NGS4THAL, a One-Stop Molecular Diagnosis and Carrier Screening Tool for Thalassemia and Other Hemoglobinopathies by Next-Generation Sequencing

Abstract

Thalassemia is one of the most common genetic diseases and a major health threat worldwide. Accurate, efficient, and scalable analysis of next-generation sequencing (NGS) data is much needed for its molecular diagnosis and carrier screening. We developed NGS4THAL, a bioinformatics analysis pipeline analyzing NGS data to detect pathogenic variants for thalassemia and other hemoglobinopathies. NGS4THAL realigns ambiguously mapped NGS reads derived from the homologous Hb gene clusters for accurate detection of point mutations and small insertions/deletions. It uses a combination of complementary structural variant (SV) detection tools and an in-house database of control data containing specific SVs to achieve accurate detection of the complex SV types. Detected variants are matched with those in HbVar (A Database of Human Hemoglobin Variants and Thalassemia Mutations), allowing recognition of known pathogenic variants, including disease modifiers. Tested on simulation data, NGS4THAL achieved high sensitivity and specificity. For targeted NGS sequencing data from samples with laboratory-confirmed pathogenic Hb variants, it achieved 100% detection accuracy. Application of NGS4THAL on whole genome sequencing data from unrelated studies revealed thalassemia mutation carrier rates for Hong Kong Chinese and Northern Vietnamese that were consistent with previous reports. NGS4THAL is a highly accurate and efficient molecular diagnosis tool for thalassemia and other hemoglobinopathies based on tailored analysis of NGS data and may be scaled for population carrier screening.

Thalassemia is one of the most common genetic diseases, with high prevalence and carrier rates in many tropical and subtropical regions.^1,2 It is a group of heterogeneous inherited disorders with a broad spectrum of mutations that often differs among geographic regions and ethnic groups.³ More than 400 thalassemia pathogenic variants have been documented in HbVar (A Database of Human Hemoglobin Variants and Thalassemia Mutations)⁴ up to date, which is a comprehensive database on pathogenic and benign Hb genetic variants.

Deletions in the α-globin gene locus are the underlying causes of most of the α-thalassemia,⁵ whereas point mutations and small insertions/deletions (InDels) in the β-globin gene region explain most β-thalassemia.⁶ Although most of the pathologic mutations involve the coding regions of the Hb genes, there are pathogenic variants in the regulatory elements in the intergenic regions or introns, affecting expression or splicing of the Hb genes.⁷ In addition, some disease modifiers,8, 9, 10 both in the Hb gene regions and in other regions of the genome, impact disease severity of β-thalassemia and sickle cell disease (Mendelian Inheritance in Man number 603903), mainly through regulation of the expression of fetal Hb.

Traditionally, molecular diagnosis of thalassemia (as well as other hemoglobinopathies) starts with hematological tests (such as mean corpuscular Hb and mean corpuscular volume) and Hb electrophoresis, followed by genetic tests to identify the exact pathogenic variant(s).^11,12 Although it works well in most cases, the process requires cumbersome laboratory work and step-by-step decision making based on the circumstances, which makes it difficult to scale up. In some rare cases, certain disease mutations may not show positive hematological features and might be missed by hematological screening.¹²

Next-generation sequencing (NGS) technology provides a possibility of detecting almost all the pathologic variants and disease modifiers in a single experimental setting and has been used to various degrees in the molecular diagnosis of thalassemia and other hemoglobinopathies.13, 14, 15 However, none of these attempts has resulted in an accurate, efficient, and scalable solution for Hb gene variant detection, and as a result, NGS technology has not been widely applied in clinical laboratories and population carrier screening up to date. As genomic sequencing is getting cheaper and more readily available, an accurate, efficient, and universally applicable thalassemia and Hb genetic variant detection platform analyzing NGS data is greatly needed to screen relevant Hb genetic variants and disease modifiers for this group of diseases.

The α-globin gene clusters are located in 16p13.3, a region with features like homologous sequences, interspersed repeats, and various types of genetic variants.¹⁶ Short NGS reads derived from this region are often ambiguously mapped to multiple positions because of sequence homology. This not only causes missing detection of point mutations and InDels, but inaccurate read depth estimation, which is one of the features required for detection of structural variants (SVs). Furthermore, compound forms of Hb SVs are common in thalassemia endemic regions. For example, --^SEA/–α^3.7, a compound heterozygous genotype that causes Hb H disease, is common in Hong Kong and some other Southeast Asian regions. Being able to detect complex forms of genotypes is essential for analysis tools aiming at comprehensive thalassemia molecular diagnosis and carrier screening.

Herein, we present NGS4THAL, a one-stop bioinformatics analysis pipeline aiming at detecting Hb genetic variants through analyzing various forms of NGS data. NGS4THAL recovers and realigns NGS reads with multiple alignments in the Hb regions for accurate detection of point mutations and small InDels. It uses SV callers with complementary algorithms, tailored parameters, and an in-house database of relevant controls of specific SVs to improve SV detection. Validation of NGS4THAL on simulation data and targeted sequencing data on samples with globin gene variants previously confirmed from clinical laboratories achieved high sensitivity and specificity. It can also detect thalassemia pathogenic variants located in regulatory regions, as well as known disease modifiers. NGS4THAL is implemented in Python language (https://github.com/JavenCao/Realign_BAM, last accessed December 23, 2021).

Materials and Methods

Detection of Point Mutations and Small InDels

A schematic illustration of NGS4THAL is presented in Figure 1. NGS4THAL tries to improve the detection of single-nucleotide variants (SNVs) and small InDels through the recovery and realignment of NGS sequencing reads with multiple alignments (RMAs), which are common in the two Hb genomic regions (16p13.3 and 11p15.4, respectively).

Figure 1.

