ABSTRACT
Purpose By identifying pathogenic variants across hundreds of genes, expanded carrier screening (ECS) enables prospective parents to assess risk of transmitting an autosomal recessive or X-linked condition. Detection of at-risk couples depends on the number of conditions tested, the diseases’ respective prevalences, and the screen’s sensitivity for identifying disease-causing variants. Here we present an analytical validation of a 235-gene sequencing-based ECS with full coverage across coding regions, targeted assessment of pathogenic noncoding variants, panel-wide copy-number-variant (CNV) calling, and customized assays for technically challenging genes.
Methods Next-generation sequencing, a customized bioinformatics pipeline, and expert manual call review were used to identify single-nucleotide variants, short insertions and deletions, and CNVs for all genes except FMR1 and those whose low disease incidence or high technical complexity precludes novel variant identification or interpretation. Variant calls were compared to reference and orthogonal data.
Results Validation of our ECS data demonstrated >99% analytical sensitivity and >99% specificity. A preliminary assessment of 15,177 patient samples reveals the substantial impact on fetal disease-risk detection attributable to novel CNV calling (13.9% of risk) and technically challenging conditions (15.5% of risk), such as congenital adrenal hyperplasia.
Conclusion Validated, high-fidelity identification of different variant types—especially in diseases with complicated molecular genetics—maximizes at-risk couple detection.
INTRODUCTION
There are more than 1000 recessive single-gene conditions that vary in both severity and age of onset.1 Each is uncommon in the general population, yet collectively these Mendelian diseases account for approximately 20% of infant mortality and 10% of infant hospitalizations.2,3 Screening for carriers of such conditions in the preconception or prenatal period informs couples about both the risk of having a child with a serious disease and the available family-planning options. The risk assessment and patient autonomy provided by carrier screening can have substantial impact, as nearly 2 million women give birth to their first child each year in the United States.4
Due to the rising quality and falling cost of genomic technologies, it is now possible to perform pan-ethnic carrier screening on a large number of conditions simultaneously (referred to as “expanded carrier screening”, or “ECS”). Our recent retrospective study of carrier rates in 346,790 patients showed that an ECS panel is expected to identify more pregnancies at risk for severe or profound conditions than ethnic-based panels spanning far fewer genes.5 Citing this work, the American College of Obstetricians and Gynecologists (ACOG) issued new guidelines in March 2017, recognizing expanded carrier screening as an acceptable strategy for prepregnancy and prenatal carrier screening for patients and their partners.6
An ECS must have a high detection rate for each disease on the panel to identify at-risk couples and to minimize the residual risk in couples where only one partner has tested positive. Indeed, detection rate is particularly important for the recessive diseases that predominate ECS panels because the odds of detecting an at-risk couple scales as the square of the rate for finding an individual carrier (e.g., 80% detection rate for one parent means only a 64% detection rate for an at-risk couple). Relative to the targeted genotyping approaches used for classical carrier screening, next-generation sequencing (NGS) has enabled ECS panels to achieve very high per-disease detection rates, as both common and rare variants can be identified across the entire coding region and in relevant noncoding positions of the disease gene.1,7–11 Underscoring the at-risk couple detection gain afforded by NGS, a study of 11,691 individuals screened for 15 genes by NGS revealed that approximately one quarter carried mutations not typically included in targeted panels.8
To maximize detection rates, novel copy number variants (CNVs) must be identified, yet most ECS offerings—even those using NGS—report only single nucleotide variants (SNVs), indels, and, at most, a small handful of common CNVs with known breakpoints. However, CNVs can vary in size and position, encompassing everything from single exons (which account for 29% of CNVs for Mendelian conditions12) to the entire gene. A diversity of pathogenic CNVs has been observed in cystic fibrosis carriers, accounting for 1.6% of carriers13, meaning that the single-carrier detection rate without CNV detection is <98.4%, which in turn makes the at-risk couple rate <96.8%. The inverse is noteworthy: including novel CNV detection can boost at-risk couple detection for cystic fibrosis to nearly 100%. To our knowledge, the impact of CNVs across all genes on an ECS has not yet been characterized.
