Variant annotation across homologous proteins (“Paralogue Annotation”) identifies disease-causing missense variants with high precision, and is widely applicable across protein families

Technical Summary: Paralogue Annotation for Variant Interpretation

Problem Statement
A primary challenge in clinical genetics is distinguishing pathogenic missense variants from rare but benign ones, particularly for variants not previously observed or characterized in humans. While in vitro and in vivo functional assays exist, they are resource-intensive and may not accurately predict human disease phenotypes. Existing in silico prediction tools often lack the necessary precision for clinical decision-making. Although the ACMG/AMP framework provides guidelines for variant interpretation, it lacks a formalized method for inferring pathogenicity based on equivalent substitutions in evolutionarily related proteins (paralogues), despite the biological plausibility that a disease-causing substitution in one protein family member might cause disease in a paralogous protein via a similar mechanism.

Methodology
The authors developed a framework termed Paralogue Annotation (PA) to transfer pathogenicity annotations from characterized variants in homologous proteins to uncharacterized variants in a gene of interest. The methodology proceeds as follows:

1. Alignment and Homology Mapping: For a query variant in a gene of interest, the system identifies paralogues of that gene using Ensembl Compara. A Multiple Sequence Alignment (MSA) is used to map the query variant's position to equivalent residues in paralogous proteins.
2. Annotation Transfer: The system searches ClinVar (version 20190114) for known Pathogenic/Likely Pathogenic (P/LP) variants at these equivalent positions in the paralogues. If a P/LP variant exists at an aligned position, the annotation is transferred to the query variant, inferring it is likely pathogenic.
3. Conservation Stringency Filters: To increase confidence in functional equivalence, the authors implemented a hierarchy of conservation filters:
   • Baseline: No conservation filter (all aligned positions).
   • Conserved Ref AA: Requires the reference amino acid to be identical between the query gene and the source paralogue.
   • Fully Conserved Ref AA: Requires the reference amino acid to be conserved across the entire protein family.
   • Conserved Alt AA: Requires the alternate (substituting) amino acid to be identical between the query and source.
   • Para Z Score: A quantitative measure of conservation used as an alternative to categorical filters.
4. Extension to Protein Domains: To increase sensitivity, the method was extended beyond direct paralogues to homologous Pfam domains. Using Pfam metadomain alignments, the system identifies equivalent positions across more distantly related proteins that share specific functional domains.
5. Implementation: A Perl plugin for Ensembl's Variant Effect Predictor (VEP) was developed to automate this process for VCF files.

Key Contributions and Results

• High Precision: When applied to 30,334 ClinVar P/LP missense variants, PA correctly annotated 4,328 as pathogenic with only 245 false positives (Benign/Likely Benign variants incorrectly annotated), yielding a precision of 0.95.
  • Applying stricter conservation filters (requiring conserved reference amino acids) increased precision to 0.97, and requiring full family-wide conservation increased it to 0.99, though at the cost of sensitivity.
  • PA demonstrated higher precision than established in silico tools (AlphaMissense, EVE, REVEL, SIFT, M-CAP, CADD) across diverse test sets, including the full ClinVar dataset and specific genes like SCN5A.
• Validation via Orthogonal Approaches:
  • Functional Assays: In 64 genes with Multiplex Assays of Variant Effect (MAVE) data, PA showed precision comparable to or exceeding established methods, particularly when conservation filters were applied.
  • Case-Control Association: In cohorts for Long QT Syndrome (LQTS), Brugada Syndrome (BrS), and Neurodevelopmental Disorders with Epilepsy (NDD+E), variants prioritized by PA showed significantly stronger disease associations (higher Odds Ratios) than unstratified rare variants. For example, in NDD+E genes (SCN1A, SCN8A, KCNQ2), PA increased the mean Odds Ratio from 1.2 to 12.7.
  • De Novo Mutations (DNM): In Autism Spectrum Disorder (ASD) trios, PA-annotated variants were highly enriched in cases compared to controls. While PA identified fewer variants than other predictors (e.g., REVEL), the enrichment rate was significantly higher (Rate Ratio up to 11.4 for fully conserved variants).
• Scalability and Applicability:
  • Approximately 71% of human disease-associated genes have at least one paralogue.
  • Extending the method to Pfam domains increased the number of annotated P/LP variants from 4,328 to 7,543, maintaining high precision (0.89–0.95 depending on filters) while significantly increasing sensitivity.
  • The authors estimate that PA could annotate approximately 855,552 additional missense variants (at 335,243 distinct genomic positions) based on current ClinVar data, representing a 30-fold increase over the currently annotated variants.

Significance and Claims
The paper claims that Paralogue Annotation provides a high-precision line of evidence for clinical variant interpretation that is complementary to existing approaches. The authors emphasize that in clinical genetics, precision (minimizing false positives) is often prioritized over sensitivity to avoid incorrect genetic diagnoses.

Key claims regarding the significance of the work include:

1. Complementarity: PA is not a replacement for functional assays or other in silico tools but offers a distinct, high-confidence evidence source based on observed human variation in related proteins.
2. ACMG/AMP Integration: The approach aligns with existing ACMG/AMP criteria (specifically PS1 and PM5) and can be formally integrated into variant interpretation workflows. The ClinGen RASopathy Expert Panel has already adopted similar rules for specific protein groups.
3. Future Potential: The authors note that the sensitivity of PA is currently limited by the sparsity of characterized variants in reference databases like ClinVar. They assert that as the volume of well-characterized variants increases, the sensitivity and applicability of PA will naturally improve without compromising its high precision.
4. Broad Applicability: The method is not restricted to closely related ion channels but is applicable across a wide range of protein classes, including enzymes, transporters, and transcription factors, provided they have paralogues or shared Pfam domains with characterized disease variants.

The authors conclude that while the absence of a paralogue annotation should be treated as an absence of evidence rather than evidence of benignity, the presence of such an annotation provides robust, high-precision evidence for pathogenicity.