Published online 23 February 2021 NAR Genomics and Bioinformatics, 2021, Vol. 3, No. 1 1 doi: 10.1093/nargab/lqab008 TMSNP: a web server to predict pathogenesis of missense mutations in the transmembrane region of membrane proteins Adrián Garcia-Recio 1,2,† , José Carlos Gómez-Tamayo3,† , Iker Reina1 , Mercedes Campillo1 , Arnau Cordomı́ 1,4,* and Mireia Olivella 2,4,* 1 Laboratori de Medicina Computacional, Facultat de Medicina, Universitat Autònoma de Barcelona, 08193 Bellaterra, Spain, 2 Bioinformatics and Medical Statistics Group, Facultat de Ciències i Tecnologia, UVIC-UCC, 08500 Vic, Barcelona, Spain, 3 Pharmacoinformatics Group, Research Program on Biomedical Informatics (IMIM/UPF), 08003 Barcelona, Spain and 4 Bioinformatics Department, ESCI-UPF, 08003 Barcelona, Spain Received May 18, 2020; Revised December 16, 2020; Editorial Decision January 25, 2021; Accepted January 27, 2021 ABSTRACT variants within candidate variants. For this purpose, vari- ant prioritization tools are widely used to predict the effect The massive amount of data generated from genome of mutations. These are mostly based on evolutionary con- sequencing brings tons of newly identified mu- servation and expected impact on structure and function tations, whose pathogenic/non-pathogenic effects using evolutionary conservation parameters and physico- need to be evaluated. This has given rise to sev- chemistry properties of amino acids from sequence data eral mutation predictor tools that, in general, do [SIFT (2), Provean (3), MutationTaster (4)], while some not consider the specificities of the various protein tools such as Polyphen-2 (5) also incorporate features re- groups. We aimed to develop a predictor tool dedi- lated to structural data [see (6) for a review]. cated to membrane proteins, under the premise that Membrane proteins represent 25% of all human proteins their specific structural features and environment (7) and perform essential roles in cellular functions (8). Con- would give different responses to mutations com- sequently, they are the target for 50% of drugs in the market (9). Moreover, 90% of membrane proteins present disease- pared to globular proteins. For this purpose, we cre- associated missense mutations that may affect protein fold- ated TMSNP, a database that currently contains in- ing, stability and/or function (10). Some of them have been formation from 2624 pathogenic and 196 705 non- related to various diseases, including cardiopathies, neuro- pathogenic reported mutations located in the trans- logical diseases, cystic fibrosis and cancer (11,12). In fact, membrane region of membrane proteins. By comput- mutations in membrane proteins are more likely to cause ing various conservation parameters on these mu- diseases than in globular proteins (13). Membrane proteins tations in combination with annotations, we trained differ from globular proteins in terms of amino acid com- a machine-learning model able to classify mutations position, distribution, inter-residue interactions and struc- as pathogenic or not. TMSNP (freely available at http: ture (14,15). The main differences are in the transmembrane //lmc.uab.es/tmsnp/) improves considerably the pre- (TM) region of the proteins because of the different surface diction power of commonly used mutation predictors environments, that is lipid exposed versus water exposed. Current variant prioritization tools, which are mainly based trained with globular proteins. on data from globular proteins, present low reliability for predicting the pathogenicity of mutations in membrane pro- INTRODUCTION teins (13). Thus, there is a need for computational tools spe- cific for membrane proteins to understand its relation be- Whole genome and exome sequencing have revealed that tween sequence and structure (16). This is especially impor- Mendelian rare disease-causing missense mutations are tant given the scarce number of membrane protein struc- more frequent than previously thought and collectively af- tures compared to globular proteins due to experimental fect millions of patients worldwide (1). Thus, there is an limitations (17). Mutation prediction tools and databases urgent need to understand the relation