The prevalence of protein misfolding as a mechanism for hereditary deafness Rose A. Gogal, Genevieve M. Cox, Diana L. Kolbe, Amanda M. Odell, Chloe E. Ovel, Katherine I. McCormick, Brian Hong, Hela Azaiez, Thomas L. Casavant, Richard J. H. Smith, Terry A. Braun, Michael J. Schnieders, David Balding Abstract Hearing loss is the most common sensory deficit impacting ~5% of the world’s population. The Deafness Variation Database (DVD) is a public resource of deafness variants, containing 381,924 missense variants across 224 genes, with 303,577 classified as a variant of uncertain significance (VUS). To address the challenge of evaluating each deafness associated VUS, we evaluate a family of probabilistic frameworks to quantify the strength of computational evidence based on ACMG/AMP recommendations. First, CADD and REVEL are compared using Bayesian models parameterized using either a ClinVar 2019 dataset or labeled DVD variants. The REVEL model built using the DVD dataset demonstrates the best accuracy, sensitivity, and specificity. Incorporation of (in)tolerance to missense variation based on sorting each gene into three bins (tolerant, average, intolerant) shows that intolerant DVD genes are consistent with a higher prior probability of being pathogenic (25.7%) than average (10.7%) or tolerant (8.7%) genes. Finally, the impact of protein folding stability was incorporated into the Bayesian model and surpassed the simpler versions while also offering a biophysical rationale for the disease mechanism. The 28,866 VUSs that reach a posterior probability of pathogenicity above 98% based on the protein-folding informed Bayesian model were prioritized as likely to be pathogenic. Overall, 54,752 missense variants (14.3% of 381,924) are predicted to cause modest protein folding destabilization of greater than 1.0 kcal/mol, while 18,706 of these are prioritized VUSs (34% of 54,752). 22,237 missense variants (6.2% of 381,924) are predicted to cause more severe protein folding destabilization of at least 2 kcal/mol with 12,424 being prioritized VUSs (55.8% of 22,237). From these VUSs, we identify twelve probands where the patient’s genetic diagnosis is upgraded to likely pathogenic/pathogenic. We highlight two variants that cause clear structural disruption, demonstrating the impact of biophysical characterization on variant evaluation. We investigate the impacts of single amino acid changes on protein structure and folding in the context of hearing loss. Hearing loss is the most common impairment of the main senses affecting nearly 5% of the world’s population. About 45% of people with hearing loss receive a diagnosis after targeted genetic testing. Here, we integrate biophysical data that quantifies the effect of a change to protein sequence on protein folding in combination with genetic data to improve our ability to identify protein amino acid changes that are likely to impact hearing. Our work leads to 12 patients receiving an upgraded diagnosis with their variant disrupting protein stability. Although the method is applied to hearing loss, it can be used for interpreting protein sequence changes in other disease contexts. Introduction Hearing loss is the most common sensory deficit. It affects ~5% of the world’s population, impacting people of all ages and exacting significant personal and societal costs []. Amongst newborns, 1–3 of every 1000 babies have hearing loss, with about 60% of cases due to genetic factors, and by the age of 80, 50% of octogenarians will required some form of auditory amplification for meaningful communication [,]. Although a Mendelian inheritance pattern is common for genetic hearing loss, extreme genetic and phenotypic heterogeneity drive an underlying complexity that makes the genetic diagnosis of hearing loss challenging []. Typically, establishing a genetic diagnosis necessitates multidisciplinary expertise, with input from otolaryngologists, human geneticists, genetic counselors, bioinformaticians, biomedical engineers, and auditory scientists []. A genetic diagnosis guides the clinical care of persons with hearing loss. For that reason, the OtoSCOPE sequencing panel was developed at the University of Iowa in 2012 to provide comprehensive genetic testing for hearing loss. Version 9 targets 224 genes associated with non-syndromic and syndromic hearing loss [] and has a diagnostic rate across all patients of 45% []. Variants identified by OtoSCOPE testing are classified following guidelines established by the American College of Medical Genetics and Genomics (ACMG) and Association for Molecular Pathology (AMP) into one of five categories: benign, likely benign, likely pathogenic, pathogenic, and variant of uncertain significance (VUS) [] by combining varying levels of evidence (sequencing data, in vivo and/or in vitro experiments, family history, etc.) into a comprehensive framework. The Deafness Variation Database [] (DVD) is a publicly available database of 10,586,514 variants for OtoSCOPE genes that have been collected from dbSNP [], gnomAD [], 1000 Genomes Project [], ClinVar [] and HGMD []. Of these variants, 668,489 are within exons or impact canonical splice sites, including 381,924 missense variants. These missense variants are classified as benign (0.75%, 2,895), likely benign (17.2%, 65,585), likely pathogenic (0.83%, 3,205), and pathogenic (1.74%, 6,662), with 79.5% (303,577 variants) being a VUS. Due to the intractable challenge of generating experimental evidence to assess the impact of VUSs, here we explore computational methods to identify missense VUSs that exhibit features consistent with pathogenicity. Two popular in silico methods for large dataset variant evaluation are the Combined Annotation Dependent Depletion (CADD) and Rare Exome Variant Ensemble Learner (REVEL) scores. CADD evaluates the deleteriousness of all genetic variants based on a set of 63 genomic features that include conservation and regulatory annotations [,]. REVEL is an ensemble method, assessing the deleteriousness of missense variants by combining outputs of 13 individual tools, several of which use features like those in CADD []. While CADD is a single model applicable to all variant types, REVEL’s ensemble approach is optimized for pathogenicity prediction in missense variants. Importantly, neither CADD nor REVEL model the structural or energetic consequences