MissenseHMM: state-based annotations for missense variants through joint modeling of pathogenicity scores

We present MoCHI, a tool to fit interpretable models using deep mutational scanning data. MoCHI infers free energy changes, as well as interaction terms (energetic couplings) for specified biophysical models, including from multimodal phenotypic data. When a user-specified model is unavailable, global nonlinearities (epistasis) can be estimated from the data. MoCHI also leverages ensemble, background-averaged epistasis to learn sparse models that can incorporate higher-order epistatic terms. MoCHI is freely available as a Python package (https://github.com/lehner-lab/MoCHI) relying on the PyTorch machine learning framework and allows biophysical measurements at scale, including the construction of allosteric maps of proteins.

ObjectiveDetermining the pathogenicity of missense variants in clinical genetic tests for individuals with epilepsy is crucial for guiding personalized treatment. However, achieving a definitive pathogenic classification remains challenging, with most missense variants still classified as variants of uncertain significance (VUS) and with the availability of many computational tools which may provide conflicting predictions. Here, we aim to evaluate the performance of state‐of‐the‐art computational tools in pathogenicity prediction of missense variants in epilepsy‐associated genes. This will assist in selecting the most appropriate tool and critically assess their use in clinical setting.MethodsWe assessed the performance of nine in silico pathogenicity prediction tools for missense variants in epilepsy‐associated genes on three carefully curated data sets. The first two data sets comprise missense variants in epilepsy associated genes that have been uploaded to ClinVar in the last year and were, therefore, not part of the training set of any of the nine considered tools. These two data sets are based on two different lists of epilepsy‐associated genes and comprise ~700 and ~ 250 missense variants, respectively. The third data set includes ~400 missense variants within epilepsy‐associated genes for which the functional effects have been determined experimentally and are therefore used here to infer pathogenicity. These three data sets represent the best available approximation to blind and independent test sets.ResultsAmong the nine assessed tools, AlphaMissense (area under the curve [AUC]: .93, .88, and .95) and REVEL (AUC: .93, .88, and .93) showed the best classification performance, also outperforming other tools in the number of classified variants.SignificanceWe show which recently developed prediction tools achieve higher performance in epilepsy‐associated genes and should be integrated, therefore, into the American College of Medical Genetics and Genomics/Association of Molecular Pathology (AGMC/AMP) variant classification process. Periodic reevaluation of genetic test results with newly developed or updated tools should be incorporated into standard clinical practice to improve diagnostic yield and better inform precision medicine.

Deep mutational scanning (DMS) experiments provide a powerful method to measure the functional effects of genetic mutations at massive scales. However, the data generated from these experiments can be difficult to analyze, with significant variation between experimental replicates. To overcome this challenge, we developed popDMS, a computational method based on population genetics theory, to infer the functional effects of mutations from DMS data. Through extensive tests, we found that the functional effects of single mutations and epistasis inferred by popDMS are highly consistent across replicates, comparing favorably with existing methods. Our approach is flexible and can be widely applied to DMS data that includes multiple time points, multiple replicates, and different experimental conditions. popDMS is implemented in Python and Julia, and is freely available on GitHub at https://github.com/bartonlab/popDMS.

To assess the joint effects of SARS-CoV-2 viral traits (ACE2 binding, immune escape, and cell entry) and socio-demographic context on viral lineage fitness across countries and to develop predictive models of lineage fitness using these factors. A total of 5,092,686 SARS-CoV-2 spike protein sequences from nine representative countries, spanning the least developed, developing, and advanced economies, were collected. Viral traits for individual sequences were calculated using deep mutational scanning data. Lineage-specific growth rates (i.e., viral fitness) were estimated with multinomial logistic regression. To examine the link between the socio-demographic index (SDI) and viral fitness changes, we first evaluated the effects of viral traits on lineage growth rates using a Gamma regression model. Then we assessed the linear relationship between SDI and these effects. Machine learning models combining viral traits and SDI were trained to predict viral fitness, with performance assessed by regression and classification metrics. Immune escape exhibited the strongest association with viral fitness changes in the least developed countries (rate ratio [RR] = 1.90, 95% CI: 1.72-2.09), with this effect declining as economic development increased (developing countries: RR = 1.50, 95% CI: 1.42-1.58 and advance countries: RR = 1.20, 95% CI: 1.14-1.26). After model comparison, XGBoost showed the best predictive performance using viral traits, when SDI was included (RMSE = 16.269, R² = 0.786, Accuracy = 0.923). Viral traits and socio-economic context jointly shaped SARS-CoV-2 variant fitness. Immune escape was more influential for lineage growth in low-resource settings, emphasizing the need for equitable vaccine access and targeted interventions against the spread of immune-escape variants.