between genotype specific for membrane proteins are starting to emerge, such and phenotype in order to identify disease-causing genetic * To whom correspondence should be addressed to Mireia Olivella. Tel: +34 932954710; Email: mireia.olivella@uvic.cat Correspondence may also be addressed to Arnau Cordomı́. Tel: +34 935813812; Email: arnau.cordomi@uab.cat † The authors wish it to be known that, in their opinion, the first two authors should be regarded as Joint First Authors. C The Author(s) 2021. Published by Oxford University Press on behalf of NAR Genomics and Bioinformatics. This is an Open Access article distributed under the terms of the Creative Commons Attribution-NonCommercial License (http://creativecommons.org/licenses/by-nc/4.0/), which permits non-commercial re-use, distribution, and reproduction in any medium, provided the original work is properly cited. For commercial re-use, please contact journals.permissions@oup.com 2 NAR Genomics and Bioinformatics, 2021, Vol. 3, No. 1 as Mut-HTP (18), Pred-MutHTP (10) or BorodaTM (19), Feature extraction. Multiple sequence alignments for the which are all based on evolutionary conservation parame- different families of the proteins in our dataset were taken ters, the former including structure descriptors and the lat- from the Pfam database (26). For each missense mutation in ter focusing on regions with known structure. the balanced dataset, we computed four variables related to With the aim of contributing to emerging mutation pre- evolutionary conservation and the likelihood that an amino dictor servers for the TM region of membrane proteins, here acid change is tolerated in a position: (i) frequency of the we present TMSNP (accessible at http://lmc.uab.es/tmsnp/), wild-type amino acid in the Pfam alignment, (ii) frequency a database of TM missense mutations (pathogenic and non- of the mutated amino acid, (iii) substitution matrix score (as pathogenic) and a predictor server trained using evolution- a measure of similar physicochemical properties between ary conservation parameters. wild-type and mutated amino acid) and (iv) entropy of the position (as a measure of sequence variability or informa- tion content) (27). For the substitution matrix score, we MATERIALS AND METHODS used the PHAT 75/73 matrix, which is specific for mem- brane proteins (28). The entropy of the position i is max- Database of pathogenic and non-pathogenic mutations in imal (= 1) if all 20 amino acids at the position i present transmembrane proteins equal frequencies and is minimal ( = 0) if only 1 amino acid Our selected set of membrane proteins consisted of all hu- has been observed at this position. Four additional vari- man membrane proteins tagged as reviewed in the UniProt ables: type of the reference and the mutated amino acids, database (20,21). For each protein, we retrieved all disease- Pfam and UniProt accession codes were included through causing/pathogenic mutations associated with Mendelian one-hot-encoding of the qualitative variables. In the final disorders as reported in ClinVar (22) and SwissVar (23). dataset, each missense mutation had information encoded We only kept in mutations occurring in the TM helices be- in eight variables contributing to a total of 569 features (20 cause these are the regions that mostly differ from glob- features for reference and 20 for the mutated amino acids, ular proteins. The ranges of TM segments were taken 358 features for the UniProt accession codes and 167 for the from the UniProt database (20,21). We also retrieved non- Pfam accession codes). pathogenic missense mutations and their population al- lele frequency from GnomAD (24) and ClinVar (22). The Machine-learning models. We built three datasets from database (accessible at http://lmc.uab.es/tmsnp/tmsnpdb) the 5408 missense mutations (see http://lmc.uab.es/tmsnp/ resulted in 196 705 non-pathogenic, 2624 pathogenic and datasets): (i) 8V dataset, containing all variables (569 fea- 437 likely pathogenic mutations in the TM region of mem- tures); (ii) 6V dataset, lacking UniProt and Pfam accession brane proteins. code variables, which are informative of the tendency of a protein or a protein family to pathogenesis, but