¹ Genomics Coordination Center, University Medical Center Groningen, Groningen, 9713AV, The Netherlands
² Department of Genetics, University Medical Center Groningen, Groningen, 9713AV, The Netherlands
³ Department of Human Genetics, Amsterdam University Medical Centres, Groningen, The Netherlands

Tim Niemeijer
Roles: Conceptualization, Formal Analysis, Writing – Original Draft Preparation, Writing – Review & Editing

René Mulder
Roles: Conceptualization, Writing – Review & Editing

Helga Westers
Roles: Conceptualization, Writing – Review & Editing

Jan D. H. Jongbloed
Roles: Conceptualization, Writing – Review & Editing

Bart Charbon
Roles: Writing – Review & Editing

Birgit Sikkema-Raddatz
Roles: Writing – Review & Editing

Lennart F. Johansson
Roles: Writing – Review & Editing

Mariëlle E. van Gijn
Roles: Writing – Review & Editing

Cleo C. van Diemen
Roles: Writing – Review & Editing

Dennis Hendriksen
Roles: Writing – Review & Editing

Kristin M. Abbott
Roles: Writing – Review & Editing

Willem T. K. Maassen
Roles: Writing – Review & Editing

Morris A. Swertz
Roles: Writing – Review & Editing

K. Joeri van der Velde
Roles: Conceptualization, Software, Writing – Review & Editing

Despite decades of active research, the average diagnostic yield of next-generation sequencing (NGS) in genome diagnostic analyses rarely exceeds 35% (Pandey et al., 2025; Henkel et al., 2023; Earle et al., 2024). A major challenge in clinical genetics is the large proportion of DNA variants classified as variants of uncertain significance (VUS) (Chen et al., 2023). Of these VUS, missense variants particularly can be challenging to interpret. These variants, which alter the amino acid sequence of proteins, can alter structure and function. According to established clinical guidelines (Richards et al., 2015), determining whether a missense variant is pathogenic or benign instead of VUS requires multiple strong lines of evidence and supportive data. The decisive evidence to up- or downgrade the classification of VUS is often provided by functional studies. Yet, such studies are laborious, variant specific, and, given the sheer volume of variants identified in genome diagnostics, often impractical (Anderson et al., 2022).

To facilitate the prioritization and classification of missense variants, numerous pathogenicity prediction tools have been developed (Hu et al., 2019). These predictors employ various methodological approaches, including evolutionary conservation (Ng, 2003), population allele frequency data (Wu et al., 2021), and ensemble learning frameworks (Ioannidis et al., 2016). Training labels are derived from evolutionary constraints (Schubach et al., 2024) or curated variant databases (Rogers et al., 2017). Additional strategies assess the impact of variants on protein functional domains (Wiel et al., 2019) or on three-dimensional structural features (Ittisoponpisan et al., 2019). Currently, there are over 400 such predictors (Lin et al., 2024). While the field of variant effect prediction has started to suffer from data leakage (Bernett et al., 2024), imbalance (Rastogi et al., 2025), bias, and circularity (Grimm et al., 2015), few methods combine genomics with less biased physical modeling (Hollingsworth and Dror, 2018). The ones that do, lack explainability (Banerjee et al., 2025; Lai et al., 2021; Zhao et al., 2024) or are highly disease specific (Wang et al., 2025; Hamed et al., 2025).

An unbiased and high-performance missense pathogenicity predictor based on AlphaFold (Jumper et al., 2021) protein models along with evolutionary data and contextual features is AlphaMissense (Cheng et al., 2023). However, like most predictors, it produces only numerical predictions without explanations. This lack of interpretability limits our ability to investigate the underlying causal mechanisms, while interpretability is essential for translating computational predictions into clinically actionable insights and to guide functional laboratory experiments.