Workflow of NGS4THAL. CNV, copy number variation; InDel, small insertion and deletion; RMA, next-generation sequencing reads with multiple alignments; SNV, single-nucleotide variant; SV, structural variant.

NGS4THAL firstly recovers informative RMAs in the Hb gene regions from the alignment files. The following criteria are applied for recovering an RMA: i) it maps to one of the homologous regions (namely, HBZ-HBZP1, HBA2-HBA1-HBAP1, or HBG1-HBG2); and ii) it has a bwa¹⁷-based mapping quality score (MAQ) of 0 due to having an alternative mapping position with an identical mapping profile as that of the primary mapping position; or iii) it has an MAQ <30 and the difference of mismatches between the primary mapping hit and the alternative mapping hit is <3 bp under the short-read settings (eg, 100 to 300 bp). In total, NGS reads meeting these criteria account for about 95% of the RMAs in the two Hb regions. After recovering these reads, one read of the read pair, designated as R1, is realigned to either the primary mapping position or the alternative mapping position at random. The mate read, designated as R2, will be mapped nearby but within 5 SDs of the mean insert size of the NGS library from R1. The MAQs of these realigned RMAs will be adjusted to 60 in a new BAM file for downstream analysis. Using Python module pysam, an upper limit of maximum of 1000 bp between the paired NGS reads was applied for this process in order not to realign reads overlapping a large deletion. SNVs/InDels are detected on the basis of the new BAM file by GATK-HaplotypeCaller version 3.8.¹⁸

The detected genetic variants are classified into known pathogenic SNVs/InDels, disease modifiers, benign Hb variants, or novel variants, with the latter being reported for manual annotation. Candidate variants located in equivocal positions within the homologous genes, such as between HBA2 and HBA1 (eg, HBA2:c.64C>G or HBA1:c.64C>G), cannot be distinguished on the basis of available short-read NGS data. NGS4THAL will report and label the causal mutation according to entries in the HbVar database. Further laboratory experiments using traditional genetic testing methods will have to be applied to determine the exact position of these mutations.

Detection of Hb Structural Variants

Depending on the relative positions between the breakpoints of the SVs and the homologous sequences, Hb SVs were classified into four subtypes: subtype-A SVs, whose breakpoints are both located in nonhomologous regions; subtype-B and subtype-C SVs, for which one of the two breakpoints is located in a homologous region; and subtype-D SVs, whose breakpoints are both located in homologous regions (Supplemental Figure S1). NGS4THAL uses complementary tools to detect these SVs, including BreakDancer version 1.1,¹⁹ Pindel version 0.2.0,²⁰ and CoNIFER version v0.2.2,²¹ which are methods based on features including insert size estimate, mapping positions of the two fragments on splitting a read, and read depth, respectively.

Some modifications were adopted to tailor the analysis process to better detect variants in Hb gene regions. Briefly, requirement on minimum alternative MAQ was lowered to 0 (–q 0), allowing BreakDancer to consider RMAs for detecting SVs located in homologous sequences. Results from BreakDancer are used as a source of input for Pindel, which is a practice that has been shown to increase the accuracy of SV detection.²⁰ A set of in silico probes covering the two Hb loci continuously is generated to allow the use of CoNIFER in detecting duplications and deletions, with each probe at 80 bp in length and with a 25-bp overlap between probes. To further increase sensitivity, the detection threshold was lowered from 1.0 to 0.5, effectively allowing the detection of copy number variation with 50% read-depth change relative to the baseline average, instead of the default threshold of 100% depth change.

In addition, a stage-wise process was implemented in NGS4THAL to detect compound SVs (such as --^SEA/–α^3.7). Specifically, in the screening stage, all the samples are analyzed using BreakDancer, Pindel, and CoNIFER to determine the presence of SVs/copy number variations. In the fine-profiling stage, copy number variation status of samples with large SVs covering both the HBA2 and HBA1 genes (such as --^SEA) is further examined against reference data from heterozygotes deletion carriers to detect compound heterozygous SVs (such as using --^SEA as controls to detect the co-existence of –α^3.7 in the form of --^SEA/–α^3.7). The principle adopted was to fine-tune the background baseline to make it in line with one of the two types of the compound SVs, so that the coexistence of the other type of SV on the other haplotype becomes easier to detect.

Data Sets Used to Evaluate NGS4THAL

Simulated Data

A total of 69 point mutations and 19 small InDels, covering exons and introns of HBA2, HBA1, HBB, and HBD genes, as well as 25 SVs spanning HBA2 and HBA1 genes were selected for simulation (Supplemental Table S1). Sequencing reads from 10 independent heterozygous carriers for each variant were simulated, as well as sequencing reads from 10 homozygous –α^3.7 genotypes. An additional 870 independent individuals without any causal mutations were simulated as negative controls, summing up to a total of 2000 simulated individuals. VarSim²² and ART version 2016-06-05²³ were used sequentially to generate NGS reads with continuous coverage of the HBA (hg19: chromosome 16:132,744–279,999; size = approximately 147 kbp) and the HBB (hg19: chromosome 11:5,221,230–5,321,230; size = 100 kbp) regions. The simulated NGS reads were 150 bp in length, paired end, and with similar base pair quality to those from the Illumina (San Diego, CA) HiSeq 2500 platform.

Targeted Sequencing Data on Samples with Clinical Laboratory-Confirmed Mutations

DNA samples from 71 patients with Hb mutations causal to thalassemia or other hemoglobinopathies were kindly provided by the Department of Pathology in Queen Mary Hospital and Clinical Genetic Service, Department of Health, Hong Kong, and Faculty of Medicine, Chulalongkorn University, Bangkok, Thailand. These laboratory-confirmed Hb mutations included Hb Constant Spring, Hb Quong Sze, Hb Luton, HbQ-Thailand, HbE, –28 (A>G), CD 41/42 (–CTTT), IVS I-5 (G>C), IVS II-654 C>T, IVS I-1 (G>A), CD 17 (AAG>TAG), Hb Malay, Hb Hoshida, CD 35 TAC>TAA, CD71/72 (+A), --^SEA, –α^3.7, –α^4.2, and compound Hb SVs, such as --^SEA/–α^4.2, --^SEA/–α^3.7, and --^SEA/ααα^anti4.2.