still includ- ing one-hot encoded reference and mutated amino acids (60 features) and (iii) 4V dataset containing only conservation TMSNP predictor variables (wild-type and mutated frequencies, substitution Filtering, homology reduction and dataset balancing. To en- matrix score and entropy). For each dataset, five different sure that mutations used in the machine-learning models training (80%) and test (20%) sets were created by random were linked to protein function and/or structure alteration, sampling under certain restrictions for internal validation. we discarded all mutations in proteins for which no single For external-validation, in the 8V dataset and 4V datasets, pathogenic disease-causing mutations have been reported, mutations with the same UniProt code were equally split be- that is those likely involved in complex diseases or reces- tween training set and test set while for the 6V dataset Pfam sive inheritance and tagged as ‘non-pathogenic proteins’ accession codes were used to split mutations either in the (25). Thus, the obtained pathogenic and non-pathogenic training ser or in the validation set. Machine-learning mod- missense mutations were used to classify human TM pro- els were built using Flame (https://github.com/phi-grib/ teins as ‘pathogenic proteins’ (358 proteins), when at least flame; a Python modeling framework which wraps scikit- one disease-causing pathogenic mutation has ever been re- learn (http://scikit-learn.sourceforge.net)) or Keras (https: ported for this protein and as ‘non-pathogenic proteins’ //keras.io/). Various predictive models using different al- (2420 proteins), elsewhere. We next performed homology gorithm settings, applicability domain and dataset were reduction by discarding mutations for proteins belonging built and internally validated using K-fold (K = 5) cross- to the same Pfam family (26) that resulted in the same validation. In specific we used Random Forest (RF), Gra- amino acid change in the same aligned position. The dataset dient Boosting (XGBoost), Supporting Vector Machines after homology reduction contained 2704 pathogenic and (SVM) and a sequential neural network. The conformal likely pathogenic mutations and 19 292 non-pathogenic prediction was used as an applicability domain technique mutations. Data were subsequently subsampled to ob- (29) by testing our models at three different confidences: 95, tain a balanced dataset (50% pathogenic and 50% non- 90 and 80%. pathogenic mutations), by selecting the non-pathogenic mu- tations with the highest population allele frequency accord- Web server ing to GnomAD (24). The final dataset used for training in the machine-learning model presented 5408 missense mu- TMSNP web application tool was constructed using a tations, which implies a reduction of non-pathogenic muta- Python backend (v.3.7) with the Flask framework (v.1.0.2). tions by 1/6. Both the application and the associated datasets used for NAR Genomics and Bioinformatics, 2021, Vol. 3, No. 1 3 Table 1. Model statistics in cross-validation Dataset Algorithm Confidence Sensitivity Specificity MCC Coverage Accuracy 8V (569 features) RF 0.95 0.92 0.88 0.80 0.42 0.90 RF 0.90 0.89 0.83 0.72 0.62 0.86 RF 0.80 0.82 0.77 0.58 0.88 0.79 XGBOOST 0.95 0.96 0.93 0.89 0.39 0.94 XGBOOST 0.90 0.92 0.88 0.80 0.58 0.90 XGBOOST 0.80 0.85 0.81 0.66 0.86 0.83 SVM 0.95 0.96 0.93 0.89 0.46 0.95 SVM 0.90 0.92 0.90 0.82 0.63 0.91 SVM 0.80 0.88 0.84 0.71 0.86 0.85 6V (44 features) RF 0.95 0.91 0.71 0.63 0.13 0.81 RF 0.90 0.87 0.72 0.60 0.29 0.79 RF 0.80 0.79 0.67 0.46 0.58 0.72 XGBOOST 0.95 0.88 0.73 0.60 0.08 0.79 XGBOOST 0.90 0.81 0.74 0.55 0.25 0.77 XGBOOST 0.80 0.76 0.69 0.45 0.57 0.72 SVM 0.95 0.90 0.77 0.68 0.08 0.84 SVM 0.90 0.81 0.74 0.55 0.25 0.77 SVM 0.80 0.76 0.69 0.45 0.57 0.72 4V (4 features) RF 0.95 0.86 0.59 0.46 0.12 0.72 RF 0.90 0.77 0.64 0.42 0.29 0.71 RF 0.80 0.71 0.63 0.34 0.57 0.67 XGBOOST 0.95 0.87 0.72 0.59 0.09 0.79 XGBOOST 0.90 0.82 0.72 0.54 0.21 0.76 XGBOOST 0.80 0.73 0.67 0.40 0.48 0.70 SVM 0.95 0.93 0.30 0.29 0.15 0.71 SVM 0.90 0.90 0.41 0.37 0.34 0.69 SVM 0.80 0.73 0.65 0.38 0.67 0.69 The table shows quality metrics (5-fold) for the machine-learning models created using 8V, 6V and 4V