We hypothesize that the gap between traditional predictive modeling and mechanistic evidence can be bridged by breaking down prediction output into comprehensible feature contributions that allow further functional interpretation. This requires selecting biological features for the predictor, derived from protein stability, hydrophobicity, electrostatics, and interaction with ligands, DNA, RNA, and other proteins. Combining predictions with insight into the model’s decision making provided by SHapley Additive exPlanations (SHAP) (Lundberg and Lee, 2017) quantifies the contribution of each feature to individual predictions, guiding targeted digital or experimental follow-up tests to confirm or refute pathogenicity and help prioritize missense variants of interest. This hypothesis is tested through the development of Digital Approximation of Variant Effects, part of the MOLGENIS software family (MOLGENIS DAVE) to predict pathogenicity on carefully selected functional features, trained on pathogenic and benign variants from accredited Dutch genome diagnostic laboratories and providing predictions for the VUS. We compared the predictions with a more recent release of the same dataset and classifications in ClinVar (Landrum et al., 2013). Lastly, we break down the selected predictions and highlight three examples in which explainability assists in their mechanistic evaluation.

Functional features were computed for a selection of missense variants from the April 2024 release of fully de-identified and publicly accessible variant classifications provided by all Dutch genome diagnostic laboratories through the Datashare working group of the Dutch Clinical Genetics Laboratory Society (VKGL) (Fokkema et al., 2019). From these, we selected 12 biophysically grounded features: Folding energy, net electric charge, Hydrophobic moment, Hydrophobicity, Isoelectric point, No. ligand binding pockets, SAS points of the top ligand binding pocket, DNA binding affinity, RNA binding affinity, DNA binding residues, RNA binding residues and protein binding residues. These features were derived from five complementary tools: P2Rank (Krivák and Hoksza, 2018), FoldX 5 (Delgado et al., 2019), GLM-Score (Dias and Kolazckowski, 2015), Peptides R package (Osorio et al., 2014) and GeoNet (Han et al., 2024), as summarized in Table 1.

The DAVE model was trained on the feature differences (i.e., delta) between wild-type and protein structures containing an amino acid substitution caused by benign (LB/B) and pathogenic (LP/P) missense variants split into a train and test dataset. We determined an ROC AUC of 86% on the test set and a probability of 0.286 as the most effective binary classification threshold to identify pathogenic variants.

The trained DAVE model was applied on VUS missense variants from the VKGL April 2024 data release. Using the binary classification threshold, 3,801 variants were characterized as pathogenic and 7,420 as benign. To compare DAVE’s classification performance with real-world reclassification of VUS, we used a more recent release of the VKGL dataset and variant classifications from ClinVar. In July 2025, compared to April 2024, only 12 VUS had been reclassified as likely pathogenic/pathogenic or likely benign/benign on the VKGL data-sharing platform. These variants are listed in Table 2. Of these, nine DAVE predictions were in concordance with the reclassification, of which four true positives (i.e., correctly identified as pathogenic) and five true negatives (i.e., correctly identified as benign). Two were false positives (i.e., incorrectly predicted as pathogenic), and one was false negative (i.e., incorrectly predicted as benign). Additionally, cross-referencing of the VKGL VUS with ClinVar revealed that 478 variants were classified as benign/likely benign, while 176 were considered likely pathogenic/pathogenic. Figure 1 illustrates the distribution of DAVE pathogenicity probabilities and the corresponding reclassification labels from VKGL and ClinVar.

To demonstrate how DAVE can guide interpretation of functional consequences, we performed a more detailed analysis of three VUS that were reclassified in the July 2025 VKGL data-sharing release compared to the April 2024 release or had a different classification in ClinVar. These analyses are enabled by DAVE’s ability to break down the final pathogenicity probability into separate feature contributions as SHAP values (Lundberg and Lee, 2017), capturing interactions unique to each prediction. SHAP values of contributing features, including their delta and unit, are shown in a cumulative plot per variant. ChimeraX (Meng et al., 2023) was used to visualize the predicted effects on protein structures.

