Autism spectrum disorder (ASD) comprises a collection of heterogeneous neurodevelopmental disorders that share two behavioral characteristics—difficulties in social communication and restricted, repetitive behaviors and interests1,2. The etiology of ASD has a significant genetic component, as is evident from multiple twin and family studies3,4,5,6. Yet, over the years, very few genetic causes of ASD have been discovered; thus, today, despite extensive research, an understanding of the overall genetic architecture of ASD remains obscure6,7.

The emergence of next-generation sequencing (NGS) approaches in the past decade has transformed the genetic research of complex traits8. These NGS technologies have facilitated high-throughput DNA sequencing for large cohorts of patients, allowing the comparison of multiple variants that includes single-nucleotide variants (SNVs) and short insertions/deletions (INDELs) between large groups of patients9,10,11,12. In this realm, whole-exome sequencing (WES) is particularly suitable for studying the genetics of heterogenous traits such as ASD, as it focuses on a relatively limited number of protein-coding variants9,10,13,14,15,16,17,18.

However, understanding the functional consequences of coding variants is not a trivial task. In 2015, the American College of Medical Genetics and Genomics (ACMG) and the Association for Molecular Pathology (AMP) published standards and guidelines to generalize sequence variant interpretation and to address the issue of inconsistent interpretation across laboratories8. The resulting system for classifying variants recommends 28 criteria (16 for pathogenic and 12 for benign variants) and provides a set of scoring rules based on variant population allele frequency, variant functional annotation, variant familial segregation, etc.8,19; Variants are classified as pathogenic (P), likely pathogenic (LP), variants of uncertain significance (VUS), likely benign (LB) or benign (B). Subsequently, multiple in-silico tools were developed to implement these ACMG/AMP criteria for annotating the prospective pathogenicity of variants detected in WES studies.

While the ACMG/AMP scoring approach is highly effective for detecting de-novo highly penetrant mutations for rare Mendelian disorders, it is less suitable for detecting inherited partially penetrant variants20. Such variants, usually annotated as VUS in terms of the ACMP/AMP criteria, are expected to contribute significantly to the risk of developing neurodevelopmental conditions, including ASD9,17,18,21,22. Thus, relying solely on the ACMG/AMP criteria for variant annotation in WES studies of ASD may result in an under-representation of susceptibility variants, which will lead to a lower diagnostic yield for ASD. To overcome this potential limitation, we have developed “Psi-Variant,” a pipeline to detect different types of likely gene-disrupting (LGD) variants, including protein truncating and deleterious missense variants. We applied Psi-Variant – in comparison with InterVar and TAPES, two variant interpretation tools that use the ACMG/AMP criteria – to a large WES dataset of ASD to evaluate the concordance between these tools to detect variants and to assess their effectiveness in detecting ASD susceptibility variants.

Methods

Study sample

Initially, the study sample comprised 250 children diagnosed with ASD who are registered in the National Autism Database of Israel (NADI)23,24 and whose parents gave consent for participation in this study. Based on our clinical records, none of the parents in the study has been diagnosed with ASD, intellectual disability, or any other type of neurodevelopmental disorder. Genomic DNA was extracted from saliva samples from participating children and their parents using Oragene®•DNA (OG-500/575) collection kits (DNA Genotek, Canada).

Whole exome sequencing

WES analysis was performed on the above-mentioned samples with Illumina HiSeq sequencers, followed by the Illumina Nextera exome capture kit at the Broad Institute as part of the Autism Sequencing Consortium, described previously11. Sequencing reads aligned to Genome Reference Consortium Human Build 38 and aggregated into BAM/CRAM files were analyzed using the Genome Analysis Toolkit (GATK)25 to generate a joint variant calling format (vcf) file for all subjects in the study. We excluded data for 30 probands from the raw vcf file due to incomplete pedigree information or low-quality WES data. Thus, WES data for 220 ASD trios were analyzed in this study (Fig. 1).

Figure 1
figure 1

Analysis workflow for detecting LP/P/LGD variants from the WES data. InterVar and TAPES detected LP/P variants by implementing ACMG/AMP criteria. Psi-Variant detected LGD variants by utilizing in-house criteria.

Full size image

Data analysis

The variant detection process in this study is outlined in Fig. 1. and explained below.