of variants on protein structure. AlphaMissense is another variant effect predictor based on the deep learning model AlphaFold2 []. It provides a score between 0 and 1 as well as a predicted classification, although it does not fully adhere to ACMG criteria. Ideally, its use of AlphaFold2 structural information should improve interpretation of structural changes due to a missense variation that would be missed by CADD and REVEL. All three provide a numerical score to indicate the relative likelihood that a variant is deleterious without assigning an ACMG variant classification code. Work by Pejaver et al. [] has derived score thresholds for genetic effect predictors using Bayesian statistics. These score thresholds are consistent with varying levels of evidence following ACMG PP3/BP4 criteria where the lowest level of evidence (supporting) reaches a posterior probability of pathogenicity of 10% and the highest level of evidence (very strong), reaches a posterior probability of pathogenicity of 98% []. Using the ClinVar 2019 dataset, Pejaver et al. found that no tool was able to provide a very strong evidence designation. The structure and dynamics of a protein dictates its function and phenotype. When evaluating the impact of a missense variant that causes a downstream amino acid change, it is critical to access the impact on protein folding stability []. Folding free energy difference ([pgen.1012085.e001.webp]ΔΔGFold) quantifies the change in stability of a folded protein state upon mutation. Positive folding free energy differences indicate a destabilizing mutation thereby impacting function. With the introduction of AlphaFold3 [] (an improved protein structure prediction tool), the Bayesian framework introduced by Pejaver et al., and more than 380,000 missense variants in the DVD, we can combine genetic tools with protein folding information to help classify variants of uncertain significance. Previous work investigating the impact of disease-causing missense variants on protein folding found that ~60% to ~83% cause destabilization [–]. We hypothesize that combining genetic tools with protein folding information for the specific phenotype of hearing loss will improve the predictive power of computational tools. Here, we re-evaluate the thresholds proposed by Pejaver et al. in the context of hearing loss to improve the predictive power of CADD, REVEL, and AlphaMissense (provided in the ). We use scores describing the tolerance of a gene to mutation from gnomAD to refine our classification methods by organizing genes into tolerance-based groups with specific priors []. We combine the Bayesian framework for genetic effect predictors (e.g., CADD and REVEL) with predicted folding free energy differences to define a protein folding-informed classification approach with improved predictive power for prioritizing missense VUSs in the DVD as deleterious [,]. The clinical impact of this work is highlighted by the identification of 12 variants identified by OtoSCOPE that have very strong evidence for being pathogenic based on protein folding-informed modeling; two variants are described in detail where the combination of genetic tools and protein folding stability demonstrate how the prediction of variant effect can be missed with genetic tools alone. Results Monomer structures We predicted structures for 258 proteins across 224 genes with AlphaFold3. AlphaFold3 [] is an updated version of the deep learning structure prediction model AlphaFold2 [] that contains a diffusion-based architecture to predict protein monomer structures. Previous work using the AlphaFold2 network predicted protein structures for all DVD genes and their corresponding isoforms referred to as OtoProteinV2 []. AlphaFold3, however, offers improved accuracy over AlphaFold2, motivating updated protein models []. AlphaFold3 performs a local minimization in the Amber fixed-charge force field as part of the prediction [,]. Even with Amber optimization, the structures have atomic van der Waals clashes and high-energy backbone torsions that can be improved using local backbone optimization together with global side chain optimization [,] based on the Atomic Multipole Optimized Energetics for Biomolecular Applications (AMOEBA) polarizable force field [,]. We optimized all 258 protein structures and used MolProbity, a tool that scores protein structures based on a database of well understood conformational features (e.g., equilibrium bond lengths, protein backbone geometry, and side chain rotamers). A MolProbity score of 1.0 implies the expected quality of a structural model from X-ray diffraction at a resolution of 1.0 Å [,], whereas higher scores indicate lower quality models consistent with lower quality diffraction data. With optimization, the average MolProbity score across all DVD proteins decreased from 1.85 to 1.05 after optimization with the AMOEBA force field (referred to as OtoProteinV3). The MolProbity scores for OtoProteinV3 are summarized in . Most structures have a MolProbity score between 0.5 and 1.0, as compared to the majority AlpaFold3 structures that fall between 2.0 and 2.5. The van der Waals clash score was reduced from 5.29 clashes per 1000 atoms to 0.15 clashes per 1000 atoms for OtoProteinV3. A comparison of OtoProteinV2 and OtoProteinV3 alpha tectorin structure (TECTA) is shown in together with MolProbity statistics, reflecting the reported differences in accuracy for AlphaFold2 and AlphaFold3 []. shows the significant (4-fold) increase in DVD missense variants investigated from OtoProteinV2 (2023) and OtoProteinV3 (2026). Tolerance in deafness-associated genes The tolerance score, missense o/e, is a ratio of the observed-to-expected number of variants in a particular gene. Tolerance-based bins of DVD genes as compared to genome-wide tolerance were roughly comparable, with median missense o/e scores being 0.89 and 0.88, respectively. However, the o/e score distribution for the DVD is slightly negatively skewed towards tolerant genes (skewness of DVD -0.36 vs of genome -0.31) shown in Fig A in . We highlight three genes across the range of o/e scores in the DVD: ACTG1, COCH, and USH1C. ACTG1 () is the second most intolerant gene in the DVD with an o/e of 0.43 (the most tolerant being WFS1). Gamma actin is highly conserved across species, and mutations in actin proteins are often lethal