2.3.1 WNT7B T111P

The variant in WNT7B NM_058238.3:c.331A>C causes the amino acid substitution Thr111Pro in the Wnt-7b protein. It was selected for having the highest absolute folding energy change of the variants with an updated classification in the VKGL dataset in Table 2. It is predicted to have a substantial increase in ∆G folding energy, see Figure 2A. A higher ∆G causes protein folding to occur less spontaneous, affecting correct folding and stability. Molecular visualization shows a change in the hydrogen bond integral in the secondary structures and their connecting elements (Figure 2B). The impact on protein structural integrity can be a good mechanistic explanation of the effect of this missense variant.

2.3.2 SLCO2A1 G554R

The SLCO2A1 variant NM_005630.2:c.1660G>A, causes the amino acid substitution Gly554Arg in the solute carrier organic anion transporter family member 1A2 protein. It was selected for having the largest absolute change in solvent-accessible surface (SAS) points within its top-ranking ligand-binding pocket. Although currently still classified as a VUS in the VKGL dataset, it is reported as pathogenic in ClinVar (Xu et al., 2020). DAVE predicted the variant to be pathogenic, see Figure 3A. The predicted differences in the electrostatic surface potential of the ligand-binding pocket in the simulated structure of the G554R versus wild-type protein, are shown in Figure 3B. Moreover, recent functional characterization of variants in the transmembrane domain of SLCO2A1 replacing glycine with residues with larger side chains underscored this, as it revealed a loss of helical dynamics, accompanied by potential disruption of the helical packing (Xia et al., 2025).

2.3.3 NKX2–5 L153P

The variant NKX2–5 NM_004387.3: c.458T>C, causes the amino acid change, Leu153Pro in the homeobox Nkx-2.5 protein. It was selected for having largest absolute change in protein binding site residues. This variant is currently classified as VUS in the VKGL dataset while reported as pathogenic in ClinVar. DAVE predicts that this variant is pathogenic, shown in Figure 4A. The protein encoded by NKX2–5 is a transcription factor, and the substitution L153P, lies in an important homeodomain (AA 138–197) responsible for the binding of DNA as well as other transcription factors (Pradhan et al., 2012). Interestingly, despite the location of the variant, the most contributing feature to the predicted pathogenicity of this variant is not the change in binding site residues, but the change in folding energy ∆G. The predicted structural change of the L153P variant protein compared to the wild-type is visualized in Figure 4B.

We developed MOLGENIS DAVE, a supervised learning model that predicts pathogenicity of missense variants and provides interpretable insights via contributions of functionally relevant features based on protein modeling. Unlike methods such as AlphaMissense (Chen et al., 2023) and REVEL (Ioannidis et al., 2016), DAVE prioritizes physical explainability over predictive power, providing complementary evidence for variant classification and guidance for targeted experimental validation of VUS either in dry or wet laboratory.

The selected features of DAVE represent multiple disease-causing mechanisms (Sen et al., 2022) and thus provide a starting point for interpretation. However, no single model can fully capture the complexity and unique biology of all proteins. Manual follow-up will remain essential to accurately determine the molecular consequences of prioritized variants. By integrating more diverse and context-dependent features, future versions of DAVE could better capture the nuanced mechanisms underlying variant effects, ultimately improving both accuracy and interpretability across a broader range of proteins and variant types. For instance, variants that have a gain-of-function or dominant-negative consequence have different, often milder, molecular mechanisms compared to loss-of-function variants (Gerasimavicius et al., 2022; Backwell and Marsh, 2022). Other examples of such effects include allostery, post-translational modifications (PTMs), toxicity, structural flexibility, folding rate, kinetic effects, aggregation propensity, membrane association, immunogenicity, and cooperativity.