Data cleaning

The raw vcf file contained 1,935,632 variants. From this file, we removed variants with missing genotypes and/or variants in regions with low read coverage (≤ 20 reads) and/or with low genotype quality (GQ ≤ 50). In addition, we removed all common variants (i.e., those with a population minor allele frequency > 1% as per the Genome Aggregation Database (gnomAD))26,27 as well as those that did not pass the GATK’s “VQSR” and “ExcessHet” filters. Thereafter, we used an in-house machine learning (ML) algorithm to remove other potentially false-positive variants. The details of this ML algorithm and its efficiency in classifying true positive and false positive variants are summarized in the supplementary file S1. Finally, we used the pedigree structure of the families to identify proband-specific genotypes, including de-novo variants, recessively inherited variants, and X-linked variants (in males). Recessively inherited variants occur in the same loci of both copies of a gene in autosomes (where both the parents are carriers). Whereas one altered copy of the gene in chromosome X among males is defined as X-linked (males). We removed multiallelic variants from these genotypes and those classified as “de-novo” that appeared in more than two individuals in the sample. In this study, we haven’t considered compound heterozygote variants (in cis/trans).

Identifying ASD candidate variants

We searched for candidate ASD Variants using three complementary approaches. First, we applied InterVar19 and TAPES28, two commonly used publicly available command-line tools that use ACMG/AMP criteria8, to detect LP/P Variants. In addition, we assigned the ACMG/AMP PS2 criterion to all the de-novo Variants to detect additional LP/P Variants from the list of VUS. Since InterVar and TAPES are less sensitive tools for detecting recessive possible gene disrupting (LGD) variants20, we developed an integrated in-house tool, Psi-Variant, to detect LGD variants. The Psi-Variant workflow starts using Ensembl’s Variant Effect Predictor (VEP)26 to annotate the functional consequences for each variant in a multi-sample vcf file. Then, all frameshift indels, nonsense, and splice acceptor/donor variants are further analyzed by the LoFtool29 with scores of < 0.25 are annotated as intolerant variants. In addition, it applies six different in-silico tools to all missense substitutions to determine their likelihood to be “deleterious/damaging” using the following cutoffs: SIFT30 (< 0.05), PolyPhen-231 (≥ 0.15), CADD32 (> 20), REVEL33 (> 0.50), M_CAP34 (> 0.025) and MPC35 (≥ 2) as recommended by the dbNSFP database36.

Comparison between InterVar, TAPES, and Psi-variant

We compared the number of variants detected by the three tools and the percentages of variants detected by different combinations. Thereafter, we used the list of ASD genes (n = 1031) from the SFARI Gene database37 (accessed on 11 January 2022) as the gold standard to compute the odds ratio (OR) and positive predictive value (PPV) for detecting candidate ASD variants in SFARI genes. In addition, we assessed the diagnostic yield (%) for each tool combination by computing the proportion of children with detected candidate ASD variants in SFARI genes. Diagnostic yield (%) was computed as the proportion of ASD probands with at least one ASD candidate variant out of the total affected ASD probands with complete pedigree.

Software

Data storage, management, and analysis were conducted on a high-performing computer cluster in a Linux environment using Python version 3.5 and R Studio version 1.1.456. All the statistical analyses and data visualizations were incorporated into R Studio.

Ethics approval and consent to participate

Informed consent was obtained from all the families involved in the study.

Institutional review board statement

The study was conducted according to the guidelines of the Declaration of Helsinki and approved by the Ethics Committee of Soroka University Medical Center (SOR-076-15; 17 April 2016).

Results

Determining the detection threshold for Psi-Variant

Psi-Variant integrates information from six in-silico tools that assess the functional consequence of missense variants. We examined the positive predictive value (PPV) and diagnostic yield for different numbers of in-silico tools annotating missense variants as “deleterious/damaging” (Fig. 2). There was a strong and statistically significant negative linear correlation (r = −0.98, p < 0.01) between the PPV and diagnostic yield for different detection thresholds with no noticeable optimal threshold (e.g., an increase in PPV without a comparable decrease in yield). Thus, we decided to use a conservative threshold of ≥ 5 tools for this study since it detected a similar number of variants as detected by both InterVar and TAPES (see below).