during development (selected against) due to their role in the structural framework of cells []. Consequently, the structure of gamma actin is well-defined, which leads to more regions of high AlphaFold3 confidence (). COCH encodes for the protein cochlin and has an o/e of 0.86. While cochlin is essential for hearing function, mutations are not as strongly selected against during development as those in gamma actin, and it has both well-folded regions and disordered regions, which contribute to the average tolerance (). USH1C () encodes for harmonin and pathogenic variants can be associated with autosomal recessive non-syndromic hearing loss or autosomal recessive Usher Syndrome Type I []. USH1C has an o/e of 1.0, and its recessive inheritance pattern is consistent with its above average o/e. The structure for USH1C has the lowest level of predicted AlphaFold3 confidence. Tolerance-based priors vs a DVD prior Pejaver reported the prior probability of pathogenicity as 4.41% (prior odds of 0.046) based on the ClinVar 2019 dataset, which falls below the ClinGen Structural Variant Interpretation working group estimation of a prior probability of pathogenicity of ~10% in human variants []. The prior probability of pathogenicity for the DVD based on currently classified variants (32,895 benign, 65,585 likely benign, 3,205 likely pathogenic, and 6,662 pathogenic) is 12.6% (prior odds of 0.144). The higher prior is consistent with selection of OtoSCOPE genes with known disease associations and their relative (in)tolerance to missense variants (as discussed further in the next section on gene tolerance). For each gene in the DVD, we collected a missense tolerance score from gnomAD to create groups (bins) of genes whose missense variants were hypothesized to have similar prior probabilities of being pathogenic []. The observed variant count is the number of unique single nucleotide variants in the transcript with minor allele frequency of less than 0.1% and a median depth in exome samples greater than 30. The expected variant count uses a model that corrects for local sequence context as described in Karczewski et al []. The final ratio provides a continuous measure of gene tolerance to mutation where higher scores are more tolerant and lower scores are less tolerant (or more intolerant). We hypothesized that genes with a lower o/e (more intolerant) score would have a higher prior probability of pathogenicity []. For each bin, shows the o/e score range, the number of genes and variants (total and labeled), and the tolerance-based prior probability of pathogenicity/benignity. The bin for average tolerance contains most of the DVD genes with its center near the average missense o/e (± 0.1) for both the entire genome and subset of genes in the DVD. The tolerant and intolerant bins contain genes on the extreme ends of the DVD o/e range. The prior in the intolerant bin is much higher than that for the DVD prior, while that for the tolerant bin is lower. This supports the hypothesis that missense variants in genes highly intolerant to mutation are more likely to be pathogenic than are missense variants in genes tolerant to mutation. Genetic tool analysis We determine the CADD, REVEL, and AlphaMissense score threshold for the posterior to reach supporting (10%), moderate (20%), strong (60%), and very strong (98%) evidence []. However, due to limitations of AlphaMissense (e.g., unreleased training weights, poor performance on DVD labeled variants, and no assessment for ~25% of the DVD labeled variants), we have elected to include this analysis in the Supplementary Information. It is important to note variants that were included in the training set for CADD and REVEL were removed from the calculation of thresholds in the ClinVar 2019 study. For this study, we included all variants in the DVD to fully understand each tool’s predictive performance on this well-curated dataset. We used the calculated disease-specific score thresholds and compared them to those from ClinVar 2019 for their ability to predict labeled variants in the DVD. We calculated the posterior associated with REVEL or CADD and determined the score threshold for each evidence level (supporting, moderate, strong, or very strong). The threshold curve for the genetic tools and the posterior along the range of scores using a DVD likelihood and prior for both CADD and REVEL are shown in . and show the REVEL posterior curve for benign and pathogenic variants, respectively. and show the CADD posterior curve for benign and pathogenic variants, respectively. We derived DVD and tolerance-based thresholds as displayed in . Using strong levels of evidence, we calculated accuracy, sensitivity (ability to predict pathogenicity), and specificity (ability to predict benignity) for labeled DVD variants as summarized in . This analysis was repeated for ClinVar 2019 score thresholds and a DVD derived likelihood with a DVD prior or tolerance-based priors. Labeled benign and pathogenic variants that fall between the benign and pathogenic thresholds into what would be considered an uncertain range were counted as incorrectly labeled variants. Because the posterior for CADD did not reach a strong (posterior probability of 60%) evidence level for benign prediction in the analysis or in intolerant genes, a value of 0 was used in place of a lower threshold to calculate accuracy and specificity. Overall, accuracy was highest when using the DVD likelihood and tolerance-based priors to define score thresholds with strong evidence from CADD at 54.9%, with similar accuracy from REVEL at 53.1%. With a DVD prior, REVEL (52.1% accurate) outperforms CADD (9.5% accurate) due to the low specificity caused by the lack of strong evidence for benign prediction from CADD. The improved accuracy, sensitivity, and specificity from our method affirms the benefits of recalculating score thresholds using disease-specific datasets and binning genes by tolerance. While CADD accuracy is slightly higher than REVEL in the case of the tolerance-based prior, CADD cannot provide Moderate, Strong, or Very Strong evidence for benignity in intolerant genes. Therefore, when using genetic tools alone, tolerance-based thresholds for REVEL have more consistent performance for strong evidence. The protein folding-informed analysis with CADD is included only in the supplemental data (Table