Some AlphaFold structures are predicted with low confidence, particularly for disordered regions or poorly conserved domains (Luppino et al., 2025; Brotzakis et al., 2025). In these regions, predictive accuracy can be significantly lower, especially if models rely heavily on structural data (Kong et al., 2025). Another limitation of current models is that they represent the protein without their biological context, and that models larger than 2700 residues (<2% of all models) are chunked into smaller fragments due to computational limitations, making them unsuitable for meaningful predictions. To increase applicability across diverse protein types and missense effect types, structure confidence metrics and more biological context should be integrated to refine functional predictions. We expect that future versions of AlphaFold and similar efforts will produce unfragmented models for all proteins.

While DAVE could certainly be configured to run as part of DNA interpretation pipelines, currently it is not offered an easy to install standalone product. The reasons are the integration with external software, one of which used under academic personal license, as well as a prohibitive computational burden, averaging around 20 minutes per variant on commodity hardware. This hinders its applicability in clinical genetics workflows where rapid turnaround time is essential. For future versions, we will prioritize improving computational efficiency and ensuring permissive licensing of underlying methods.

In summary, DAVE combines predictive accuracy with meaningful information on how its predictions are made. By complementing numerical predictions with mechanistic insights, we demonstrate the potential to transform in silico variant effect predictions into interpretable molecularly grounded explanations. The insights provide a practical route to generate testable hypotheses for the resolution of VUS in functional follow-up studies, thereby bridging the gap between prediction, explanation, and clinical utility. Although DAVE captures multiple protein-related features, it does not currently account for variant effects through other mechanisms, such as altered gene expression or splicing. Consequently, pathogenicity arising from these processes is not assessed by the model. In such cases, the use of an ensemble of specialized tools covering complementary variant effects is recommended. This limitation is exemplified by our false negative prediction for CACNA1A Leu622Gln, see Table 2. This variant was classified in diagnostics as likely pathogenic due to a predicted impact on splicing, which explains why DAVE did not predict a significant pathogenic effect at the protein level. Looking ahead, we envision a framework that integrates DAVE with complementary evidence such as population frequency, conservation, inheritance, and phenotypes as well as specialized tools that cover effects on transcription and translation to systematically classify missense variants and reduce VUS. Integrating DAVE into pipelines such as MOLGENIS VIP (Maassen et al., 2025) would enable the inclusion of these complementary tools within a scalable and modular infrastructure for variant interpretation. Such an integrative approach enables large scale reanalysis efforts, to prioritize variants for experimental validation, and ultimately accelerate the route from variant discovery to informed clinical decision making.

The main resource is the fully de-identified and publicly accessible DNA variant classifications dataset collected in April 2024 by the Datashare working group of the nine genome diagnostic labs in the Netherlands, organized in the VKGL (Vereniging Klinisch Genetische Laboratoriumdiagnostiek) (Fokkema et al., 2019). We downloaded the GRCh38 liftover version from the MOLGENIS VIP (Maassen et al., 2025) resources. If variants within the same gene with the same protein change had differences in classification, they were marked as conflicting. The July 2025 version of the VKGL release, was collected similarly to the April 2024 release. For protein structures, we used pdb’s from the AlphaFold v4 release. AlphaFold structures are represented by UniProt and mappings to HGNC gene symbols are provided. The GRCh38 VCF file for ClinVar (Landrum et al., 2013) release 2025-09-23 was used as-is by matching on chromosome, position, reference and alternative allele.

The VKGL variant dataset was annotated with Ensembl VEP (McLaren et al., 2016) version 112. Based on protein localization data from the Protein Atlas (Uhlén et al., 2015) we randomly selected 1007 intracellular, 1028 membrane, and 692 secreted proteins for which we also had genomic variants in the dataset. This selection of proteins was used to subset the VKGL variants to balance the data on localization. Additionally, we annotated whether chaperones were involved in the folding of these proteins, for this we used a dataset of 194 manually curated chaperones (Shemesh et al., 2021) and expanded these genes with UniProt/SwissProt IDs using Ensembl BioMart. These known chaperones were connected to their interaction protein partners using BioGRID (Oughtred et al., 2020).

