Welcome to DeMAG web server!

We present DeMAG (Deciphering Mutations in Actionable Genes) a supervised and specialized VEP (Variant Effect Predictor) to interpret any missense mutations in 59 actionable genes (ACMG SF v2.0 genes[1]).

DeMAG web server allows scientists, health professionals or anyone curious to:

  • access and download predictions for all ~1.3 million missense mutations for the ACMG SF v2.0 genes (ACMG59 genes),
  • investigate the features of the model for any mutation to better understand DeMAG's pathogenicity score,
  • download the high-quality training set to fully reproduce our results,
  • download the validation sets we used to benchmark DeMAG against popular VEPs (REVEL, EVE, …).

DeMAG is a joint collaboration between the Max Planck Insitute of Molecular Cell Biology and Genetics (MPI-CBG) and Harvard Medical School (HMS). You can read our paper on bioRxiv:

DeMAG predicts the effect of variants in clinically actionable genes by integrating structural and evolutionary epistatic features

Federica Luppino1,2, Ivan A. Adzhubei4,5, Christopher A. Cassa4, Agnes Toth-Petroczy*1,2,3

1. Max Planck Institute of Molecular Cell Biology and Genetics, Dresden 01307, Germany. 2. Center for Systems Biology, Dresden 01307, Germany. 3. Cluster of Excellence Physics of Life, TU Dresden, 01062 Dresden, Germany. 4. Brigham and Womenʼs Hospital Division of Genetics, Harvard Medical School, Boston, MA, 02115 USA. 5. Department of Biomedical Informatics, Harvard Medical School, Boston, MA 02115.


[1] Kalia, S. S. et al. (2017). Recommendations for reporting of secondary findings in clinical exome and genome sequencing, 2016 update (ACMG SF v2.0): a policy statement of the American College of Medical Genetics and Genomics. Genet. Med. 19, 249–255

ACMG SF v2.0 genes

Table content downloaded from UniProt and ClinVar

Here you can search for your variant!



Variants with a pathogenicity probability greater or equal than 0.5 are predicted as pathogenic

Mutation landscape

The plot shows all missense substitutions for the selected gene. The black dot indicates the selected mutation


Features Plot

The violinplots show the distribution of the selected features for the benign and pathogenic class as in our training set. The white line corresponds to the specific variant's feature-value.

AlphaFold2 3D model

Features definition

Feature Definition Source
Score1 PSIC score for wild type amino acid residue (aa1) pph2
Score2 PSIC score for wild type amino acid residue (aa2) pph2
dScore difference of PSIC scores for two amino acid residue variants (Score1-Score2) pph2
DistQmin minimum distance (sum of branch lengths) along the phylogenetic tree across all substitution types encountered at the substitution position pph2
phylop conservation scoring by phyloP (phylogenetic p-values) from the PHAST package for multiple alignments of 99 vertebrate genomes to the human genome pph2
BaRE Bayesian Rate Estimator for scoring evolutionary conservation, D.M. Jordan (2015) pph2
Nres number of unique residues observed at the substitution position in multiple alignment (without gaps) pph2
Nsubs number of residues different from reference residue (aa1) observed at the substitution position in multiple alignment pph2
EVmutation A log-odds ratio between the sequence probability of the wild-type and the mutant sequence T. A. Hopf (2017)
Partners score residue's posterior probability of pathogenicity given the residue score of its co-evlving and spatially close residues positions See Methods of the manuscript
NormASA normalized accessible surface pph2
IUPred2A disorder score sequence-based disorder propensity predictor IUPred2A
AlphaFold2 pLDDT per-residue confidence metric on a scale from 0-100 AlphaFold2 FAQ

Training data

The training data is available with the exception of HGMD variants as they are license protected.
Data URL Date
PrimateAI (DataFileS1.csv) The user has to create an account at this link (https://basespace.illumina.com/s/yYGFdGih1rXL)
KRGDB project email the author to ask for the data
3.5KJPNv2 project https://humandbs.biosciencedbc.jp/files/hum0015/tommo-3.5kjpnv2-20181105-af_snvall-autosome.zip
NCBI ALFA https://ftp.ncbi.nih.gov/snp/population_frequency/latest_release/freq.vcf.gz 24.11.20
HGMD Licence version 2020.03. HGMD data were available to the authors under a subscription data use agreement which prohibits sharing variant data from HGMD Professional (QIAGEN). Users and developers may not make HGMD data publicly available. (https://www.hgmd.cf.ac.uk/docs/disclaimer.html)
ClinVar https://ftp.ncbi.nlm.nih.gov/pub/clinvar/vcf_GRCh37/archive_2.0/2021/clinvar_20210501.vcf.gz
HumOrtho and UniProt HumVar: http://genetics.bwh.harvard.edu/downloads/demag/training/
gnomAD https://storage.googleapis.com/gcp-public-data–gnomad/release/2.1.1/vcf/exomes/gnomad.exomes.r2.1.1.sites.vcf.bgz.

Testing data

We benchamrked DeMAG on clinical, functional and population data. The clinical and functional data can be found below, while the population data will be available on request. The population data refers to putatively benign common variants from the Estonia Biobank and as it is not public yet we can't upload the data on web server. Please refer to the methods section of the manuscript for a detailed description of the analysis of the testing set.

ClinVar testing set

The clinical validation set consists of variants submitted to ClinVar after 2017. This ensures that none of the predictors we are benchmarking with trained on those variants, as the last supervised tool was published in 2016.

Functional testing set

The functional validation set consists of variants evaluated with DMS experiments for four available genes (MSH2, BRCA1, TP53 and PTEN).

Common variants testing set

The population data refers to benign variants from the Estonia Biobank


Here you can download DeMAG predictions for all possible missense amino acids substitutions for the ACMG SF v2.0 genes
Here you can download DeMAG predictions for all possible missense amino acids substitutions for the additional 257 genes