This website shows the results of analysing the distribution of ClinVar variants at certain CADD-Score Thresholds.


What to do here?

  1. You can look at the results of the calculations with the whole dataset here .

  2. If you are interested in the comparison of the different CADD versions and genome releases you can look here .

  3. If you have a specific use case and know the genes for the variants, you are looking at you can look here .


What is CADD?

CADD (Combined Annotation Dependent Depletion) is a tool that is used for scoring the deleteriousness of single nucleotide variants, multi nucleotide substitutions and insertions/deletions variants in the human genome.
When using CADD there are two scores. The raw and the PHRED-score. For the PHRED-score all potential single nucleotide variants (SNVs) in the genome (~9 billion) are sorted by their pathogenicity in comparison to all others. Each SNV then gets assigned a PHRED score depending on their rank. This means a variant that ranks in the top 10 percent of potentially pathogenic variants receives a PHRED score of 10 or higher. Variants in the top 1 percent receive a score of 20 or higher. PHRED scores are less resolved than Raw scores but are often used as they can be compared better with other scores.
It might seem useful to have a universal cut-off value that clearly seperates pathogenic from benign variants. However, the CADD authors advise against this, as the threshold depends on the specific analysis and use case. Applying a single universal cut-off would risk a considerable loss of valuable information.
Still, it is useful to see how variants are spread across different thresholds and to understand which factors affect what might be a good cut-off. The score distribution of known benign and pathogenic variants has been analysed and made usable on this website to help with finding a good cut-off for specific use cases.

For more information and reference please refer to the [CADD Website] (https://cadd.bihealth.org/).
You may also look at these publications:

The most recent manuscript describes CADD v1.7, an extension to the annotations included in the model. Most prominently, this version improves the scoring of coding variants with features derived from the ESM-1v protein language model as well as the scoring of regulatory variants with features derived from a convolutional neural network trained on regions of open chromatin:

Schubach M, Maass T, Nazaretyan L, Röner S, Kircher M.
CADD v1.7: Using protein language models, regulatory CNNs and other nucleotide-level scores to improve genome-wide variant predictions.
Nucleic Acids Res. 2024 Jan 5. doi: 10.1093/nar/gkad989.
PubMed PMID: 38183205.

Then there is CADD-Splice (CADD v1.6), which specifically improved the prediction of splicing effects:

Rentzsch P, Schubach M, Shendure J, Kircher M.
CADD-Splice—improving genome-wide variant effect prediction using deep learning-derived splice scores.
Genome Med. 2021 Feb 22. doi: 10.1186/s13073-021-00835-9.
PubMed PMID: 33618777.

Our third manuscript describes the updates between the initial publication and CADD v1.4, introduces CADD for GRCh38 and explains how we envision the use of CADD. It was published by Nucleic Acids Research in 2018:

Rentzsch P, Witten D, Cooper GM, Shendure J, Kircher M.
CADD: predicting the deleteriousness of variants throughout the human genome.
Nucleic Acids Res. 2018 Oct 29. doi: 10.1093/nar/gky1016.
PubMed PMID: 30371827.

Finally, the original manuscript describing the method was published by Nature Genetics in 2014:

Kircher M, Witten DM, Jain P, O'Roak BJ, Cooper GM, Shendure J.
A general framework for estimating the relative pathogenicity of human genetic variants.
Nat Genet. 2014 Feb 2. doi: 10.1038/ng.2892.
PubMed PMID: 24487276.


Which dataset was used and how?

The variants used for the calculations were taken from ClinVar (accessed February 28, 2025). The original file had 6.806.227 entries.
To only use qualitative variants, only variants with the rating of “criteria provided, multiple submitters, no conflicts”, “reviewed by expert panel”, or “practice guideline” were kept. After removing the other entries 1.135.635 entries were left. Also, only variants with the clinical classification “benign”, “likely benign”, “pathogenic”, and “likely pathogenic” are usable so only these were kept. Now 668.455 entries were left. Because ClinVar has both reference genomes GRCh37/hg19 and GRCh38/hg38, these had to be separated too. In the end we were left with 334.246 entries for GRCh37 and 334.209 entries for GRCh38.
All the variants that were left were now scored with CADD version 1.6 and 1.7 including annotations. CADD does not score InDels with more than 50 base pairs, variants where the reference allele does not fit with the reference allele of the reference genome and mitochondrial variants. So, CADD did not score 4.085 variants for GRCh37 and 4.196 variants for GRCh38.
It might be interesting to note that CADD sometimes assigns more than one annotation to one variant. As the score for each annotation stays the same, one entry per variant is enough, so all duplicates were randomly deleted. That means for the table in the bab "Genes" only one annotation is included.
GRCh37 has 252.785 benign and 77.3776 pathogenic variants while GRCh38 has 252.626 benign and 77.387 pathogenic variants.


Used Metrics
Metric Meaning
True Negatives (TN) Negative values were correctly identified as negative
True Positives (TP) Positive values were correctly identified as positive
False Negatives (FN) Positive values were incorrectly identified as negative
False Positives (FP) Negative values were incorrectly identified as positive
Precision TP / (TP + FP): proportion of correctly positive predictions among all predicted positives
Recall (Sensitivity) TP / (TP + FN): proportion of correctly positive predictions among all actual positives
False Positive Rate (FPR) FP / (FP + TN): proportion of false positive predictions among all actual negatives
Specificity TN / (TN + FP): proportion of correct negative predictions among all actual negatives
F1 Score 2 * (Precision * Recall) / (Precision + Recall): harmonic mean of precision and recall
F2 Score Same as F1 Score but recall is weighted more heavily: 5 * (Precision * Recall) / (4 * Precision + Recall)
Accuracy (TP + TN) / (TP + FP + FN + TN): proportion of correct predictions
Balanced Accuracy (Recall + Specificity) / 2: useful for unbalanced classes

Results of the calculations with the whole dataset

  • You can choose a genome release in combination with a CADD version and then choose the metrics you would like to look at. Then a line graph will load. You can hover over the lines to see specific data or zoom in, as well as change the range of the x-axis.

Distributions

  • You can also look at the distribution of the variants for the different thresholds for your chosen CADD version and genome release. It is possible to adjust the x-axis for the more small-scaled bar chart.
  • If you want to know the consequences of all the pathogenic variants at different threshold, you may look at the last bar chart. (the likely pathogenic variants have a lower opacity)

Metrics Calculation for specific genes

  1. Upload a list of your genes (as csv, txt, tsv file) or write them in the text field.
  2. Choose your genome release and CADD version and then click on the “Generate metrics” button.
  3. Now all the metrics will load in one line graph. (If you want to see one metric, double click on the name on the legend. If you want to see more than one metrics, deselect all others b clicking once on the name on the legend.)
  • If you want to know which variants were used for calculating, together with their annotations, you can look at the table. You may choose if you want to look at the ClinVar or CADD annotations or both. For ClinVar only these annotations were kept: 'AlleleID', 'Type_x', 'Name', 'GeneID_x', 'GeneSymbol', 'Origin', 'OriginSimple', 'Chromosome', 'ReviewStatus', 'NumberSubmitters', 'VariationID', 'PositionVCF', 'ReferenceAlleleVCF', 'AlternateAlleleVCF', 'ClinicalSignificance'
  • To see how many variants were used per gene and if they are pathogenic or benign you can look at the bar chart (it might not be visible if you used a lot of variants, you could still zoom in). Below the bar chart is also a table that summarizes the information from the bar chart.