Carrier status for a minority of the most prevalent serious conditions is difficult to resolve with standard NGS and bioinformatics approaches due to the challenging sequence features of the disease gene; thus, these conditions require special handling. Low complexity spans (e.g., CGG repeat expansion in FMR1 for fragile X syndrome) and highly homologous regions (e.g., SMN1 and SMN2 genes for spinal muscular atrophy) complicate variant identification, yet these hard-to-sequence genes simultaneously contribute substantially to the disease risk. For instance, fragile X syndrome, spinal muscular atrophy, 21-hydroxylase deficient congenital adrenal hyperplasia (21-OH CAH), and alpha thalassemia account for 54 affected fetuses per 100,000 pregnancies.14
Here, we validate and describe an ECS (Counsyl Foresight™ Carrier Screen) leveraging NGS to identify SNVs, indels, novel CNVs (deletions for nearly all genes and both deletions and duplications for CFTR and DMD), and hard-to-sequence targeted variants. Following recommendations of the College of American Pathologists (CAP)15 and the American College of Medical Genetics and Genomics (ACMG)16, we measured the analytical accuracy, precision, sensitivity, and specificity of the test. Further, on a cohort of 15,177 patients, we performed a preliminary assessment of the impact of panel-wide CNV calling on detection of at-risk pregnancies. Our data demonstrate the accurate detection of the most common types of genomic alterations found in reference cell lines and routine clinical specimens.
MATERIALS AND METHODS
Institutional Review Board approval
The protocol for this study was approved by Western Institutional Review Board (IRB number 1145639) and complied in accordance with the Health Insurance Portability and Accountability Act (HIPAA). The information associated with patient samples was de-identified in accordance with the HIPAA Privacy Rule. A waiver of informed consent was requested and approved by the IRB.
Test description
We compiled a panel (Counsyl Foresight™ Carrier Screen) of 235 genes responsible for 234 clinically important autosomal recessive and X-linked diseases (Table S1). This panel consists of a “Universal” sub-panel (176 diseases) for routine ECS and an opt-in panel (234 diseases) aimed at specific high-risk populations. The design of the Universal panel is described in Beauchamp et al14 and prioritized diseases that are prevalent, with serious and highly penetrant phenotypes that would likely affect clinical counseling for preventative measures and family planning. Some genes for high-prevalence diseases with moderate but lifelong impact are also included (e.g., MEFV and GJB2). The gene-specific methodologies and types of variants reported are summarized in Table 1.
Next generation sequencing, bioinformatics processing, and variant interpretation
Next generation sequencing
The molecular workflow of our NGS pipeline was previously described.17 Briefly, DNA is fragmented to 200-1000 bp by sonication and then converted to a sequencing library by end repair, A-tailing, and adapter ligation. Samples are then amplified by PCR with barcoded primers, multiplexed, and subjected to hybrid capture-based enrichment with 40-mer oligonucleotides. Sequencing of the selected targets is performed on the Illumina HiSeq 2500 instrument.
Bioinformatics processing
Sequencing reads are aligned to the hg19 human reference genome using the BWA-MEM algorithm.18 Novel SNVs and indels are identified and genotyped using GATK 1.6 and FreeBayes19,20, and nine known-pathogenic sites involving complex indels are detected with custom genotyping software. Copy number variants are determined using custom software that leverages read-depth values.17 A combination of targeted genotyping and read depth-based copy number analysis is used to determine the number of functional gene copies and/or the presence of selected mutations in technically challenging genes, such as SMN1, GBA, HBA1/2, and CYP21A2 (Supplementary Methods).
Quality control metrics
Ancillary quality-control (QC) metrics are computed on the sequencing output and used to exclude and re-run failed samples. QC metrics include the fraction of sample contamination (<5%), extent of GC bias, read quality (percent Q30 bases per Illumina specifications), depth of coverage (mean coverage of >50x), and region of interest (ROI) coverage (>99% per base minimum coverage >=20x) (Table S2). Calls that do not meet QC criteria are set to “no-call”.