datasets with different conformal significance. MCC stands for Matthews correlation coefficient, which is a measure that combines sensitivity and specificity. Coverage stands for the percentage of samples inside the applicability domain. training and testing the predictor were built automati- (type of reference and mutated residues) included in the 6V cally using Python/Bash scripts that collected the required dataset increase the average performance up to ∼80% ac- data and stored it in a MySQL database (v.8.0.18), fa- curacy (at the maximum confidence) except for SVM which cilitating regular updates. All scripts could be found in keeps at ∼70%. 8V dataset clearly shows the best perfor- a GitHub repository (https://github.com/adriangarciarecio/ mance both in internal cross validation and external val- TMSNP). idation. SVM provides the best models, which reach 94% accuracy with 46% coverage (95% confidence), or 85% with 86% coverage (80% confidence). XGBoost and RF mod- RESULTS AND DISCUSSION els follow closely although with slightly worse performance We initially constructed a database of missense muta- (<5%). tions in human membrane proteins that exclusively fo- In order to check for possible overfitting of the models cused on TM helices (accessible at http://lmc.uab.es/tmsnp/ using the largest (8V) dataset, we performed feature selec- tmsnpdb; see ‘Materials and Methods’). The database cur- tion for the three different algorithms. Supplementary Table rently contains 2624 pathogenic, 437 likely pathogenic S1 compares the performance of the models with the orig- and 196 705 non-pathogenic mutations. We used a sub- inal and the reduced features using K-best feature selection set of this database (see ‘Materials and Methods’) to de- performed with K = 60 or 30 (number of variables reduced velop machine-learning models able to classify mutations to 10% and ∼5%, respectively). The mean accuracy loss for as pathogenic or not. We assessed three different algorithms RF, XGBoost and SVM was, respectively, 2%, -2% and 3% (RF, SVM and XGBoost) and three different datasets (8V, for K = 60, and 4%, 0% and 6% for K = 30. These small dif- 6V and 4V). All combinations showed good performance in ferences suggest lack of overfitting and also point out that both internal (5-fold cross-validation; Table 1) and external SVM is less robust than RF or XGBoost algorithms. In or- validation (independent 20% test set; Table 2) with none of der to assess the presence of bias due to dataset balancing them clearly outperforming the other two. RF showed the (see ‘Materials and Methods’), we generated an additional best performance on the 4V dataset, while XGBoost and dataset containing the first 3000 non-pathogenic mutations SVM were the best on 6V and 8V, respectively. Although we following those used in training. This bias might lead to un- also tested a sequential neural network, we could not find realistic predictions, possibly translated to an excess of false advantages of using this method despite performance being positives. Supplementary Table S2 shows prediction results close to the other algorithms used in this study. Models that using 8V models at 95%, 90% and 80% confidence. Mini- use only four features reach a maximum accuracy of ∼70%, mum true negatives/false positives ratio is ∼4, being most without the increase in the confidence of the model bring- of the predictions either negatives or out of the applicabil- ing additional improvement. The two additional features ity domain and always sticking to the confidence restraints. 4 NAR Genomics and Bioinformatics, 2021, Vol. 3, No. 1 Table 2. Model statistics at external validation Dataset Algorithm Confidence Sensitivity Specificity MCC Coverage Accuracy 8V (569 features) RF 0.95 0.90 0.86 0.76 0.38 0.88 RF 0.90 0.86 0.82 0.68 0.58 0.84 RF 0.80 0.81 0.75 0.56 0.86 0.78 XGBOOST 0.95 0.90 0.88 0.78 0.30 0.89 XGBOOST 0.90 0.85 0.84 0.70 0.48 0.85 XGBOOST 0.80 0.79 0.78 0.57 0.77 0.78 SVM 0.95 0.91 0.89 0.80 0.39 0.90 SVM 0.90 0.86 0.85 0.71 0.55 0.85 SVM 0.80 0.77 0.79 0.56 0.81 0.78 6V (44 features) RF 0.95 0.86 0.66 0.54 0.78 0.74 RF 0.90 0.81 0.69 0.51 0.76 0.73 RF 0.80 0.76 0.66 0.42 0.71 0.70 XGBOOST 0.95 0.87 0.69 