Compendium of Clinical Variant Classification for 2,246 Unique ABCA4 Variants to Clarify Variant Pathogenicity in Stargardt Disease Using a Modified ACMG/AMP Framework

Biallelic variants in ABCA4 cause Stargardt disease (STGD1), the most frequent heritable macular disease. Determination of the pathogenicity of variants in ABCA4 proves to be di ﬃ cult due to (1) the high number of benign and pathogenic variants in the gene; (2) the presence of many rare ABCA4 variants; (3) the presence of complex alleles for which phasing data are absent; (4) the extensive variable expressivity of this disease and (5) reduced penetrance of hypomorphic variants. Therefore, the classi ﬁ cation of many variants in ABCA4 is currently of uncertain signi ﬁ cance. Here, we complemented the ABCA4 Leiden Open Variation Database (LOVD) with data from ~ 11,000 probands with ABCA4 -associated inherited retinal diseases from literature up to the end of 2020. We carefully adapted the ACMG/AMP classi ﬁ cations to ABCA4 incorporating ClinGen recommendations and assigned these classi ﬁ cations to all 2,246 unique variants from the ABCA4 LOVD to increase the knowledge of pathogenicity. In total, 1,248 variants were categorized with a likely pathogenic or pathogenic classi ﬁ cation, whereas 194 variants were categorized with a likely benign or benign classi ﬁ cation. This uniform and improved structured reclassi ﬁ cation, incorporating the largest dataset of ABCA4 -associated retinopathy cases so far, will improve both the diagnosis as well as genetic counselling for individuals with ABCA4 -associated retinopathy

Furthermore, in determining the pathogenicity of an ABCA4 variant, multiple other factors should be considered.Several missense and synonymous variants are known to cause splicing defects in ABCA4 [11][12][13]; therefore, missense variants at positions that are not conserved and synonymous variants cannot simply be dismissed as likely benign.In addition, reduced penetrance has been reported for multiple hypomorphic variants in ABCA4 [14,15] meaning that variants with a relatively high allele frequency can still be pathogenic.One might expect that when two ABCA4 variants are detected in a person with a STGD1-like phenotype, these must be biallelic and causal.However, it is not uncommon to find variants to be in cis in ABCA4, and multiple complex alleles have been described [16][17][18].Several studies indicate that in cases with a single variant in ABCA4 or a single complex ABCA4 allele, the disease-causing variant(s) can be found in other genes [19][20][21].Moreover, the effect of severity on protein function varies enormously among ABCA4 pathogenic variants.Two deleterious loss of function or null alleles can cause legal blindness before the age of 10 [22-25], whereas mild variants usually only cause disease when present in trans with a null or severe allele.These mild variants are associated with foveal sparing [14,26], resulting in a late age at onset.Furthermore, population statistics and family studies revealed reduced penetrance for some of these variants [15].The variant c.5603A>T (p.(Asn1868Ile)) has been shown to cause visual impairment in ~5% of individuals when in trans with a deleterious allele [14,26], rendering it a clear hypomorphic variant.This further illustrates the complexity of ABCA4-AR and the difficulty of classifying ABCA4 variants.
Consequently, many individuals with ABCA4-AR are currently not genetically diagnosed, as only one pathogenic variant or allele or biallelic variants of uncertain significance have been identified.For these individuals, it is important to know whether the variants they have are pathogenic.Currently, determining the pathogenicity of ABCA4 variants is crucial as many clinical trials for gene-specific therapies require individuals to have biallelic pathogenic alleles to be eligible for participation.Moreover, as the carrier rate of pathogenic ABCA4 variants in the general population is relatively high [6,27], carrier analysis is performed frequently to determine the risk for future offspring, and in these cases, it is important to know whether identified variants are pathogenic.
Databases, such as the ABCA4-Leiden Open (source) Variation Database (LOVD) [28] and ClinVar [29], provide a wealth of information for ABCA4 variants including pathogenicity classifications.However, there is discordance between databases.The ABCA4-LOVD reports that up to 85% of the variants could be pathogenic, whereas in ClinVar, ~40% of the variants are reported to be likely pathogenic, which might be because many novel variants of uncertain significance are reported in ClinVar.Both the ABCA4-LOVD and ClinVar also allow for variable classifications as they are submitter-reported.Therefore, due to the variety of classification methods, it is difficult to truly assess and compare the pathogenicity of different variants.
A broadly used pathogenicity classification system is the ACMG/AMP classification as described by Richards et al. in 2015 [30].This classification method incorporates information such as type of variant, cis/trans criteria, variant fre-quency, phenotype, functional studies, segregation, and in silico predictions.All information can be collected per variant and can easily be combined to a final pathogenicity classification with five tiers: benign, likely benign, variant of uncertain significance (VUS), likely pathogenic, or pathogenic.Since it is consistently used worldwide, it allows easy interpretation and comparison of pathogenicity levels.
In order to increase the knowledge on the pathogenicity of genetic variants in ABCA4, we collected all the data on ABCA4-AR cases that have been published up to 31 December 2020 and uploaded these data into the ABCA4-LOVD.
Here, we adapted ACMG/AMP classifications incorporating ClinGen recommendations specifically for ABCA4 and applied them to all 2,246 variants present in the ABCA4-LOVD.