Variant review and interpretation
To ensure clinical calling accuracy, all calls and no-calls for pathogenic, likely pathogenic, and uncurated variants are manually reviewed by laboratory personnel and are subject to override if warranted, based on a pre-established protocol. Identified variants are classified according to the ACMG Standards and Guidelines for the Interpretation of Sequence Variants21 as described previously.17 Final variant classifications are regularly uploaded to ClinVar (National Center for Biotechnology Information; https://www.ncbi.nlm.nih.gov/clinvar/).22
FMR1 CGG repeat sizing
CGG trinucleotide expansions of the FMR1 promoter are measured by PCR amplification and capillary electrophoresis as previously described.23
Analytical validation
Samples and reference data. Samples and reference data are compiled from different sources (Tables S3 and S4). Purified DNA for 91 cell lines from the 1000 Genomes (1KG) Project24 and 70 cell lines with known pathogenic variants in specific genes were purchased from the Coriell cell line repository (Camden, NJ) (Tables S5A-E). In addition, 115 mutation-positive patient blood and saliva samples tested with a previous version of the Counsyl carrier test (a 94-gene panel) were included in the validation. Relevant variants in all mutation-positive patient samples were confirmed orthogonally by PCR/Sanger sequencing, quantitative PCR, or multiplex ligation-dependent probe amplification (MLPA). Further details on confirmatory testing can be found in the Supplementary Methods. Note that NA06896 was dropped from FMR1 accuracy analysis due to inconsistent reference data.25,26
1KG analysis. Ninety-one 1KG samples were sequenced to measure the accuracy of SNV and short indel calls in 229 genes (Table S5A). Ninety samples passed QC and manual review. Simulation of synthetic CNVs. For every region reportable for CNVs, we simulated a single-copy deletion and tested calling sensitivity; we also simulated single-copy duplications for DMD and CFTR regions (detailed description of CNV simulations in the Supplementary Methods). Statistical analysis. Validation metrics were defined as: Accuracy = (TP + TN) / (TP + FP + TN + FN); Sensitivity = TP / (TP + FN); Specificity = TN / (TN + FP); FDR = FP / (TP + FP), where TP - true positives, TN - true negatives, FP - false positives, FN - false negatives, and FDR - false discovery rate. The confidence intervals (CIs) were calculated by the method of Clopper and Pearson.27 Reproducibility within and between runs was calculated as the ratio of concordant calls to total calls.
RESULTS
We developed an NGS-based ECS (Counsyl Foresight™ Carrier Screen) covering 220 autosomal recessive and 14 X-linked conditions, including technically challenging diseases (Figure 1, Tables 1 and S1). For nearly all genes on the Universal panel, SNVs and indels are detected via NGS data acquired from regions that could impact gene function (e.g., padded coding exons and known or potentially pathogenic intronic variants; Figure 1, top left). Large CNVs are identified at single-exon resolution panel-wide using relative sample-to-sample changes in sequencing depth (Figure 1, top right; see Methods). The test has been validated (described in detail below) and used in a clinical production setting on 15,177 patient samples (tested between Nov. 2016 and Aug. 2017). Carrier rates on this cohort enable calculation of the expected fraction of US pregnancies whose affected status would be identified by the Universal panel of the ECS on a per-disease level (Figure 1 middle): the estimate is that 1 in 300 pregnancies would be affected by at least one serious disease on our ECS. For a handful of prevalent diseases that comprise 8.8% of the total panel disease risk, high carrier sensitivity requires customized CNV analysis due to the genes’ complicated technical features (Figure 1, bottom).
Relative to our prior version of the ECS characterized in a recent study of 346,790 patients5,14, this updated ECS differs in two ways that collectively boost the assessed fetal disease risk. The updated Universal panel ECS probes SNVs and indels for 82 more diseases (Figure 2A,B), and it additionally detects deletions ranging in size from a single exon to the entire gene (Figure 2B,C). Panel-wide CNVs contribute approximately 14% of the assessed fetal disease risk: CNVs in DMD alone represent ∼10% of the risk, with the remaining genes accounting for ∼4%, largely consistent with our estimate of CNV-attributable disease risk estimated from the previous 94 disease panel.14
Validation approach
To assess the analytical performance of the ECS panel prior to launching it in a clinical setting, we measured the accuracy of identifying variants that are small (e.g., SNVs and small indels), technically nuanced (e.g., large indels and CNVs), and in hard-to-sequence genes (e.g., CYP21A2) (Table S4). Our validation approach builds upon the broad recommendations from the College of American Pathologists (CAP)15 and the American College of Medical Genetics and Genomics (ACMG)16 for validation of targeted, germline testing using NGS.