Enrichment probes were designed and manufactured by Roche (Basel, Switzerland) with near continuous coverage of the two Hb gene loci, including the known regulatory elements (Supplemental Table S2). NGS library preparation and paired-end sequencing were performed for the above-mentioned samples on the Illumina HiSeq 1500 platform by the Centre of Genomic Sciences of the University of Hong Kong. Around 1500-fold coverage depth was achieved for the targeted regions for these samples. The data set was used to test the performance of NGS4THAL in detecting the clinically confirmed mutations.

WGS Data Set Used to Test the Performance of NGS4THAL on Population Data

Whole genome sequencing (WGS) data on a total of 954 samples from a Hirschsprung disease study^24,25 conducted in the University of Hong Kong were also used to test the performance of NGS4THAL, including 447 Hirschsprung disease cases and 507 healthy controls. The ethnic origins of the samples include Hong Kong Chinese (N = 375), mainland Chinese (N = 161), and Northern Vietnamese (N = 418). They were paired-end NGS reads of 150 bp in length and with an average coverage of 30-fold. The distinct mutation types detected from this cohort were rendered for orthogonal validation by traditional experimental methods, including Sanger sequencing (Table 1) and multiplex ligation-dependent probe amplification [MLPA; namely, the MRC Holland (Amsterdam, the Netherlands) SALSA MLPA P140 HBA probemix for α-thalassemia in vitro diagnosis].

Table 1.

Oligonucleotide Sequences Used for Variant Confirmation in Sanger Sequencing

Group	Target variants	Oligonucleotide sequence	Annealing temperature, °C	Product size, bp
1	Hb Westmead (HBA2:c.369C>G)∗	F = 5′-CCTCTCAGGGCAGAGGATCAC-3′ R = 5′-TGCAGAGAGGTCCTTGGTCT-3′	61.6460.18	416, 2246†
2	-28 (A>G) (HBB:c.-78A>G) Hb E (HBB:c.79G>A) CD 17 (AAG->TAG)(HBB:c.52A>T) -50 G>A (HBB:c.-100G>A) CD 41/42 (-CTTT) (HBB:c.126_129delCTTT)	F = 5′-TGCTTAGAACCGAGGTAGAGTTT-3′ R = 5′-AGGTGCCCTTGAGGTTGTC-3′	59.1759.54	915
3	IVS-II-654 (C->T) (HBB:c.316–197C>T)	F = 5′-GTGAGTCTATGGGACGCTTGA-3′ R = 5′-TTGGACAGCAAGAAAGCGAG-3′	59.5258.77	1000
4	-30 T>C (HBD:c.-80T>C) -77 (T->C) (HBD:c.-127T>C)	F = 5′-TTCCAACCTCTCAAAATTCCCTTG-3′ R = 5′-CTGAGGGTAGGAAAACAGCCCAA-3′	59.4162.40	711

F, forward; R, reverse.

^∗Hb Constant Spring (HBA2: c.427T>C) was confirmed by multiplex ligation-dependent probe amplification (Holland SALSA MLPA P140 HBA probemix for α-thalassemia in vitro diagnostic assay).

^†Extension time of the PCR amplification was limited to 40 seconds to avoid unintended products longer than 500 bp.

Results

Detection of Point Mutations and Small InDels

RMAs (NGS reads with multiple alignment) occur within HBZ-HBZP1 and within HBAP1-HBA2-HBA1 (Supplemental Figure S2) in the α-globin gene locus, and between HBG1 and HBG2 in the β-globin gene locus (Supplemental Figure S3). NGS4THAL identifies these RMAs and realigns the variant-informative RMAs to equally likely positions in the reference genome at random. This process was shown to significantly improve variant detection located in the homologous regions (Supplemental Figure S4).

Testing of NGS4THAL on simulation data showed that, without the process of RMA realignment, Hb SNVs located in homologous regions were often missed, including variants located in exons 1 and 2 of HBA1 and HBA2 genes (Figure 2, A and B). This was because NGS reads derived from these regions often failed to meet mapping quality threshold and were not contributing to the variant detection process (Supplemental Figure S5). In contrast, on recovery and realignment of the informative RMAs by NGS4THAL, detection of variants in these regions was significantly improved. The realignment process didn't affect detection of SNVs and InDels in HBB (Figure 2C) and HBD (Figure 2D) regions. For SNVs in the simulation data, NGS4THAL achieved a 97.87% sensitivity, compared with 68.09% in the absence of NGS read realignment process. Considering SNVs with small InDels altogether, NGS4THAL achieved a sensitivity of 98.48%, compared with a 77.27% sensitivity without the RMA realignment process.

Figure 2.

Detection of point mutations and small insertions and deletions (InDels) by NGS4THAL on simulated data. A–D: Simulated variants and detected variants by NGS4THAL are shown for HBA2, HBA1, HBB, and HBD, respectively. Point mutations are shown as thin lines, and small InDels are shown as color blocks proportional to their sizes. SNV, single-nucleotide variant.

NGS4THAL does not treat variants detected on equivocal positions between the homologous genes (such as HBA2:c.64C>G or HBA1:c.64C>G) as false positives. Under the scheme of short-reads sequencing (eg, 100 to 300 bp in read length and 300 to 500 bp in insert size), the exact position of the point mutations between HBA2 and HBA1 cannot be determined with certainty, particularly for exons 1 to 2 of these two genes that share nearly 2-kb sequences of complete identity. NGS4THAL reports them while labeling the potential causal ones noted in HbVar database. Acknowledging this caveat, a 100% specificity was considered achieved.