We applied the following five protein analysis tools on the selected variants, P2Rank (Krivák and Hoksza, 2018), FoldX 5 (Delgado et al., 2019), GLM-Score (Dias and Kolazckowski, 2015), Peptides R package (Osorio et al., 2014), and GeoNet (Han et al., 2024). Complete data processing was successful for 23,417 variants, of which the classifications, localizations, and chaperoned folding characteristics are shown in Table 3.

To obtain the most informative and independently contributing set of features, the set was minimized by removing correlations and selecting the most informative and explainable features. Among the remaining features, pairwise correlations were minimal, with the exception of correlation between‘Net electric charge at pH 7’ and ‘Isoelectric point’. Despite this, both features were deemed valuable to retain due to their individual relevance. The R function ‘cor’ from the stats (r-core) package was used to calculate the correlation of features as the Pearson correlation coefficient (PCC), shown in Figure 5.

A subset of the full annotated dataset, consisting of 12,048 variants classified as likely pathogenic (LP) or likely benign (LB), was split into training and testing sets using an 80/20 ratio. The DAVE model was implemented using the randomForest package in R (version 4.7–1.1) using the standard Random Forest algorithm with default parameters. To determine the contribution of each selected feature to individual predictions, SHAP values were calculated using the R package fastshap (version 0.1.1) with 10-fold Monte Carlo sampling. It is important to note that SHAP values do not correlate with feature values but instead capture feature contributions based on the interactions among all features uniquely for this prediction.

For binary classification, the VKGL April 2024 ‘test’ set was used to determine the optimal threshold with the R package cutpointr (version 1.1.2). From this, a threshold of 0.286 was established as the cutoff point using Youden’s J statistic expressed as:

J=TPTP+FN+TNTN+FP−1

TP = True Positive

TN = True Negative

FP = False Positive

FN = False Negative

To evaluate the utility of DAVE in the reclassification of the 11,221 variants of uncertain significance (VUS), we compared variant classifications between the April 2024 and July 2025 VKGL datasets. Additionally, we compared variants classified as VUS in the VKGL that had a B/LB or LP/P classification in ClinVar (September 23, 2025, with at least a one-star review status and no conflicting classifications) with the predictions of DAVE. For variant specific structural comparison, we substituted wild-type amino acid with the variant in the AlphaFold (Jumper et al., 2021) structure, PDBs using FoldX5 (Delgado et al., 2019) with the BuildModel and RepairPDB command. These structures were then visualized with ChimeraX (Meng et al., 2023) and compared with wild-type structures.

We would like to thank our colleagues at Genomics Coordination Center for engineering support and access to the UMCG high-performance compute cluster. We are grateful to our colleagues from all Dutch genome diagnostic laboratories for their continuing efforts to share variant classifications through the VKGL Datashare working group. We would also like to thank our previous and current heads of department Prof. Nine Knoers and Prof. Anke-Hilse Maitland-van der Zee as well as the UMCG Board of Directors, in particular Prof. Stephanie Klein Nagelvoort Schuit and Prof. Wiro Niessen for their encouragement and support to develop and use new AI methods at the University Medical Center Groningen. Lastly, we would like to express our gratitude to the exquAIro AI bootcamp organization and board members Prof. Gerard Koppelman, Dr. Martin Smit, Ilya Petoukhov, Dr. Marnix Bügel, as well as Geerte Koster of REWIRE and Dr. Kai Yu Ma of UMCG for their training, coaching and valuable input.