Validation of a new assay requires a reference set for comparison. For smaller-sized variant types, reference data and samples are readily available; however, for technically challenging genes and variants, reference material is scarce. To gather reference data that tests such challenging variants, we identified relevant patient samples tested with a previous version of the Counsyl ECS (a 94-gene panel) and orthogonally confirmed each positive variant by a Sanger, TaqMan, or MLPA assay, as appropriate. Since we also wanted to establish CNV-calling proficiency in regions for which no reference samples exist, we implemented in silico simulations of CNVs.
Collectively, the validation reference dataset establishes a high standard for validating ECS panels by spanning both the scope of clinical sample types (Table S3) and the range of variant types (Table S4).
Accuracy and reproducibility for calling SNPs and small indels in 229 genes
Validation samples were tested using our standard operating procedure (SOP), which includes in- and post-process QC at the batch, sample, and variant-call level (Table S2). Furthermore, consistent with our SOP for clinical samples, licensed experts, who were blind to the validation sample set, performed manual review of the sequencing data using our custom review interface. Samples that failed QC and manual review were excluded from further analysis.
We compared our ECS data for the reference sample NA12878 with data from the Genome in a Bottle Consortium28, which includes high-confidence calls for >97.5% of the regions covered by our test. We tested NA12878 across five batches and in duplicate within three batches for a total of eight tests, and the test results were highly accurate (>99.99%, Table S6). As NA12878 is one of our routine controls within every production batch, since the validation of our test we have further measured accuracy across 207 batches that span reagent lots and instruments, consistently observing high calling accuracy across the panel (Figure S1).
To measure SNV and indel calling accuracy across a diverse set of samples, we performed our ECS on 1KG samples. For 90 samples that we tested and passed QC, we compared genotypes across all exonic regions with sufficient coverage and quality in the 1KG data (248,490 calls in all); 52 discordant calls were adjudicated with Sanger sequencing (Table S7). Our ECS identified 36,032 true-positive calls and 212,139 true-negative calls, resulting in >99.99% accuracy, sensitivity, and specificity (Figure 3A).
In addition to establishing the analytical accuracy of the ECS using reference DNA from cell lines, we measured intra- and inter-run reproducibility using different sample types by comparing the equivalence of genotyping calls starting from separate aliquots of DNA. Overall, the test achieved >99.9% intra- and inter-assay reproducibility (Table S8).
Technically challenging variants
Larger indel detection performance
Although only 5% of indels are ≥5bp29, sensitivity falls as indel size grows. Thus, to ensure high analytical sensitivity for detecting indels, we built a cohort of 52 patient samples with 49 unique technically-challenging, larger (>5bp) deletions, insertions, or complex indels in 42 different genes (Table S9A). All of the expected indel calls (52/52), including a 33bp deletion and 21bp insertion, were observed (Figure 3B).
CNV detection performance
To overcome the limitation of scarce reference materials for CNV calling, we supplemented available reference material with orthogonally confirmed positives identified retrospectively and additionally used in silico simulated CNVs to measure sensitivity systematically across the panel. In the empirical analysis, the reference set included 11 Coriell cell lines with a known CNV (Table S5B) and 33 clinical samples with CNVs that we confirmed by MLPA (Table S9B). The 44 CNV-positive samples included 41 deletion variants in 13 different genes and 3 duplication variants in CFTR and DMD. Notably, 23 samples had a single-exon or two-exon CNV, which can be technically challenging for a NGS-based assay (Table S4). We assessed CNV calling performance separately for DMD and CFTR, for which we optimized read depth to ensure high sensitivity for both deletions and duplications at the single-exon level. As shown in Figure 3B, we detected all 44 CNVs, demonstrating high sensitivity (100%; 95% CI, 92%-100%; reproducibility data in Table S10).
In the in silico CNV simulations, we introduced a synthetic deletion or duplication spanning at least one coding exon in the background of empirical sample data from four validation flowcells (see Supplementary Methods). To assess our ability to call single-exon and multi-exon CNVs, three categories of synthetic CNVs (one-, two-, and four-exon blocks) were tested, with each synthetic CNV size and position being simulated independently using 20 different samples as background (yielding >250k total simulations). Results are summarized in Figure 3C,D. To assess sensitivity of clinically relevant deletions and duplications in CFTR and DMD, we scaled the sensitivity for each CNV size by its population frequency cataloged in public databases (The Clinical and Functional TRanslation of CFTR (CFTR2), https://www.cftr2.org; the Leiden Open Variant Database (LOVD)30, http://www.dmd.nl), yielding an aggregate 99.9% sensitivity for CNVs in each gene. Across the rest of the panel, for which only deletions are reported, our simulations revealed 81.8% sensitivity for single-exon deletions and 98.3%-100% sensitivity for multi-exon deletions. Taken together, the simulation results suggest our ECS has high proficiency in identifying CNVs with exon-level resolution.