Testing of NGS4THAL on the 71 samples sequenced by the targeted sequencing platform, 100% sensitivity was achieved for the pathogenic Hb point mutations and small InDels, which were previously confirmed by traditional laboratory methods (Supplemental Figure S6). Of note, the Hb Q-Thailand and Hb_Luton mutations, located in exons 1 and 2 of HBA2 and HBA1, would have been missed without the RMA realignment process.

Applying NGS4THAL on WGS data from the Hirschsprung disease study,^24,25 several thalassemia pathogenic point mutations and small InDels common to local populations were detected, including Hb Constant Spring, Hb Westmead, CD41/42 (-CTTT), and -28 (A>G) for Hong Kong Chinese, and HbE, CD 17 (AAAG>TAG), and -50 G>A for Northern Vietnamese. The detected carrier rates for these variants were consistent with those from previous epidemiologic reports^26,27 (Table 2 and Supplemental Table S3). Furthermore, all the unique mutations detected above were orthogonally confirmed by PCR amplification followed by Sanger sequencing (Supplemental Figure S7).

Table 2.

Summary Table of Detected Thalassemia Pathogenic Variants by NGS4THAL from Whole Genome Sequencing Cohorts

Hong Kong Chinese population (N = 375)			Northern Vietnamese population (N = 161)
Pathogenic variants	Carrier counts	%	Pathogenic variants	Carrier counts	%
Total thalassemia mutation∗	44	11.7	Total thalassemia mutation	24	14.9
α-Thalassemia mutation	36	9.6	α-Thalassemia mutation†	12	7.5
--SEA	17‡	4.5	--SEA	6	3.7
–α3.7	12	3.2	–α3.7	5(4)	3.1
–α4.2	2	0.5	Hb Westmead	1	0.6
Hb Constant Spring	2‡	0.5
Hb Westmead	3	0.8
β-Thalassemia mutation	7	1.9	β-Thalassemia mutation	11	6.8
CD 41/42 (-CTTT)	6	1.6	Hb E	5§	3.1
-28 (A>G)	1	0.3	CD 17 (AAG>TAG)	3	1.9
			-50 G>A	3	1.9
δ-Thalassemia mutation	1	0.3	δ-Thalassemia mutation	1	0.6
-30 T>C	1	0.3	-77 (T>C)	1	0.6

^∗--^SEA: NG_000006.1: g.26264_45564del19301, –α^3.7: NG_000006.1: g.34164_37967del3804, –α^4.2: NC_000016.9:g.219817_(223755_224074)del, Hb CS: HBA2: c.427T>C, Hb Westmead: HBA2: c.369C>G, CD 41/42 (-CTTT): HBB:c.126_129delCTTT, -28 (A>G): HBB: c.-78A>G, -30 T>C: HBD: c.-80T>C, HbE: HBB:c.79G>A, CD 17 (AAG>TAG): HBB:c.52A>T, -50 G>A: HBB:c.-100G>A, -77 (T>C): HBD:c.-127T>C. Variants can be found (https://www.ncbi.nlm.nih.gov/nuccore, last accessed July 8, 2022).

^†Include a carrier with both α- and β-thalassemia mutations, and each of the mutations was counted.

^‡Include a sample from compound heterozygote of --^SEA/Hb CS, and each of the mutation haplotypes was counted as 1.

^§A homozygote was counted as 2.

Detection of Structural Variants

Of the four subtypes of structural variants (SVs) defined in Materials and Methods, subtype-A deletions (such as –α^2.7), often with definitive characteristics from paired-end sequencing, are usually detected without ambiguity (Figure 3A).

Figure 3.

Detection of Hb structural variants (SVs) by NGS4THAL using simulated data. A–D: The next-generation sequencing signatures and detection of subtype-A to subtype-D Hb SVs, the classification of which is based on relative positions between the homologous regions and the SV breakpoints. For each SV subtype: Top panels: the alignment landscape in the format of Integrative Genomics Viewer screenshot, followed by the simulated SV and the detection results from BreakDancer, Pindel, and CoNIFER (with cutoffs of 1.0, 0.8, 0.5, and 0.3, respectively). CoNIFER detections with cutoff of 0.5 are used in NGS4THAL and are shown in green. Bottom panels: Read-depth changes detected by CoNIFER. Black lines are read-depth changes for all the simulated individuals. Red curves are read-depth changes for the SV carriers in question. The y axis is the relative copy number with reference to the baseline defined by all input samples, and the vertical dashed green lines illustrate the start and end of the CoNIFER variant detection using a cutoff of 0.5. The simulated SVs are 2.7-kb deletion (NG_000006.1:g.36664_39364del2701), representing SV subtype A; 5.3-kb deletion (NG_000006.1:g.28684_33930del5246), representing subtype B; 3.5-kb deletion (NG_000006.1:g.37464_40964del3501), representing subtype C; and 3.7-kb (type I) deletion (NG_000006.1:g.34164_37967del3804), representing subtype D. Variant details can be found at https://www.ncbi.nlm.nih.gov/nuccore (last accessed July 8, 2022). Chr, chromosome.

For subtype-B (Figure 3B) and subtype-C (Figure 3C) deletions, with one of the two breakpoints located in homologous sequences, such as –α^5.3, tailored software BreakDancer detected two candidate deletions: one that was consistent with the simulated –α^5.3 deletion and the other that was a longer deletion with one of the ends incorrectly assigned because of ambiguous mapping (Figure 3B). The issue of breakpoint ambiguity is resolved by matching the results with HbVar database, and the one matching with the known variants in the database is reported as the likely variant. For algorithms based on mapping of split reads in this case, such as Pindel, the 5′ breakpoint could be correctly determined with supporting reads that span the breakpoint. Because the split fragments could also be mapped to the paralogous HBA1 gene, this causes uncertainty in defining the 3′ breakpoint. NGS4THAL detects all the plausible SVs supported by the data, and calls the variants based on matching all the candidates with the known Hb SVs from the HbVar database. In addition, simultaneously applying CoNIFER also detected subtype-B and subtype-C deletions, analyzing changes in read depth.