Figure 2
figure 2

Performance of Psi-Variant using different detection thresholds. The positive predictive value (PPV; X-axis) and diagnostic yield (Y-axis) for different number of in-silico tools annotating missense variants as “deleterious/damaging”. Increasing the minimal number of detection tools increases the PPV but proportionally decreases the detection yield.

Full size image

Detection of candidate variants by the different tools

A total of 372 variants in 168 probands (highlighted in the supplementary Table S2) were detected by at least one of InterVar (n = 220), TAPES (n  = 199), or Psi-Variant (n = 187) from a dataset of 2,035 high-quality, ultra-rare variants with proband-specific genotypes (Fig. 1). Of these, 55 variants (14.8%) were detected by all three tools. The highest concordance in detected variants was observed between InterVar and TAPES (64.3%), followed by TAPES and Psi-Variant (23.1%) and InterVar and Psi-Variant (22.9%).

The characteristics of the detected variants are shown in Table 1. As expected, missense variants comprised the majority of detected variants, with 58.8%, 53.5%, and 51.4% of the variants detected by TAPES, Psi-Variant, and InterVar, respectively. Notably, a higher rate of frameshift variants was detected by Psi-Variant than by InterVar and TAPES (31.5% vs. 17.7% and 11.1%, respectively).

Table 1 Characteristics of the detected variants by InterVar, TAPES, and Psi-Variant from the WES data.
Full size table

Almost all (≥ 95%) variants detected by either InterVar or TAPES were de-novo variants, while de-novo variants comprised only 54.5% of the variants detected by Psi-Variant, which also detected a high portion of X-linked and autosomal recessive variants (17.6% and 27.8%, respectively). Examination of the distribution of the detected variants in genes associated with ASD according to the SFARI Gene database37 revealed a two-fold enrichment of variants distributed in ASD genes (for all detection tools) compared to their portion in the clean vcf file and even a higher enrichment of variants distributed in high-confidence ASD genes (P < 0.001).

Effectiveness of ASD candidate variants detection

To assess the effectiveness of the different tools in detecting ASD candidate variants, we calculated the PPV and the OR for detecting ASD genes (i.e., those listed in the SFARI Gene database37) for different combinations of utilization of the three tools. The results of these analyses are shown in Fig. 3. Utilization of any of the three tools resulted in a significant enrichment of ASD genes, with the highest enrichment being observed in variants detected by both InterVar and Psi-Variant (PPV = 0.178; OR = 4.10, 95% confidence interval (C.I.) = 2.77–5.90; and PPV = 0.176; OR = 4.02, 95% C.I. = 2.63–5.95) followed by TAPES (PPV = 0.158; OR = 3.53, 95% C.I. = 2.28–5.27). Notably, the best performance in detecting ASD candidate variants was obtained at the intersection of the detected variants between InterVar and Psi-Variant (I ∩ P) (PPV = 0.274; OR = 7.09, 95% CI = 3.92–12.22). The I ∩ P combination was also the most effective in detecting SNVs in high-confidence ASD genes (i.e., those with a score of 1 in the SFARI Gene database37 (Fig. 3A-B). However, the I ∩ P combination had a relatively low diagnostic yield of 7.3% for SFARI genes. On the other hand, the union of InterVar and Psi-Variant (I U P) achieved a diagnostic yield of 20.5% (Fig. 3C), but its effectiveness in detecting variants in SFARI genes was lower (PPV = 0.159; OR = 3.59, 95% C.I. = 2.57–4.91) (Fig. 3A-B).

Figure 3
figure 3

Effectiveness of InterVar (I), TAPES (T), Psi-Variant (P), and their combinations in detecting candidate variants in ASD genes. (A) Positive predictive value (PPV) of detecting candidate variants in SFARI 1 and all SFARI genes. (B) Odds Ratios (ORs) of detecting candidate variants in SFARI 1 and all SFARI genes. (C) Diagnostic yield (%) achieved by the different tool combinations for detecting candidate variants in SFARI 1 and all SFARI genes.