Variant detection performance using NGS in the technically challenging genes CYP21A2, HBA1/2, GBA, and SMN1
Several diseases of clinical importance result from mutations in genes that have a paralog or pseudogene that complicates molecular analysis. Such diseases include spinal muscular atrophy (SMN1 and SMN2 encode the same protein, but SMN2 harbors a splicing variant that results in ∼10% of functional SMN protein relative to SMN131), alpha-thalassemia (HBA1 and HBA2 have identical coding sequences and few distinguishing noncoding bases), 21-OH deficient CAH (the CYP21A2 coding sequence is >99% identical to its pseudogene CYP21A1P), and Gaucher disease (GBA has a nearby pseudogene GBAP1 with which it shares high sequence identity in certain exons). Recombination and gene conversion is frequent among these genes and their homologs, which can result in copy number changes. To detect deleterious variants in these technically challenging genes, we implemented custom variant-calling algorithms combining depth-based copy number and specific mutation analyses for the disease genes and their homologs (see Supplementary Methods). Below, we describe results for measuring sensitivity; a summary of reproducibility can be found in Table S10.
CYP21A2 CNV analysis
Mutations in CYP21A2 account for 3.6% of the risk assessed by our ECS (Figure 1), with approximately 65-70% arising from gene conversion events with a pseudogene and 25-30% from large gene rearrangements.32 To assess the accuracy of our NGS-based assay to detect large and small CYP21A2 rearrangements, we tested 14 specimens previously called positive on the Counsyl 94-gene panel (Table S9C). We confirmed the variants using MLPA or long-range PCR and Sanger sequencing (see Supplementary Methods). All 14 validation samples—whose variants account for >95% of deleterious CYP21A2 mutations32—were genotyped correctly (see Figure 4A).
HBA1/2 common-variant analysis
Alpha thalassemia represents 1.2% of risk assessed by our ECS (Figure 1), and much of that risk arises from deletions among HBA1 and HBA2, which have identical coding sequences. Screening for alpha thalassemia requires detection of the total copy number of HBA genes, as well as their phasing on the chromosome. We identified 10 patient samples that had been tested on our previous panel and were found to have mutations in the HBA1 and/or HBA2 genes (Table S9C). Heterozygous and homozygous deletions for single genes, cis deletions of both genes, and combinations thereof were confirmed during validation (Figure 4A).
GBA CNV analysis
Gaucher disease, which accounts for 0.8% of the total panel disease risk, can arise from gene-conversion events where pseudogenic sequence in GBAP1 recombines into the GBA locus. To validate that our ECS can identify deleterious gene-conversion alleles, six patient samples and one Coriell DNA sample with the pathogenic mutations L444P, D448H, and IVS2+1G>A in GBA were successfully identified (Figure 4A).
SMN1 copy number and g.27134T>G SNP analyses
Risk for spinal muscular atrophy—3.1% of ECS panel disease risk (Figure 1)—is mostly influenced by the copy number of SMN1, which is distinguished from the highly homologous SMN2 gene by an intronic variant that influences splicing of exon 7. To determine the accuracy of NGS-based SMN1 copy-number calling, 128 unique (234 total with replicates) DNA samples with 0, 1, 2, or 3 copies of SMN1 were analyzed by NGS (Tables S4 and S5D). Carrier (samples with 0 or 1 copy) versus non-carrier (samples with 2 or more copies) identification accuracy by NGS was 100% (95% CI, 98.4%-100%). NGS copy-number accuracy was 233/234, 99.6% (95% CI, 97.6%-100%) (Figure 4B), where one non-carrier patient sample had three copies by NGS and two by TaqMan.