The –α^3.7 is a subtype-D deletion, caused by crossing over between misaligned homologous X, Y, and Z boxes during meiosis.²⁸ As both breakpoints are located in homologous sequences, subtype-D variants cannot be detected by evaluating insert size aberrations by tools such as BreakDancer (Figure 3D). Similarly, they cannot be detected by split-read mapping methods such as Pindel either. Nevertheless, subtype-D deletions could still be detected by read depth–based algorithms, such as CoNIFER, when parameters are properly adjusted. In the case of –α^3.7 (hg19: chromosome 16:223,299–227,103), NGS4THAL detected a deletion spanning hg19:chromosome 16:222,879–226,799. Matching the result with variants in HbVar database reveals that this is likely the –α^3.7 deletion (Figure 3D).

Detecting compound heterozygous forms of SVs, which are not uncommon in endemic regions for thalassemia, can be complicated. Taking --^SEA/–α^3.7 as an example, in the screening stage, although the --^SEA deletion could be identified by both BreakDancer and Pindel, the loss of another copy of ΗΒΑ2 on top of --^SEA (eg, –α^3.7) could not be correctly determined by any of the SV tools, including CoNIFER with modified parameters. Considering these scenarios, in the fine-profiling stage, NGS4THAL groups these samples with several heterozygous --^SEA controls with similar sequencing depth for further evaluation using CoNIFER. In so doing, the loss of the second copy in HBA genes from the --^SEA/–α^3.7 genotype could be clearly detected (Figure 4). The same principle holds true for other compound heterozygous SVs with position overlaps, such as --^CANT/–α^3.7 and --^SPAN/–α^3.7 (Supplemental Figure S8).

Figure 4.

Detection of a compound structural variant (SV) (--^SEA/–α^3.7) by NGS4THAL through a stage-wise process. (1) A: Detection of --^SEA/–α^3.7. B: Detection of –α^3.7/–α^3.7 as comparison. (2) Integrative Genomics Viewer (IGV) screenshot panel shows the alignment landscape and the detections from BreakDancer, Pindel, and CoNIFER; screening stage panel shows the read-depth change in the initial screening stage when all the samples are pooled together for analysis by CoNIFER; fine-profiling stage panel shows the read-depth change in the fine-profiling stage when pooling all the --^SEA deletion carriers together to serve as baseline controls (from in-house control database if necessary). Black lines are read-depth changes for all the simulated individuals. Red curves are read-depth changes for the SV carrier in question, and the vertical dashed green lines illustrate the start and end of the CoNIFER variant callings using a cutoff of 0.5. (3) Detection of homozygous –α^3.7, shown in B, does not need the fine-profiling stage.

From simulated data, NGS4THAL achieved 98.40% sensitivity and 100% specificity for the detection of the Hb SVs, including deletions, duplications, and compound rearrangements (Supplemental Table S4). NGS4THAL determines whether a detected SV is known in the HbVar database by matching the coordinates of the SV boundaries. Because of sequencing depth fluctuations and issues with RMAs, the boundaries of SVs determined by read-depth method are not always accurate. As a result, a few SVs were not accurately reported by NGS4THAL, accounting for the 1.6% of missed detections.

Testing NGS4THAL on the targeted sequencing data of the 71 samples with confirmed Hb mutations led to fully accurate detection of the Hb SVs, including --^SEA, –α^3.7, and –α^4.2 (Supplemental Figure S9), as well as compound rearrangements, such as --^SEA/–α^4.2, --^SEA/–α^3.7, and --^SEA/ααα^anti4.2 (Supplemental Figure S10), without false-positive detection. In addition, applying NGS4THAL on WGS data from the Hirschsprung disease study led to detection of common Hb SVs in samples from Hong Kong and North Vietnam (Table 2), such as common deletions of --^SEA, –α^3.7, and –α^4.2 (Supplemental Figure S11), and duplications of HBA2 or HBA1 (Supplemental Figure S12). The carrier rates of the thalassemia pathogenic deletions detected were consistent with those from epidemiology reports.^26,29, 30, 31 A total of 19 samples with representative Hb SVs, including --^SEA, –α^3.7, –α^4.2, and α-region duplications, were selected for verification by MLPA assay, and all were confirmed experimentally.

Population Difference in Thalassemia Mutation Carrier Rates

On the basis of thalassemia screening on WGS cohorts, thalassemia mutation carrier rate was estimated to be 11.7% for Hong Kong Chinese population, with α-thalassemia mutation carrier rate estimated at 9.6% and β-thalassemia mutation carrier rate estimated at 1.9% (Table 2). The most common thalassemia mutations for Hong Kong Chinese population include --^SEA deletion (4.53%), –α^3.7 deletion (3.20%), –α^4.2 deletion (0.53%) for α-thalassemia, and CD 41/42 (-CTTT; 1.60%) for β-thalassemia. The estimated thalassemia mutation carrier rate for Northern Vietnamese was 14.91%, with high frequencies of Hb E (3.10%), --^SEA (3.70%), and –α^3.7 deletion (3.10%). Because the α-duplication may exacerbate disease severity in β-thalassemia, carrier rates for duplications were also calculated. In Hong Kong Chinese population, αα/ααα^anti3.7(2/375) and αα/ααα^anti4.2(1/375) together account for 0.80% of the population. In the Northern Vietnamese population, the carrier rate for αα/ααα^anti3.7 was estimated at 3.11% (5/161).