0.58 0.09 0.80 XGBOOST 0.90 0.82 0.70 0.53 0.24 0.77 XGBOOST 0.80 0.77 0.68 0.45 0.54 0.73 SVM 0.95 0.90 0.77 0.68 0.08 0.84 SVM 0.90 0.81 0.74 0.55 0.25 0.77 SVM 0.80 0.76 0.69 0.45 0.57 0.72 4V (4 features) RF 0.95 0.85 0.60 0.47 0.13 0.75 RF 0.90 0.79 0.66 0.46 0.29 0.73 RF 0.80 0.71 0.67 0.37 0.56 0.69 XGBOOST 0.95 0.79 0.65 0.45 0.09 0.73 XGBOOST 0.90 0.76 0.67 0.43 0.21 0.72 XGBOOST 0.80 0.69 0.66 0.35 0.47 0.68 SVM 0.95 0.92 0.38 0.37 0.14 0.74 SVM 0.90 0.84 0.55 0.41 0.35 0.73 SVM 0.80 0.72 0.68 0.40 0.67 0.70 The table shows performance metrics in external validation (20% of the original dataset) for the machine-learning models created using 8V, 6V and 4V datasets with different conformal significance. MCC and coverage are described in Table 1. These results demonstrate lack of sampling bias. Interest- Pfam codes as features, as they are related to different vul- ingly, the RF model at 95% confidence correctly predicts nerability of proteins and/or protein families to amino acid 658 non-pathogenic mutations with only 6 false positives change in their transmembrane region. out of 3000 non-pathogenic mutations. TMSNP returns the unambiguous class prediction at the Random Forest 8V was selected as the final model to be highest confidence possible. Predictions with a confidence implemented in the web application. Although the perfor- below 0.75 are considered outside the domain of appli- mance of RF was not the best, it demonstrated to be the cability. Table 3 shows the comparison between TMSNP more robust algorithm at different conditions. While XG- models (8V dataset) generated at three levels of significance Boost performed on average better than RF, it provides together with the results of SIFT, Polyphen-2 and Pred- more importance to protein classification features rather MutHTP. Pred-MutHTP is also specific for membrane pro- than sequence conservation (amino acid frequencies, sub- teins, whereas both SIFT and Polyphen-2 are not. As re- stitution matrix score and entropy), questioning its ability flected by the equilibrated sensitivity/specificity and the to generalize the predictions (Supplementary Table S3). On higher accuracy at the different confidence levels, TMSNP the other hand, SVM was discarded because it was less ro- not only provides a more balanced model but also performs bust towards feature reduction and also because the imple- very well when tested against a full non-pathogenic dataset mentation of the SVM algorithm using radial basis function of 3000 mutations. Importantly, TMPSNP brings higher kernel did not allow to inspect feature importances. specificity compared to the other methods. The robustness Feature importance analysis (Supplementary Table S4) of the algorithm relies on the quality of the extracted con- shows that for the original RF model as well as for the 60 servation features, which are the most important according and 30 best feature-reduced models, conservation features to the feature contribution analysis of the model. Notewor- contribute the most clearly driving the predictive power of thy, TMSNP is using conformal prediction as an applicabil- the algorithms (30%, 60% and 75% of the total contribu- ity domain and uncertainty framework, providing predic- tion, respectively). Pfam PF00520 (ion channel family) and tions under confidence restraints. Our models demonstrated UniProt P35498 (sodium channel protein) accession codes to be good at all confidence limits, which translates into the follow in contribution (∼2% each), probably indicating high corresponding accuracy (i.e. a confidence of 0.8 sets to ∼0.2 sensitivity of this family of receptors to become pathogenic the maximum error rate). The higher specificity and accu- upon a mutation. Mutation to proline and to arginine (P m racy of TMSNP compared to SIFT and PolyPhen-2, which and R m features) appear as the next features in impor- also rely on similar conservation parameters, might be re- tance (∼1% each). This is compatible with the known dis- lated to using a dataset for training specific for membrane torting effects of these amino acids when present at TM he- proteins, as the used evolutionary conservation parame- lices (13). Accordingly, the loss