Full size image

Discussion

In this study, we assessed the concordance and effectiveness of three bioinformatics tools in the interpretation of variants detected in the WES of children with ASD. There was better agreement in variant detection between InterVar and TAPES than between Psi-Variant and each of these two tools, probably because both InterVar and TAPES are based on the ACMG/AMP guidelines8, while Psi-Variant uses the interpretation of seven in-silico tools in assessing the functional consequences of LGD variants. In addition, most (94%) of the variants detected by either InterVar or TAPES were de-novo variants, compared to only 54.5% of the variants detected by Psi-Variant. This difference may be attributed to the fact that ACMG/AMP guidelines are particularly designed to detect de-novo highly penetrant variants, while inherited variants (autosomal recessive and X-linked) are usually classified as VUS20. Of note, seven of these variants were detected by both InterVar and TAPES as de-novo variants in more than one proband from different families, thus raising the possibility for false positive findings. On the other hand, Psi-Variant detected more homozygote recessive variants than InterVar and TAPES (52 vs. 10 and 5 respectively) and more than estimated in a previously large study18 suggesting some of them may be false positive. However, it also could be partially explained by the ethnic mixture of the study sample, that included many children from consanguineous Bedouin families. Indeed, most of these recessive variants were found in Bedouin children. Importantly, such rare inherited variants have been found to be associated with a variety of neurodevelopmental conditions, including ASD9,17,18,21,22. Another major difference between these tools lies in the detection of in-frame insertions/deletions that comprised ~ 20% of the variants detected by either InterVar or TAPES, while such SNVs were discarded by Psi-Variant. We decided to exclude these variants from Psi-Variant because their clinical relevance has been demonstrated in several genetic disorders38,39 but not in ASD40,41,42.

Another important factor that could affect the concordance between the three tools is the annotation tools they use. Specifically, both InterVar and TAPES use AnnoVar43 for their variant annotation, while Psi-Variant uses Ensembl’s VEP26. It has already been shown that AnnoVar and VEP have a low concordance in the classification of LoF variants44. In addition, each tool, InterVar, TAPES, and Psi-Variant, utilizes a different set of in-silico tools for the classification of missense variants, with SIFT30 alone being shared by all three tools.

Today, there are no accepted guidelines for detecting ASD susceptibility variants from WES data. Many genetic labs use the ACMG/AMP guidelines8, leading to a relatively low diagnostic yield45,46. Our findings suggest that different combinations of bioinformatics tools for variant interpretation may improve the detection of ASD susceptibility variants. Furthermore, combining these tools provides more flexibility in selecting the desired proportion between the detection yield and false positives. Thus, future guidelines for the detection of ASD susceptibility variants should consider the integration of different variant interpretation criteria.

Of note, many of the variants detected by the integrative pipeline affect genes with no known association with ASD, according to the SFARI Gene database37. This finding highlights the capability of the integrative pipeline to detect novel ASD genes. In addition, the combination of these tools could be used for detection and prioritization of susceptibility genetic variants of other medical conditions and human traits. Obviously, the validity of our findings and the applicability of these tools for other conditions should be assessed in additional studies.

The results of this study should be considered under the following limitations. First, the effectiveness assessments of the different tools and their combinations were based on ASD genes from the SFARI Gene database37. While this is the most commonly used database for ASD genes and is continuously updated, it is based on data curated from the literature and may thus include genes falsely associated with ASD. Second, the variant detection analyses were performed on WES data of a cohort from the Israeli population, which may not necessarily be representative of the genetic architecture of ASD. Third, the tools used in this study were designed to detect only extremely rare variants with relatively large functional effects. Thus, a more effective approach for the detection of ASD susceptibility variants should also include the interpretation of other types of genomic variations, such as copy-number and compound heterozygote variants47,48,49,50,51,52, as well as other variants with milder functional effects17,53,54. Finally, it should be noted that there are many other approaches for variant interpretation from WES data. Thus, it is possible that combinations of other approaches will be more effective in the detection of ASD susceptibility variants from WES data than the approaches investigated in this study.

Conclusions

Our findings suggest that combination of different bioinformatics tools is more effective in the detection of ASD candidate variants from WES data than each of the examined tools alone. Future guidelines for the detection of ASD susceptibility variants should consider integrating different variant interpretation approaches to improve the effectiveness of ASD candidate variants detection from whole exome sequencing data.