Demag Hc 810

[Edie Staniszewski]

Thankyou for visiting nature.com. You are using a browser version with limited support for CSS. To obtain the best experience, we recommend you use a more up to date browser (or turn off compatibility mode in Internet Explorer). In the meantime, to ensure continued support, we are displaying the site without styles and JavaScript.

Assessing the pathogenicity of genetic variants remains a significant challenge in research and clinical translation. The American College of Medical Genetics and Genomics (ACMG) recommends the reporting of secondary findings in clinically actionable genes (e.g., ACMG SF lists1,2) when patients undergo sequencing3. Knowledge of a pathogenic variant in such a gene might improve clinical management, diagnosis, and prevention. Given insufficient epidemiological, functional, or other supportive evidence, over three quarters of variants which have been submitted to ClinVar4 are classified as Variants of Uncertain Significance (VUSs, Supplementary Fig. 1). The uncertainty about the pathogenicity of a variant may pose a psychological burden for patients5,6, left without guidance, and can lead to potential morbidity and health costs associated with under and overdiagnosis7.

Unsupervised methods, such as DeepSequence16, EVmutation17, and EVE18 are agnostic to variant labels as they infer functional effects from multiple sequence alignment (MSA). These methods rely on the availability of high quality MSA data, which is often missing in disordered and low-complexity regions, and poorly conserved regions19. Unsupervised methods characterize the fitness effects of mutations independently from reported disease-causing variants, and do not provide an interpretation of pathogenicity17,20. An exception is EVE, which provides two gene-specific unsupervised thresholds for pathogenic and benign variants respectively, however it leaves the most uncertain variants without annotation18. The method relies on labeled clinical data to identify the uncertain class. While this is useful for clinical applications, it suffers from labeling biases of supervised tools that use publicly available variants databases21.

Here, we extend the traditional conservation paradigm to assess variant effects with novel protein sequence- and structure-based features. We designed an epistatic feature, the partners score, which defines epistatic residue pairs based on co-evolutionary and 3D structural partnership of residues as defined by AlphaFold239 models. The partners score is informed by the clinical labels of partner residues, taking advantage of the wealth of existing clinical knowledge. Based on their medical importance and the abundance of clinical diagnostic data, we focused on interpreting missense variants in 59 clinically actionable disease genes in the ACMG SF v2.0 list, which we refer to as ACMG SF genes2.

We developed DeMAG (Deciphering Mutations in Actionable Genes) a supervised classifier to assess the pathogenicity of mutations in 59 clinically actionable disease genes (ACMG SF v2.0 list) and support clinical decision making. First, we carefully curated pathogenic and benign variants used for training the model (Fig. 1 and Supplementary Fig. 2). For those variants, we then tested several sequence- and structure-based features and selected those that discriminated between variants with high confidence pathogenic and benign classifications (Fig. 1 and Supplementary Table 1). We designed the partners score, which is based on evolutionary and structural partnerships of residues as estimated by AlphaFold2 structural models (Figs. 2 and 3 and Supplementary Fig. 3). Overall, DeMAG used only 13 features, 8 derived from sequence conservation, and 5 from 3D structural models, disorder scores and epistatic relationships (Supplementary Table 2).

a On the top left, co-evolving residue positions are shown for the DNA mismatch repair protein MSH6. The innermost circle indicates residues that are co-evolving with at least one other residue whose phenotypic effect is known (i.e., annotated in the training set). Among such residues, 8 are pathogenic, 7 benign, and 255 lack prior assessment. The outer circle shows the partners score resulting from co-evolutionary partnership of residues, which gives a score to 255 residue positions without known annotations. The outermost circle indicates the Pfam domains of MSH6. On the right, AlphaFold 3D model of the protein is shown. Residue positions are colored based on the partners score derived from spatially close partnerships of residues, which inform 750 previously unannotated residue positions (55% of the protein length). Below, the same representation is shown for the cellular tumor antigen P53 protein. The circle plot shows correlation between partners score and domain annotation: the residues that belong to the DNA-binding domain have high partners scores while in the low-complexity region have low partners scores (Supplementary Fig. 7). In total 208 (53%) previously unlabeled residue positions are now annotated with partner scores. b A structural example, where MSH6 ATP binding site residues, have been previously shown50 to be critical for DNA mismatch repair (MMR) function, have a high partners score.