Learning protein fitness landscapes with deep mutational scanning data from multiple sources.

Evolutionary information is the primary tool for detecting functional conservation in nucleic acid and protein. This information has been extensively used to predict structure, interactions and functions in macromolecules. Pathogenicity prediction models rely on multiple sequence alignment information at different levels. However, most accurate genome-wide variant deleteriousness ranking algorithms consider different features to assess the impact of variants. Here, we analyze three different ways of extracting evolutionary information from sequence alignments in the context of pathogenicity predictions at DNA and protein levels. We showed that protein sequence-based information is slightly more informative in the annotation of Clinvar missense variants than those obtained at the DNA level. Furthermore, to achieve the performance of state-of-the-art methods, such as CADD and REVEL, the conservation of reference and variant, encoded as frequencies of reference/alternate alleles or wild-type/mutant residues, should be included. Our results on a large set of missense variants show that a basic method based on three input features derived from the protein sequence profile performs similarly to the CADD algorithm which uses hundreds of genomic features. As expected, our method results in ~ 3% lower area under the receiver-operating characteristic curve (AUC). When compared with an ensemble-based algorithm (REVEL). Nevertheless, the combination of predictions of multiple methods can help to identify more reliable predictions. These observations indicate that for missense variants, evolutionary information, when properly encoded, plays the primary role in ranking pathogenicity.

Genetic variants in the genes GRIN1, GRIN2A, GRIN2B, and GRIN2D, which encode subunits of the N-methyl-D-aspartate receptor (NMDAR), have been associated with severe and heterogeneous neurologic and neurodevelopmental disorders, including early onset epilepsy, developmental and epileptic encephalopathy, intellectual disability, and autism spectrum disorders. Missense variants in these genes can result in gain or loss of the NMDAR function, requiring opposite therapeutic treatments. Computational methods that predict pathogenicity and molecular functional effects of missense variants are therefore crucial for therapeutic applications. We assembled 223 missense variants from patients, 631 control variants from the general population, and 160 missense variants characterized by electrophysiological readouts that show whether they can enhance or reduce the function of the receptor. This includes new functional data from 33 variants reported here, for the first time. By mapping these variants onto the NMDAR protein structures, we found that pathogenic/benign variants and variants that increase/decrease the channel function were distributed unevenly on the protein structure, with spatial proximity to ligands bound to the agonist and antagonist binding sites being a key predictive feature for both variant pathogenicity and molecular functional consequences. Leveraging distances from ligands, we developed two machine-learning based predictors for NMDA variants: a pathogenicity predictor which outperforms currently available predictors and the first molecular function (increase/decrease) predictor. Our findings can have direct application to patient care by improving diagnostic yield for genetic neurodevelopmental disorders and by guiding personalized treatment informed by the knowledge of the molecular disease mechanism.