of performance for the 6V ters differ between globular and membrane proteins. When or 4V dataset might be related to not using UniProt and compared to the previously reported membrane-specific NAR Genomics and Bioinformatics, 2021, Vol. 3, No. 1 5 Table 3. Sensitivity, specificity, Matthews correlation coefficient (MCC) and coverage of TMSNP model (8V dataset) and comparison to Pre-MutHTP, SIFT and Poyphen-2 Predictor types Method Sensitivity Specificity MCC Coverage Accuracy Specific for membrane proteins TMSNP (0.95 confidence) 0.90 0.86 0.76 0.38 0.88 TMSNP (0.90 confidence) 0.86 0.82 0.68 0.58 0.84 TMSNP (0.8 confidence) 0.81 0.75 0.56 0.86 0.78 Pre-MutHTP (0.95 confidence) 0.96 0.54 0.56 0.76 0.64 Pre-MutHTP (0.90 confidence) 0.96 0.53 0.55 0.76 0.67 Pre-MutHTP (0.80 confidence) 0.96 0.53 0.56 0.76 0.71 Non-specific for membrane proteins Polyhen-2 0.93 0.35 0.35 1 0.64 SIFT 0.88 0.52 0.42 1 0.70 Data are shown at various levels of significance in external validation. MCC and coverage are described in Table 1. predictor Pred-MutHTP, the higher specificity and accu- REFERENCES racy of TMSNP might be related to the better curated non- 1. Chong,J.X., Buckingham,K.J., Jhangiani,S.N., Boehm,C., pathogenic TM variants as the result of (i) only considering Sobreira,N., Smith,J.D., Harrell,T.M., McMillin,M.J., the highest allele frequencies in GnomAD and (ii) discard- Wiszniewski,W., Gambin,T. et al. (2015) The genetic basis of ing non-pathogenic mutations in proteins for which no sin- Mendelian phenotypes: discoveries, challenges, and opportunities. Am. J. Hum. Genet., 97, 199–215. gle causative disease mutation has been identified that might 2. Sim,N.-L., Kumar,P., Hu,J., Henikoff,S., Schneider,G. and Ng,P.C. be affecting the structure and function of the protein with- (2012) SIFT web server: predicting effects of amino acid substitutions out being related to pathogenesis (recessive inheritance and on proteins. Nucleic Acids Res., 40, W452–W457. complex diseases). Compared to MutHTP database, TM- 3. Choi,Y. and Chan,A.P. (2015) PROVEAN web server: a tool to predict the functional effect of amino acid substitutions and indels. SNP (i) contains mutations in the TM segments of mem- Bioinformatics, 31, 2745–2747. brane proteins but discards mutations in the extracellular 4. Schwarz,J.M., Cooper,D.N., Schuelke,M. and Seelow,D. (2014) and intracellular regions of membrane proteins, as these re- MutationTaster2: mutation prediction for the deep-sequencing age. gions and domains of the proteins are exposed to an en- Nat. Methods, 11, 361–362. vironment similar to globular proteins; (ii) is based on a 5. Adzhubei,I., Jordan,D.M. and Sunyaev,S.R. (2013) Predicting functional effect of human missense mutations using bigger dataset of non-pathogenic variants as MutHTP does PolyPhen-2. Curr. Protoc. Hum. Genet., 7, Chapter 7:Unit7.20. not include variants from GnomAD database (24); and (iii) 6. Niroula,A. and Vihinen,M. (2016) Variation interpretation does not include somatic mutations. predictors: principles, types, performance, and choice. Hum. Mutat., 37, 579–597. 7. Dobson,L., Langó,T., Reményi,I. and Tusnády,G.E. (2015) CONCLUSIONS Expediting topology data gathering for the TOPDB database. Nucleic Acids Res., 43, D283–D289. TMSNP is a regularly updated web server that presents two 8. Gromiha,M.M. and Ou,Y.-Y. (2014) Bioinformatics approaches for main functionalities: on the one hand, it brings a search- functional annotation of membrane proteins. Brief. Bioinform., 15, 155–168. able database of reported pathogenic and non-pathogenic 9. Overington,J.P., Al-Lazikani,B. and Hopkins,A.L. (2006) How many mutations in TM segments of membrane proteins; on the drug targets are there? Nat. Rev. Drug Discov., 5, 993–996. other hand, it provides a mutation prediction tool able to 10. Kulandaisamy,A., Priya,S.B., Sakthivel,R., Frishman,D. and predict pathogenicity for previously non-reported TM mis- Gromiha,M.M. (2019) Statistical analysis of disease-causing and neutral mutations in human membrane proteins. Proteins, 87, sense mutations. The predictive model developed specif- 452–466. ically for membrane proteins allows to improve the pre- 11. Kulandaisamy,A., Zaucha,J., Sakthivel,R., Frishman,D. and Michael diction power compared to unspecific mutation predictor Gromiha,M. (2020) Pred-MutHTP: prediction of disease-causing and servers. neutral mutations in human transmembrane proteins. Hum. Mutat., 41, 581–590. 