Approximately 1 to 3 of every 1,000 newborns in the U.S. are born with a detectable and permanent hearing loss, and by the age of 70, more than 50% of Americans have disabling hearing loss. To understand its genetic underpinnings, the deafness variation database (DVD) lists 224 genes, variations in which are definitively implicated in hearing loss. Included in the DVD are more than 400,000 missense variants with over 320,000 labeled as variants of unknown significance (VUS). As new, more accurate tools emerge to predict protein structure, we can include biophysical evidence in our genetic profiling of missense variants to better understand and classify phenotypes, potentially enabling reclassification of a VUS to a pathogenic category. To test this hypothesis, we applied AlphaFold3 to predict protein structures for 320 unique isoforms of the 224 genes in the DVD. To resolve steric clashes, repack sidechains, and improve backbone conformations, we optimized raw AlphaFold3 structures with the AMOEBA polarizable force field in the Force Field X software. With the optimized protein structures referred to as OtoProteinV3, we calculate folding free energy differences ( ΔΔGFold ) to quantify the impact each missense variant has on protein folding. We then constructed a combined likelihood ratio to determine the necessary ΔΔGFold value and the genetic effect predictor score from REVEL to reach an 1,124:1 likelihood (“very strong” evidence) of implicating a variant as causally related to hearing loss, with a false positive rate of 0.15%. Based on these metrics, over 30,000 missense VUS are pathogenic for hearing loss. In sum, we have used biophysical evidence from deep learning predicted protein structures to enhance our understanding of hearing loss. While this work focuses on hearing loss, the methods used are broadly applicable to other diseases.
Gogal et al. (Sun,) studied this question.