Thalassemia is not endemic in North China. The samples collected from North China have a thalassemia carrier rate of 2.7% (7/254) (Supplemental Table S3), most of which (71.43%; five of seven) were –α^3.7 deletion. And the detected carrier rate for αα/ααα^anti3.7 was 1.57% (4/254), suggesting that these mutations are likely recurrent. As expected, --^SEA and other thalassemia mutations common to Hong Kong Chinese and Northern Vietnamese were absent in the Northern Chinese samples. The mutations in Northern Chinese are expected to have little functional and clinical impact in the absence of other prominent causal mutations, such as --^SEA.

In addition, it seems that both targeted sequencing data of the Hb regions with in-depth and continuous coverage and WGS with the widely adopted 30-fold coverage depth led to mutation detections that are accurate or in line with expectations from epidemiology reports, for both point mutation/small InDels and SVs.

Discussion

Thalassemia is a major public health burden in many countries, especially those in the tropical and subtropical regions. Global population migration, however, has led to spread of the disease to historically nonendemic regions, such as Northern Europe and the Americas.³² It has been shown that establishing nation- or territory-wide thalassemia screening programs can significantly reduce the incidence rates of thalassemia or other hemoglobinopathies, moving forward toward the goal of controlling these ancient diseases.³³ It is commonly believed that preconception knowledge of carrier status, which can be achieved by population screening or screening for married couples who are planning to have children, could offer initiatives to at-risk subjects to take necessary measures in avoiding having children with thalassemia. This is especially valuable in regions where termination of pregnancies is difficult for religious, cultural, or legal reasons.

The current laboratory diagnosis methods for thalassemia start with detection of reduced levels of mean corpuscular volume/mean corpuscular Hb, which could miss certain mutations for various reasons. For example, silent heterozygous α-thalassemia and β-thalassemia carriers, compound heterozygotes with both an α-globin gene deletion and a duplication (α-/ααα), heterozygous Hb E carriers, and double heterozygotes of β-thalassemia and α-thalassemia mutations may not show reductions in mean corpuscular volume/mean corpuscular Hb levels. Diagnosis methods based on Hb electrophoresis could also be problematic, especially for some silent heterozygous mutation carriers. For example, coexistence of β-thalassemia mutations may significantly reduce HbH levels. Furthermore, downstream genetic tests, such as Gap-PCR and MLPA, which are the gold standard for the diagnosis of certain mutations, are often cumbersome and are targeted toward certain genetic mutations. Many of these techniques are only available to some tertiary hospitals, preventing their widespread applications.

NGS sequencing of Hb regions has the potential to reveal nearly the complete spectrum of genetic mutations. However, data analysis poses a big challenge because of sequence complexity of the Hb regions. This is aggravated by the fact that bioinformatic expertise is still not widely available in clinical laboratories. Thus, it is necessary to develop an accurate, comprehensive, and efficient platform to aid the molecular diagnosis of thalassemia and other hemoglobinopathies. NGS4THAL, an NGS data analysis pipeline, aims at serving this purpose. NGS4THAL particularly deals with issues caused by homologous sequences among the Hb genes and pseudogenes for accurate detection of point mutations and InDels. Using a combination of complementary SV callers with tailored parameters and aided by specific in-house controls on one of the two types of heterozygous SV mutations, NGS4THAL achieved excellent performance in Hb SV detections as well. Tested by both simulated and real-world thalassemia patient data, NGS4THAL achieved high sensitivity and specificity.

Homologous DNA sequences in the human genome cause challenges for mutation detection by NGS, featuring short reads and small genomic fragment sizes. Different approaches have been proposed to deal with the alignment issues (namely, unique strategy, best match strategy, and all matches strategy).³⁴ They refer to keeping only the uniquely mapped reads, reads with the best matches, or all reads, including those with lower mapping quality scores, respectively. To our experience, most of the analysis tools by default remove RMAs before variant calling, akin to the unique strategy. In a genome-wide perspective, this is an acceptable approach that balances sensitivity and specificity. However, when it comes to a specific region, such as the HBA gene locus, this approach leads to the removal of most of the NGS reads derived from HBA1 and HBA2, resulting in failure of detection of mutations in or across these regions.

NGS4THAL makes use of a combination of three approaches: i) the reads with MAQ >30 were used as they were, akin to the unique strategy; ii) the reads with MAQ = 0 were reassigned to one of the equally likely mapping positions at random, which follows the best match strategy; and iii) for the reads with 0 < MAQ < 30 and with differences of mismatches <3 bp between the primary and the alternative hits, they are realigned as being variant informative, which is akin to the all matches strategy. Tested by both simulated data and real-world sequencing data, the use of a combination of different strategies was shown to handle the multiple alignments issue well from homologous Hb genes and pseudogenes, significantly increasing sensitivity for point mutation detection. Many false detections are caused by the choice of simultaneous reporting of mutations from HBA1 and HBA2, an issue that cannot be solved by short-read sequencing. Some other false-positive callings caused by RMA realignments are from upstream of HBA2 and between HBA2 and HBA1. As they are not involved in pathogenesis for thalassemia and other hemoglobinopathies, they were treated as likely false detections and were not reported.

Currently, there is no widely accepted best practice for SV detection. Most SV detection tools are specialized for a certain type of SVs by making use of limited sequencing signatures. Comparing mapping positions/orientations of read pair and split read, evaluation of read depth and de novo assembly of relevant reads are among the most used approaches. Integrating results from complementary SV callers has been suggested to improve SV detection.³⁵ Classifying the Hb SVs into four subtypes helped us select relevant and complementary tools to increase detection sensitivity.

Compound SVs (such as --^SEA/–α^3.7) are difficult to handle for any available tools based on a single algorithm. Making use of heterozygous controls on one of the two types of the SVs allows the software to re-estimate read depth baselines and to facilitate the detection of the co-existing, overlapping SVs. Compound SVs are common in thalassemia endemic regions. Fortunately, types of compound SVs are predictable according to local epidemiology data, allowing us to prepare this specific control data set to aid the detection of compound heterozygous SVs, such as --^SEA in Southeast Asia and --^THAI in Thailand. There are still genotypes that cannot be detected by NGS4THAL, such as –α^3.7 in combination with ααα^anti3.7. In this case, evaluating parental data might help, and considering family data together might be preferable whenever these data are available.