Simple SummaryHearing loss is the most common sensory impairment in humans. Globally, GJB2 is the most common responsible gene, with missense variants being the most frequent variant type. Pathogenic missense variants in the GJB2 gene cause isolated hearing loss and hearing loss combined with skin diseases. However, the mechanism by which different missense variants cause different hearing loss phenotypes is unknown. We have summarized 84 functionally studied missense variants and reviewed the clinical phenotypes and the molecular mechanisms that affected connexon and gap junction functions. We have also compiled a comprehensive dataset of GJB2 missense variants from public databases and published literature and found that over 2/3 of the GJB2 missense variants are currently classified as variants of uncertain significance. With the development of cutting-edge deep mutational scanning technology, we predict that the molecular mechanisms by which all missense variants of GJB2 lead to different clinical phenotypes will be elucidated. These gained insights will facilitate the use of genetic testing in the prevention and control of hearing loss.The GJB2 gene is the most common gene responsible for hearing loss (HL) worldwide, and missense variants are the most abundant type. GJB2 pathogenic missense variants cause nonsyndromic HL (autosomal recessive and dominant) and syndromic HL combined with skin diseases. However, the mechanism by which these different missense variants cause the different phenotypes is unknown. Over 2/3 of the GJB2 missense variants have yet to be functionally studied and are currently classified as variants of uncertain significance (VUS). Based on these functionally determined missense variants, we reviewed the clinical phenotypes and investigated the molecular mechanisms that affected hemichannel and gap junction functions, including connexin biosynthesis, trafficking, oligomerization into connexons, permeability, and interactions between other coexpressed connexins. We predict that all possible GJB2 missense variants will be described in the future by deep mutational scanning technology and optimizing computational models. Therefore, the mechanisms by which different missense variants cause different phenotypes will be fully elucidated.

Ciliopathies are a broad range of inherited developmental and degenerative diseases associated with structural or functional defects in motile or primary non-motile cilia. There are around 200 known ciliopathy disease genes and whilst genetic testing can provide an accurate diagnosis, 24–60% of ciliopathy patients who undergo genetic testing do not receive a genetic diagnosis. This is partly because following current guidelines from the American College of Medical Genetics and the Association for Molecular Pathology, it is difficult to provide a confident clinical diagnosis of disease caused by missense or non-coding variants, which account for more than one-third of cases of disease. Mutations in PRPF31 are the second most common cause of the degenerative retinal ciliopathy autosomal dominant retinitis pigmentosa. Here, we present a high-throughput high-content imaging assay providing quantitative measure of effect of missense variants in PRPF31 which meets the recently published criteria for a baseline standard in vitro test for clinical variant interpretation. This assay utilizes a new PRPF31+/– human retinal cell line generated using CRISPR gene editing to provide a stable cell line with significantly fewer cilia in which novel missense variants are expressed and characterised. We show that high-content imaging of cells expressing missense variants in a ciliopathy gene on a null background can allow characterisation of variants according to the cilia phenotype. We hope that this will be a useful tool for clinical characterisation of PRPF31 variants of uncertain significance, and can be extended to variant classification in other ciliopathies.

SummaryThe tumor suppressor CHEK2 encodes the serine/threonine protein kinase CHK2 which, upon DNA damage, is important for pausing the cell cycle, initiating DNA repair, and inducing apoptosis. CHK2 phosphorylation of the tumor suppressor BRCA1 is also important for mitotic spindle assembly and chromosomal stability. Consistent with its cell-cycle checkpoint role, both germline and somatic variants in CHEK2 have been linked to breast and other cancers. Over 90% of clinical germline CHEK2 missense variants are classified as variants of uncertain significance, complicating diagnosis of CHK2-dependent cancer. We therefore sought to test the functional impact of all possible missense variants in CHK2. Using a scalable multiplexed assay based on the ability of human CHK2 to complement DNA sensitivity of Saccharomyces cerevisiae cells lacking the CHEK2 ortholog, RAD53, we generated a systematic “missense variant effect map” for CHEK2 missense variation. The map reflects known biochemical features of CHK2 while offering new biological insights. It also provides strong evidence toward pathogenicity for some clinical missense variants and supporting evidence toward benignity for others. Overall, this comprehensive missense variant effect map contributes to understanding of both known and yet-to-be-observed CHK2 variants.