B, Figs C and D in the ). Protein folding-informed analysis We predicted folding free energy differences ([pgen.1012085.e002.webp]ΔΔGFold) for all missense variants in the DVD with DDGun3D to quantify their impact on protein stability. DDGun3D is a high throughput predictor for [pgen.1012085.e003.webp]ΔΔGFold that uses a linear regression of biochemical and structural features to determine the thermodynamic effects of missense variants []. DDGun3D has a reported root-mean-square-error (RMSE) of ~1.5 kcal/mol. The advantages of DDGun3D include its ability to be applied to a dataset of our scale and protein size range, low compute cost, and the anti-symmetry of the method. Perfect anti-symmetry constrains the free energy change of mutating from residue A → B to be equal and opposite in sign to the free energy change of mutating from B → A. A [pgen.1012085.e004.webp]ΔΔGFold greater than 1.0 kcal/mol for a protein leads to a five-fold relative increase in the ratio of unfolded to folded protein, making it a reasonable starting point for identifying variants that cause protein destabilization. For example, there are 54,752 variants in the DVD that have a [pgen.1012085.e005.webp]ΔΔGFold above 1.0 kcal/mol, including 2,464 likely benign/benign, 2,884 likely pathogenic/pathogenic, and 49,404 VUSs. Ultimately, our variant classification model does not depend on any arbitrary [pgen.1012085.e006.webp]ΔΔGFold cutoff. We combined DVD derived likelihood ratios for [pgen.1012085.e007.webp]ΔΔGFold and REVEL to determine the probability of a variant being pathogenic using both tolerance-based priors and a DVD prior. A posterior of 98% or greater is considered very strong evidence for pathogenicity []. Using MATLAB contour and curve fitting functions, polynomial models were fit to define the combined scores to reach the posterior probability needed for very strong evidence (see provided code S1_threshold_postprocessing.mlx and for specifics). The application of these evidence thresholds for a DVD prior is shown in with containing only labeled variants and including VUSs (with VUSs above 98% probability colored red). The number of labeled variants and prioritized VUSs per tolerance-based bin for the ClinVar 2019 prior, the DVD prior, and the tolerance-based prior are summarized for REVEL x [pgen.1012085.e012.webp]ΔΔGFold in . Results and figs for CADD x [pgen.1012085.e013.webp]ΔΔGFold are outlined in the S1_Appendix (Table B in the ). Assuming all variant labels in the DVD are correct (see Discussion for sources of possible label bias), we calculate a false positive rate as the number of variants with a posterior probability of pathogenicity >98% that are labeled likely benign/benign over the total number of variants labeled likely benign/benign. Application of the ClinVar 2019 and DVD priors have false positive rates of 0.04% and 0.14% for REVEL x [pgen.1012085.e019.webp]ΔΔGFold, respectively. The tolerance-based DVD prior method also has a false positive rate of 0.14% (1 additional variant). Although the difference in false positive rate is minimal between all the priors, the tolerance-based prior correctly labels 2,842 or 115 additional pathogenic variants compared to the ClinVar 2019 and the DVD derived priors, respectively. It is difficult to calculate sensitivity (the ability of a model to predict pathogenic variants) for the protein folding-informed thresholds since [pgen.1012085.e020.webp]ΔΔGFold only tests one hypothesis for pathogenicity, and variants can be pathogenic due to other mechanisms (i.e., protein binding disruption). However, if we make the very liberal assumption that all labeled pathogenic variants not selected by our protein folding-informed thresholds are incorrectly labeled benign, the sensitivity for the ClinVar 2019, DVD, and tolerance-based protein folding-informed score thresholds are 23%, 50.7%, and 51.8%, respectively. The sensitivity for the very strong REVEL threshold for the DVD is 32.0%, and the tolerance-based sensitivity is 34.9%. Our ability to predict pathogenic variants to a very strong evidence level is increased with protein folding-informed, tolerance-based priors with an increase of only 0.1% in benign variants being mislabeled when compared to the ClinVar 2019 prior and no increase compared to the DVD prior. The tolerance-based protein folding-informed posterior assigns PP3 very strong evidence of pathogenicity to 28,886 VUS. The DVD protein folding-informed and ClinVar 2019 posteriors assign PP3 very strong evidence to 26,947 or 12,095 VUSs, respectively. The tolerance-based protein folding-informed posterior captures more VUSs than the DVD or ClinVar 2019 protein folding-informed methods with better sensitivity and a minimal increase in false positive rate. shows the score thresholds for REVEL x [pgen.1012085.e021.webp]ΔΔGFold in the three bins. In , and are the variants in the intolerant bin, and are those in the average tolerance range, and and are those in the tolerant bin colored by classification with VUS at or above a posterior of 98% colored red - see ). We evaluated the performance of our protein-folding informed thresholds against experimental data for 4085 KCNQ4 variants [], which is included in the and Table A in . The study used a whole-cell patch clamp technique to determine the current across the membrane for wildtype KCNQ4 and all 4085 variants []. The resulting scores indicated loss, neutral, or gain of function. The tolerance-based protein-biophysics informed thresholds outperformed both AlphaMissense and REVEL in their ability to select pathogenic variants to a posterior probability of pathogenicity of 98% when compared to either the experimental data or DVD labels for KCNQ4 variants. We find that REVEL x [pgen.1012085.e026.webp]ΔΔGFold best concords with variants reclassified from VUS to pathogenic when the experimental evidence (PS3 supporting) is included in ACMG criteria. Due to the significant increase in sensitivity, increase in VUSs with PP3 very strong evidence of pathogenicity and low false positive rate, we recommend using the REVEL x [pgen.1012085.e027.webp]ΔΔGFold tolerance-based posterior for clinical use. Protein misfolding in deafness We investigated the prevalence of protein misfolding in deafness associated genes by evaluating [pgen.1012085.e028.webp]ΔΔGFold values for labeled DVD variants. First, we determined