Methods
We collected all papers published until 31 December 2020, which contain likely pathogenic ABCA4 variants in individuals with retinopathy by searching the following search terms in PubMed: (ABCA4[All Fields] OR (("Stargardt disease"[All Fields] OR "Macula Lutea"[All Fields]) AND ("Genetics"[All Fields] OR "mutation"[All Fields] OR "Sequence Analysis"[All Fields] OR "gene panel"[TiAb]))) OR ("Retinal Dystrophies"[All Fields] AND ("mutation"[All Fields] OR "Sequence Analysis"[All Fields] AND "gene panel"[TiAb])) Reported variants were collected per patient as well as additional available data such as gender, type of vision impairment, ethnicity, geographical origin, age at onset, phenotype at onset, segregation data, consanguinity status, and other remarks.The data were supplemented with data from 412 persons with ABCA4-AR from PreventionGenetics (a division of Exact Sciences).All data have been uploaded into the ABCA4-LOVD [28].

2.2.
Nomenclature.The annotation of all variants was done according to Human Genome Variation Society (HGVS) nomenclature guidelines where possible and is based on the GRCh37 hg19 genomic coordinates, gene location NM_ 000350.3.All variants were checked using the Batch Validator of the online VariantValidator tool [31].Throughout the article, the c. notation of variants is used, supplemented with the p. notation, if available, when mentioned for the first time.
2.3.ACMG/AMP Classification.All variants were classified according to ACMG/AMP variant classification guidelines described by Richards et al. [30], the updated recommendations from ClinGen [32], and the naturally scaled ACMG/AMP point system of Tavtigian et al. [33].Our project group, consisting of experts on ABCA4 genetics and ophthalmology, decided how best to apply the ACMG/AMP categories and ClinGen recommendations to ABCA4 variants, which is summarized in Figure 1 and can be found in more detail in the Supplemental Materials & Methods (available here).

Results
3.1.Published ABCA4-AR data Cohort.After the removal of likely duplicates, the collection of data contained variants from 10,391 likely ABCA4-AR individuals, of which 3,411 were already reported in our 2017 study [34].The cohort contained 6,240 likely biallelic cases and 4,151 monoallelic cases.Among these were 943 nonconsanguineous homozygous cases and 127 consanguineous homozygous cases.A total of 2,094 unique variants were identified in the cohort.All data were uploaded to the ABCA4-LOVD database [28].variants from LOVD based on ACMG/AMP classification.For each step, a number of points are awarded and the sum of the points is used to determine the initial ACMG/AMP classification according to Tavtigian et al. [33].GAM BAP gnomAD data are described in Cornelis et al. [27].(b) Based on the initial ACMG/AMP classification, subsequent and iterative classification steps were executed.#The missense classification steps were not iterated to avoid circular reasoning.(c) The total sum of points lead to the final ACMG/AMP classification: benign (≤−7), likely benign (−6 to −1), VUS (0 to 5), likely pathogenic (6 to 9), and pathogenic (≥10).