For targeted sequencing data, it is likely that certain deletions are even larger than the entire targeted region, which applies to some known thalassemia deletions. In this case, we suggest examining runs of homozygosity on high-quality SNVs in the region and violation of Mendelian inheritance if parental data are also available. NGS4THAL not only detects known thalassemia mutations, it is also capable of detecting novel Hb variants, disease modifiers, and variants causal to other hemoglobinopathies. Looking up the detected variants in the HbVar database allows easy reporting of known causal mutations. Thus, we categorized all the variants into known thalassemia causal, known hemoglobinopathy causal, known noncausal Hb variants, and novel variants for downstream follow-ups.

NGS4THAL is implemented in Python and can be easily accessed (https://github.com/JavenCao/Realign_BAM, last accessed December 23, 2021), with detailed specifications. The RMA realignment steps implemented in NGS4THAL can be completed in minutes. The major time-consuming steps for NGSTHAL are variant callings by GATK-HaplotypeCaller and Pindel. Future improvements are needed to ramp up the efficiency when processing large amount of sequencing data, such as in a scenario of population screening.

Author Contributions

W.Y. conceived and supervised the project; Y.C. developed NGS4THAL; C.C.S., M.-f.T.T., P.S., and N.H. provided thalassemia patient samples used in this study; Y.L.L., B.H.-Y.C., C.S.-m.T., and M.-M.G.-B. provided the whole genome sequencing data; Y.C., J.Y., H.Z., and R.L. performed bench experiments, validating findings from NGS4THAL; Y.L.L., G.C.-F.C., C.C.S., E.S.-K.M., and S.-y.H. gave suggestions on the design and execution of the project; and Y.C. and W.Y. wrote and revised the article.

Supplemental Data

Supplemental Figure S1.

Illustration of characteristics of Hb structural variants (SVs) from next-generation sequencing (NGS). Hb SV (eg, deletion) classification is based on relative positions between the homologous regions and SVs, and their characteristics from paired-end short-read NGS sequencing, including read pair, split read, and read depth, are shown. Blue ovals show regions where the expected read-depth change should be observed.

Supplemental Figure S2

Positioning of next-generation sequencing (NGS) reads with multiple alignments (RMAs) in HBA gene clusters. A: Distribution of RMAs in HBA gene clusters. Black curves show all the alignments; red curves show RMAs. The y axis is the depth of coverage from NGS sequencing with different read length. B: RMAs happen between homologous genes and pseudogenes. Primary mapping positions and alternative mapping positions for RMAs are shown on x and y axes, respectively.

Supplemental Figure S3

Positioning of next-generation sequencing (NGS) reads with multiple alignments (RMAs) in HBB gene clusters. A: Distribution of RMAs in HBB gene clusters. Black curves show all the alignments; red curves show RMAs. The y axis is the depth of coverage from NGS sequencing with different read length. B: RMAs happen between homologous genes and pseudogenes. Primary mapping positions and alternative mapping positions for RMAs are shown on x and y axes, respectively.

Supplemental Figure S4

Comparison of the alignments before and after the realignment of informative reads with multiple alignments. A: Comparison of mapping quality score (MAQ) of the alignments before (red) and after (cyan) the realignment process. The y axis is average MAQ, calculated with a window size of 100 bp and a step size of 50 bp from 1500-fold 100-bp paired-end targeted sequencing with continuous coverage of HBA gene locus. B: Comparison of candidate variant loci before and after the realignment process. (1) Red curves in top track show all the aligned next-generation sequencing (NGS) reads for HBA2 and HBA1 regions before the realignment process, and the orange curves are the candidate variant loci. Cyan curves in bottom track show all the aligned NGS reads for HBA2 and HBA1 regions after the realignment process, and the orange curves are the candidate variant loci. (2) Candidate variant loci are defined by the re-assembled haplotypes from GATK-HaplotypeCaller (--bamOutput option).

Supplemental Figure S5

Annotation distribution for the simulated single-nucleotide variants before realignment of reads with multiple alignments. (1) Blue dashed lines are thresholds to determine PASS/FILTERED_OUT status. (2) PASS status is defined as follows: variant confidence normalized by unfiltered depth of variant samples (QD) > 2.0, and strand bias estimated using Fisher's exact test (FS) < 60.0, and rank sum test for mapping qualities of REF versus ALT reads (MQRankSum) > −12.5, and root mean square of the mapping quality of reads across all samples (MQ) > 40.0, and strand bias estimated by the symmetric odds ratio test (SOR) > 3.0, and rank sum test for relative positioning of REF versus ALT alleles within reads (ReadPosRankSum) > −8.0.

Supplemental Figure S6

Detected mutation spectrum by NGS4THAL from real-word next-generation sequencing cohorts. (1) Human Genome Variation Society names for the listed mutations are as follows: Hb Luton: HBA2:c.269A>T (or HBA1), Hb Chesapeake: HBA2:c.278G>T (or HBA1), Hb Owari: HBA2: c.364G>A (or HBA1), Hb Westmead: HBA2: c.369C>G, Hb quong sze: HBA2:c.377T>C, Hb Constant Spring: HBA2: c.427T>C, Hb Q-Thailand: HBA1:c.223G>C, IVS II-654 C>T: HBB:c.316-197C>T, CD 71/72 (+A): HBB:c.217dupA, Hb Hoshida: HBB:c.130G>C, CD 41/42 (-CTTT): HBB:c.126_129delCTTT, CD 35 TAC>TAA: HBB:c.108C>A, IVS I-5 (G>C): HBB:c.92+5G>C, IVS I-1 G>A: HBB:c.92+1G>A, Hb E: HBB:c.79G>A, Hb Malay: HBB:c.59A>G, CD 17 (AAG>TAG): HBB:c.52A>T, -28 (A>G): HBB: c.-78A>G, -50 G>A: HBB:c.-100G>A, -30 T>C: HBD: c.-80T>C, -77 (T>C): HBD:c.-127T>C. (2) Hb Luton, Hb Chesapeake, Hb Owari, Hb Q-Thailand, and Hb Hoshida are Hb variants.

