Investigating the use of SpliceAI in Variant Reanalysis of a Cohort of Fetuses with Congenital Brain Abnormalities
Prenatal Genetics
-
Primary Categories:
- Prenatal Genetics
-
Secondary Categories:
- Prenatal Genetics
Introduction:
Splice-altering variants have been proposed as an important potential class of variants to explain the “missing heritability” of cases where a genetic etiology is strongly suspected but not revealed by current analysis. With improvements in machine learning, the ability to identify and characterize splice site variants has improved remarkably. Specifically, one such model powered by artificial intelligence, SpliceAI, has substantially improved accuracy compared to prior models and has started to be incorporated into research and clinical bioinformatic pipelines. The current study used SpliceAI to reanalyze a cohort of fetuses with brain abnormalities. We hypothesized that a reanalysis with SpliceAI would lead to increased diagnostic yield. Additionally, we sought to further characterize splicing variants in our cohort and the additional effort required to incorporate SpliceAI into analysis.
Methods:
This is a retrospective study examining exome and genome sequencing from a cohort of fetuses with congenital brain abnormalities. Data on the number of variants identified and the rate of negative results were collected and standardized. Possible pathogenic splicing variants were further analyzed and annotated in conjunction with patient-specific phenotypic information. Variants were classified according to ClinGen/ACMG/AMP criteria. To focus on genes with the highest likelihood of relationship to indication for testing, we used the PanelApp fetal anomalies gene list. Variants that met our criteria to be examined had a Groupmax filtering allele frequency of 0.01 or lower in gnomAD v4.0, were located within a gene on the PanelApp fetal anomalies gene list, and had a spliceAI score of 0.2 or greater. We used SpliceAI to analyze the exome and genome sequence data from 91 fetal probands.
Results:
From 91 DNA samples extracted from the amniocytes or chorionic villi of fetuses with brain abnormalities, a total of 490 variants met our criteria to be examined. Of the 490, five variants were considered possibly disease causing. Four of the five had already been identified with prior analysis and three out of the five variants were missense, frameshift, or canonical splice site. Notably, two of the five variants were previously identified potentially pathogenic variants located at the +5 intronic splice donor site. Further analysis of our dataset revealed a high rate of variants with a SpliceAI score of 0.2 or greater that were classified in ClinVar as benign or likely benign (950/1312 (72%)) or had a small predicted impact on protein function by SnpEff 5.2c (2718/3815 (71%)). Finally, we found that 214 canonical splicing variants in our dataset had a donor or acceptor gain score of 0.2 or greater, 15 of which had a predicted alternative splice site located 3, 6, 9, 12, or 15 base pairs away from the canonical splice site.
Conclusion:
The results of our study suggest that SpliceAI adds little additional sensitivity in variant re-analysis. Our results favor the use of SpliceAI in more narrow applications, such as the interpretation of variants just outside canonical splice sites, investigating cryptic splice sites, or in predicting the impact of de novo variants.
Splice-altering variants have been proposed as an important potential class of variants to explain the “missing heritability” of cases where a genetic etiology is strongly suspected but not revealed by current analysis. With improvements in machine learning, the ability to identify and characterize splice site variants has improved remarkably. Specifically, one such model powered by artificial intelligence, SpliceAI, has substantially improved accuracy compared to prior models and has started to be incorporated into research and clinical bioinformatic pipelines. The current study used SpliceAI to reanalyze a cohort of fetuses with brain abnormalities. We hypothesized that a reanalysis with SpliceAI would lead to increased diagnostic yield. Additionally, we sought to further characterize splicing variants in our cohort and the additional effort required to incorporate SpliceAI into analysis.
Methods:
This is a retrospective study examining exome and genome sequencing from a cohort of fetuses with congenital brain abnormalities. Data on the number of variants identified and the rate of negative results were collected and standardized. Possible pathogenic splicing variants were further analyzed and annotated in conjunction with patient-specific phenotypic information. Variants were classified according to ClinGen/ACMG/AMP criteria. To focus on genes with the highest likelihood of relationship to indication for testing, we used the PanelApp fetal anomalies gene list. Variants that met our criteria to be examined had a Groupmax filtering allele frequency of 0.01 or lower in gnomAD v4.0, were located within a gene on the PanelApp fetal anomalies gene list, and had a spliceAI score of 0.2 or greater. We used SpliceAI to analyze the exome and genome sequence data from 91 fetal probands.
Results:
From 91 DNA samples extracted from the amniocytes or chorionic villi of fetuses with brain abnormalities, a total of 490 variants met our criteria to be examined. Of the 490, five variants were considered possibly disease causing. Four of the five had already been identified with prior analysis and three out of the five variants were missense, frameshift, or canonical splice site. Notably, two of the five variants were previously identified potentially pathogenic variants located at the +5 intronic splice donor site. Further analysis of our dataset revealed a high rate of variants with a SpliceAI score of 0.2 or greater that were classified in ClinVar as benign or likely benign (950/1312 (72%)) or had a small predicted impact on protein function by SnpEff 5.2c (2718/3815 (71%)). Finally, we found that 214 canonical splicing variants in our dataset had a donor or acceptor gain score of 0.2 or greater, 15 of which had a predicted alternative splice site located 3, 6, 9, 12, or 15 base pairs away from the canonical splice site.
Conclusion:
The results of our study suggest that SpliceAI adds little additional sensitivity in variant re-analysis. Our results favor the use of SpliceAI in more narrow applications, such as the interpretation of variants just outside canonical splice sites, investigating cryptic splice sites, or in predicting the impact of de novo variants.