We trained a machine learning model (Fig. 4), and validated it with 3 different ground-truth test sets: clinical (Fig. 5 and Table 1), functional (deep mutational scanning, Supplementary Fig. 4), and benign variants from population data (Fig. 6 and Table 2). We further evaluated its performance on an additional set of 257 clinically relevant genes, which have sufficient numbers of variants with high quality diagnostic interpretations (Supplementary Table 4 and Supplementary Data 1). Finally, we computed DeMAG pathogenicity scores for all missense variants in the ACMG SF genes and additional 257 clinically relevant genes.

We designed a novel feature called the partners score based on the observation that partner residues that are connected, either because they are close in 3D proximity or because they are co-evolving, share the same phenotypic effect (Supplementary Fig. 6a). We used the AlphaFold2 3D protein structural models to identify residues in spatial proximity (

Each residue position can be associated with only pathogenic, only benign, both pathogenic and benign (mixed), or not being associated with any known variant (Fig. 2a). Each residue has a score (residue score) based on the type and number of connections it has (Fig. 2b). We used a mixture-based discriminant analysis46 approach to define the partners score: first, the density of the residue score is estimated independently for the pathogenic and benign class in the training set assuming a gaussian mixture distribution (Fig. 2b). Then, each variant is assigned a posterior probability of belonging to either class, given the residue score and the prior probability of both classes (i.e., frequency). The posterior probability of pathogenicity defines the partners score (Fig. 2c and Methods section) which highlights how mutations with the same phenotypic effects cluster both in linear and 3D space of the protein.

While only 13% of ACMG SF residues have annotated variants in the training set, we can inform 74% of positions with the partners score by making use of epistatic relationships (Fig. 2d). For example, the DNA mismatch repair protein MSH6 has only 8 pathogenic and 7 benign residue positions that are also co-evolving with other positions. With the partners score, we annotated 255 positions whose clinical significance has not been assessed yet. The same trend applies to positions in spatial proximity (Fig. 3a): amongst the spatially close residue positions, only 53 have annotated variants (33 pathogenic and 20 benign). With the partner score based on spatial proximity, we annotated 750 positions. Overall, if we consider both evolutionary and spatial partnerships, the partners feature assigns a score to 55% of all MSH6 residues (750 positions). For the cellular tumor protein P53, we observed a clear correlation between the partners score and Pfam47 protein domain annotations, e.g., residue positions of the low-complexity region and disordered region are characterized by low partners scores, while residue positions of the DNA-binding domain has overall high scores (Fig. 3a and Supplementary Fig. 7). In addition, we observed that the MSH6 ATP binding site has a partners score >0.6 (Fig. 3b). The role of the ATP binding site of the MSH2-MSH6 heterodimer is crucial for DNA mismatch repair (MMR) competency: mutations of the lysine residue in the MSH6 Walker A motif are complete loss of function mutations in vivo in S. cerevisiae48. Moreover, all 14 mutations (G1134[A,R,E,V], P1135A, N1136D, M1137[T,V], G1138R, G1139[D,C,V], S1141[C,P]) in this site are ClinVar VUSs, with no definitive clinical interpretation, while they are predicted pathogenic by DeMAG.

Several existing VEPs, such as M-CAP and SIFT4G have high sensitivity but low specificity23. Their recommended thresholds (M-CAP10 0.025 and SIFT4G 0.0524) are set to reach high sensitivity in variant interpretation, while tolerating a high misclassification rate for benign variants. This imbalance increases the number of potentially false positive variants (benign variants predicted incorrectly to be pathogenic). To address this issue, we made extensive efforts to improve training set balance by expanding the number of available benign mutations (Fig. 1 and Supplementary Fig. 2d). We selected only 13 features with balanced performance in discriminating between pathogenic and benign classes (Supplementary Table 1a and Methods section), including 8 derived from sequence conservation, and 5 from 3D structural models disorder scores, and epistatic relationships (Supplementary Table 2). DeMAG was trained with a gradient-boosting tree method49,50 (see Methods section) and it yielded high accuracy (87%) and AUC-ROC (92%) values that correspond to high sensitivity (87%) and specificity (85%), as well as high precision (90%) (Fig. 4c). Overall and at the single gene level, DeMAG has a balanced sensitivity and specificity (Fig. 4a, c), which corresponds to setting the threshold to 0.5 to interpret a variant as pathogenic.

3a8082e126