ACMG/AMP
variants from the published cohort as well as other ABCA4 variants from the ABCA4 LOVD (Table S1) and were annotated with the predicted pathogenicity severity according to Cornelis et al. [27] when available.Results of each ACMG/AMP classification step can be found in Tables S2-S10.After applying the noniterative classification steps, 1,276 variants could be categorized as benign, likely benign, likely pathogenic, or pathogenic.The iterative classification steps increased this number to 1,442 (Table 1 and Figure 2(a)).Of note, at the end of the analysis, 49 null variants reach "pathogenic" without the PVS1 criterium and >10% of variants associated with ABCA4-AR are loss of function confirming that the use of PVS1 flowchart is correct [35].

Null Variants.
In total, 752 null variants were reported.Interestingly, 24 of these variants (3%) are still classified as VUS after the application of the ACMG/AMP classification system (Figure 2).Compared to other null alleles, the splice predictions for these 24 variants indicated that alternative in-frame splicing could occur or in-frame deletions were predicted.Furthermore, their occurrence in the dataset was too low to reach significance in the frequency analysis.

Missense Variants.
A group of variants that are very difficult to interpret without an ACMG/AMP classification are missense variants.Here, we were able to classify 431 missense variants as (likely) pathogenic and 33 missense variants as likely benign.At the end of the analyses, 627 variants were classified as VUS (Figure 2).

Synonymous Variants.
Of the 86 synonymous variants, five variants could be classified as (likely) pathogenic and 67 variants as (likely) benign (Figure 2).

Frequency Analysis.
In the frequency analysis, 856 variants reached significant enrichment in the likely biallelic dataset compared to the earlier described [27] genetic ancestry matched (GAM) to biallelic affected persons (BAP) gnomAD control dataset after correction for multiple testing with the Benjamini-Hochberg method [36].Interestingly, two variants, c.5603A>T and c.4253+43G>A ((p.[=,Ile1377Hisfs * 3])), were significantly enriched but had an odds ratio close to 1 (1.10 and 1.49, respectively) without having 1 in the confidence interval.This is a smaller odds ratio than 3, which is generally considered by Richards et al. [30] as the minimal odds ratio for variants with a modest Mendelian effect size.This effect is likely due to the reduced penetrance of these var-iants.Furthermore, 1,125 variants were not enriched in the likely biallelic dataset but had an allele frequency below 0.0001 in all gnomAD populations and therefore got "PM2_ Supporting" evidence.
3.7.In Silico Analyses.SpliceAI scores were given to all variants < 50 nucleotides apart from indels.When possible, indels were given a CI-SpliceAI score.In total, 472 variants had a (CI-)SpliceAI score > 0 2 and received PP3 as minimal evidence.In parallel, REVEL and CADD scores were given to missense variants and other variants, respectively.In total, 1,141 variants received "PP3_Moderate" evidence, of which 658 variants were missense variants.On the lower end of the spectrum, 194 variants received "BP4_Moderate."Interestingly, of those 194 variants, only 6 were missense variants, and 80 were other exonic variants (71 synonymous variants, 6 point deletions, and 3 in-frame duplications).Finally, 180 additional variants received PP3, and 48 variants received BP4.Of note, when comparing CADD and REVEL scores for missense variants based on the cut-offs as described by Pejaver et al. [37], it was noticed that only 7 variants received "BP4_Moderate" based on the cut-off of ≤0.183 for REVEL, while 124 variants would have received "BP4_Moderate" if the CADD score (cut-off of ≤17.3) would have been used (Figure S1).

