Recent gnomAD versions include a "filtering allele frequency" which tells you when a variant can be safely adjudged not to be disease-causing. Unfortunately, I'm having trouble making sense of the definition, which is

"Technically, this is the highest disease-specific maximum credible population AF for which the observed AC is not compatible with pathogenicity. More practically, If the filter allele frequency of a variant is above the maximum credible population AF for a condition of interest, then that variant should be filtered (ie not considered a candidate causative variant). See http://cardiodb.org/allelefrequencyapp/ and Whiffin et al. 2017 for additional information."

I interpret this statement as "If

gnomAD_WG_AF_{population} > gnomAD_WG_FAF95_{population},

then the variant should be filtered."

(Here "FAF95" stands for "filtering allele frequency at 95% confidence".) However, the above inequality seems to hold for every variant in gnomAD, so I'm clearly missing something. On the other hand, perhaps the statement "maximum credible population AF for a condition of interest" means "proportion of people in the population who have the condition," but in that case the use of the term AF doesn't make sense. Does anyone have a better understanding of this FAF95 number than I do?


1 Answer 1


My understanding is from Whiffin et al. 2017.

Imagine a dominant mendelian disease A with complete penetrance, caused only by one allele, which is currently unknown. In your search for that unknown allele, a "credible" candidate variants is one whose population allele frequency is lower than the prevalence of that disease in the population. In that sense,

$$\text{max. credible. AF}_\text{A} = \text{prevalence}$$

In a more realistic scenario, a similar disease B can be caused by more than one allele. A credible candidate variant would therefore have a lower expected allele frequency, because only some proportion of diseases in the population should be attributable to that variant (the "allelic contribution" of that variant),

$$\text{max. credible AF}_\text{B} = \text{prevalence} \times \text{max. allelic contrib.}$$

Finally, consider a similar disease C with incomplete penetrance. Intuitively, a credible candidate variant would be allowed to have higher population allele frequency than the disease prevalence, due to the presence of unaffected carriers,

$$\text{max. credible AF}_\text{C} = \text{prevalence} \times \text{max. allelic contrib.} \times 1/\text{penetrance}$$

Thus far, we have talked about the maximum credible population allele frequency. Population reference databases like gnomAD only have a sample of the population. Whiffin and colleagues observe that one can set a value for the maximum credible allele frequency, and subsequently compute by modelling a Poisson distribution, the one-tailed 95% confidence interval for the expected allele count in a sample of a given size.

Conversely, they suggest, you could work backwards from an allele count and find the highest possible value for the maximum credible population allele frequency that gives the allele count as the 95% confidence interval. I believe this is the FAF95 recorded in gnomAD, or at least its definition should be similar.

In other words: find a reasonable estimate for $\text{max. credible AF}_\text{C}$, and filter FAF95 against that.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.