the number (left axis, solid lines) and fraction (right axis, dashed lines) of labeled variants that are LP/P and LB/B for a [pgen.1012085.e029.webp]ΔΔGFold range from 0 to 5 kcal/mol (). At 0 kcal/mol, the fraction of labeled variants that are LP/P (0.126) and LB/B (0.874) is equal to their overall fraction in the DVD. As the [pgen.1012085.e030.webp]ΔΔGFold increases, the number of both LB/B and LP/P variants decreases. However, the fraction of LP/P variants increases while the fraction of LB/B decreases. The number of LB/B and LP/P for at least 1 kcal/mol is approximately equal with a fraction of 0.5. For labeled variants with a [pgen.1012085.e031.webp]ΔΔGFold of at least 2 kcal/mol, 80% are LP/P. The fraction of labeled variants asymptotes around 4 kcal/mol with 0.04 LB/B and 0.96 LP/P. We then calculated the number (left axis, solid lines) and fraction (right axis, dashed lines) of VUSs values that reach a posterior of 60% (“strong”) and 98% (“very strong”) ≥[pgen.1012085.e032.webp]ΔΔGFold values in the range from 0 to 5 kcal/mol (). As [pgen.1012085.e033.webp]ΔΔGFold increases, the number of VUSs decreases while the fraction of VUSs that reach “strong” and “very strong” thresholds increases. 80% of VUSs with a [pgen.1012085.e034.webp]ΔΔGFold of at least 2 kcal/mol reach “strong” evidence and 60% reach “very strong” evidence which concords the trend for labeled LP/P variants (). We evaluated the percentage of variants in each label whose [pgen.1012085.e035.webp]ΔΔGFold was greater than or equal to a threshold of 1 up to 3 kcal/mol in steps of 0.5 kcal/mol (). Notably, ~ 5% of all labeled LB/B variants have a [pgen.1012085.e036.webp]ΔΔGFold ≥1.0 kcal/mol, and as expected, this percentage decreases as [pgen.1012085.e037.webp]ΔΔGFold increases. For LP/P variants, ~ 30% have a [pgen.1012085.e038.webp]ΔΔGFold of ≥1.0 kcal/mol, reducing to ~10% by ≥3.0 kcal/mol. 20% of all VUSs have a [pgen.1012085.e039.webp]ΔΔGFold of at least 1 kcal/mol while 65% of VUSs above a posterior of 98% reach at least the same value, indicating the VUSs selected by REVEL x [pgen.1012085.e040.webp]ΔΔGFold are enriched for protein misfolding. We ordered genes by enrichment for protein misfolding as the mechanism for pathogenic variants (). We filtered for genes with at least 10 labeled LP/P variants and calculated the percentage of LP/P variants that have a [pgen.1012085.e041.webp]ΔΔGFold of at least 1 kcal/mol. Of the 224 genes, 23 have at least 50% of their LP/P with a [pgen.1012085.e042.webp]ΔΔGFold ≥ 1 kcal/mol. All 23 genes are included in in ascending order of percentage of LP/P with [pgen.1012085.e043.webp]ΔΔGFold ≥1 kcal/mol. Genes with less than 50% are included in the (Fig E in the ), and genes with less than 10 LP/P variants are excluded. As novel variants are discovered in these genes, this ordering can guide clinicians toward protein misfolding as a possible mechanism for hearing loss. Proband variant analysis VUSs with a 98% posterior probability of being pathogenic (based on the REVEL x [pgen.1012085.e048.webp]ΔΔGFold tolerance-based posterior) were filtered against variants identified by OtoSCOPE v9 sequencing in probands with hearing loss to identify 12 variants. All 12 variants (outlined in the Table C in the ) received an upgraded genetic diagnosis to likely pathogenic/pathogenic with the inclusion of protein folding data, and 11 of the 12 variants have a clear structural mechanism to cause protein misfolding. Two cases detailed below illustrate the impact of these variants on protein structure and misfolding (). MYO6 p.(Leu1086Pro) Variants in MYO6 are associated with autosomal dominant and recessive non-syndromic hearing loss, although MYO6-related hearing loss is likely semidominant [–]. The encoded protein, myosin VI (MyoVI), is an unconventional actin-based motor important for endocytosis in cochlear hair cells []. A sibling pair with congenital mild-to-moderate bilateral sensorineural hearing loss were heterozygous for the MYO6 variant, p.(Leu1086Pro). This variant has a REVEL score of 0.847 and causes a destabilizing [pgen.1012085.e050.webp]ΔΔGFold of 4.1 kcal/mol. Proline is unique among amino acids due to the covalent bond its side chain forms with the protein backbone, which makes it relatively rigid and prevents formation of secondary structure due to the inability of the backbone nitrogen atom to donate a hydrogen bond. As residue Leu1086 is in the middle of an alpha helix (), the helix is broken by the introduction of a proline, significantly impacting the structure and stability of the protein OTOF p.(Met966Arg) OTOF encodes for otoferlin, a calcium binding transmembrane protein in synaptic vesicles important for signal transmission in inner hair cells [,]. Mutations in this gene lead to autosomal recessive severe-to-profound bilateral hearing loss. Accurate classification of variant effect is of prime importance as there are now several gene therapy options for persons with OTOF-related hearing loss [–]. OtoSCOPE identified the heterozygous OTOF variant, p.(Met966Arg) in a proband with early childhood onset profound hearing loss and a family history of hearing loss in a sibling. Importantly, this proband was heterozygous for the known pathogenic OTOF variant, p.(Ile1573Thr). OTOF has a missense o/e score of 0.94, falling within average tolerance. The variant p.(Met966Arg) has a REVEL score of 0.785 and a [pgen.1012085.e051.webp]ΔΔGFold of 2.1 kcal/mol indicating a destabilizing effect on the protein folding. Methionine is a hydrophobic residue, and the residue Met966 (orange in ) is surrounded by other hydrophobic residues (cyan in ). The hydrophobic effect describes the thermodynamically unfavorable interaction of hydrophobic protein residues with water, which drives protein folding by the favorable burial of hydrophobic residues into a hydrophobic protein core as seen in []. Arginine is positively charged and hydrophilic, which disrupts the stabilizing hydrophobic core consistent with the high [pgen.1012085.e052.webp]ΔΔGFold. Discussion Although genome sequencing offers the promise of personalized precision medicine, variant classification remains a challenge. In the domain of hearing loss, for example, there are 381,924 missense variants in the 224 genes represented on the OtoSCOPE v9 comprehensive genetic testing panel – 303,577 of these variants (79.5%) are