Segregating Complex Alleles.
The dataset contained a total of 65 unique alleles containing two or more variants (also known as complex alleles) for which segregation had been reported (Table S13).Of note, many individuals with multiple ABCA4 variants were not reported to have undergone segregation analysis, resulting in a relatively low number of known complex alleles.The three most frequently reported complex alleles were c.S13).Of note, the pathogenicity scores of alleles with two variants do not have to reflect that both single variants have the same scores.Therefore, for reference, ACMG/AMP classifications of the single variants contained in the complex allele are given in Table S13 as well.
3.9.Frequent Pathogenic Variants.Two previously known pathogenic variants met the BS1 criterium-an allele frequency of >0.0163 in any gnomAD population-while also reaching a (likely) pathogenic classification.The  For each gnomAD population, the three most frequent (likely) pathogenic variants are reported in Table S11.Interestingly, three relatively less well-known variants, c.2791G>A (p.(Val931Met)), c.2971G>C (p.(Gly991Arg)), and c.6320G>A, have a very high frequency (0.004-0.021) in the gnomAD African population, while they are less frequent in other populations.Similarly, in the Latino/ Admixed American population, variant c.872C>T (p.(Pro291Leu)) with an allele frequency of 0.0038 was found to be likely pathogenic.Given their high allele frequencies, it will be of interest to investigate if they show reduced penetrance, which might not be unexpected as all of these variants were predicted to be mild [27,38].Previously reported frequent pathogenic variants and their updated ACMG/AMP classification can be found in Table S12.Interestingly, the variant c.2588G>C, earlier described as "North European," has a high frequency in the gnomAD South Asian population as well.

Discussion
In this study, we assigned ACMG/AMP classifications other than VUS to 1,442 of 2,246 ABCA4 variants based on the point system of Tavtigian et al. [33] and ClinGen recommendations [32], which is more than twice as much as our 2017 classification [34].Compared to the 2017 classification, 1,419 new variants were analyzed.From the previously analyzed variants, 210 variants got a more severe pathogenicity score, and 142 got a more benign pathogenicity score.In total, 93 variants are now classified as VUS, while the 2017 study assigned a different classification to them.These differences are likely the result of an increase in available knowledge, such as information from new functional studies and improved prediction software, as well as a more genespecific approach of applying the ACMG/AMP rules.In total, 1,248 variants were classified as either likely pathogenic or pathogenic, 804 variants were classified as VUS, and 194 variants were classified as either likely benign or benign.Furthermore, Table S1 provides a framework that can easily be adjusted to improve the AMP/ACMG classification of ABCA4 variants when additional information becomes available.This will be of important value to clinical geneticists, individuals affected by ABCA4-AR, their family members, and ongoing clinical trials for gene-specific therapies.