Supplemental Figure S7

Sanger sequencing confirmation for representative pathogenic point mutations detected by NGS4THAL. The targeted variant in each experiment is positioned by a blue arrow.

Supplemental Figure S8

Detection of two compound rearrangements by NGS4THAL through stage-wise process in simulation. (1) A: Detection of --^CANT/–α^3.7. B: Detection of --^SPAN/–α^3.7. (2) Integrative Genomics Viewer (IGV) screenshot panel shows the alignment landscape and the detections from BreakDancer, Pindel, and CoNIFER; screening stage panel shows the read-depth change in the initial screening stage when all the samples are pooled together for analysis by CoNIFER; fine-profiling stage panel shows the read-depth change in the fine-profiling stage when pooling all the --^CANT or --^SPAN deletion carriers together. Vertical dashed green lines illustrate the start and end of the CoNIFER variant calls using a cutoff of 0.5. SV, structural variant.

Supplemental Figure S9

Detection of deletions from the targeted sequencing data by NGS4THAL. Top panels: Integrative Genomics Viewer screenshot for the alignment landscape and the detections from BreakDancer, Pindel, and CoNIFER. Bottom panels: The read-depth changes in the screening stage when all the samples are pooled together for analysis by CoNIFER. Vertical dashed green lines illustrate the start and end of the CoNIFER variant calls using a cutoff of 0.5. Chr, chromosome; SV, structural variant.

Supplemental Figure S10

Detection of compound rearrangements from the targeted sequencing data by NGS4THAL. (1) A: Detection of –^SEA/–α^3.7. B: Detection of –^SEA/–α^4.2. C: Detection of –^SEA/–ααα^anti4.2. (2) Integrative Genomics Viewer (IGV) screenshot panel shows the alignment landscape and the detections from BreakDancer, Pindel, and CoNIFER; screening stage panel shows the read-depth change in the initial screening stage when all the samples are pooled together for analysis by CoNIFER; fine-profiling stage panel shows the read-depth change in the fine-profiling stage when pooling all the –^SEA deletion carriers together. Vertical dashed green lines illustrate the start and end of the CoNIFER variant calls using a cutoff of 0.5. SV, structural variant.

Supplemental Figure S11

Detection of thalassemia pathogenic deletions from whole genome sequencing data by NGS4THAL. A–D: examples of detected deletions: αα/--^SEA (A), αα/–α^3.7 (B), –α^3.7/–α^3.7 (C), and αα/–α^4.2 (D). Top panels: The alignment landscape in the format of Integrative Genomics Viewer (IGV) screenshot, followed by the reported pathogenic structural variants (SVs) and detection results from BreakDancer, Pindel, and CoNIFER (with cutoffs of 1.0, 0.8, 0.5, and 0.3, respectively). CoNIFER detections with cutoff of 0.5 are used in NGS4THAL and are shown in green. Middle panels: Copy number variation (CNV) detection panel shows read-depth changes detected by CoNIFER. Black lines are read-depth changes for all the whole genome sequencing individuals. Red curves are read-depth changes for the SV carriers in question. The y axis is the relative copy number with reference to the baseline defined by all input samples, and the vertical dashed green lines illustrate the start and end of the CoNIFER variant callings using a cutoff of 0.5. Bottom panels: Multiplex ligation-dependent probe amplification (MLPA) test panel shows the confirmation of the SVs by assay of MLPA. The x axis shows positions of the MLPA probes, and the y axis shows the probe ratios compared with the wild-type samples. The probe ratios had a normal range of 0.7 to 1.3 for nondeletion samples. Probe ratio of 0.5 indicated a single copy deletion, except for probes of 172, 214, and 220 nucleotides (nt), for which a ratio of 0.75 indicates a single copy deletion, because these three probes hybridize to both HBA1 and HBA2. Besides, the 135-nt probe (the one on the most right-hand side of the purple regions) would only give a signal on samples containing the Hb Constant Spring point mutation. Probes in the gray shading area are quality control probes. Chr, chromosome.

Supplemental Figure S12

Detection of an α^3.7 duplication from whole genome sequencing data by NGS4THAL. A: The alignment landscape in the format of Integrative Genomics Viewer screenshot, followed by the reported structural variants (SVs) and detection results from BreakDancer, Pindel, and CoNIFER (with cutoffs of 1.0, 0.8, 0.5, and 0.3, respectively). CoNIFER detections with cutoff of 0.5 are used in NGS4THAL and are shown in green. B: Read-depth changes detected by CoNIFER. Black lines are read-depth changes for all the whole genome sequencing individuals. Red curves are read-depth changes for the SV carriers in question. The y axis is the relative copy number with reference to the baseline defined by all input samples, and the vertical dashed green lines illustrate the start and end of the CoNIFER variant callings using a cutoff of 0.5. C: The confirmation of the SVs by assay of multiplex ligation-dependent probe amplification (MLPA). The x axis shows positions of the MLPA probes, and the y axis shows the probe ratios compared with the nondeletion reference samples. The probe ratios had a normal range of 0.7 to 1.3 for nondeletion samples. Probe ratio of 1.5 indicated a single copy gain, except for probes of 172, 214, and 220 nucleotides, for which a ratio of 1.25 indicates a single copy gain, because these three probes hybridize to both HBA1 and HBA2. Probes in the gray shading area are quality control probes. Chr, chromosome.