Assessing variants of uncertain significance implicated in hearing loss using a comprehensive deafness proteome.
Level 5 - mechanism / opinion, no new human data
In silico computational modeling and retrospective database analysis without direct functional assay validation.
PubMed 37086329 · doi:10.1007/s00439-023-02559-9
What was done
Researchers used AlphaFold2 to model protein structures for 224 hearing loss-associated genes curated in the Deafness Variation Database (DVD), refined the structures with the AMOEBA force field, and calculated folding free energy differences (ΔΔG Fold) for 128,167 missense variants using DDGun3D. Variants of uncertain significance (VUS) were evaluated for pathogenicity by combining predicted thermodynamic destabilization with CADD scores (> 25.7), and diagnostic utility was assessed across 119 patients sequenced with the OtoSCOPE panel.
What was found
Thermodynamic modeling identified 5,772 VUSs with large destabilizing ΔΔG Fold shifts consistent with pathogenicity. Filtering with CADD scores classified 3,456 VUSs across 166 genes (74% of evaluated genes) as likely pathogenic at a 99.0% probability. Applying these reclassifications to 119 patients (~3% of the OtoSCOPE cohort) harboring prioritized variants resolved inconclusive reports to provide new genetic diagnoses for six patients.
Why it matters
Applying deep learning-driven structural modeling to proteome-wide variant curation offers an automated, high-throughput pipeline to resolve variants of uncertain significance. This approach directly assists clinical geneticists in interpreting inconclusive sequencing panels in highly heterogeneous disorders like hearing loss.
Limits
The classifications are derived from in silico structural and energetic modeling rather than direct functional wet-lab experiments. The method only detects loss-of-function caused by protein destabilization and misses other mechanisms such as binding site disruption or regulatory changes. The diagnostic yield resolved only six inconclusive cases.
Cited by
- context Machine learning and AI tools developed at Stanford in collaboration with Google can classify genetic variants of unknown significance in deafness genes, raising genetic diagnostic yield from 50% to 80%.