The Limitation of ACMG/AMP Classification for ABCA4
Variants.The ACMG/AMP classification is designed in a way that likely pathogenic and likely benign variants are classified with ≥90% certainty.Although this is usually interpreted as a very reliable classification, it should be mentioned that when a large group of variants is classified, some variants will receive an incorrect classification.Furthermore, it is important to mention that the dichotomous pathogenicity framework that the ACMG/AMP classification system is based on currently categorizes a variant as pathogenic if it can cause disease, even if it does not always cause disease in trans with another pathogenic variant, such as for the reduced penetrant variants that have been reported in ABCA4 [15].Recently, multiple studies have shown that using the Mendelian model of traits being either recessive or dominant limits the understanding of the role genetic variants have in disease mechanisms that show a gradual or varying effect [39,40].It has, therefore, been suggested to expand the ACMG/AMP classification to a seven-tier system including "predisposing" and "likely predisposing" as additional classifications [39].The term "predisposing" may indeed be a better classification for variants with reduced penetrance than the dichotomous term "pathogenic."However, for most variants, it will be difficult to determine whether they show reduced penetrance, leading to a "Predisposing" classification, or not, leading to a "pathogenic" classification.Here, we annotated variants that have been reported to show reduced penetrance (Table S1).Reduced penetrance is reflected in both discordance in families with ABCA4-AR as well as in a low odds ratio in enrichment studies, e.g., below 28.1 in this study (Table S4).It may, therefore, be advisable to be cautious for pathogenic and likely pathogenic variants with an odds ratio < 28 1 in particular, as these might be variants with reduced penetrance.Interestingly, the well-known variant c.5603A>T, which shows a very low penetrance when in trans with a severe or null allele (approximately 5%) [14] is classified as a VUS.This variant was long believed to be benign as it was found to cooccur with the less frequent variant c.2588G>C which was thought to be pathogenic [41].However, in 2017, Zernant et al. identified that c.5603A>T is disease causing, and c.2588G>C without c.5603A>T might not be pathogenic [26].It was clear that c.5603A>T was underreported in the data from before 2017 because of this, since c.2588G>C has often been reported without c.5603A>T in the literature.This means that the pathogenicity of c.2588G>C is likely overclassified here, while that of c.5603A>T is likely underclassified.Similarly, the mild variant c.4253+43G>A, which shows a splice defect in vitro [16] and shows reduced penetrance [42], is classified as VUS.A possible explanation for this is that deep intronic variants-including the ones reported by Braun et al. in 2013, although to a lesser extent [43]-are underreported in literature, since targeted ABCA4 exon sequencing or WES were the norm for a long time and additionally because their interpretation can be challenging.Therefore, they are less likely to reach significant enrichment in the dataset in step PS4 and may be excluded from the PM3 criterium in trans classification when they are not reported.Of note, new and more affordable techniques are now increasing the number of identified deep intronic variants [16,44,45], which will likely improve the knowledge on their pathogenicity.However, since those techniques are not yet available everywhere, it will be challenging to identify the pathogenicity of deep intronic variants in all populations.
This illustrates the difficulty of recognizing frequent mild pathogenic variants that show reduced penetrance.Therefore, variants with an odds ratio between 1 and 3 should particularly be treated with caution, although reduced penetrance has been predicted for variants with an odds ratio up to 28.1 [15].Larger studies are necessary to identify whether these variants show reduced penetrance or that instead, they may be in linkage disequilibrium with an unknown pathogenic variant.The distinction between these may in part be predicted by the variants identified in trans; if the majority of those are severe, the variant or a variant in linkage disequilibrium with it is likely mild and might show reduced penetrance.However, if the variants in trans are not consistently severe, then the variant is likely in linkage disequilibrium with an unidentified (moderately) severe variant.
Modifiers seem to play a role in ABCA4-AR and may explain the occurrence of variants with reduced penetrance.A sex imbalance has been reported for individuals having mild likely reduced penetrant variants [15], and common PRPH2 variants and rare ROM1 variants have been reported to act as modifiers of ABCA4-AR [46].In three Dutch families with biallelic sibling pairs carrying c.5603A>T in trans with another severe ABCA4 variant, not all siblings were affected by ABCA4-AR [14].Kjellström and Andréasson 7 Human Mutation may also have found two unaffected male individuals over 50 that had both c.5603A>T and a severe variant [47].Modifiers may similarly explain the reported high variety in the disease course of ABCA4-AR [48].Furthermore, modifiers might also aggravate the ABCA4-AR phenotype of individuals; Leber's congenital amaurosis, causing severe vision loss in the first year of life, is usually not associated with variants in ABCA4, but Panneman et al. identified probands in which two ABCA4 null alleles are hypothesized to cause Leber's congenital amaurosis [49] (Panneman, Koenekoop, Cremers, unpublished data).
Another important point to raise for recessive disease is that variants leading to a protein with reduced but not abolished expression and/or function may not always be disease causing, depending on the variant in trans.For example, if a variant reduces protein expression to 45% compared to WT, then its occurrence next to a null variant is likely disease causing.A homozygous occurrence of this variant, however, will lead to expression only just below that of an individual with a null allele in addition to a WT allele.This may cause a situation to a combination of a hypomorphic variant next to a null allele, where modifiers may determine whether an individual will be affected or not.In other words, considering that all ABCA4 variants are on a spectrum based on residual protein function and resulting cellular dysfunction, it is likely that the combined severity of ABCA4 variants and modifiers together determine disease penetrance and severity.
Furthermore, several studies indicate that pathogenic variants in other genes can be responsible for ABCA4-AR even when one likely pathogenic ABCA4 variant is present in the patient.Disease-causing variants can sometimes be found in genes like PRPH2 and PROM1, but also in less common genes associated with ABCA4-like diseases like BEST1, CDHR1, CERKL, CNGA3, CRX, ROM1, and RPE65 [19,20,50].