•  Anderson CL, Munawar S, Reilly L, et al.: How functional genomics can keep pace with vus identification. Frontiers in Cardiovascular Medicine. 2022; 9. Publisher Full Text
•  Backwell L, Marsh JA: Diverse molecular mechanisms underlying pathogenic protein mutations: Beyond the loss-of-function paradigm. Annual Review of Genomics and Human Genetics. 2022; 23(1): 475–498. PubMed Abstract | Publisher Full Text
•  Banerjee A, Bogetti AT, Bahar I: Accurate identification and mechanistic evaluation of pathogenic missense variants with rhapsody-2. Proceedings of the National Academy of Sciences. 2025; 122(18): e2418100122. PubMed Abstract | Publisher Full Text | Free Full Text
•  Bernett J, Blumenthal DB, Grimm DG, et al.: Guiding questions to avoid data leakage in biological machine learning applications. Nature Methods. 2024; 21(8): 1444–1453. PubMed Abstract | Publisher Full Text
•  Brotzakis ZF, Zhang S, Murtada MH, et al.: Alphafold prediction of structural ensembles of disordered proteins. Nature Communications. 2025; 16(1): 1632. PubMed Abstract | Publisher Full Text | Free Full Text
•  Chen E, Facio FM, Aradhya KW, et al.: Rates and classification of variants of uncertain significance in hereditary disease genetic testing. JAMA Network Open. 2023; 6(10): e2339571. PubMed Abstract | Publisher Full Text | Free Full Text
•  Cheng J, Novati G, Pan J, et al.: Accurate proteome-wide missense variant effect prediction with alphamissense. Science. 2023; 381(6664): eadg7492. PubMed Abstract | Publisher Full Text
•  Delgado J, Radusky LG, Cianferoni D, et al.: Foldx 5.0: working with rna, small molecules and a new graphical interface. Bioinformatics. 2019; 35(20): 4168–4169. PubMed Abstract | Publisher Full Text | Free Full Text
•  Dias R, Kolazckowski B: Different combinations of atomic interactions predict protein-small molecule and protein-dna/rna affinities with similar accuracy. Proteins: Structure, Function, and Bioinformatics. 2015; 83(11): 2100–2114. PubMed Abstract | Publisher Full Text | Free Full Text
•  Earle NJ, Winbo A, Crawford J, et al.: Genetic testing yield and clinical characteristics of hypertrophic cardiomyopathy in understudied ethnic groups: Insights from a new zealand national registry. Circulation: Heart Failure. 2024; 17(3).
•  Fokkema IFAC, Velde KJ, Slofstra MK, et al.: Dutch genome diagnostic laboratories accelerated and improved variant interpretation and increased accuracy by sharing data. Human Mutation. 2019; 40(12): 2230–2238. PubMed Abstract | Publisher Full Text | Free Full Text
•  Gerasimavicius L, Livesey BJ, Marsh JA: Loss-of-function, gain-of-function and dominant-negative mutations have profoundly different effects on protein structure. Nature Communications. 2022; 13(1): 3895. PubMed Abstract | Publisher Full Text | Free Full Text
•  Grimm DG, Azencott C, Aicheler F, et al.: The evaluation of tools used to predict the impact of missense variants is hindered by two types of circularity. Human Mutation. 2015; 36(5): 513–523. PubMed Abstract | Publisher Full Text | Free Full Text
•  Han J, Zhang S, Guan M, et al.: Geonet enables the accurate prediction of protein-ligand binding sites through interpretable geometric deep learning. Structure. 2024; 32(12): 2435–2448.e5. PubMed Abstract | Publisher Full Text
•  Henkel J, Laner A, Locher M, et al.: Diagnostic yield and clinical relevance of expanded germline genetic testing for nearly 7000 suspected hboc patients. European Journal of Human Genetics. 2023; 31(8): 925–930. Publisher Full Text
•  Hollingsworth SA, Dror RO: Molecular dynamics simulation for all. Neuron. 2018; 99(6): 1129–1143. Publisher Full Text