VUSs. Generating experimental evidence to phenotype these variants is implausible, highlighting the importance of optimizing variant analysis using in silico tools. Here, we construct a Bayesian framework for evaluating VUSs to define deafness-specific thresholds augmented with protein folding modeling to identify VUSs in the DVD that have a 98% posterior probability of being pathogenic []. We refined the structure of 258 unique protein isoforms for 224 genes in the DVD predicted by AlphaFold3 with the AMOEBA force field to ensure local structural minimization and global side chain optimization. Using these optimized structures (OtoProteinV3), we predicted folding free energy differences ([pgen.1012085.e053.webp]ΔΔGFold) for all missense variants in the DVD. We then evaluated REVEL, CADD, and [pgen.1012085.e054.webp]ΔΔGFold by applying a Bayesian framework with a tolerance-based prior specific for hearing loss genes []. We found that a prior probability of pathogenicity of 12.6% (specific for hearing loss) increased accuracy, sensitivity, and specificity for REVEL scores when compared to ClinVar 2019 derived thresholds (ClinVar 2019 CADD thresholds could not provide strong or very strong evidence of pathogenicity; [pgen.1012085.e055.webp]ΔΔGFold was not predictive on its own with a 4.3% accuracy). When using tolerance-based thresholds, accuracy and specificity were improved for both CADD and REVEL, although sensitivity decreased when compared to using the DVD prior. The choice of prior probability has a profound effect on deleteriousness predictions: a low prior probability of 4.41% (prior odds 0.046) as used by Pejaver et al. for predictions on the ClinVar 2019 dataset under calls pathogenic variants in the context of the DVD. The ClinVar 2019 prior is also significantly lower than previous estimates including those defined by other work on deafness phenotypes []. Use of a DVD prior of 12.6% (prior odds 0.144) is more consistent with labeled DVD variants based on improved accuracy, sensitivity and specificity. Our tolerance-based bins suggest that highly constrained genes such as ACTG1 require an even larger prior (25.7%) to explain the observed data. We next calculated likelihoods for CADD and REVEL with [pgen.1012085.e056.webp]ΔΔGFold to combine biophysical evidence with genetic tools. Using a tolerance-based prior and REVEL x [pgen.1012085.e057.webp]ΔΔGFold likelihood resulted in score thresholds with a false positive rate of 0.14%, a sensitivity of 51.8% and identified 28,886 of 303,577 (9.5%) variants with a 98% posterior probability of being pathogenic. Approximately 50% of labeled variants with a [pgen.1012085.e058.webp]ΔΔGFold of at least 1 kcal/mol are likely pathogenic/pathogenic; this increases to 80% by at least 2 kcal/mol. A total of 18,706 (60%) of the VUS identified have a protein folding free energy difference greater than 1 kcal/mol and also have a posterior probability of pathogenicity of at least 98%, which indicates around two thirds of the VUS captured by the protein folding-informed thresholds are implicated due to their downstream effects on protein structure and stability. This is consistent with previous studies that conclude approximately ~60% to ~83% of disease-causing missense variants result in loss of protein stability [–]. When looking at already labeled likely pathogenic/ pathogenic variants, 2,884 of 9,865 variants have a predicted protein folding free energy difference above 1.0 kcal/mol. Practically, adopting 10% as a conservative minimum as suggested by the ClinGen Structural Variant Interpretation working group, approximately 38,000 variants are expected to be LP/P (currently labeled LP/P values together with VUS above 98% posterior probability), i.e., the roughly 28,000 upgraded VUS together with 9,865 already labeled LP/P are consistent with expectations []. With a false positive rate of 0.1–0.2%, we would expect fewer than 100 erroneous promotions—arguably an acceptable tradeoff for greater sensitivity. The DVD is regularly updated, experiencing a nearly 4-fold increase in missense variants from 2023 to 2026. Calculating the prior probability rate of pathogenicity for the previous instance of the DVD (v8) yields nearly 30%, which is much higher than the value for the current version of 12.6%. As more variants are labeled by the field, the prior rate will converge toward the true value, including additional support for tolerance-based or gene specific models. The classifications in the DVD will also continue to improve as the ACMG/AMP criteria evolve and the internal evaluation of variants advances. The clinical impact of these data is significant. In investigating variants in probands sequenced with OtoSCOPE, we used the protein folding-informed thresholds to upgrade 12 VUS to likely pathogenic/pathogenic. As an example, we highlight two variants identified in persons with hearing loss – MYO6 p.(Leu1086Pro), and OTOF p.(Met966Arg) – and show how close structural analysis of protein folding contributes to our understanding of variants. MYO6 p.(Leu1086Pro) causes a disruption to fold of the alpha helix due to proline rigidity; OTOF p.(Met966Arg) disrupts the hydrophobic pocket secondary to arginine positive charge, thereby affecting the overall fold. There are important limitations to this work: 1) the possibility of overestimating the performance due to circularity in variant classification data, 2) the limited ability of the methods to predict benign variants, and 3) limitations inherent to each computational tool (CADD, REVEL, AlphaFold3 and DDGun3D). First, some DVD variants were included in the original training sets used for CADD and REVEL, which could artificially inflate performance. Many of the labels in the DVD are collected from ClinVar which can artificially inflate the prior probabilities due to its bias toward pathogenic variants. Additionally, the variant classification of likely benign is assigned to some variants in the DVD using a custom bioinformatics pipeline. This pipeline utilizes computational tools to classify variants and introduces artificial inflation of the benign estimates in the DVD. We also introduce some overestimation when combining the likelihoods of CADD or REVEL with [pgen.1012085.e059.webp]ΔΔGFold. Bayesian statistics stipulate that to combine likelihoods by multiplication the variables must