Study Limitations.
Apart from the aforementioned limitation that the ACMG/AMP guidelines have when considering variants with a gradual versus a dichotomous pathogenicity effect, there are a few more limitations to this study.The first one is associated with PS4, the allele frequency analysis.First, the GAM BAP control dataset used is based on the reported ethnicity, which may not correspond with the gnomAD population that individuals were matched with, since ethnicity is a social construct and gno-mAD populations are for a big part based on principal component analyses [51], and it is unknown to what extent those overlap.Second, for those patients without reported ethnicity, the GAM BAP gnomAD has incorporated estimated ethnicity based on population statistics.However, there may be a bias in the ethnicity of patients who are able to either afford healthcare, who have the option to take part in a study, or who feel safe to take part in a study since, for example, historic transgressions have been made against Black research participants [52].Finally, as labs that study ABCA4-AR and report individuals with ABCA4-AR are unequally distributed over the world, a bias in the genetics of subpopulations of those regions could occur.Therefore, it is likely that population stratification will have affected the results of the allele frequency test.In order to improve healthcare for everyone, it is, therefore, important that rare genetic variants in individuals with underreported genetic ancestry in literature are studied more to improve the knowledge on all genetic variants, and that variants with a high-frequency difference between populations are investigated more closely to study their effect since differences in identified variants and numbers of ABCA4-AR cases between different ethnicities have been reported [53].Currently, both genetic and disease data of white individuals with ABCA4-AR are overrepresented, creating an imbalance in understanding of the genetic cause of ABCA4-AR and treatment options between these individuals and individuals of color with ABCA4-AR.
Moreover, based on the final classification, 298 total variants from the biallelic dataset (2.3%) are likely benign or benign.This means that up to 298 cases from this dataset are not actually known biallelic, which might indicate that those cases are not actually ABCA4-AR cases, which could create a bias in the enrichment analysis for the variants present in trans.
Furthermore, the ACMG/AMP guidelines and recommendations warn for the use of functional studies.We indeed encountered that a variant, c.4539+2028C>T (p.[=,Arg1514Leufs * 36]), which likely is a pathogenic variant based on genotype-phenotype correlations [43], shows a higher percentage of WT RNA in patient-derived retinallike cells than expected [54].We decided to remove this data point from the dataset as an outlier in the classification based on functional studies in step BS3.In addition, other variants, e.g., c.4539+2001G>A (p.[=,Arg1514Leufs * 36]) and c.1937+435C>G (p.[=,Ser646Serfs * 25]), show a relatively high percentage of WT RNA production, 75% and 55%, respectively, in patient-derived retinal-like cells [54] and midigene assays, respectively [16], while genotypephenotype correlations show that these variants are likely pathogenic [43,55].This indicates that other variants could show a similar pattern, meaning that intronic variants causing a relatively high amount of WT RNA may nevertheless be pathogenic.Therefore, results from midigene assays and patient-derived retinal-like cells should be interpreted with caution.However, since most variant results seem to correlate with their pathogenicity, BS3_Supporting was deemed to be of proper evidence strength.
In addition, in the use of in silico predictions, the CADD score was used for nonmissense variants.However, the applied cut-offs were based on the study of Pejaver et al. [37] in which only missense variants were studied.Furthermore, when comparing REVEL and CADD scores for ABCA4 missense variants, CADD scores seem to lead to a more benign category (Figure S1).Only one variant that got PP3_Moderate because of its REVEL score would have gotten BP4_Moderate if CADD would have been used.However, since REVEL is specialized in scoring missense variants and as it outperformed CADD in two studies [37,56], the REVEL scores were considered to be more trustworthy.Therefore, we decided to increase the range of CADD scores leading to no in silicoevidence score from 8 Human Mutation 22.7-25.3 to 20-25.3 to avoid incorrectly classifying in-frame insertions/deletions, noncanonical splice variants, and synonymous variants as benign since 20 is often used as a cut-off between a benign and a pathogenic indication.In the final categorization, this led to four variants being categorized