•  Hu Z, Yu C, Furutsuki M, et al.: Vipdb, a genetic variant impact predictor database. Human Mutation. 2019; 40(9): 1202–1214. PubMed Abstract | Publisher Full Text | Free Full Text
•  Ioannidis NM, Rothstein JH, Pejaver V, et al.: Revel: An ensemble method for predicting the pathogenicity of rare missense variants. The American Journal of Human Genetics. 2016; 99(4): 877–885. PubMed Abstract | Publisher Full Text | Free Full Text
•  Ittisoponpisan S, Islam SA, Khanna T, et al.: Can predicted protein 3d structures provide reliable insights into whether missense variants are disease associated?. Journal of Molecular Biology. 2019; 431(11): 2197–2212. PubMed Abstract | Publisher Full Text | Free Full Text
•  Jumper J, Evans R, Pritzel A, et al.: Highly accurate protein structure prediction with alphafold. Nature. 2021; 596(7873): 583–589. Publisher Full Text
•  Hamed KA, Naghinejad M, Amirfiroozy A, et al.: A random forest-based predictive model for classifying brca1 missense variants: a novel approach for evaluating the missense mutations effect. Journal of Human Genetics. 2025; 70(7): 341–348.
•  Kong SW, Lee I-H, Collen LV, et al.: Discordance between a deep learning model and clinical-grade variant pathogenicity classification in a rare disease cohort. npj Genomic Medicine. 2025; 10(1). Publisher Full Text
•  Krivák R, Hoksza D: P2rank: machine learning based tool for rapid and accurate prediction of ligand binding sites from protein structure. Journal of Cheminformatics. 2018; 10(1): 39. PubMed Abstract | Publisher Full Text | Free Full Text
•  Lai J, Yang J, Gamsiz Uzun ED, et al.: Lyrus: a machine learning model for predicting the pathogenicity of missense variants. Bioinformatics Advances. 2021; 2(1).
•  Landrum MJ, Lee JM, Riley GR, et al.: Clinvar: public archive of relationships among sequence variation and human phenotype. Nucleic Acids Research. 2013; 42(D1): D980–D985.
•  Lin Y-J, Menon AS, Hu Z, et al.: Variant impact predictor database (vipdb), version 2: trends from three decades of genetic variant impact predictors. Human Genomics. 2024; 18(1): 90. PubMed Abstract | Publisher Full Text | Free Full Text
•  Lundberg SM, Lee S-I: A unified approach to interpreting model predictions. Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS’17. Red Hook, NY, USA: Curran Associates Inc.; 2017; page 4768–4777.
•  Luppino F, Lenz S, Chow CFW, et al.: Deep learning tools predict variants in disordered regions with lower sensitivity. BMC Genomics. 2025; 26(1): 367. PubMed Abstract | Publisher Full Text | Free Full Text
•  Maassen WTK, Johansson LF, Charbon B, et al.: Molgenis vip: an end-to-end dna variant interpretation pipeline for research and diagnostics configurable to support rapid implementation of new methods. NAR Genomics and Bioinformatics. 2025; 7(2). PubMed Abstract | Publisher Full Text | Free Full Text
•  McLaren W, Gil L, Hunt SE, et al.: The ensembl variant effect predictor. Genome Biology. 2016; 17(1). Publisher Full Text
•  Meng EC, Goddard TD, Pettersen EF, et al.: ucsf chimerax: Tools for structure building and analysis. Protein Science. 2023; 32(11): e4792. PubMed Abstract | Publisher Full Text | Free Full Text
•  Ng PC: Sift: predicting amino acid changes that affect protein function. Nucleic Acids Research. 2003; 31(13): 3812–3814. PubMed Abstract | Publisher Full Text | Free Full Text
•  Osorio D, Rondon-Villarreal P, Torres R: Peptides: Calculate indices and theoretical physicochemical properties of protein sequences.2014.
•  Oughtred R, Rust J, Chang C, et al.: The biogrid database: A comprehensive biomedical resource of curated protein, genetic, and chemical interactions. Protein Science. 2020; 30(1): 187–200. PubMed Abstract | Publisher Full Text