We also measured detection of the g.27134T>G SNP associated with 2+0 SMA carrier status.33 The analysis included 98 (92 unique) 1KG cell line samples containing the g.27134T>G SNP (Table S5D). We additionally confirmed a subset of the g.27134T>G SNP calls in 16 Coriell samples (n=14 lacking the g.27134T>G SNP, n=2 harboring the g.27134T>G SNP) via SMN1-specific PCR and Sanger sequencing (see Supplementary Methods), yielding 114 reference samples total (106 unique samples). The NGS results were 100% (114/114) concordant with the reference data (Figure 4).
FMR1 CGG-repeat analysis
Fragile X syndrome (FXS), the most common cause of inherited intellectual disability, arises from a trinucleotide CGG repeat expansion in the 5’ untranslated region of FMR1.34 FMR1 alleles are categorized as normal (NL; 5-44 CGG repeats), intermediate (IM; 45-54 CGG repeats), premutation (PM; 55-200 CGG repeats), and full-mutation (FM; >200 CGG repeats).35 We validated our assay using a sample set enriched for expansions of various sizes in both female and male samples. A total of 39 Coriell samples (Table S5C) were classified correctly by our assay (Figure 4C). Further, the identified CGG repeat allele sizes closely matched the literature consensus sizes (Figure 4D).
DISCUSSION
ECS is gaining widespread clinical adoption—recently receiving support from medical societies6,36—because it can provide reliable and affordable risk assessment for many serious recessive and X-linked diseases simultaneously. Genomics technologies like NGS have enabled dramatic growth in ECS panel size without incurring a corresponding rise in testing cost. Coupled with this increase in the achievable panel size, however, is the need to be judicious in panel construction and painstaking in the effort to validate performance. For these reasons, we recently published a systematic process for ECS panel design14 and here present both a comprehensive validation study of our updated ECS’s performance and a preliminary analysis of its use on patient samples. Our study of variants in hundreds of genes across hundreds of samples demonstrates high sensitivity, specificity, and accuracy of genotype calls across coding regions and in technically challenging genes (Figures 3 and 4).
We introduced panel-wide CNV calling to our NGS-based ECS to maximize the chance of finding couples at risk for children with serious conditions. Though our previous 94-gene ECS could detect six of the most-common CNVs, the updated ECS validated here can identify novel CNVs that span at least one exon in 218 genes. CNVs can be identified in a production workflow via orthogonal technologies (e.g., MLPA) rather than in a single NGS assay (as we do), but MLPA testing is not affordably scalable to hundreds of genes and incurs additional handling steps in the lab that can introduce operator error. Using known-positive samples from biorepositories, retrospectively identified CNV-positive samples, and in silico simulations, we demonstrated high sensitivity for novel CNV identification via NGS. Though still preliminary, our analysis of 15,177 samples with the new panel suggests that the ability to identify novel CNVs has a large impact on the efficacy of an ECS, accounting for ∼14% of the total assessed disease risk. We expect that simulation analyses, as we used here, will become increasingly important during NGS-panel validation, where performance needs to be evaluated even when clinical samples are rare or nonexistent.
Not all genes contribute the same amount to the risk assessed by an ECS (Figure 1). Indeed, our updated ECS contains more than twice as many genes as the previous version, yet the risk resolved is not twice as great (Figure 2). This phenomenon is driven by the disparate incidence of diseases and highlights that it is critical to have high detection rates for the most-common serious conditions, many of which pose screening challenges due to complicated molecular genetics. For several special cases, we have fine-tuned CNV calling to capture single-base differences (SMA), phased and overlapping rearrangements (alpha thalassemia), and very complicated gene conversions (CAH and GBA). In sum, though our risk estimates will become further refined with the addition of more screened patients, we expect the collective risk of these four diseases to rival that of >100 of the least-common diseases on the panel.
Based on the successful validation of the Counsyl Foresight™ Carrier Screen described here, we now broadly perform the test on samples from prospective parents in our clinical laboratory, which is CLIA certified (05D1102604), CAP accredited (7519776), and NYS permitted (8535).
CONFLICT OF INTEREST
All authors are current or former employees of Counsyl, Inc.
ACKNOWLEDGEMENTS
We would like to thank our Counsyl colleagues for help and advice: members of the CLIA team, including Thi Tran and Jeanette Wong; Dan Davison; Genevieve Haliburton; Andrew Horn; Eerik Kaseniit; Brandon Lee; Matt Leggett; Jeff Tratner; and Xin Wang.