be independent. CADD or REVEL and [pgen.1012085.e060.webp]ΔΔGFold are not entirely independent with correlation coefficients 0.08 and 0.16, respectively. However, the relatively small correlation coefficients together with the low false positive rates support the conclusion that the correlation is minor and does not hinder our protein folding-informed models. Second, CADD does not reach strong evidence for pathogenic or benign variant classification. When coupled with the overall lower specificity observed for all methods of predicting labeled benign variants, these recalibrated thresholds lack sufficient power to reliably detect benign variants. Third, the training set for AlphaFold3 is from experimental structures available in the Protein Databank. For approximately 60% of the proteins in this work, no experimental or homology-related structural information was available for AlphaFold2 or 3 [,,]. There remain inaccuracies for a subset of the predicted structural models that could impact the accuracy of predicted [pgen.1012085.e061.webp]ΔΔGFold values. We expect more accurate protein structure predictors (Boltz-2 [,], OpenFold [,]) that will overcome the limitations of AlphaFold3. While DDGun3D was used in this study due to its combination of accuracy and efficiency, alternative methods for the prediction of protein folding free energy differences are emerging []. Ideally, we would calculate [pgen.1012085.e062.webp]ΔΔGFold with rigorous molecular dynamics simulations, however, this option is currently not feasible at scale for more than 380,000 missense variants. We expect future machine learning and/or physics-based methods to emerge that will approach the accuracy of in vitro folding experiments. As more accurate prediction methods for both protein structure and [pgen.1012085.e063.webp]ΔΔGFold become available, we will continue to investigate all variants in the DVD []. Molecular simulation methods for quantifying [pgen.1012085.e064.webp]ΔΔGFold will become faster and allow for thorough assessment of biophysical phenotypes [,]. Our current method for computing changes in relative protein folding stability does not include the full context of DVD proteins because many (at least 200) function by binding to another protein, a ligand, or RNA/DNA. Structure prediction methods will continue to evolve and allow for the creation of protein-protein complexes, protein-ligand interactions, and the inclusion of RNA/DNA. Each of these structures has different quantifiable evidence (e.g., like ΔΔGFold for misfolding) that can be used to evaluate the mechanism of disease for missense variants [,]. It is realistic to expect future work to include binding free energy differences for protein-protein, protein-ligand, and protein-DNA interactions. In the future, we believe in silico tools will be able to provide adequate evidence for variant classification. It is important to note that ACMG/AMP guidelines stipulate that no in silico tool provides adequate evidence to classify a single variant as benign or pathogenic. Therefore, computational evidence must be integrated with other lines of evidence, such as functional or population data, before a final classification can be made []. For that reason, the DVD will include the posterior probability for all missense variants, which contributes PP3 evidence of various strengths (i.e., very strong when the posterior is greater than 98%) for clinical interpretation. This work also indicates there is additional evidence in mutation tolerance scores from gnomAD that can inform our investigation of gene-specific variants in the DVD. The nature of this impact is indicated by the increased prior odds for intolerant genes and the reduced prior odds for tolerant genes. ACMG PP2 criteria state that missense variants have additional evidence for pathogenicity if they occur within genes where few missense variants are benign and missense variation is a common mechanism of disease. Our intolerant prior captures and quantifies this information to improve our prediction of pathogenic variants. While this could lead to over-counting evidence if both PP2 and our evidence is used when evaluating variants, our intolerant priors also allow for additional information in genes where there are too few labeled variants to apply PP2. Tolerance information will aid clinicians in evaluating the impact of novel missense variants as the prior probability of pathogenicity varies greatly (e.g., from more than 25% in intolerant genes down to below 10%) []. In conclusion, deafness-specific Bayesian models improve the accuracy of REVEL and CADD for missense variant classification in hearing loss by increasing sensitivity for pathogenicity prediction. While limitations remain for the classification of benign variants, these results underscore the importance of disease-specific thresholds. Combining genetic-based predictors of variant effects that rely heavily on evolutionary conservation with biophysical evidence allows us to better prioritize VUSs as pathogenic. Use of predicted protein folding free energy differences directly tests a mechanistic hypothesis for the underlying biophysical impact of each missense variant, unlike CADD or REVEL that fail to provide a phenotypic rationale for their score. We define a Bayesian framework using a tolerance-based prior and likelihood that combines REVEL with [pgen.1012085.e065.webp]ΔΔGFold to assign over 28,000 VUS with PP3 very strong evidence (i.e., 98% posterior probability in favor of pathogenicity). Finally, while the methods described here are applied to hearing loss, the pipeline for variant evaluation could be applied in other disease contexts. Materials and methods Dataset Genes and variants associated with hearing loss or deafness have been compiled in the publicly available Deafness Variation Database (DVD) (https://deafnessvariationdatabase.org)(7). Variants are collected from dbSNP [], gnomAD [], the 1000 Genomes Project [], ClinVar [] and HGMD []. In total, the DVD contains more than 10.5 million variants. Here, we focus on 381,924 missense variants across 216 DVD genes that cause a downstream single amino acid change (the remaining 8 genes do not have missense variants with MT referring to mitochondrial genes: MT-RNR1, MT-TH, MT-TI, MT-TK, MT-TL1, MT-TS1, MT-TS2, MIR96). AlphaFold3 We model all isoforms in the DVD with