•  Pandey R, Brennan NF, Trachana K, et al.: A meta-analysis of diagnostic yield and clinical utility of genome and exome sequencing in pediatric rare and undiagnosed genetic diseases. Genetics in Medicine. 2025; 27(6): 101398. PubMed Abstract | Publisher Full Text
•  Pradhan L, Genis C, Scone P, et al.: Crystal structure of the human nkx2.5 homeodomain in complex with dna target. Biochemistry. 2012; 51(32): 6312–6319. PubMed Abstract | Publisher Full Text | Free Full Text
•  Rastogi R, Chung R, Li S, et al.: Critical assessment of missense variant effect predictors on disease-relevant variant data. Human Genetics. 2025; 144(2–3): 281–293, 293. PubMed Abstract | Publisher Full Text | Free Full Text
•  Richards S, Aziz N, Bale S, et al.: Standards and guidelines for the interpretation of sequence variants: a joint consensus recommendation of the american college of medical genetics and genomics and the association for molecular pathology. Genetics in Medicine. 2015; 17(5): 405–424. PubMed Abstract | Publisher Full Text | Free Full Text
•  Rogers MF, Shihab HA, Mort M, et al.: Fathmm-xf: accurate prediction of pathogenic point mutations via extended features. Bioinformatics. 2017; 34(3): 511–513.
•  Schubach M, Maass T, Nazaretyan L, et al.: Cadd v1.7: using protein language models, regulatory cnns and other nucleotide-level scores to improve genome-wide variant predictions. Nucleic Acids Research. 2024; 52(D1): D1143–D1154. PubMed Abstract | Publisher Full Text | Free Full Text
•  Sen N, Anishchenko I, Bordin N, et al.: Characterizing and explaining the impact of disease-associated mutations in proteins without known structures or structural homologs. Briefings in Bioinformatics. 2022; 23(4). PubMed Abstract | Publisher Full Text | Free Full Text
•  Shemesh N, Jubran J, Dror S, et al.: The landscape of molecular chaperones across human tissues reveals a layered architecture of core and variable chaperones. Nature Communications. 2021; 12(1): 2180. PubMed Abstract | Publisher Full Text | Free Full Text
•  Uhlén M, Fagerberg L, Hallström BM, et al.: Tissue-based map of the human proteome. Science. 2015; 347(6220).
•  van der Velde J : Dave codebase and resources. Zenodo. 2026. Publisher Full Text
•  Wang J, Yang M, Zong C, et al.: protphemut: An interpretable machine learning tool for classification of cancer and neurodevelopmental disorders in human missense mutations. Journal of Chemical Information and Modeling. 2025; 65(15): 8375–8384. PubMed Abstract | Publisher Full Text
•  Wiel L, Baakman C, Gilissen D, et al.: Metadome: Pathogenicity analysis of genetic variants through aggregation of homologous human protein domains. Human Mutation; 2019. Publisher Full Text
•  Wu Y, Liu H, Li R, et al.: Improved pathogenicity prediction for rare human missense variants. The American Journal of Human Genetics. 2021; 108(10): 1891–1906. PubMed Abstract | Publisher Full Text | Free Full Text
•  Xia Z, Lu G, Wu D, et al.: Structure and transport mechanism of the human prostaglandin transporter slco2a1. Nature Communications. 2025; 16(1). Publisher Full Text
•  Xu Y, Zhang Z, Yue H, et al.: Monoallelic mutations in slco2a1 cause autosomal dominant primary hypertrophic osteoarthropathy. Journal of Bone and Mineral Research. 2020; 36(8): 1459–1468. PubMed Abstract | Publisher Full Text
•  Zhao H, Du H, Zhao S, et al.: Sigma leverages protein structural information to predict the pathogenicity of missense variants. Cell Reports Methods. 2024; 4(1): 100687. Publisher Full Text

Current Reviewer Status:

AWAITING PEER REVIEW

AWAITING PEER REVIEW

?