Genetic testing is a cornerstone in the assessment of many cardiac diseases. However, variants are frequently classified as variants of unknown significance, limiting the utility of testing. Recently, the DeepMind group (Google) developed AlphaMissense, a unique artificial intelligence-based model, based on language model principles, for the prediction of missense variant pathogenicity. We aimed to report on the performance of AlphaMissense, accessed by VarCardio, an open web-based variant annotation engine, in a real-world cardiovascular genetics center. All genetic variants from an inherited arrhythmia program were examined using AlphaMissense via VarCard.io and compared with the ClinVar variant classification system, as well as another variant classification platform (Franklin by Genoox). The mutation reclassification rate and genotype-phenotype concordance were examined for all variants in the study. We included 266 patients with heritable cardiac diseases, harboring 339 missense variants. Of those, 230 (67.8%) were classified by ClinVar as either variants of unknown significance or nonclassified. Using VarCard.io, 198 variants of unknown significance (86.1%, 95% CI, 80.9-90.3) were reclassified to either likely pathogenic or likely benign. The reclassification rate was significantly higher for VarCard.io than for Franklin (86.1% versus 34.8%, P<0.001). Genotype-phenotype concordance was highly aligned using VarCard.io predictions, at 95.9% (95% CI, 92.8-97.9) concordance rate. For 109 variants classified as pathogenic, likely pathogenic, benign, or likely benign by ClinVar, concordance with VarCard.io was high (90.5%). AlphaMissense, accessed via VarCard.io, may be a highly efficient tool for cardiac genetic variant interpretation. The engine's notable performance in assessing variants that are classified as variants of unknown significance in ClinVar demonstrates its potential to enhance cardiac genetic testing.

Accurate classification of genetic variants is critical for precision medicine, particularly hereditary diseases such as breast cancer. However, widely used tools like MutPred and Combined Annotation Dependent Depletion (CADD) offer genome-wide pathogenicity predictions that often overlook disease-specific variant behavior, limiting their clinical utility. This study addresses that gap by training and benchmarking nine machine learning (ML) models-including ensemble and baseline classifiers-on a breast cancer gene-specific dataset rich in conservation scores, functional annotations, and allele frequency features. Among all models, the Extra Trees model achieved the highest performance, with an accuracy of 0.999 and a 95% confidence interval of (0.998–1.000). recursive feature elimination identified the most informative genomic features, enhancing model efficiency. To ensure clinical transparency, we applied interpretability techniques including Local Interpretable Model-Agnostic Explanations and permutation feature importance, which highlighted the key drivers of each prediction. The calibration curve further confirmed the reliability of predicted probabilities, supporting their potential use in clinical decision-making. On an independent ClinGen dataset, Extra Trees achieved 99.1% accuracy and outperformed widely used predictors confirming its robustness and clinical applicability. This is the first comprehensive benchmarking study to apply ML models specifically to breast cancer-related missense variants using disease-gene-specific training data and integrated interpretability. Our results show that disease-specific ML approaches outperform general predictors, offering improved reliability, transparency, and relevance to clinical genomics. By bridging the gap between broad genome-wide tools and tailored clinical prediction, this study lays the foundation for implementing ML-driven pathogenicity prediction in breast cancer diagnostics and precision medicine, with potential expansion to other disease contexts.

Hereditary cancers frequently arise from germline pathogenic variants, yet only a small proportion of reported variants have been clinically classified, leaving most missense variants unresolved as variants of uncertain significance (VUS). Although recent machine-learning approaches have explored disease-specific or gene-specific contexts to improve pathogenicity prediction, these models remain fundamentally limited by the scarcity of labeled data and the underutilization of abundant VUS. We propose a deep-learning framework that integrates autoencoder pretraining with a deep ensemble strategy to improve variant pathogenicity prediction, effectively leverage unlabeled VUS during pretraining, and reduce uncertainty arising from limited training samples. To validate each component of our framework, we evaluated its performance under both disease-specific and gene-specific training setups. Experiments on ClinVar variants from BRCA1, BRCA2, MLH1, and MSH2 showed that our framework achieved the best performance in the gene-specific setup for BRCA1—likely because BRCA1 contains substantially more gene-specific training data than the other genes—whereas the disease-specific setup yielded superior results for the remaining genes, which had comparatively limited gene-specific samples. Overall, our method significantly outperformed existing approaches. We also introduce an interpretability approach that provides variant-level importance profiles across pathogenicity classes, thereby enhancing transparency and clinical applicability. Moreover, by projecting feature-level importance scores into a two-dimensional space, we demonstrate that pretraining enables the model to learn distinctly different feature representations, illustrating how pretraining and ensemble learning synergistically contribute to improved predictive performance. Our framework preserves the specificity of disease- and gene-specific approaches, overcomes data scarcity through VUS-guided pretraining and ensembling, and offers interpretable outcomes that may be helpful for clinical decision support. Moreover, our results suggest a promising direction for pathogenicity prediction of rare missense variants and indicate that the proposed framework may be extendable to additional genes under appropriate data and modeling conditions.