the AlphaFold3 webserver, with the exception of those larger than AlphaFold3’s 5000 amino acid limit. Three genes (ADGRV1, KMT2D, USH2A) in the DVD encode extremely large proteins (above the 5,000 amino acid limit of AlphaFold3). For these proteins, we transferred the OtoProteinV2 model into this dataset but improved on the original optimization. All DVD genes and their relevant isoforms have full-length protein models. Optimization of monomer structures Initial local minimization was performed with the L-BFGS optimization algorithm to a convergence criterion of 0.8 kcal/mol. Protein models then underwent global side chain optimization before a final local minimization was performed to a convergence criteria of 0.1 kcal/mol. We evaluated the structures before and after AMOEBA optimization in Force Field X [] with MolProbity. The final optimized protein structures are available on GitHub (https://github.com/SchniedersLab/OtoProtein3) and through the DVD website. Predicting ΔΔGFold We used a regression based [pgen.1012085.e066.webp]ΔΔGFold predictor called DDGun3D(40) to evaluate missense variants in protein structures. DDGun3D was used in a previous study on the DVD [] and applied here on the updated 381,924 missense variants in the current DVD instance. Bayesian analysis To determine score thresholds consistent with ACMG/AMP evidence levels, we used an established Bayesian framework that focused on evaluating separate tools for their ability to predict variant pathogenicity []. ClinVar 2019 derived score thresholds for REVEL and CADD are reported with varying levels of evidence (supporting, moderate, strong, and very strong) to indicate that a variant is pathogenic or benign. To determine disease-specific score thresholds, we applied this framework to a dataset of well-curated variants from the DVD. We next calculated a pathogenic and benign likelihood ratio for each genetic tool (CADD, REVEL, and [pgen.1012085.e067.webp]ΔΔGFold). [pgen.1012085.e068.webp]ΔΔGFold has limited ability to predict classifications on its own, however, we include [pgen.1012085.e069.webp]ΔΔGFold applied in isolation in the . First, the prior odds of pathogenicity for the DVD were calculated. where [pgen.1012085.e071.webp]v is a variant. We then calculate the likelihood ratio for a given score using . where [pgen.1012085.e073.webp]s is the score (CADD, REVEL, or [pgen.1012085.e074.webp]ΔΔGFold). The likelihood ratio for each metric is calculated for 10 discrete bins that span the scoring range of each metric and ensures there are enough classified variants within each bin to calculate a reliable likelihood. The likelihood ratios for each tool are in Table D in the . The likelihood for CADD or REVEL was combined with the likelihood for [pgen.1012085.e075.webp]ΔΔGFold to define protein folding-informed (PFI) likelihoods where[pgen.1012085.e076.webp] lrPFI=lrCADD/REVEL*lrΔΔGFold. is applied to [pgen.1012085.e077.webp]lrCADD, [pgen.1012085.e078.webp]lrREVEL, and [pgen.1012085.e079.webp]lrPFI. where [pgen.1012085.e081.webp]P(v is pathogenic | s) is the posterior probability a variant is pathogenic given some score. Posterior values of 98% correspond to PP3 very strong evidence for pathogenicity []. To calculate the protein folding-informed score thresholds for REVEL in combination with [pgen.1012085.e082.webp]ΔΔGFold, we calculated the posterior for all combinations of REVEL and [pgen.1012085.e083.webp]ΔΔGFold likelihoods (i.e., for 10 x 10 = 100 bins). Variants from REVEL x [pgen.1012085.e084.webp]ΔΔGFold bins above the 98% posterior threshold (very strong evidence) were added to a prioritized list for the protein folding-informed score threshold. We performed the same analysis with CADD and [pgen.1012085.e085.webp]ΔΔGFold with results outlined in the (Table B, Figs C and D in ). Checking for independence between tools We performed a regression analysis on DDGun3D values vs CADD/REVEL to determine if the variables are independent, which is necessary to justify creation of a combined likelihood as the product of each tool’s individual likelihood. The R-squared coefficient between DDGun3D and CADD/REVEL was just 0.08 and 0.16, respectively. For two scores to be completely independent, their correlation would be 0. Although the R-squared coefficient is slightly above 0, the low false positive rate of the combined likelihoods supports the approximation of treating CADD/REVEL and DDGun3D as independent in defining their joint likelihood ratio. Prior probability from genes grouped by tolerance We used missense o/e scores to bin genes in the DVD based on their tolerance and calculated priors based on the genes/variants in each bin. Genes in the DVD had missense o/e values ranging from 0.1 (intolerant) to 1.45 (tolerant) and were divided into three separate bins. The missense o/e and genes/variants in each bin are shown in along with their prior probability of pathogenicity and benignity. The bins are not evenly distributed by o/e score; the first and last bins are significantly larger than the middle bin to capture the less common extremes of the o/e score. The average tolerance (middle) bin is defined by using the mean missense o/e as an approximate center and includes the central distribution of the DVD. The two outer bins (tolerant and intolerant) contain the extreme ends of the distribution. There are 16 genes that are X-linked or mitochondrial (X-linked/MT) and do not have an o/e score per gnomAD. The X-linked/MT genes were assigned the overall DVD prior. We compared accuracy, sensitivity, and specificity when using tolerance-based priors to using a single DVD prior with the DVD derived likelihoods to establish score thresholds and the score thresholds derived from the ClinVar 2019 dataset to select a genetic tool and protein folding-informed Bayesian framework for evaluating variants. Proband variant analysis As a proof of concept, to understand the translational impact of our protein folding-informed Bayesian framework, we identified probands sequenced on OtoSCOPE v9 panel without a definitive genetic diagnosis but with at least one VUS within the group of more than 28,000 VUSs prioritized by REVEL x [pgen.1012085.e086.webp]ΔΔGFold to be pathogenic. Matched variants received additional scrutiny (a rigorous structural and genetic analysis) to determine their likely effect on protein structure/function.