as likely pathogenic or pathogenic instead of a lower category.Finally, it should be mentioned that variant classification is, and should be, dynamic.Since the initial ACMG/AMP guidelines were published in 2015 [30], new insights have led to many recommendations to improve the classification system [32,33,37,39].With the increasing knowledge on genetic disease and improving strategies to understand variant effects, it is, therefore, important to regularly incorporate evolving variant classification strategies.4.3.Future Scope.Finally, the ABCA4 variant dataset analyzed here mostly stems from 421 peer-reviewed publications as well as data from PreventionGenetics (a division of Exact Sciences).In the future, it would be very valuable to include variant data collected in all academic and nonacademic diagnostic centers worldwide.This will be challenging as privacy rules may prevent data sharing and differences between rules in different countries likely create a bias in the data.Furthermore, data currently existing in different online databases may show overlap and are not all curated.
With the advent of novel therapies, it is essential to have an accurate genetic diagnosis, which emphasizes the importance of the classification of variants and proper guidelines.ABCA4 variant classification is challenging due to the higher mutation frequency, presence of complex alleles, and hypomorphic variants with reduced penetrance.The adapted ACMG/AMP classifications provided in this study, in combination with the earlier established severity assessments for ABCA4 variants, will facilitate the interpretation of diagnostic results for ABCA4-AR, the most common recessive retinal disease.classifications are based on the point system as described by Tavtigian et al. [33].In short, supporting, moderate, strong, and very strong evidence is combined into a score where each type of evidence gives a score of 1, 2, 4, or 8, respectively, where pathogenic evidence gives a positive score and benign evidence gives a negative score.The resulting total score per variant results in a benign (<-6), likely benign (-1--6), VUS (0-5), likely pathogenic (6)(7)(8)(9), or pathogenic (>9) classification.Table S2: ACMG/AMP classification step PVS1 Null variants.Table S3: ACMG/AMP classification step PM6 de novo variants.Table S4: ACMG/AMP classification step PS4 variant frequency and use of control populations.Table S5: ACMG/ AMP classification step PM4 protein length changes due to in-frame deletions/insertions and stop losses.Table S6: ACMG/AMP classification steps PP3 and BP4 computational (in silico) data.Table S7: ACMG/AMP classification step BP7 synonymous variants.Table S8: ACMG/AMP classification steps BS1 and PM2 variant frequency and use of control populations.Table S9: ACMG/AMP classification steps PS1 and PM5 Same amino acid change and novel missense at the same position.Table S10: ACMG/AMP classification steps PS3 and BS3 functional studies.Table S11: three most frequent (likely) pathogenic variants per gnomAD population.Table S12: previously reported frequent pathogenic variants based on literature.Table S13: published segregating complex alleles.Figure S1: in silico comparison of CADD and REVEL for missense variants in ABCA4.In silico comparison of ABCA4 missense variants.CADD PHRED values are plotted against REVEL values.Cut-off values between "BP4_Moderate" and "BP4," and between "PP3" and "PP3_Moderate" are shown in green and red, respectively, vertically for CADD scores and horizontally for REVEL scores.Overall, it can be observed that the variants reach a higher category for REVEL than for CADD. Figure S2: correlation between variant F-indexes and age of onset in patients.Correlations between ABCA4 variants' Findex and the corresponding log 10 and square root transformed age at onset for homozygous configuration (A) and compound heterozygous configuration with a severe variant (B).(Supplementary Materials)

Figure 1 :
Figure 1: Applied ACMG/AMP classification steps incorporating ClinGen recommendations.(a) Initial steps taken to classify ABCA4variants from LOVD based on ACMG/AMP classification.For each step, a number of points are awarded and the sum of the points is used to determine the initial ACMG/AMP classification according to Tavtigian et al.[33].GAM BAP gnomAD data are described in Cornelis et al.[27].(b) Based on the initial ACMG/AMP classification, subsequent and iterative classification steps were executed.#The missense classification steps were not iterated to avoid circular reasoning.(c) The total sum of points lead to the final ACMG/AMP classification: benign (≤−7), likely benign (−6 to −1), VUS (0 to 5), likely pathogenic (6 to 9), and pathogenic (≥10).

Table 1 :
Overview of the number of ABCA4 variants from LOVD per ACMG/AMP categorization.