Deafness is the most frequent sensory disorder. With over 90 genes and 110 loci causally implicated in non-syndromic hearing loss, it is phenotypically and genetically heterogeneous. Here, we investigate the genetic etiology of deafness in four families of Iranian origin segregating autosomal recessive non-syndromic hearing loss (ARNSHL). We used a combination of linkage analysis, homozygosity mapping, and a targeted genomic enrichment platform to simultaneously screen 90 known deafness-causing genes for pathogenic variants. Variant segregation was confirmed by Sanger sequencing. Linkage analysis and homozygosity mapping showed segregation with the DFNB57 locus on chromosome 10 in two families. Targeted genomic enrichment with massively parallel sequencing identified causal variants in PDZD7: a homozygous missense variant (p.Gly103Arg) in one family and compound heterozygosity for missense (p.Met285Arg) and nonsense (p.Tyr500Ter) variants in the second family. Screening of two additional families identified two more variants: (p.Gly228Arg) and (p.Gln526Ter). Variant segregation with the hearing loss phenotype was confirmed in all families by Sanger sequencing. The missense variants are predicted to be deleterious, and the two nonsense mutations produce null alleles. This report is the first to show that mutations in PDZD7 cause ARNSHL, a finding that offers addition insight into the USH2 interactome. We also describe a novel likely disease-causing mutation in CIB2 and illustrate the complexity associated with gene identification in diseases that exhibit large genetic and phenotypic heterogeneity.

BackgroundRare genetic variants contribute to the etiology of both autism spectrum disorder (ASD) and schizophrenia (SCZ). Most genetic studies limit their focus to likely gene-disrupting mutations because they are relatively easier to interpret their effects on the gene product. Interpretation of missense variants is also informative to some pathophysiological mechanisms of these neurodevelopmental disorders; however, their contribution has not been elucidated because of relatively small effects. Therefore, we characterized missense variants detected in NRXN1, a well-known neurodevelopmental disease-causing gene, from individuals with ASD and SCZ.MethodsTo discover rare variants with large effect size and to evaluate their role in the shared etiopathophysiology of ASD and SCZ, we sequenced NRXN1 coding exons with a sample comprising 562 Japanese ASD and SCZ patients, followed by a genetic association analysis in 4273 unrelated individuals. Impact of each missense variant detected here on cell surface expression, interaction with NLGN1, and synaptogenic activity was analyzed using an in vitro functional assay and in silico three-dimensional (3D) structural modeling.ResultsThrough mutation screening, we regarded three ultra-rare missense variants (T737M, D772G, and R856W), all of which affected the LNS4 domain of NRXN1α isoform, as disease-associated variants. Diagnosis of individuals with T737M, D772G, and R856W was 1ASD and 1SCZ, 1ASD, and 1SCZ, respectively. We observed the following phenotypic and functional burden caused by each variant. (i) D772G and R856W carriers had more serious social disabilities than T737M carriers. (ii) In vitro assay showed reduced cell surface expression of NRXN1α by D772G and R856W mutations. In vitro functional analysis showed decreased NRXN1α-NLGN1 interaction of T737M and D772G mutants. (iii) In silico 3D structural modeling indicated that T737M and D772G mutations could destabilize the rod-shaped structure of LNS2-LNS5 domains, and D772G and R856W could disturb N-glycan conformations for the transport signal.ConclusionsThe combined data suggest that missense variants in NRXN1 could be associated with phenotypes of neurodevelopmental disorders beyond the diagnosis of ASD and/or SCZ.