot study of familial LRS of simplex autism discovered de novo STR and VNTR variation missing from SRS [13]. As costs decrease and throughput increases, not only will more genomes be sequenced but so too will higher quality phased genomes of near T2T status providing access to the largest and most identical repeats. We anticipate the discovery of many new disease associations as LRS of unsolved Mendelian disorders and deeply phenotyped and diverse cohorts such as GTEx samples [185], autism families [186], tumor-normal cancer studies [187] and biobanks (e.g., All of Us) begin over the next few years.

Another important area of future development will be the application of new LRS platforms to investigate somatic variation and expression at the single-cell level. This will require the development of cost-effective methods to generate long-read sequences from low-input DNA and mRNA materials. New methods such as MAS-ISO-seq [188] are beginning to emerge, which allow quantitative sequencing of longer transcripts (~1 kbp) providing tissue-specific spliceform characterization. In light of the extensive mosaicism already associated with tandem repeat disorders, the characterization of somatic variability at the DNA will be critical and clinically relevant. For example, somatic instability has been observed in Huntington’s disease, ALS [90], and Fragile X syndrome [189]. In these cases, the severity and age of onset of the disease can depend on the number of repeat units, as well as the extent of somatic instability in affected tissues. Some tissues may have a higher propensity for somatic instability than others, as has been shown forFMR1 where the instability of the CGG repeat expansion can vary, with higher instability observed in brain tissues compared to blood [190]. Similarly, instability of the CAG repeat expansion in the HTT gene can depend on genetic modifiers in other genes, such as DNA repair enzymes [191]. A better understanding of repeat structure, expression of genes known to be associated with somatic instability, and response of individual cells to repeat expansions will expand our understanding of the pathogenesis, uncover previously unknown associations [192], and potentially guide novel treatment approaches [193].

Summary Points.

• Tandem repeat sequences span multiple classes of DNA including short tandem repeats, variable-number tandem repeats, satellite DNA, and segmental duplications.

• Long-read sequencing is rapidly improving our understanding of tandem repeat organization and variation in human genomes.

• Current advancements in the scale and diversity of populations undergoing long-read sequencing have unveiled nearly 150,000 polymorphic VNTRs where 33.5% are located in the subtelomeric regions of the genome.

• Many different computational approaches are required to discover tandem repeat variation with short reads, depending on the class and scale of the variation.

• Standard approaches for variant analysis with long reads can detect variation in tandem repeats, however specific methods are required to organize and visualize tandem repeat variation.

• Patterns within the complex nature of tandem repeat variation becomes more clear when length and motif variation are visualized.

• Over 56 diseases linked to tandem repeat variation have been identified.

• The pace that tandem repeat disease loci are being discovered is increasing due to long-read sequencing.

• Detailed resolution of tandem repeats using new technologies may lead to novel therapeutic approaches