12. Hauser,A.S., Chavali,S., Masuho,I., Jahn,L.J., Martemyanov,K.A., DATA AVAILABILITY Gloriam,D.E. and Babu,M.M. (2018) Pharmacogenomics of GPCR drug targets. Cell, 172, 41–54. TMSNP is available at http://lmc.uab.es/tmsnp/. 13. Zaucha,J., Heinzinger,M., Kulandaisamy,A., Kataka,E., Salvádor,Ó.L., Popov,P., Rost,B., Gromiha,M.M., Zhorov,B.S. and Frishman,D. (2020) Mutations in transmembrane proteins: diseases, SUPPLEMENTARY DATA evolutionary insights, prediction and comparison with globular proteins. Brief. Bioinform., bbaa132. Supplementary Data are available at NARGAB Online. 14. Olivella,M., Gonzalez,A., Pardo,L. and Deupi,X. (2013) Relation between sequence and structure in membrane proteins. Bioinformatics, 29, 1589–1592. FUNDING 15. Mayol,E., Campillo,M., Cordomı́,A. and Olivella,M. (2019) Inter-residue interactions in alpha-helical transmembrane proteins. Spanish Ministerio de Ciencia, Innovación y Universidades Bioinformatics, 35, 2578–2584. 16. Almeida,J.G., Preto,A.J., Koukos,P.I., Bonvin,A.M.J.J. and [SAF2015-74627-JIN, SAF2016-77830-R, PI19/00348]. Moreira,I.S. (2017) Membrane proteins structures: a review on Funding for open access charge: Ministerio de Ciencia e computational modeling tools. Biochim. Biophys. Acta Biomembr., Innovación y Universidades [PI19/00348]. 1859, 2021–2039. Conflict of interest statement. None declared. 6 NAR Genomics and Bioinformatics, 2021, Vol. 3, No. 1 17. Burley,S.K., Berman,H.M., Christie,C., Duarte,J.M., Feng,Z., 23. Mottaz,A., David,F.P.A., Veuthey,A.-L. and Yip,Y.L. (2010) Easy Westbrook,J., Young,J. and Zardecki,C. (2018) RCSB Protein Data retrieval of single amino-acid polymorphisms and phenotype Bank: sustaining a living digital data resource that enables information using SwissVar. Bioinformatics, 26, 851–852. breakthroughs in scientific research and biomedical education. 24. Karczewski,K.J., Francioli,L.C., Tiao,G., Cummings,B.B., Alföldi,J., Protein Sci., 27, 316–330. Wang,Q., Collins,R.L., Laricchia,K.M., Ganna,A., Birnbaum,D.P. 18. Kulandaisamy,A., Binny Priya,S., Sakthivel,R., Tarnovskaya,S., et al. (2020) The mutational constraint spectrum quantified from Bizin,I., Hönigschmid,P., Frishman,D. and Gromiha,M.M. (2018) variation in 141,456 humans. Nature, 581, 434–443. MutHTP: mutations in human transmembrane proteins. 25. Eilbeck,K., Quinlan,A. and Yandell,M. (2017) Settling the score: Bioinformatics, 34, 2325–2326. variant prioritization and Mendelian disease. Nat. Rev. Genet., 18, 19. Popov,P., Bizin,I., Gromiha,M., A,K. and Frishman,D. (2019) 599–612. Prediction of disease-associated mutations in the transmembrane 26. El-Gebali,S., Mistry,J., Bateman,A., Eddy,S.R., Luciani,A., regions of proteins with known 3D structure. PLoS One, 14, Potter,S.C., Qureshi,M., Richardson,L.J., Salazar,G.A., Smart,A. e0219452. et al. (2019) The Pfam protein families database in 2019. Nucleic 20. McGarvey,P.B., Nightingale,A., Luo,J., Huang,H., Martin,M.J., Acids Res., 47, D427–D432. Wu,C. and UniProt Consortium (2019) UniProt genomic mapping 27. Pei,J. and Grishin,N.V. (2001) AL2CO: calculation of positional for deciphering functional effects of missense variants. Hum. Mutat., conservation in a protein sequence alignment. Bioinformatics, 17, 40, 694–705. 700–712. 21. Consortium, T.U.The UniProt Consortium (2019) UniProt: a 28. Ng,P.C., Henikoff,J.G. and Henikoff,S. (2000) PHAT: a worldwide hub of protein knowledge. Nucleic Acids Res., 47, transmembrane-specific substitution matrix. Bioinformatics, 16, D506–D515. 760–766. 22. Landrum,M.J., Lee,J.M., Riley,G.R., Jang,W., Rubinstein,W.S., 29. Norinder,U., Carlsson,L., Boyer,S. and Eklund,M. (2014) Church,D.M. and Maglott,D.R. (2014) ClinVar: public archive of Introducing conformal prediction in predictive modeling. A relationships among sequence variation and human phenotype. transparent and flexible alternative to applicability domain Nucleic Acids Res., 42, D980–D985. determination. J. Chem. Inf. Model., 54, 1596–1603.